Nikhil Sharma

The fast lane to nowhere? What FERC's June orders actually fix

Fri, 17 Jul 2026 00:00:00 GMT

On June 18, 2026, the Federal Energy Regulatory Commission did something it almost never does: it moved fast. FERC directed each of the six regional grid operators it regulates, along with their member transmission owners, to justify or reform the terms of their tariffs related to interconnections for data centers and other large energy users. The show cause orders, issued under Section 206 of the Federal Power Act, arrived with deadlines measured in weeks, not years. For an agency whose signature reforms typically gestate through multi-year rulemakings, this was a statement.

The statement was necessary. The United States is trying to build the physical layer of the AI economy on a grid whose front door has been jammed for a decade. The constraint on new AI capacity is no longer chips, capital, or land; it is the queue of projects waiting for permission to plug in. FERC’s orders are the most consequential attempt yet to clear that queue.

But there is a distinction that much of the coverage has missed. The orders accelerate process: study timelines, tariff clarity, filing deadlines. They do not, and cannot, conjure generation. A faster queue that leads to the same insufficient supply of electrons is a fast lane to nowhere. The reform succeeds only if the generation pipeline moves through the widened gate. That is the test, and the early evidence suggests it will be a close call.

Exhibit 1. The queue waiting to plug in is now larger than the entire fleet already plugged in

Gigawatts of capacity: active interconnection queue against total US installed generating capacity.

Year	Interconnection queue	Installed generating fleet
2021	1400	1300
2024	2000	1300
2025	2061	1300

Lawrence Berkeley National Laboratory, Queued Up 2026.

The backlog is structural, not administrative

Start with the scale of the problem FERC is trying to solve. The interconnection queue is not a waiting room; it has become larger than the building it leads into. As of the end of 2025, roughly 8,200 projects were actively seeking grid interconnection in the United States, representing 1,312 GW of generation and approximately 749 GW of storage. Combined, that is more than 2,000 GW of capacity in line, a figure that exceeds the entire installed generating fleet of the country.

A queue of that size is not a symptom of slow paperwork; it is a symptom of a system designed for a different era. The interconnection process was built for a world in which a handful of large, predictable plants requested connection each year. It now faces thousands of applicants, and it responds the way any overloaded system does: with attrition. Historical completion rates are brutal. Of all capacity that submitted interconnection requests between 2000 and 2019, only 13% had reached commercial operation by the end of 2024. The overwhelming majority of projects that enter the queue never leave it as operating assets.

This is the context in which FERC acted, and it explains why the Commission concluded that incremental tinkering had run its course. When a process fails 87% of its participants, the process itself is the product. No AI buildout timeline, and no national competitiveness argument built on one, survives contact with a five-year median wait.

Data centers turned a generation problem into a load problem

For most of the past decade, the queue crisis was a story about supply: solar farms, wind projects, and batteries waiting to connect. What changed in the past two years is the direction of the pressure. The crisis has flipped to the demand side, and the flip is most visible in Texas.

The Electric Reliability Council of Texas is tracking more than 438 GW of large-load interconnection requests, and nearly 90% of them come from data centers. To put that number in perspective, it is several multiples of ERCOT’s current peak load. The momentum is accelerating rather than stabilizing: 198 GW of large load applied for interconnection in ERCOT in the first quarter of 2026 alone.

The demand-side flip matters because grid operators have decades of machinery for studying new generation and almost none for studying new load at this scale. A 500 MW data center campus stresses the system differently than a 500 MW power plant: it consumes rather than contributes, it wants firm service on an aggressive schedule, and it concentrates in clusters that strain local transmission. Texas responded first; the Public Utility Commission of Texas approved rules on June 18, 2026 for ERCOT to begin reviewing an initial set of large-load interconnection requests, known as Batch Zero. It is telling that the state regulator and the federal one acted on the same day. Both had reached the same conclusion: the existing rulebook simply does not contemplate the customer now knocking on the door.

Exhibit 2. Nearly ninety per cent of ERCOT’s large-load queue is data centres

ERCOT large-load interconnection queue, gigawatts.

Load type	Queued capacity
Data centres	394
Other large loads	44

ERCOT and PUC Texas, June 2026.

The cost of delay is already priced in

The bottleneck is often framed as a future risk: capacity that will not arrive, growth that will not happen. That framing is too generous. Consumers are paying for the queue today, and the bill is visible in the one market mechanism designed to reveal scarcity: the capacity auction.

The trajectory in PJM, the largest US grid operator, is stark. PJM capacity auction prices rose from about $29 per MW-day for the 2024-25 delivery year to about $270 per MW-day for 2025-26, and prices for the 2026-27 delivery year hit the FERC cap of $329 per MW-day across the entire PJM footprint. That is an eleven-fold increase in two auction cycles, halted only by a regulatory ceiling. Prices did not rise because generators became more expensive to run; they rose because new supply could not get through the queue while new demand kept arriving.

The counterfactual sharpens the point. Analysis commissioned by GridLab and executed by Aurora Energy Research concluded that if just 10% of the 107 GW of land-based renewables in the queue before 2024 had been built in time for the 2026-27 auction, it would have added 1.5 GW of net supply, with material savings for ratepayers. The queue, in other words, is not a neutral waiting line; it is an active transfer mechanism, converting administrative delay into higher electricity bills. The failure to connect new, cheaper generation in the PJM market cost consumers an estimated $7 billion in a single capacity auction. Every month the reform saves is a month of that transfer avoided.

Exhibit 3. PJM capacity prices rose elevenfold in two cycles, stopping only at the cap

PJM capacity auction clearing price, dollars per MW-day.

Auction	Clearing price
2024-25	29
2025-26	270
2026-27	329

PJM capacity auction results, 2024 to 2026.

FERC chose speed over uniformity

Against that backdrop, the design of the June orders makes sense. FERC faced a choice between two reform architectures. The first was a comprehensive national rulemaking: uniform, durable, and slow. FERC had opened a rulemaking docket that accumulated an administrative record of over 3,500 pages of comments from nearly 175 industry stakeholders, the standard raw material for a multi-year proceeding. The second was targeted enforcement: narrower, regionally fragmented, and fast. FERC chose the second.

The mechanics are aggressive by regulatory standards. Each RTO has 60 days from June 18, 2026 to demonstrate that its existing tariff remains just and reasonable or to submit revisions addressing FERC’s concerns, and grid operators and transmission owners must submit informational reports within 30 days describing how they plan to ensure sufficient generation resources are available to serve both existing and new large-load customers. Requests to pause are actively discouraged: FERC stated that it will heavily scrutinize abeyance requests, that any abeyance will be limited to 90 days, and that it will disfavor requests to extend the abeyance period.

Note the second requirement carefully. FERC is not only demanding faster interconnection process; it is demanding that grid operators show their homework on generation adequacy. That is the Commission implicitly acknowledging the fast-lane-to-nowhere problem: a shortened queue is worthless if there is nothing behind it. The regional approach also has a defensive logic. By deploying targeted, region-specific orders restricted to FERC’s jurisdiction rather than a broad national rulemaking, FERC maximizes the legal durability of its actions against potential challenges from state-level stakeholders. Speed was purchased with fragmentation; six dockets will now evolve six answers.

The unresolved fight is who pays

Timelines can be ordered. Cost allocation must be negotiated, and it is here that the reform will be won or lost.

The federal position has been unusually blunt. The orders trace back to a 2025 request submitted by Department of Energy Secretary Chris Wright aimed at accelerating grid interconnection for large-load energy users, and the DOE framework did not dodge the distributional question: it recommended standardizing study deposits, studying load concurrently with generation, and assigning 100% of network upgrade costs to the interconnecting load.

That last principle is the fault line. If data centers bear the full cost of the transmission upgrades their connections require, the reform is politically sustainable; ratepayers are shielded, and the AI industry pays for its own on-ramp. If cost responsibility blurs, as it historically has in transmission planning, the PJM auction dynamic repeats in a new form: the public funds private infrastructure through opaque channels, and the backlash eventually reaches the reform itself. The six RTO compliance filings due this summer will reveal where each region lands, and the variance between them will define the investment map for the next phase of the buildout. Watch the cost allocation provisions, not the timeline provisions; the timelines are the headline, but the money is the story.

Conclusion

FERC’s June orders deserve the significance attached to them. They compress a decade of drift into a 60-day clock, they force grid operators to answer for generation adequacy rather than just process speed, and they arrive while the cost of inaction is compounding in capacity auctions. As a piece of regulatory engineering, the show cause architecture is shrewd: fast, legally defensible, and regionally adaptive.

But the honest verdict has to be conditional. The orders fix the queue, not the grid. The 2,000 GW backlog, the 13% historical completion rate, and the $329/MW-day price cap all describe a system whose deeper constraints are physical: transmission that takes a decade to build, turbines with multi-year lead times, and a generation fleet that retires faster than it is replaced. Process reform was necessary, and FERC has now delivered the most serious version of it in years. Whether it proves sufficient will be decided not in the six dockets, but in what gets built behind them. The paperwork is now moving at AI speed. The electrons are not, yet.

Views are my own and do not represent my employer.

The $725 billion question: why the AI bubble debate is about balance sheets, not valuations

Thu, 16 Jul 2026 00:00:00 GMT

Every bubble debate eventually collapses into the wrong argument. With AI, the argument has fixated on whether equity valuations are stretched; whether a price-to-earnings multiple of 25 or 30 on the technology complex is defensible. This is the least interesting question available. Valuations are an opinion. Cash flows are a fact. And the facts now describe the widest gap between capital deployed and revenue generated in modern economic history.

The thesis of this article is simple: the AI buildout is running an unprecedented capex-to-revenue spread, and the way that spread is being financed, through circular deals, debt, and passive household exposure, is what determines whether the eventual correction is a contained equity repricing or a systemic credit event. The bubble question is not “are AI stocks expensive.” It is “who is holding the funding gap, and what happens to them when the music slows.”

1. Capex has decoupled from any current revenue base

Start with the raw numbers, because they have stopped behaving like corporate budgeting and started behaving like national industrial policy.

The four largest US hyperscalers are collectively guiding to roughly $725 billion in capital expenditure for 2026, the bulk of it directed at AI data centers, GPUs, and power. Compared with about $410 billion in 2025, that is a 77 percent year-over-year jump, with analysts already projecting more than $1 trillion in 2027. Amazon alone has guided for around $200 billion in capex this year, more than doubling its 2025 outlay.

For context, this level of spending exceeds the annual GDP of most countries. No revenue line inside these companies justifies it on conventional payback math; the spending is justified by a forecast, specifically the belief that agentic AI will convert infrastructure into recurring consumption. That may prove correct. But it means the largest capital deployment cycle since the railway era is underwritten by a projection, not by demonstrated unit economics.

Exhibit 1. Amazon alone is guiding to more capital spending this year than any of its rivals, and the whole group is up three quarters in twelve months

Capital expenditure, billions of US dollars. 2025 actual against 2026 guidance.

Company	2025 actual	2026 guidance
Amazon	95	200
Alphabet	85	180
Microsoft	90	155
Meta	70	135

Company investor disclosures and guidance, mid-2026.Combined spending rises from about $410bn to about $725bn, a 77 per cent jump in one year.

2. The required-revenue math keeps outrunning actual AI sales

The canonical framework for sizing the gap comes from a widely circulated venture analysis, and its trajectory is the story. The math takes projected data center chip revenue, doubles it to account for the full cost of ownership (power, buildings, networking, cooling), then doubles it again to reflect the roughly 50 percent gross margin operators need for the spend to pencil. In late 2023, the annual AI revenue required to justify the capex was $200 billion. By mid-2024, it was $600 billion. In 2026, with chip revenue running substantially higher and gigascale buildouts in flight, the implied number is meaningfully larger, and the gap between that requirement and what the AI industry actually generates has widened, not closed.

This is the crucial dynamic that bulls tend to wave away: the target is not static. Every quarter of accelerating capex raises the revenue bar that future AI sales must clear. The industry is not chasing a fixed finish line; it is chasing a finish line that its own spending pushes further away. One estimate places the annual revenue gap between hyperscaler infrastructure spending and actual AI ecosystem sales at approximately $600 billion, and widening in 2026 as capex accelerates faster than revenue projections.

3. We have passed the telecom benchmark, not approached it

Bubble skeptics love the telecom analogy; the surprise is that the data says we are already beyond it. The divergence between AI capital expenditure and revenue growth is running at roughly 46 percent, already exceeding the 32 percent divergence observed during the 2001 telecom excess cycle, a period that preceded a brutal multi-year market correction in tech.

The standard rebuttal is that this time the spenders are profitable giants, not leveraged startups laying fiber on junk bonds. That rebuttal is partially right and entirely beside the point. The telecom comparison is not about who goes bankrupt; it is about the mechanics of overbuild. When capital formation runs 14 points further ahead of revenue than it did at the peak of the last great infrastructure mania, the question is no longer whether excess capacity exists. It is who absorbs the writedowns when depreciation schedules meet reality, particularly for hardware that becomes technologically obsolete far faster than fiber ever did.

Exhibit 2. AI capital formation is running fourteen points further ahead of revenue than telecom did at its peak

Capital expenditure as a share of revenue, per cent.

Period	Capex as share of revenue
Telecom, 2001 peak	32
AI, 2026	46

Allianz Research.

4. The value is real; it is just being captured by the wrong people

Here is where this article departs from the standard bear case. The problem with AI economics is not that the technology fails to create value. It is that the value is leaking straight past the firms paying for it.

Estimated US consumer surplus from generative AI reached $172 billion annually by early 2026, up from $112 billion a year earlier, with the median value per user tripling over the same period. Most of these tools remain free or close to it. Generative AI reached 53 percent population adoption within three years, faster than the personal computer or the internet.

Read those two facts together and the shape of the problem emerges. Adoption is historically fast, the utility is measurably large, and almost none of it is monetized at the point of use. Consumer surplus is wonderful for households and terrible for payback periods. The dot-com era had exactly this profile: enormous, genuine value creation (e-commerce, search, email) whose commercial capture arrived years after the infrastructure investors who funded it had been wiped out. Being right about the technology and wrong about the timing of monetization is the oldest way to lose money in markets.

Exhibit 3. Almost none of the value generative AI creates is monetised at the point of use

US consumer surplus from generative AI, billions of dollars per year.

Period	Consumer surplus
Early 2025	112
Early 2026	172

Stanford HAI AI Index 2026.

5. The financing structure turns an equity story into a systemic one

This is the argument that should worry policymakers more than any valuation metric. The funding gap is not being carried on venture balance sheets that can fail quietly. It is being distributed through channels that touch the broader financial system.

The IMF has flagged that circular investment and procurement arrangements, in which firms invest in each other while securing future orders, among large AI players create opacity and concentration risk, making ownership structures and valuations harder to assess. Rising reliance on debt financing, reflected in high debt ratios and widening credit default spreads of some firms, raises additional concerns. Meanwhile, the equity exposure has been socialized: a major portion of rising household exposure runs through benchmark indices, particularly the S&P 500, largely via 401k retirement accounts and passive investment vehicles, making household balance sheets vulnerable to sharp corrections and prolonged declines in the index.

Chip vendors investing in model labs that commit to buying compute from cloud providers that the chip vendors also supply; this is vendor financing wearing a trench coat. It inflates reported revenue on the way up and synchronizes losses on the way down. Combine that with record passive household exposure to a historically concentrated index, and the transmission mechanism from AI disappointment to household wealth is shorter than at any point in the dot-com era, when index concentration was lower and retirement savings were less equity-heavy.

The honest caveat

A fair reading requires acknowledging what the bears get wrong. The IMF itself notes meaningful differences between current AI investment and the dot-com period; today’s spenders are established firms, and the risk channels differ. AI revenue is not imaginary: one major hyperscaler’s AI business has surpassed a $37 billion annual run rate, up 123 percent year over year, and global corporate AI investment reached $581.7 billion in 2025, up 130 percent year over year, which signals conviction from sophisticated capital, not retail mania. Efficiency gains may also trigger the demand paradox, where cheaper inference drives higher total consumption and ultimately justifies the capacity.

But note what this caveat concedes: the bull case now rests on second-order effects (future agentic demand, paradoxical consumption growth) rather than first-order arithmetic. When the defense of a capital cycle requires appealing to demand that does not yet exist, the cycle is speculative by definition. Speculative is not the same as wrong. It is the same as fragile.

Conclusion

The AI bubble debate has been framed as a referendum on the technology. It should be framed as an audit of the financing. The technology is delivering value at historic speed; the capex is running 46 percent ahead of the revenue that must eventually service it; the gap is bridged by circular deals, rising debt, and the retirement accounts of households who never chose the exposure. That combination does not guarantee a crash. It guarantees that if a correction comes, it will not stay politely contained inside the technology sector. The right question for 2026 is not whether AI is overvalued. It is whether the funding structure can survive the years between now and monetization, and history’s answer to that question is rarely kind.

Views are my own and do not represent my employer.

The new billable hour: how AI governance is rebuilding the consulting model it broke

Wed, 15 Jul 2026 00:00:00 GMT

For three decades, the consulting industry sold a simple product: judgment that clients could not produce internally. AI has attacked that product from both ends. Clients now run analysis in-house that once justified seven-figure engagements, and the firms themselves are automating the junior work that fed the pyramid. The mainstream story is that consulting is being hollowed out. The more interesting story is what is quietly filling the hollow.

The same technology that is eroding consulting’s traditional advisory revenue is simultaneously manufacturing its replacement product. Governance and compliance work, the unglamorous business of classifying AI systems, documenting models, standing up oversight committees, and preparing for regulatory audits, is becoming the industry’s new billable hour. The disruption and the recovery are the same event, viewed from different sides of the invoice.

This is not a small pivot. It is a structural repricing of what consulting is for. And the data suggests it is already well underway.

Regulation turned governance from virtue into obligation

The trigger is regulatory, and it is dated. The EU AI Act becomes fully applicable on 2 August 2026, with penalties for the most serious violations reaching 35 million euros or 7 percent of global annual turnover, whichever is higher. High-risk systems embedded in regulated products get a transition runway to 2028, but the core obligations, including documentation, risk classification, and human oversight requirements, now carry enforcement teeth.

The significance is not the fine itself; few companies will ever pay the maximum. The significance is what a 7 percent turnover penalty does to board psychology. Risk of that magnitude cannot be delegated to a mid-level compliance function and forgotten. It demands documented process, external validation, and someone to blame if things go wrong. Those three demands describe the consulting value proposition almost perfectly.

There is historical precedent. GDPR did not shrink the advisory market; it created a privacy-industrial complex of assessments, gap analyses, and data-protection-officer services that persists years after the compliance deadline passed. The AI Act is GDPR with broader scope and higher stakes, because it regulates not a data practice but a general-purpose technology being wired into every core process. Every system a client deploys is a future compliance artifact, and someone has to produce the paperwork.

Spending is scaling faster than the ability to oversee it

Regulation supplies the obligation; spending supplies the exposure. Executives surveyed for BCG’s AI Radar 2026, roughly 2,400 of them, expect corporate AI spending to double in 2026, from around 0.8 percent of revenue to approximately 1.7 percent. For a 10-billion-dollar company, that is a jump from 80 million to 170 million dollars in a single budget cycle.

Doubling spend in twelve months means doubling the surface area of things that can go wrong: more models in production, more vendors in the stack, more automated decisions touching customers, employees, and regulators. Oversight capacity does not double on the same schedule. Governance expertise is scarce, internal risk teams were sized for a pre-AI world, and the technology itself changes faster than any annual review cycle.

That widening gap between exposure and oversight is precisely where external advisors live. It is the same dynamic that made cybersecurity consulting a permanent industry: spending on the underlying technology grew faster than the internal capacity to secure it, and the difference was outsourced. AI governance is running the identical play, compressed into a shorter timeline.

Exhibit 1. Corporate AI spend doubles in a single budget cycle

Corporate AI spending as a share of revenue, per cent.

Year	Share of revenue
2025	0.8
2026 (expected)	1.7

BCG AI Radar 2026, survey of approximately 2,400 executives.

Boards built the committees before they found the expertise

The demand signal is already visible in corporate structure. A 2025 Gartner poll of more than 1,800 executive leaders found that 55 percent of organizations report having an AI board or dedicated oversight committee in place. On paper, that looks like maturity. In practice, it is scaffolding without a building.

A committee is easy to charter; a functioning governance capability is not. Most of these bodies were stood up quickly, often in response to board pressure or a regulatory deadline, and they now face questions their members are not equipped to answer. Which of our systems qualify as high-risk under the classification rules? Is our model documentation audit-ready? Who is accountable when an agentic workflow makes a decision no human reviewed?

This is the sweetest spot in professional services: a client that has already accepted responsibility for a problem but lacks the capability to discharge it. The committee’s existence is itself a demand generator. Every quarterly meeting produces action items, and a substantial share of those action items become statements of work. Consulting firms did not have to create this market; corporate governance created it for them, one charter at a time.

The firms are already rebuilding their revenue engine around it

None of this is hypothetical for the supply side. The Management Consultancies Association’s member survey, published in January 2026, found that 66 percent of UK consulting firms cite AI services as their greatest driver of revenue growth. The industry that AI is supposedly dismantling is reporting that AI is its best-performing product line.

The composition of that work is what matters. Early AI consulting was strategy-flavored: roadmaps, use-case prioritization, operating-model decks. That work commoditizes quickly, because clients learn it and tools absorb it. Governance work behaves differently. It is recurring rather than one-off, since compliance is never finished; it is regulator-anchored, so demand does not depend on client enthusiasm; and it carries liability transfer value, because an external assessment is worth more in an enforcement proceeding than an internal one.

In other words, the industry is trading a cyclical, commoditizing revenue stream for an annuity. The billable hour is not disappearing; it is migrating from the strategy deck to the risk register. That migration also favors firms with audit and assurance DNA, which is why the most aggressive moves are coming from firms that already know how to sell recurring compliance relationships rather than episodic advice.

The gap between ambition and returns keeps the meter running

There is a final, quieter driver: disappointment. Deloitte’s State of AI in the Enterprise 2026 survey of 3,235 leaders found that only 20 percent of organizations are currently growing revenue from their AI initiatives, while 74 percent aspire to. That 54-point gap between ambition and realized return is the most underpriced statistic in the AI economy.

When returns lag ambition, boards do not simply cancel programs; they demand explanations, controls, and proof of value. Underperformance triggers exactly the machinery that governance consulting monetizes: ROI gates, value-tracking frameworks, portfolio reviews, and independent assessments of why the promised transformation has not arrived. Failure is billable. So is the fear of failure.

This creates an unusual hedge for the consulting industry. If AI delivers, spending scales and governance scales with it. If AI disappoints, the post-mortem and remediation work scales instead. Either branch of the outcome tree produces engagements. Very few industries get paid on both sides of a technology bet; consulting has engineered itself into one of them.

Exhibit 2. A fifty-four point gap sits between AI ambition and realised return

Share of organisations surveyed, per cent.

Position	Share of organisations
Aspire to grow revenue via AI	74
Currently growing revenue via AI	20

Deloitte State of AI in the Enterprise 2026, 3,235 leaders surveyed.

The uncomfortable equilibrium

Put the five pieces together and the shape of the new model is clear. Regulation manufactured a legal obligation with existential penalties. Spending growth widened the gap between exposure and oversight. Boards built committees that generate demand faster than they build capability. Firms repositioned their revenue engines around the work. And the persistent gap between AI ambition and AI returns guarantees the meter keeps running regardless of outcomes.

There is an obvious tension in this equilibrium, and clients should name it. The firms selling AI governance are often the same firms that sold the AI transformation now requiring governance, and in some cases the same firms embedding AI so deeply into their own delivery that their advice is partially machine-generated. The industry is charging to install the engine, charging to inspect it, and charging to explain why it has not yet reached full speed. That is not necessarily cynical; the expertise genuinely concentrates in the same places. But it means procurement teams should treat governance engagements with the same scrutiny they apply to any vendor whose incentives run in both directions.

The deeper lesson is about what consulting actually sells. It was never really analysis; analysis was the delivery mechanism. The durable product is institutional reassurance: the ability of a board to say that a credible third party looked at the risk and signed off. AI has automated the delivery mechanism while multiplying the demand for reassurance. Seen that way, the industry is not being disrupted by AI so much as repriced by it, out of insight and into assurance.

The pyramid may not survive. The invoice will.

Views are my own and do not represent my employer.

The sovereign AI boom is a subsidy race in disguise

Tue, 14 Jul 2026 00:00:00 GMT

Every few decades, governments rediscover the same idea: if a technology is strategic, the state should pay to own it at home. In the 2010s and early 2020s, the technology was semiconductors, and the result was the largest coordinated industrial subsidy push since the Cold War. In 2026, the technology is artificial intelligence, and the vocabulary has changed to “sovereign AI.” The economics have not changed at all.

Strip away the language of sovereignty, autonomy, and national champions, and what remains is a familiar structure: capital grants, tax credits, state co-investment vehicles, and subsidized access to scarce inputs, all deployed to localize a capital-intensive technology stack inside national borders. That is the fab subsidy playbook, reprinted with a new cover. And it carries the same risks the fab race produced: escalating state outlays, chronic underdelivery against headline numbers, and a looming overhang of subscale, fragmented capacity.

This is not an argument against public investment in AI. It is an argument that policymakers are repeating a specific, recent, well-documented mistake, and that the data from the semiconductor precedent tells us how this movie ends.

The playbook is identical to chip subsidies

The most striking feature of sovereign AI programs is how little institutional innovation they contain. The instruments are the ones governments built for fabs: direct grants for facility construction, investment tax credits for equipment, state equity stakes in national champions, and preferential access to state-controlled resources, whether land, power, or compute.

The semiconductor precedent shows exactly where this instrument mix leads. OECD analysis of firms in its MAGIC database records a notable increase in government grants in 2023, reflecting the introduction by several governments of support schemes aimed at encouraging construction of new fabrication facilities; contract foundries received the most subsidies relative to their revenue, followed by integrated device manufacturers. In other words, once governments started competing to attract fabs, subsidy intensity climbed fastest exactly where capital intensity was highest, and support concentrated on a handful of anchor firms rather than lifting the broader ecosystem.

Sovereign AI is reproducing that concentration dynamic in real time. State money flows to national compute champions, flagship data center projects, and a small cohort of anointed model developers. The subsidy escalation curve that took semiconductors a decade to trace is being compressed into a much shorter window, because every government can see every other government’s announcements, and no one wants to be the capital that underbid.

The money is already at chip-race scale

If sovereign AI were a modest complement to private investment, the subsidy-race framing would be overwrought. It is not modest. The commitments announced in the past eighteen months already rival, and in some cases exceed, the national semiconductor programs that defined the chip race.

The program-level figures speak for themselves. France has assembled commitments of roughly EUR 109 billion in AI investment pledges. Japan’s government-backed national AI consortium will receive about 387.3 billion yen, roughly $2.4 billion, in subsidies this fiscal year and approximately 1 trillion yen, around $6.1 billion, over five years. The United Kingdom has launched a GBP 500 million national unit to back domestic AI companies, alongside a GBP 282 million programme to fund shared strategic AI assets such as high-value datasets. Canada has committed approximately $890 million to build a national AI supercomputing system under its Sovereign AI Compute Strategy.

For comparison, the US CHIPS and Science Act, the single largest semiconductor intervention in the West, allocated $52.7 billion, and took years of political wrangling to pass. Sovereign AI programs have reached comparable aggregate scale in roughly two years, with far less scrutiny, because “AI leadership” currently functions as a political trump card that closes debate rather than opening it. Speed of escalation is itself a warning sign; subsidy races accelerate when governments respond to each other’s announcements rather than to underlying economics.

Exhibit 1. One country's AI pledges already exceed the entire US CHIPS Act, assembled in a fifth of the time

Announced programme commitments, billions of US dollars.

Programme	Commitment
France AI pledges	118
US CHIPS Act	52.7

French government announcements to mid-2026; US CHIPS and Science Act allocation.Japan, the UK and Canada have announced $6.1bn, $1.05bn and $0.65bn respectively. The spread across all five is too wide to hold one linear scale, so the chart carries only the chip-race comparison.

Sovereignty is being conflated with self-sufficiency

The deeper design flaw is conceptual. Sovereignty, properly defined, is about strategic control: the ability to keep operating, to govern AI according to national values, and to avoid single points of foreign dependency. Self-sufficiency is something else entirely: owning every layer of the stack, from silicon to applications, inside national borders. Most sovereign AI programs claim to pursue the first while budgeting for the second.

This is not a fringe critique. The World Economic Forum’s January 2026 analysis of AI sovereignty finds that several economies have attempted to compete by owning the entire AI value chain, from raw materials to AI-based applications, and concludes that based on investment patterns, “AI sovereignty” has been conflated with self-sufficiency; it argues explicitly that this is not the only path to AI competitiveness.

The semiconductor race made the same conflation, and it was the single most expensive mistake of that era. No country, including the United States, achieved chip self-sufficiency, because the value chain is irreducibly global: lithography from the Netherlands, design tools from the US, advanced packaging from Asia. AI is even less separable. Frontier models, accelerator hardware, training data, and talent flow across borders by default. A national program that prices full-stack ownership is buying an outcome that the structure of the industry does not permit at any price a mid-sized economy can pay.

The precedent shows underdelivery, not capability

Suppose we set aside design questions and ask a simpler one: when governments announce these numbers, what actually gets delivered? The chip race provides a clean natural experiment, and the answer is uncomfortable.

The European Chips Act, adopted in 2023, was intended to mobilize public support comparable to the US program, yet as of early 2025 only EUR 13.75 billion in state aid had been approved, while the US Department of Commerce had awarded $33.7 billion in grants plus $5.5 billion in loans; independent analysis concludes the act has underdelivered both in the scale of funding mobilized and in the strategic coordination of its deployment, aggravated by what the OECD describes as an ongoing subsidy race.

Note what this comparison shows. Even the better-performing program, the US one, delivered awards well below its headline over its first years, and the EU program delivered roughly a third of its American counterpart despite similar ambitions. Headline numbers in subsidy races are announcements, not appropriations; they are discounted by permitting delays, state-aid procedures, co-funding shortfalls, and shifting political priorities.

Exhibit 2. The chip precedent: announced ambition and delivered funding are different numbers

Funding awarded against funding allocated, per cent.

Programme	Delivery ratio
US CHIPS	74
EU Chips Act	32

US Department of Commerce and European Commission, as of early 2025.

Sovereign AI inherits every one of these frictions, plus new ones. Compute infrastructure depends on grid connections with multi-year queues, on power prices that several European governments cannot control, and on GPU supply governed by one dominant vendor’s allocation decisions. When national AI strategies are audited in 2029, expect the same pattern the chip race produced: delivered capability at a fraction of announced ambition, with the gap widest in the economies that announced loudest.

Fragmentation destroys the returns

The final problem is arithmetic. AI infrastructure economics are brutally scale-dependent: utilization rates, energy contracts, hardware refresh cycles, and model training runs all reward concentration. A subsidy race pushes in exactly the opposite direction, toward dozens of parallel national stacks, each built for political visibility rather than economic efficiency.

The concentration data makes the problem concrete. The WEF’s cumulative investment estimates, spanning electricity capacity, silicon processing, chip manufacturing, data centres, foundation model training, and application development, show spending overwhelmingly dominated by a small set of economies, with a long tail of countries grouped into “rest of world”. The national programs multiplying across that long tail are subscale by orders of magnitude relative to the leaders. A $900 million national compute program is a rounding error against private hyperscaler capex, which is measured in the hundreds of billions annually.

Subscale capacity is not merely inefficient; it is stranded-asset risk. GPU fleets depreciate fast, and a national supercomputer that runs below utilization because the domestic ecosystem cannot fill it is a monument, not an asset. The chip race left behind exactly this residue: announced fabs delayed, downsized, or quietly cancelled once subsidy terms met market reality. Sovereign AI’s equivalent will be half-utilized national clusters, still drawing subsidized power, defended in budget hearings as strategic.

What a smarter sovereignty would buy

None of this means governments should stand aside. It means the objective function is wrong. The lesson of the chip race is that indispensability beats self-sufficiency: economies won durable leverage by dominating specific chokepoints, not by replicating the whole stack. The Netherlands did not build a national semiconductor industry; it built a chokepoint in lithography.

Applied to AI, that logic points away from national full-stack replication and toward three cheaper, higher-leverage plays. First, specialize: fund the layers where the economy has genuine comparative advantage, whether sovereign datasets in regulated domains, domain-specific models, or energy-advantaged compute for others to rent. Second, pool: shared compute across allied economies captures the scale economics that national fragmentation destroys. Third, secure optionality rather than ownership: contractual guarantees, multi-vendor strategies, and trusted-partner agreements deliver most of what sovereignty actually requires at a fraction of the capital cost.

The subsidy race will not stop because this argument is right. Races have their own momentum, and no elected government wants to explain why it declined to fund the defining technology of the decade. But investors, corporate strategists, and the analysts who will eventually audit these programs should be clear-eyed now: sovereign AI, as currently constructed, is the fab race with better branding. The precedent is not encouraging, and this time we cannot claim we did not have the data.

Views are my own and do not represent my employer.

Most competitive intelligence is expensive news summarization

Mon, 13 Jul 2026 00:00:00 GMT

Walk into most CI functions and audit what actually ships. Weekly competitor newsletters. Quarterly battlecard refreshes. Alert digests forwarded with a sentence of commentary. Conference summaries. A monthly deck that lists what rivals announced, in the order they announced it.

Now ask a harder question: which decision changed because of any of it?

In most functions, the honest answer is none, or almost none. The output is accurate, current, professionally formatted, and inert. Strip away the branding and it is a news clipping service with a salary band. Companies pay six figures per analyst for a product that Google Alerts delivers for free, plus formatting.

This is not an insult to the analysts. Most of them are capable, curious people who know something is wrong. The drift into summarization is not a talent failure. It is a structural outcome, produced by five forces that push every CI function toward the same equilibrium unless someone actively pushes back. This piece is about those forces, because you cannot fix a drift you have misdiagnosed as laziness.

Force 1: Summarization is legible, judgment is not

Start with the cruelest asymmetry in the job. A summary can be evaluated instantly. Anyone can check whether the newsletter covered the week’s announcements, whether the dates are right, whether the competitor’s press release was accurately condensed. Completeness and accuracy are visible at a glance.

Judgment cannot be evaluated instantly. A genuine analytical call, this move means X, therefore we should do Y, takes months to be proven right or wrong, and even then the proof is contaminated by everything else that happened. A manager reviewing a summary knows within minutes whether the analyst did the work. A manager reviewing a judgment has to either share the judgment or wait a year.

Organizations manage what they can see. So the visible virtues, coverage, timeliness, accuracy, polish, get rewarded on every cycle, while the invisible virtue, being usefully right about what matters, gets rewarded rarely and noisily. Analysts are not stupid. They optimize for the scoreboard that exists. The scoreboard measures summarization.

Force 2: The risk asymmetry punishes exactly the work that matters

A summary cannot be wrong in any career-damaging way. If the newsletter says a competitor launched a product, and the competitor launched a product, the analyst is safe. The worst case for pure reporting is a factual error, which is embarrassing and correctable.

A judgment can be wrong in a way everyone remembers. Call a competitor’s move a strategic pivot, recommend a response, have the pivot fizzle, and your next ten reports get read with a discount. The analyst who never concludes anything never takes that hit.

So the incentive gradient inside every report points downhill, away from the conclusion. Add one more caveat. Soften the recommendation into “worth monitoring.” Present both interpretations and let the reader choose. Each individual retreat is defensible. Compound them across a hundred reports and the function has fully converted into a summarization shop, without any single person ever deciding it should.

Here is the uncomfortable part: the readers cooperate. Executives rarely punish an inconclusive report. They skim it, nod, and derive their own conclusions in the hallway afterward. The report that concludes nothing generates no friction, and no friction feels like success. Both sides of the transaction are locally satisfied while the function’s actual reason for existing quietly evaporates.

Force 3: The tooling industry sells collection, so collection becomes the job

Look at the CI software market. Almost every product in it is a collection and monitoring engine: track competitor websites, scrape pricing pages, aggregate news, alert on changes, auto-populate battlecards. The pitch is always the same: never miss anything.

Never missing anything is a summarization goal. No tool on the market sells “reach a conclusion,” because conclusions do not demo well. But the tools shape the workflow, and the workflow shapes the job. An analyst whose day is structured by an alert queue becomes a queue processor. The unit of work becomes the item, cleared, tagged, forwarded, and the backlog is infinite by design, because the internet produces more competitor-adjacent content every day than any team can process.

This is how capable analysts end up genuinely busy while producing nothing that moves a decision. The queue is always full. Clearing it feels like diligence. And the deep work, sitting with three weeks of accumulated signals and asking what they add up to, has no queue, no alert, and no tool. It has to be defended as unstructured time, and unstructured time is the first thing that dies when the queue is full.

Volume of input became the proxy for quality of function. That proxy is not just wrong; it is inverted. The functions drowning in the most feeds usually have the least time to think.

Force 4: CI is org-chartered as a service, not as a participant

Ask where CI reports in most companies. Usually into marketing, sometimes product, occasionally strategy. Almost always it is chartered as a support service: internal clients submit requests, CI fulfills them. Sales wants battlecards. Product wants a feature comparison. An executive wants a briefing before a board meeting.

A service function answers the questions it is asked. And the questions it is asked are overwhelmingly descriptive: what did they launch, what do they charge, who did they hire. Descriptive questions have summary-shaped answers. The requester was never going to ask “what should we conclude from the pattern of the last six months,” because the requester does not know there is a pattern; noticing patterns nobody asked about is precisely the job that falls to no one when CI is a ticket queue.

The analysts closest to the evidence are structurally excluded from the conversations where the evidence would matter. They find out which decisions were on the table after the decisions are made, usually by reading the same announcements they summarize for everyone else. Intelligence produced outside the decision loop can only ever be news, because relevance to a decision is the only thing that separates intelligence from news, and you cannot aim at a decision you cannot see.

Force 5: Nobody ever closes the loop

The final force is the missing feedback cycle. In functions that make real calls, forecasting, trading, underwriting, predictions get scored. Someone writes down what was claimed, waits, and checks. The pain of being visibly wrong is the mechanism that improves judgment.

CI functions almost never do this. Reports ship, land, and vanish. No one records what the report claimed would happen, so no one ever checks, so no one ever learns whether the function’s judgment is any good, so the judgment never improves, and, more corrosively, never earns trust. Executives discount CI conclusions partly because those conclusions have no track record, and they have no track record because nobody kept one.

Summarization thrives in this vacuum because it is the one product that needs no track record. Yesterday’s newsletter is not falsified by tomorrow’s events. A function that never scores itself will drift toward the only output that cannot lose.

What the drift actually costs

The obvious cost is the salary line, paying analyst rates for aggregation work. That is the small cost.

The real cost is the decisions made blind while the CI team was busy. Every company that got surprised by a competitor’s move it “knew about”, the signals were in the newsletter, in week 34, bullet four, paid this cost. The information was collected, summarized, distributed, and never converted into a conclusion anyone had to confront. Collection without conclusion produces the most dangerous state in competitive strategy: the illusion of awareness. Leadership believes it is informed because the briefings arrive on schedule. Informed is not the same as prepared, and the gap between them is exactly the analytical work the function stopped doing.

There is a third cost, quieter: the analysts themselves. The people drawn to CI wanted to be detectives and became clipping services. The good ones leave for strategy roles where conclusions are the job. The function keeps the ones who made peace with the queue, which locks the equilibrium in for another cycle.

The test, and the direction out

One diagnostic separates an intelligence function from a summarization function, and it takes thirty seconds. Take the last five deliverables the team shipped and ask of each one: does this document make a claim that could turn out to be wrong?

Not a fact that could be mistyped. A claim. A judgment about what something means or what will happen or what should be done, stated plainly enough that reality could contradict it. If none of the five contain one, the function is summarizing, whatever its charter says. Risk of being wrong is not a flaw of analysis; it is the definition of it. A deliverable that cannot be wrong contains no analysis, only arrangement.

The direction out follows from the diagnosis, force by force. Measure conclusions shipped and decisions touched, not items covered, so judgment gets a scoreboard. Make the analytical call a required section of every deliverable, so the retreat downhill has a floor. Cap collection deliberately, fewer feeds, more synthesis time, because the queue will never volunteer to shrink. Put analysts inside the decision conversations, or at minimum tell them which decisions are live, so the work has a target. And score the calls: keep a visible log of what the function claimed, revisit it on a schedule, and let the track record, including the misses, accumulate. The misses are not the risk. The absence of any record is.

None of that is complicated. All of it is uncomfortable, because every step trades safety and legibility for exposure. That is the actual price of doing intelligence rather than reporting on the news, and most functions, without ever deciding to, have chosen not to pay it.

The competitors’ announcements will be summarized either way. The only question is whether anyone in the building is paid to say what they mean.

Views are my own and do not represent my employer.

Answer first: SCQA and the Pyramid Principle, applied to research deliverables

Sun, 12 Jul 2026 00:00:00 GMT

Most research deliverables are written in the order the work was done. Methodology, then market background, then findings, then, somewhere around page nine, the point. The analyst experiences this as thoroughness. The reader experiences it as a document that hides its conclusion, and readers with authority do not go looking; they skim, form their own version of the point, and act on that instead of yours.

Barbara Minto solved this problem at McKinsey in the 1970s, and her solution, the Pyramid Principle with its SCQA opening, became the writing operating system of the entire consulting industry. This piece applies both tools to research and CI deliverables specifically, and it starts with a distinction that saves a lot of confusion.

The CI Report Pyramid and Minto’s pyramid are different tools that happen to share a shape. The CI Report Pyramid governs the thinking: whether your work actually climbs from data through insight and implication to a recommended action. Minto governs the telling: the order in which a finished argument is delivered to a reader. You climb your pyramid to build the finding. The reader descends Minto’s: conclusion first, then the support, layer by layer, only as deep as they need. Most bad deliverables are not badly thought; they are well-thought work delivered in excavation order rather than reader order.

The worked thread for this piece is one deliverable, rebuilt. The setup: you are the research lead at a European maker of industrial valves. A Chinese competitor, call it Jianhe Flow Control, has spent 18 months quietly winning European certifications, and your team has produced a 22-page assessment. The before version opens like this:

“This report examines recent developments in the European industrial valve market. Section 1 reviews market structure and sizing methodology. Section 2 profiles Jianhe Flow Control. Section 3 assesses certification activity. Findings are presented in Section 4.”

Accurate, professional, and dead. Twenty-two pages of table stakes before a stake. Rebuild it with SCQA as the spine.

S: Situation. Start where the reader already nods

The Situation states what the reader knows and agrees with, in one or two sentences. Its job is not to inform; it is to establish shared ground so the reader’s first reaction is a nod, not a question. The test: if a reasonable reader could dispute your Situation, it is not a Situation yet.

Research deliverables usually fail here by overloading. Fifteen paragraphs of market context is not a Situation; it is the analyst proving they did the reading. The reader who commissioned a competitive assessment does not need the market explained. They need to be met where they stand.

The rebuild: “We hold roughly 30 percent of the European industrial valve market, and our pricing has assumed that Asian entrants remain locked out of certified, safety-critical segments.”

One sentence. The reader nods, and notice what the sentence quietly does: it states the assumption that is about to break.

C: Complication. The thing that changed

The Complication is why this document exists now. Something moved: a signal fired, a trend crossed a threshold, an assumption stopped holding. In CI terms, the Complication is your signal cluster, and if you triaged properly, you already know what it is; it is whatever earned this deliverable its place on the calendar.

The discipline here is honesty of scale. A Complication oversold (“the market is being disrupted”) burns credibility; one undersold gets the deliverable filed. State what happened, dated and labeled, at the size the evidence supports.

The rebuild: “Over the past 18 months, Jianhe Flow Control has obtained the three certifications that gate our most profitable segments, hired a former certification-body engineer to lead its Rotterdam office, and quoted 30 to 40 percent below us in two tenders we can verify.”

Three facts, checkable, no adjectives doing the work.

Q: Question. The pivot that disciplines the whole document

The Complication forces a question in the reader’s mind, and the Question makes it explicit. This is the least glamorous element of SCQA and the most structurally important, because of one rule: the Question must be the question your deliverable actually answers. If your document answers “how exposed are we and what should we do,” but your opening implies “who is Jianhe,” the reader finishes the wrong document.

This is the telling-side twin of a thinking-side rule from market sizing: the definition is the sizing. On the writing side, the Question is the deliverable. Getting it wrong misaligns every page that follows. In practice, writing the Q is a test of the work itself: if you cannot state the single question the document answers, the document answers several, poorly, and the problem is not the writing.

The rebuild: “How much of our certified-segment revenue is exposed, how fast, and what do we do before the next tender cycle?”

A: Answer. The top of your pyramid, stated first

Now the move that research culture resists: give the answer, immediately, before the evidence. The Answer is Minto’s “governing thought,” and for a CI deliverable it is simply the top of your CI Report Pyramid, the implication and action you climbed to, restated as the document’s opening claim, with its Fact vs. Inference Ladder label attached.

The rebuild: “We assess that roughly a fifth of our European revenue becomes contestable within 24 months, [CI] corroborated inference from the certification, hiring, and tender evidence. We recommend defending the top 15 certified accounts with multi-year service contracts this quarter, and re-testing this assessment in 90 days against Jianhe’s tender activity.”

Analysts resist answer-first for an honest reason: it feels unearned, a verdict before the trial. Two replies. First, executives read openings and skim bodies; answer-first is not a style preference, it is meeting the actual reading behavior of the people who decide. Second, the Ladder label is what makes answer-first honest. A conclusion stated up front with its confidence level declared is more truthful than a conclusion smuggled in on page nine after the reader’s attention has left. Answer-first without labels is arrogance. Answer-first with labels is a briefing.

After the A: the body is Minto’s pyramid, descended

SCQA gets the reader to the governing thought in half a page. The Pyramid Principle then structures everything below it, and three of Minto’s rules carry directly into research work.

Vertical logic: every statement provokes the question the level below answers. The Answer above makes the reader ask “why contestable, and why a fifth?” The next layer answers with the three key-line claims: certifications remove the formal barrier; the tender evidence shows an executable price gap; the exposed accounts can be named and sized. Each of those provokes its own “how do you know,” which the layer below answers with the labeled evidence. If a section does not answer the question raised by the statement above it, the section is misplaced or padding. This dialogue test, statement, reader’s question, answer, is the fastest structural edit that exists.

Horizontal logic: groupings must be MECE and ordered by logic, not chronology. Ideas sitting side by side must be the same kind of thing, not overlap, and together cover the claim above them. Research drafts violate this constantly by grouping findings in discovery order. The reader does not care what you found first; they care what supports what.

Headings are claims. Every heading in the body is a full assertion someone could dispute, never a topic label. This is the action-title rule from slide work, applied to prose: a reader who reads only your headings should receive the entire argument. “Certification activity” is a filing label. “Certification is complete for our three most profitable segments” is a sentence doing its job.

Notice the symmetry that makes the two pyramids complementary rather than redundant. Building the work, you moved up: data, insight, implication, action. Writing the work, the reader moves down: action and implication first, insights as the key line, data at the base, available on demand. Same pyramid, opposite directions of travel, and the Ladder labels ride along at every level in both.

Variants and the one-page test

Two practical notes. For standing deliverables, weekly signal briefs, escalation notes, invert the order: lead with the Complication, one line of Situation only if the reader needs re-anchoring, then the Answer. Recurring readers do not need the ground re-established. And for any deliverable of any length, the full SCQA belongs on one page or one slide. If the opening needs two pages, the Question has not been found yet.

The complete system, in one line each: the CI Report Pyramid tells you whether you have an argument. The Ladder tells you how honestly it is labeled. SCQA and Minto tell you the order in which a busy, skeptical, powerful reader should meet it. The thinking earns the conclusion. The telling starts with it.

Views are my own and do not represent my employer.

Five Forces, rebuilt: where Porter breaks on a live CI desk, and what to run instead

Sat, 11 Jul 2026 00:00:00 GMT

Porter’s Five Forces is the first framework every analyst learns and the last one most ever question. It deserves better than reverence. Published in 1979, it compressed industrial-organization economics into five questions anyone could ask, and as a teaching device it remains excellent. As an operating tool on a live competitive intelligence desk, it breaks, predictably and in the same five places.

This piece walks the framework force by force against one real environment: the AI accelerator market from 2023 through 2026, the same desk the Signal Triage Matrix piece used. The choice is deliberate. This market is the most consequential competitive arena of the decade, and it violates Porter’s assumptions more clearly than any textbook case. The stance here is not demolition. Every break points at a repair, and the piece ends with the rebuilt version I actually run.

One framing note before the tour. Porter models an industry as the unit of analysis and forces as impersonal pressures on average profitability. Both assumptions fail in modern concentrated markets, and most of the five breaks below are that one failure wearing five costumes.

Force 1: Rivalry among existing competitors. Breaks on the averaging assumption

What Porter asks: how intense is competition among incumbents, and what does it do to industry profitability?

Where it breaks. The question averages an industry that has no meaningful average. In AI accelerators, one firm has held a reported share of the merchant market somewhere near 90 percent through this period, with gross margins in the seventies, while rivals fought for the remainder at a fraction of the profitability. “Industry rivalry is moderate” or “high” is an adjective about a fiction; the profit pool is not an industry attribute, it is one company’s attribute. Any CI answer at the industry level tells your leadership nothing about the only questions that matter: whose pool, defended by what, eroding at what rate.

The repair the rebuild keeps: replace the intensity adjective with named profit pools and dated erosion evidence. Rivalry becomes a claim, for example: “the incumbent’s inference-workload share is being contested by two named actors, evidence attached,” which is falsifiable, labelable on the Fact vs. Inference Ladder, and re-testable next quarter. Adjectives are not claims.

Force 2: Bargaining power of buyers. Breaks on role multiplexing

What Porter asks: can buyers force prices down or quality up?

Where it breaks. Porter assumes a buyer is a buyer. The largest AI-chip buyers, the hyperscalers, are simultaneously the incumbents’ biggest customers, their emerging competitors through custom silicon programs like TPU, Trainium, and Maia, their distribution channel through cloud rental, and in some cases their partners on networking and systems. A single actor occupies four boxes of the framework at once, and the framework has no way to say so. Analyze Google as “a buyer” and you miss that its bargaining power is not negotiation leverage; it is a credible exit into self-supply, which is a different threat with different indicators.

The repair: stop mapping forces and start mapping actors. For each named actor, list every role it plays. Role multiplexing is not an edge case in concentrated technology markets; it is the norm, and it is where the strategic action lives.

Force 3: Bargaining power of suppliers. Breaks on the bargaining abstraction

What Porter asks: can suppliers squeeze the industry on price or terms?

Where it breaks. “Bargaining power” frames supply as a negotiation. In this market, supply is a physical dependency graph with chokepoints: leading-edge fabrication concentrated overwhelmingly in one foundry, advanced packaging capacity that rationed the entire industry’s output for stretches of 2023 and 2024, and high-bandwidth memory from a three-firm oligopoly. The binding question was never “what price will the supplier demand.” It was “who gets allocated capacity at all,” which is a strategy decision by the supplier and a survival variable for everyone downstream. Porter registers a price effect and misses the quantity weapon entirely.

The repair: draw the dependency graph. Nodes are named suppliers and inputs, edges carry two annotations: single-source or multi-source, and allocation-constrained or not. A chokepoint two tiers upstream matters more than any bargaining dynamic one tier up, and only a graph shows it.

Force 4: Threat of new entrants. Breaks on where entrants come from

What Porter asks: how high are the barriers protecting incumbents from new entry?

Where it breaks. Twice. First, the dangerous entrants did not enter the industry; they seceded from it. Hyperscaler silicon programs do not need to win merchant customers, distribution, or brand; they serve a captive internal workload, which means every classical entry barrier is irrelevant to them. Porter’s entrant is a firm trying to join the industry; the real threat was firms trying to leave it. Second, the barrier that actually protected the incumbent barely registers in the framework: a software ecosystem, CUDA, two decades deep, that makes the hardware sticky through developer switching costs. Well-funded chip startups cleared every barrier Porter names, capital, technology, talent, and still struggled against the one he does not.

The repair: the entry question becomes two questions. Who can self-supply and exit the customer base, and what complement, not what asset, do defectors and startups have to replicate? Which points directly at the first missing force below.

Force 5: Threat of substitutes. Breaks on invisible substitution

What Porter asks: what alternative products could displace the industry’s product?

Where it breaks. The framework scans for substitute products. The most violent substitution shock this market received was not a product; it was an efficiency claim. In early 2025, DeepSeek’s reported training-cost figures wiped hundreds of billions from chip-sector market value in a day, on the thesis that better algorithms substitute for compute itself. Whatever one believes about those specific numbers, the mechanism is real and recurring: demand-side technique changes, model efficiency, quantization, smaller specialized models, substitute for the product without any rival product existing. A product scan returns “no credible substitute” right up until the demand curve moves.

The repair: substitution monitoring gets a second lane, technique substitution, with its own tripwires: efficiency benchmarks, cost-per-training-run disclosures, workload migration signals. This lane belongs in the watch log as standing quadrant 2 material.

The three forces Porter never had

Running the rebuilt map requires three additions the 1979 framework structurally lacks.

Complements. CUDA is the single most important competitive fact in this market and has no Porter slot; complements shape moats in every ecosystem business. The rebuild treats complement owners as first-class actors.

The state. Export controls redrew this market’s geography twice in two years, deciding who could sell what into a major market. That was the desk’s loudest quadrant 1 event, and in Porter’s model it is weather. The rebuild lists relevant states as named actors with instruments and dates, not as background “regulation.”

The clock. Five Forces produces a snapshot, and this market invalidated snapshots in quarters, not years. A force assessment without a re-run date and named tripwires is a photograph of a moving object. The rebuild attaches a cadence to the map itself, set by the Triage Matrix.

The rebuilt version: the Actor-Role Grid

What survives from Porter is the checklist instinct: five prompts that stop you from forgetting a pressure source. What gets replaced is the structure. The working version has four rules.

Rule 1: map actors, not forces. Rows are named actors: incumbents, hyperscalers, foundries, memory makers, complement owners, states, credible technique disruptors. Columns are roles: buyer, supplier, rival, self-supplier, complement owner, rule-setter, technique substitute. Most interesting actors fill multiple cells, and the multi-cell actors are the analysis.

Rule 2: cells contain claims, not adjectives. Every filled cell carries a dated piece of evidence and a Ladder label. “High supplier power” is banned; “packaging capacity allocated by supplier X constrained incumbent output in H1, [F], filing dated, re-check at next earnings” is the standard. The grid is only as honest as its labels.

Rule 3: attach the clock. The grid carries a re-run cadence and a short tripwire list per volatile cell. In fast markets the cadence is quarterly; a tripwire firing re-runs the affected cells immediately.

Rule 4: demote Porter to QA. After the grid is built, run the original five questions once, as a completeness check: did any buyer, supplier, rival, entrant, or substitute pressure get missed? Porter ends the analysis instead of beginning it. That is the correct place for a 47-year-old checklist: not wrong, just finished being the tool and ready to be the test.

The five questions were never the problem. The problem was letting a framework built for stable, boundaried, mid-century industries define the unit of analysis in markets where the biggest customer is also the next competitor, the binding constraint sits two tiers upstream, and the substitute is an equation. Map the actors, label the evidence, date the map. Keep Porter where he now belongs: at the end, checking your work.

Views are my own and do not represent my employer.

The division of labor: an AI-augmented CI workflow that keeps the judgment human

Thu, 09 Jul 2026 00:00:00 GMT

AI has changed the economics of research. Reading a hundred documents used to cost a week; now it costs a prompt. What it has not changed is where credibility comes from. A competitive intelligence deliverable is trusted because a named person verified the facts, made the inferences, and signed the labels. No model can carry that accountability, and pretending it can is how CI functions will embarrass themselves this decade.

So the useful question is not “should we use AI” or “can AI do CI.” It is a staffing question: for each step of the workflow, what gets delegated to the machine and what stays with the analyst. This guide walks the full CI cycle, triage, scoping, collection, synthesis, drafting, and QA, and draws that line at every step. One principle governs all six:

AI gets breadth, speed, and mechanical checks. The human keeps definitions, labels, inferences, and anything that will be signed.

If you are new to CI, three frameworks appear throughout and take one sentence each. The Signal Triage Matrix decides which incoming signals deserve analysis, by impact and urgency. The Fact vs. Inference Ladder labels every claim by how much weight it can bear, from Verified Fact [F] down to Speculation [SP]. The CI Report Pyramid says a deliverable must climb from data to insight to implication to a recommended action.

Step 1: Triage. AI clusters the noise; the human places the quadrants

Hand to AI: the volume problem. Feed the day’s inbound, alerts, filings, press releases, forum chatter, into a model and have it deduplicate, cluster related items, and produce one-line summaries per cluster. This is the work that used to eat the first ninety minutes of every morning, and models are genuinely good at it.

Keep human: the two triage judgments, impact if true and time-to-relevance. These depend on knowing your company’s exposure, your quarter’s decisions, and your leadership’s blind spots. None of that is in the model’s context, and a model asked to rate “importance” will rate coverage volume, which is precisely the bias triage exists to cure.

Example. A pump manufacturer’s CI inbox catches 40 items overnight. The model collapses them to nine clusters and flags that six articles about a competitor’s “major expansion” all trace to one press release. The analyst reads nine lines instead of 40 items, then places the expansion story herself: high impact, low urgency, scheduled deep dive. The machine compressed; the human decided.

Step 2: Scoping. The human owns the definition; AI attacks it

Keep human: the definition of the question. Market boundaries, units of measure, time windows, whose revenue counts. Every downstream number inherits these decisions, and they encode judgment about what the client actually needs. Delegating the definition is delegating the deliverable.

Hand to AI: the stress test. Once you have written the scope, have a model generate edge cases and ambiguities: does the definition include retrofits, does it double-count distributors, what adjacent categories could a reader assume are in. Models are excellent at producing the annoying questions a good reviewer would ask, in seconds instead of at the review.

Example. An analyst scopes “the European market for warehouse robotics, 2026, manufacturer revenue.” The model returns twelve boundary questions, including whether software subscriptions count as manufacturer revenue and whether a robot built in Asia but deployed in Poland is in scope. Two of the twelve force a real revision. Ten minutes, one avoided rebuild.

Step 3: Collection. AI finds and extracts; the human opens the source

Hand to AI: breadth. Locating candidate sources, extracting every mention of a product line from five years of transcripts, pulling segment tables out of filings, summarizing a 300-page regulatory decision to find the ten relevant pages. Extraction and location are where AI multiplies a researcher most.

Keep human: verification and tier judgment. Two hard rules. First, nothing enters the evidence base as a Verified Fact until a human has opened the underlying document; a model quoting a filing is a Reported Claim from a fallible narrator, not the filing. Models fabricate citations that look immaculate, and the failure mode is not frequent, it is quiet, which is worse. Second, the human assigns the source tier. A model will paraphrase a template-factory market report and a specialist research house in the same confident register, laundering tier 5 content into your notes with the tone of tier 2.

Example. Asked for a competitor’s stated capacity plans, the model returns five quotes with dates and document names. The analyst opens all five sources: four check out exactly, one attributes a real-sounding sentence to an earnings call where it does not appear. Four verified facts, one caught fabrication, and the check took twenty minutes. That ratio, high yield with a nonzero fabrication rate, is exactly why the opening-the-source rule is unconditional.

Step 4: Synthesis. The human makes the inference; AI red-teams it

Keep human: the inference itself, and its Ladder label. Connecting separate facts into a finding is the analyst’s core act of judgment and the thing a signature vouches for. A model handed four data points will produce a fluent narrative connecting them; fluency is not corroboration, and a model’s confidence level carries no information about truth.

Hand to AI: the adversarial pass. Give the model your facts and your draft inference and ask for competing explanations that fit the same evidence, plus what evidence would distinguish them. This is the cheapest red team ever built, and it directly generates the promotion tests that weak claims need.

Example. Four facts about a logistics competitor: warehouse leases in two new regions, a hiring spike in cold-chain roles, a partnership with a grocery chain, new temperature-monitoring patents. Draft inference: entry into grocery fulfillment. The model offers two rivals: serving one large pharma contract, or building capacity to resell. It suggests the discriminating evidence, whether the leases include food-grade certification. The analyst checks, confirms, and labels the finding a Corroborated Inference with a better foundation than before the challenge.

Step 5: Drafting. AI compresses; the human writes what gets signed

Hand to AI: structure and compression. First-pass ordering of the evidence, tightening a 90-word paragraph to 40, generating six candidate action titles for a slide, converting a report section into speaker notes. Mechanical writing work, high volume, low stakes.

Keep human: the top two Pyramid layers. Implications name what the finding means for your company specifically, which requires context the model does not have. Actions name an owner, a move, and a deadline, which requires authority the model does not carry. And every claim label in the final text is placed or confirmed by the analyst, because the labels are the signature.

Example. The analyst hands the model a verified evidence block and asks for five action-title candidates. Four are grammatical restatements of the topic; one contains a usable falsifiable sentence. She rewrites that one, sharpens the number in it, and writes the decision request herself. The model contributed a draft; the argument is hers.

Step 6: QA. AI runs the sweeps; the human owns Band A

The 10-point pre-delivery checklist splits cleanly along the same line.

Hand to AI: the mechanical bands. Sweep for template ghosts, placeholder text, and internal notes. Check that every figure carries an as-of date. Flag any chart whose axis does not start at zero. Reconcile units across the document and flag any number that appears twice with different values. Models are tireless at exactly the checks humans skip when the hour gets short.

Keep human: the career-killer band. Tracing facts to opened sources, rebuilding the arithmetic, verifying entities, auditing the claim labels. These are the checks where the checker’s name is the point; a QA step exists to attach accountability, and accountability cannot be delegated to a system that cannot be fired, deposed, or embarrassed.

Example. Before a sizing ships, the model flags that a growth figure appears as 14 percent on page 2 and 12 percent on page 9, and that one chart implies a 2025 data point the text dates to 2024. Both are real catches. The analyst then rebuilds the central calculation herself in a fresh sheet, because that check was never the machine’s to sign.

The pairing table, compressed

Across all six steps the same verbs sort cleanly. Delegate: cluster, summarize, extract, locate, stress-test, red-team, compress, sweep. Retain: define, place, verify, tier, infer, label, recommend, sign.

If a task’s verb is in the first list, automate it without guilt. If it is in the second and you catch yourself delegating it, you are not augmenting the workflow, you are outsourcing the accountability while keeping the byline.

Two closing rules make the whole system safe. First, the asymmetry rule: AI failures in the delegate column cost you minutes; undetected AI failures in the retain column cost you the credibility discount that never comes off. Price the delegation accordingly.

Second, the disclosure habit: note in your methods what the machine did, the same way you note your sources. Analysts who hide the AI use the AI badly; the ones who declare it have usually built the checks this guide describes.

The machine reads faster than you ever will. It cannot be responsible for anything. Build the workflow on both truths at once.

Views are my own and do not represent my employer.

The last hour: a 10-point QA checklist that runs before any CI deliverable ships

Mon, 06 Jul 2026 00:00:00 GMT

The most expensive errors in competitive intelligence are not analytical. They are mechanical: an untraced number, a mislabeled claim, a chart that says something the data does not. Analysis errors get debated; mechanical errors get remembered. One shipped mistake of the wrong kind, and every future deliverable you send gets read with a discount applied.

This checklist is the last hour before anything ships: a report, a slide, a market sizing, a two-paragraph escalation note. Ten checks, ordered by failure cost, not by workflow. The career-killers come first, because if time runs out, the bottom of the list is where you are allowed to cut corners. The top is not.

Each point carries a real, public example of the failure it exists to catch. None of these people were stupid. All of them skipped a check.

Band A: Career-killers. If one of these ships, the deliverable is dangerous

1. Trace every load-bearing fact to a source you opened yourself

Not a source someone cited. Not three articles that agree. The document itself. The specific failure this catches is false corroboration: multiple reports that feel independent but trace back to a single origin.

The example. The 2003 case for Iraqi mobile bioweapons labs rested heavily on one defector, codenamed Curveball. His claims appeared to be corroborated across multiple intelligence reports; in reality the “corroborating” streams traced back to the same man, relayed through different channels. The claims reached the UN Security Council as established fact. Three news articles citing the same wire story is the everyday CI version of the same disease. The check: for every fact under a conclusion, name the primary document and confirm you opened it.

2. Rebuild the arithmetic from raw inputs

Any number you calculated gets recalculated, from the source figures, ideally by a second person or at minimum by yourself in a fresh sheet. Formulas drift, ranges exclude rows, and models inherit errors silently.

The example. Reinhart and Rogoff’s 2010 paper on debt and growth anchored austerity arguments across multiple governments. In 2013, a graduate student attempting to replicate it found, among other issues, an Excel range error that excluded five countries from a key average. The headline result weakened substantially once corrected. The paper had been cited for three years. Nobody had rebuilt the arithmetic.

3. Check units and definitions at every handoff

Whenever a number crosses a boundary, between analysts, between a source and your model, between currencies, between manufacturer revenue and end-user spend, verify that both sides mean the same thing. The 48-hour sizing treated this as the first two hours of work; QA re-tests it in the last hour.

The example. NASA’s Mars Climate Orbiter was lost in 1999 because one engineering team delivered thruster data in pound-force seconds while the receiving system expected newton-seconds. Both teams’ work was internally correct. The handoff was not, and review after review failed to test the interface. A $125 million spacecraft burned up over a unit label.

4. Verify the entity

Confirm that every company, ticker, subsidiary, and product name refers to the thing you think it does. Parent versus subsidiary, similarly named firms, and stale tickers are where this fails.

The example. In early 2020, shares of Zoom Technologies, a tiny OTC company with the ticker ZOOM, surged hundreds of percent as investors piled into what they believed was Zoom Video Communications, which trades as ZM. The SEC eventually suspended trading in the lookalike. Thousands of people committed real money to the wrong entity. A CI report can do exactly the same thing with a segment figure pulled from the wrong company’s filing.

5. Audit the claim labels

Run the Fact vs. Inference Ladder over every analytical sentence. The specific check: is anything wearing a [F] that is actually a reported claim or an inference? Label inflation is the single fastest way to ship a falsehood while feeling rigorous.

The example. In April 2013, a hacked Associated Press account tweeted that explosions at the White House had injured the president. Automated and human traders treated a reported claim, from a normally reliable source, as verified fact. The S&P 500 shed roughly $130 billion in market value in about three minutes before the claim collapsed. The source’s reputation promoted the claim one rung too high, which is precisely the promotion your labels exist to block.

Band B: Credibility erosion. These survive the meeting and damage you afterward

6. Stress-test the scope behind any market number

Every market size, share, or TAM figure gets one question: what exactly was counted, and would the reader count the same things? A defensible number with a stated scope beats a large number with a flattering one.

The example. WeWork’s 2019 IPO filing presented a market opportunity ranging up to roughly $1.6 trillion, reached in its broadest cut by treating vast populations of urban desk workers as potential members. The arithmetic was fine; the scope was the fiction, and it became a public symbol of the whole prospectus’s credibility problem. Reviewers may not catch a scope inflation in the room. They catch it later, and they remember whose slide it was on.

7. Date-stamp everything and hunt for staleness

Every figure carries an as-of date, and the check asks: is newer data available, and would it change the finding? Data does not announce its own expiry.

The example. “Dewey Defeats Truman,” 1948. The famous headline traces to polling that had largely stopped weeks before the election, on the assumption that preferences were stable. The data was accurate when collected and wrong when used. A 2024 headcount figure in a 2026 deck is the same failure wearing business casual.

8. Run the chart integrity pass

Axes start where they claim to, categories run in a defensible order, baselines are honest, and the visual impression matches the underlying table. A chart is a claim; QA treats it like one.

The example. In May 2020, Georgia’s Department of Public Health published a COVID-19 chart in which dates and counties were reordered non-chronologically, producing a clean visual decline that the underlying data did not show. The department apologized and corrected it, but the chart had already done its damage in public. Most chart crimes in CI are subtler and unintentional. The pass exists because intent does not change what the reader takes away.

Band C: Finish-line polish. Cheap to fix, embarrassing to ship

9. Sweep for template ghosts and internal notes

Search the document for placeholder text, tracked changes, comments, another client’s name, and anything written for internal eyes. This is a two-minute mechanical sweep, and it catches the errors that make a deliverable look careless regardless of the analysis underneath.

The example. A 2016 peer-reviewed ecology paper shipped with an internal author note left in the published text, asking whether to cite a rival team’s “crappy” paper. It survived the authors, the reviewers, and the editors, and it is what that paper is now remembered for. Every profession has its version. The sweep costs less than the anecdote.

10. The cold read

Someone who has not seen the deliverable reads only the title and the final recommendation, then tells you what they think it says and what they would do. If their answer does not match your intent, the deliverable is not done, no matter how good the middle is. If no second person is available, the fallback is a timed break and a printed read, but the independent reader is the real check.

The example. Hawaii’s January 2018 false missile alert reached every phone in the state because the process allowed one operator to ship a state-wide message with no independent confirmation step. It took 38 minutes to correct. The lesson is not about interfaces; it is that anything important enough to send is important enough for a second pair of eyes before it goes.

Running the checklist

Three rules make it operational. First, run it top down: if the hour shrinks, Band C is sacrificed before Band B, and Band A is never sacrificed. Second, the checker signs. A named person confirming “traced, rebuilt, labeled” converts QA from a vibe into an accountability step, the same way the Ladder converts confidence into labels. Third, log what each pass catches. A running tally of near-misses tells you which checks your team actually needs, and it turns the checklist from ritual into feedback.

The checklist pairs with the rest of the system in one sentence: the Triage Matrix decides what gets analyzed, the Ladder and the Pyramid govern how the analysis is built, and this list is the gate it passes through on the way out the door. The analysis makes you smart. The last hour keeps you credible.

Views are my own and do not represent my employer.

The Source Stack: a reliability-tiered checklist for secondary research

Sat, 04 Jul 2026 00:00:00 GMT

Ask a junior analyst where they research and you get a list of websites. Ask a senior one and you get a hierarchy. The difference matters because sources are not interchangeable: each sits at a different distance from the underlying fact, and that distance determines how much weight a claim built on it can bear. This checklist organizes the full secondary research arsenal, free and paid, into five reliability tiers. The rule that makes it a system rather than a list: start at the top and descend only for what the tier above cannot answer. Most weak research is not built on bad sources; it is built on tier 4 answers to tier 1 questions.

The tiers map directly onto the Fact vs. Inference Ladder. Tier 1 material can support Verified Fact claims. Everything below it enters your report as a Reported Claim at best, until corroborated.

Tier 1: Primary disclosures. The company speaking under obligation

These are statements made under legal, regulatory, or contractual compulsion, which is what makes them the bedrock. Cost: almost all free.

Regulatory filings (SEC EDGAR, Companies House, local registries). What they answer: revenue, margins, segments, risk factors, ownership, executive pay. The trap: the MD&A section is management’s narrative wearing a filing’s credibility; separate the audited numbers from the framing around them.

Earnings call transcripts (company IR pages, free aggregators). What they answer: strategy language, guidance, and, most valuably, the Q&A, where executives answer questions they did not script. The trap: prepared remarks are positioning; treat every adjective as a Reported Claim.

Patents (USPTO, Espacenet, Google Patents). What they answer: where R&D money actually went, often 18 months before products ship. The trap: filing volume is not strategy; companies patent defensively and abandon freely. Look for citation clusters and continuation patterns, not counts.

Court and regulatory proceedings (PACER, competition authority decisions). What they answer: contract terms, supplier relationships, and internal documents that surface in discovery; some of the richest CI material in existence. The trap: litigation documents contain adversarial framing; the exhibits are gold, the pleadings are argument.

Trade and customs data (import/export records, paid via Panjiva or ImportGenius). What they answer: who ships what to whom, in what volumes; physical reality that marketing cannot spin. The trap: coverage varies sharply by country and mode of transport.

Job postings and procurement notices (company career pages, government tender portals). What they answer: capability building before it is announced; a company hiring 14 field technicians in one region is telling you its plans. The trap: postings signal intent, not execution; roles get posted and never filled.

Tier 2: Specialist data and research. People paid to be right about one thing

Named analysts, visible methodologies, franchises staked on a specific coverage area. Cost: mostly paid, and mostly worth it when the coverage matches your question.

Specialist research houses (Dell’Oro for network infrastructure, IDC and Gartner for enterprise tech, IHS Markit for industrials, Wood Mackenzie for energy). What they answer: market sizes, shares, and forecasts with a stated scope. The trap: even good houses define markets differently; never mix two houses’ numbers in one calculation without reconciling scopes.

Financial data platforms (Capital IQ, Bloomberg, FactSet, Refinitiv). What they answer: comparables, ownership, transactions, estimates, all normalized and fast. The trap: the normalization itself; a platform’s “EBITDA” may not be the filing’s EBITDA. Spot-check against the source document before a number becomes load-bearing.

Private market databases (PitchBook, Crunchbase, Tracxn). What they answer: funding, valuations, investor networks for companies that file nothing. The trap: self-reported and stale data; a two-year-old headcount figure presented as current.

Industry associations and trade bodies. What they answer: production statistics, capacity data, standards activity; often the only volume data in unglamorous B2B markets. The trap: associations exist to advocate for their members; their totals are solid, their outlooks are lobbying.

Tier 3: Quality journalism and informed commentary

Professional accountability without regulatory compulsion. Cost: subscriptions, cheap relative to value.

Wire services and financial press (Reuters, Bloomberg News, FT, WSJ). What they answer: events, deals, executive moves, with editorial verification behind them. The trap: “people familiar with the matter” is a Reported Claim by construction; the outlet’s credibility does not promote the anonymous source’s claim to fact.

Trade press (the two or three publications every industry insider actually reads). What they answer: operational detail and personnel moves the financial press ignores. The trap: small outlets depend on vendor advertising and access; watch for coverage that never criticizes anyone.

Sell-side research. What they answer: deep company models, channel checks, management access. The trap: structural bias is well documented; use the data and the questions analysts ask, discount the ratings and the price targets.

Expert networks (GLG, AlphaSights, Guidepoint). Technically primary conversation, priced like a luxury good. What they answer: the texture no document contains; how deals really get won, why customers really churn. The trap: one expert is one anecdote wearing authority; treat every call as Single-Source Inference until a second, independent expert corroborates.

Nobody here is accountable for accuracy, but the aggregate patterns are real. Cost: free.

Employee review sites (Glassdoor, Blind). What they answer: morale trajectories, reorganizations, leadership problems, months before they surface elsewhere. The trap: selection bias is extreme; read the trend across fifty reviews, never the content of five.

Product review platforms (G2, Capterra, app stores, Amazon reviews for physical goods). What they answer: why customers actually churn, feature gaps, pricing friction. The trap: vendors seed positive reviews; weight the negative and neutral ones, which nobody pays for.

Forums and communities (Reddit, Stack Overflow, industry Discords, LinkedIn commentary). What they answer: practitioner sentiment, early product problems, hiring market chatter. The trap: a loud thread is not a trend; five posts can be one motivated person.

Tier 4 material enters reports only as corroboration or as an early-warning flag for something to verify upstream. It never stands alone under a conclusion.

Tier 5: Commodity content. The tier you name so you can refuse it

SEO-optimized market reports from template factories, aggregator articles rewriting other articles, AI-generated industry summaries. The tell, as the 48-hour market sizing demonstrated: precise-looking numbers, invisible methodology, and estimates that disagree by 2x across publishers without explanation. Cost: cheap, which is the problem. Legitimate uses: discovering what vocabulary an industry uses, finding names of players to research properly, and nothing else. No number from tier 5 ever enters a model.

Running the stack

Three habits turn the tiers into practice. First, match the tier to the claim: a Verified Fact label requires tier 1; a market size deserves tier 2 plus your own bottom-up check; tier 4 alone never supports anything above Single-Source Inference. Second, descend deliberately: when you catch yourself citing a tier 3 article for a number that lives in a tier 1 filing, go get the filing; it takes ten more minutes and removes one layer of retelling. Third, log the tier with the source: a source list annotated by tier lets any reviewer audit your evidence mix in thirty seconds, the same way the Ladder lets them audit your labels.

The stack will not make research faster. It makes it defensible, and defensible is what survives the meeting.

Views are my own and do not represent my employer.

From update to argument: one competitor slide, rebuilt to MBB grade

Thu, 25 Jun 2026 00:00:00 GMT

Every junior analyst has made this slide. Most senior ones still do. It is the “Competitor Update” slide: dense, dutiful, and dead on arrival. This teardown takes one such slide, the kind produced in thousands of CI functions every Monday, and rebuilds it fix by fix. The content is fictional but the sins are drawn from life.

The before slide

Picture it. The title reads: “Competitor Update: Veltrax Flow Systems, Q2 2026.”

Below it, three zones fight for attention. Left: six bullet points. “Veltrax announced a distribution agreement with a Texas MRO supplier.” “14 field service technician openings posted across Gulf Coast.” “New price list shows 6% reduction on mid-range centrifugal pumps.” “CEO mentioned aftermarket revenue 4x on earnings call.” “Revenue grew 8% YoY in Q1.” “New CMO hired from a industrial software company.”

Right: a clustered bar chart titled “Veltrax Revenue by Segment, 2022 to 2025,” four years, five segments, twenty bars, a legend with five colors, and a source line in 6-point font.

Bottom: a text box labeled “Implications,” containing: “Veltrax continues to invest in its service capabilities. We will continue to monitor developments.”

Everything on this slide is accurate. Nothing on it is useful. Let us fix it in six moves.

Fix 1: The title must be the finding, not the topic

“Competitor Update: Veltrax, Q2 2026” is a filing label, not a message. The MBB convention is the action title: a full sentence stating the one thing this slide proves. If a reader saw only the titles of your deck, they should get the entire argument.

The test: does the title survive as a standalone sentence someone could agree or disagree with? “Competitor Update” cannot be disagreed with. Rewrite:

“Veltrax is pivoting from selling pumps to selling uptime, putting our Gulf Coast aftermarket revenue at risk within 18 months.”

Note what happened: the title is an inference, so the slide must now carry its labeling. That is not a burden; that is the point. A title that risks being wrong is a title that says something.

Fix 2: One slide, one message; everything else goes

The before slide carries six facts, a five-segment chart, and a vague implication: at least three separate messages. The revenue growth bullet and the CMO hire belong to different stories. The single-message rule is brutal: everything on the slide either supports the title or leaves the slide.

Apply it. The distribution deal, the technician hiring, the price cut, and the earnings language all support the pivot thesis; they stay. Revenue growth of 8% supports nothing specific here; it moves to a backup slide. The CMO hire is a different signal on a different timeline; it gets its own slide or the watch log. Cutting content feels like losing work. It is actually the work.

Fix 3: Kill the chart crime, choose the chart that argues

The twenty-bar segment chart answers a question nobody asked. It exists because the data existed. A chart on an MBB-grade slide is not decoration; it is a witness called to prove the title.

What would prove “pivot to services”? One comparison: service-related signals over time. Rebuild as a single, minimal visual: four converging evidence streams on one timeline, hiring, distribution, pricing, and earnings language, each plotted as a dated marker, all pointing at the same six-week window. One color for evidence, one accent for the inference they converge on. No legend needed if the labels sit on the data. The chart now performs the argument: separate facts, one direction.

The general rule: if your chart needs a legend with five entries, you have not decided what it is for.

Fix 4: Structure the body as evidence, not as a feed

Bullets ordered by arrival date are a feed. Evidence ordered by logic is an argument. The rebuilt body has three tiers, mirroring the CI Report Pyramid:

Evidence line (the four facts, each dated and source-tagged, each labeled [F]). Inference line (one sentence: “Together these point to a strategy shift toward aftermarket services, labeled [CI], corroborated inference from four independent streams”). Stake line (why we care: “Our three largest Gulf Coast accounts already run Veltrax units; regional aftermarket is roughly 30% of margin”).

Three tiers, maybe forty words total. The reader’s eye travels fact to inference to stake in one pass, which is exactly the climb the Pyramid framework demands, compressed onto one page.

Fix 5: Label the epistemics on the slide itself

MBB conventions and CI honesty meet here. The before slide presented facts and the phrase “continues to invest” with identical visual weight, letting the reader assume everything was equally solid. The after slide tags claims inline: [F] on each fact, [CI] on the inference in the title’s support line. Two-character tags, footnoted once.

This looks like humility. It functions as authority: the reader learns your untagged claims never need checking, and your tagged inferences show your reasoning instead of hiding it. Precision compounds into trust; a slide is just the smallest unit where that compounding starts.

Fix 6: Replace “monitor developments” with a decision request

“We will continue to monitor developments” is the analyst handing the work back to the reader. An MBB-grade slide ends with a so-what box that requests a decision or names a next step with an owner and a date.

Rewrite: “Decision requested: approve pre-emptive multi-year service offers to our three exposed accounts before end of quarter. CI re-tests this thesis in 90 days against technician onboarding and published contract wins.”

Now the meeting has something to do. The slide can be accepted, rejected, or amended, which means it can succeed or fail, which means it matters.

The after slide, assembled

Title: the falsifiable finding. Left two-thirds: the convergence timeline, four dated markers, one inference callout. Right third: the three-tier evidence block with Ladder tags. Bottom strip: the decision request and the 90-day re-test. Generous white space; one message; perhaps sixty words on the entire page against the before slide’s two hundred.

The before slide reported that things happened. The after slide argues that something is happening, shows its evidence, admits what is inferred, and asks for a move. Same facts. Different profession.

The 60-second slide audit

Five questions, run on any slide before it ships:

Is the title a full sentence someone could disagree with?
Does every element on the page support that sentence?
Does the chart prove the title, or merely accompany it?
Can the reader tell fact from inference without asking you?
Does the slide end by requesting a decision or naming an owner and a date?

Five yeses and the slide argues. Fewer, and you know exactly which fix to apply.

Views are my own and do not represent my employer.

The 48-hour market sizing: data-center liquid cooling, and why the first page of Google got discarded

Tue, 16 Jun 2026 00:00:00 GMT

Method note: Self-initiated, public information only, unrelated to any client work. Every claim carries a Fact vs. Inference Ladder label: [F] verified fact, [RC] reported claim, [CI] corroborated inference, [SI] single-source inference, [SP] speculation. Figures current as of early July 2026.

Market sizing has a dirty secret: most “48-hour” sizings are 30 minutes of googling and 47.5 hours of formatting. The analyst grabs the first market report a search engine serves, wraps the headline number in a chart, and ships it. This walkthrough shows what the 48 hours are actually for, using a market that is genuinely exploding: liquid cooling for data centers.

The client question, framed as a real one would be: How big is the data center liquid cooling market in 2026, and how fast is it growing? The honest answer takes two days precisely because the fast answer fails the first test any serious number must pass: knowing where it came from.

Hours 0 to 2: Define the market before you size it

Method callout: the definition is the sizing. Every downstream number inherits your scope decisions. For liquid cooling: hardware only (cold plates, CDUs, manifolds, piping), or hardware plus installation and services? Manufacturer revenue or end-user installed spend? Direct-to-chip and immersion only, or rear-door heat exchangers too?

We scope it as: global, calendar 2026, direct liquid cooling for data centers, measured two ways, manufacturer revenue and end-user spend. Holding both measures is deliberate; they answer different client questions, and conflating them is the most common sizing error in circulation.

Hours 2 to 6: The discard phase

Search the market. The first page returns a wall of near-identical reports from commodity research houses, each with a precise-looking headline figure, a CAGR to one decimal place, and a purchase link. Put them side by side and the tell appears immediately: their estimates for the same market in the same year differ by more than a factor of two, with no visible methodology to explain why.

Method callout: a number you cannot trace is not evidence, it is decoration. Run every source through three questions before a single figure enters your model. One, can you see the methodology, or at least reconstruct what was counted? Two, does the publisher have specialist depth in this domain, or does it publish the same template across four hundred markets? Three, does the source have skin in the game, analysts whose reputation rides on this specific coverage area? The commodity reports fail all three. They are not malicious; they are optimized to rank in search, not to be right. Two-times disagreement with no explanation is not a spread to average, it is a signal to discard the whole tier.

What survives the filter is a short list: Dell’Oro Group, a specialist infrastructure research house with named analysts covering this exact market; BloombergNEF for capacity data; company primary disclosures, earnings calls, capex guidance, and product specifications; and practitioner cost benchmarks, used with explicit flags. Four hours spent throwing away the obvious answer is the highest-leverage block in the entire sizing.

Hours 6 to 12: Build the trusted factual base

The specialist read. [RC] Dell’Oro expects the liquid cooling market to roughly double in 2025, reaching close to $3 billion in manufacturer revenue, and to scale toward approximately $7 billion by 2029. This is a reported claim, not a fact, but it is a claim from a house with named analysts, a stated scope (manufacturer revenue), and a franchise staked on this coverage. One credible number with a visible definition beats six incompatible numbers without one.

The physical driver. [F] NVIDIA’s GB200 NVL72 rack draws roughly 120 to 132 kW under load; air cooling is not viable at that density, and direct liquid cooling is required. [F] Microsoft confirmed in December 2025 that all future Maia accelerator deployments will use a liquid-default architecture. [RC] Dell’Oro projects leading-edge GPU thermal design power exceeding 4,000 W by 2029. The direction of thermal physics is not in dispute.

The capacity wave. [F] Over 23 gigawatts of data center IT capacity was under construction globally at the end of September 2025, per BloombergNEF, roughly three quarters of it in the US. [F] The four largest hyperscalers guided to approximately $630 billion in combined 2026 capex on their earnings calls, up from $388 billion in 2025. These come from filings and calls, checkable by anyone.

Hours 12 to 24: The bottom-up build

Method callout: never ship a top-down number you cannot rebuild from physical units. Three steps.

Step 1, new capacity. [CI] From the 23 GW under construction and typical 12-to-24-month build timelines, we estimate 12 to 18 GW of new IT capacity energized during calendar 2026. Inference from two facts; labeled.

Step 2, liquid share. [CI] With flagship AI racks liquid-mandatory and at least one hyperscaler liquid-default, we estimate 35 to 50 percent of 2026 additions ship liquid cooled. The mandatory-cooling facts beneath this are solid; the percentage is our judgment and is the model’s soft spot.

Step 3, unit cost. [SI] Practitioner deployment guides benchmark liquid cooling at $1,000 to $2,000 per kW of cooling capacity for equipment, with installed direct-to-chip costs quoted higher at the facility level. Single-source-grade benchmarks, flagged, not audited.

The arithmetic. 12 to 18 GW, times 35 to 50 percent, gives roughly 4 to 9 GW of liquid-cooled load added in 2026. At the equipment benchmark that is $4 to 9 billion in the central scenarios before retrofits, which the model deliberately ignores. The range is wide, and it should be; a bottom-up model that produces a suspiciously tight range is usually hiding its assumptions rather than lacking them.

Hours 24 to 48: Triangulate, stress, write

Now check the two independent methods against each other. Dell’Oro’s trajectory, roughly $3 billion in 2025 and doubling-era growth, implies manufacturer revenue in the $4 to 5 billion range for 2026. The bottom-up build, run at equipment-level costs and conservative liquid share, lands its lower half in the same corridor; run at installed costs, it points to total end-user spend meaningfully above that, plausibly $6 to 9 billion. [CI] The reconciled finding: roughly $4 to 5 billion in 2026 manufacturer revenue, with total end-user installed spend materially higher, and a credible path to Dell’Oro’s ~$7 billion manufacturer figure by 2029. Two methods, independently sourced, failing to disagree. Honest caveat: this triangle has two legs, not three; a supply-side check against vendor segment revenue is the named gap, and closing it is the first task of any follow-on week.

[SP] The labeled speculation for the watch list: NVIDIA’s roadmap points toward 600 kW to 1 MW racks in the Rubin era. If that ships on schedule, cooling stops being a procurement line and becomes co-designed infrastructure, and every current forecast is too low. Watch indicator: two-phase cooling moving from pilots to volume purchase orders.

The deliverable, built as a Pyramid: Data, the trusted base above, every claim labeled. Insight, one sentence: the market is $4 to 5 billion in 2026 manufacturer revenue, growing toward a doubling by 2029, and the first page of search results was off by up to 2x because it measures different things without saying so. Implication and Action, written for whoever commissioned it, with a 90-day re-test against Dell’Oro’s next update and hyperscaler capex revisions.

Total time: 48 hours. Time spent finding numbers: about four. Time spent deciding which numbers deserved to exist in the model: about eight. That second block is the craft, and it is the one the 30-minute version skips.

Views are my own and do not represent my employer.

CI teardown: Spotify in 2026, a full-spectrum read from public information only

Sat, 06 Jun 2026 00:00:00 GMT

Method note: This analysis is self-initiated, uses only public information, and is unrelated to any client work. Every claim is labeled using the Fact vs. Inference Ladder: [F] verified fact, [RC] reported claim, [CI] corroborated inference, [SI] single-source inference, [SP] speculation. The report itself follows the CI Report Pyramid: data, insight, implication, action. Facts current as of early July 2026; primary sources are Spotify’s quarterly filings and press releases, plus dated wire reporting.

Why this signal cluster earned a deep dive

Run the Signal Triage Matrix first. Spotify in mid-2026 is a quadrant 2 case: high impact, moderate urgency. The company just crossed three quarters of a billion users while changing its leadership structure, its pricing, and its relationship with the live music business, all within about twelve months. No single one of those signals demands a same-day brief. Together they justify a scheduled deep dive, which is what this is.

Layer 1: The factual base

Scale and money. [F] Spotify ended Q1 2026 with 761 million monthly active users, up 12 percent year over year, and 293 million Premium subscribers, up 9 percent. [F] Quarterly revenue was 4.53 billion euros, gross margin hit a Q1 record of 33.0 percent, and operating income reached a record 715 million euros. [F] The prior quarter, Q4 2025, added a record 38 million MAU against guidance of 32 million. [F] The company held 8.8 billion euros in cash and short-term investments at the end of Q1 2026, after buying back 306 million euros of shares. This is a structurally different company from the one that spent most of its life losing money.

Leadership. [F] Founder Daniel Ek moved to executive chairman, with Alex Norström and Gustav Söderström serving as co-CEOs from the start of 2026. [RC] Norström has framed 2026 as the “Year of Raising Ambition,” following what the company called the “Year of Accelerated Execution.” That is the company’s own narrative about itself; treat it as positioning, not as evidence.

Pricing. [F] Spotify raised US subscription prices in early 2026, its third increase in four years. [F] Subscriber growth continued through the hike: the company added 3 million Premium subscribers in Q1 2026 despite the recent US price increases.

Product moves. [F] In September 2025, Spotify launched lossless audio as a standard perk for existing subscribers, ending years of speculation that hi-fi would anchor a premium tier. [F] In June 2026, Spotify launched “Reserved by Spotify,” a system that holds two concert tickets for selected superfans of an artist before general sales open, in partnership with Live Nation, US-only at launch. [F] Selection is algorithmic, based partly on streams and shares, and Spotify is deliberately not disclosing the full criteria.

The tier that has not shipped. [RC] Bloomberg and the Financial Times reported in early 2025 that Spotify was preparing a “Music Pro” superfan add-on at up to 5.99 dollars per month, featuring early ticket access, AI remix tools, and higher-fidelity audio. [F] As of mid-2026, no such tier has launched. [RC] Ek told analysts in 2025 that the company still needed “partners to come to the table” on the super-premium offering.

Layer 2: The insight, what the facts add up to

Insight 1: Spotify has dismantled its own rumored premium tier and is shipping the pieces separately. [CI] Look at the sequence. The reported Music Pro bundle had three pillars: lossless audio, ticket access, and remix tools. Lossless shipped in September 2025 as a free upgrade for everyone. Ticket access shipped in June 2026 as Reserved, an engagement-gated perk rather than a paid one. Two of the three pillars have now been released outside any paid tier. The corroborated inference: the original bundle, as reported, is dead or fundamentally redesigned, most likely because [CI] giving lossless away removed the tier’s most legible feature, and because label negotiations, the “partners at the table” problem, made a rights-heavy bundle slow to assemble.

Insight 2: The strategic center of gravity is shifting from selling access to music toward owning the fan relationship. [CI] Multiple independent streams point the same way. Reserved inserts Spotify between fans and Live Nation’s ticketing machine, using listening data as the allocation mechanism. The algorithmic, undisclosed selection criteria make engagement on Spotify the currency that buys concert access. Meanwhile [F] the company rolled out a more personalized free experience that it credits with users “listening and watching more days per month” in key markets, and [RC] co-CEO Söderström describes the platform’s edge in terms of its engaged user base, creator relationships, and years of personalization infrastructure. The pattern: every major 2025-2026 move deepens the data and dependency loop between fan, artist, and platform, rather than simply adding content.

Insight 3: Pricing power is now demonstrated, not theoretical. [CI] Three US price increases in four years, with subscriber additions continuing through the latest one and gross margin expanding to records, is about as clean a natural experiment as public data offers. The inference that Spotify can reprice faster than churn punishes it is now corroborated by repeated trials, not one.

One deliberately flagged weak point. [SI] The read that the free tier’s personalization push is primarily a funnel-widening move for future conversions rests mainly on the company’s own commentary in one earnings cycle. It is plausible, and management says engagement is up, but the conversion claim has a single source: Spotify. Promotion test: two more quarters of premium net adds at or above guidance in markets where the new free experience launched first.

Layer 3: Implications, read from a competitor’s seat

Take the chair of a CI lead at a rival streaming service, a label, or a ticketing player. Three implications follow, each with your company as the subject.

For streaming rivals. [CI] The battleground has moved. If Spotify’s differentiation is now the fan relationship layer, ticket access, engagement-based perks, personalization depth, then matching on catalog and audio quality no longer closes the gap; lossless became table stakes the day Spotify gave it away. A rival whose roadmap still centers on content parity is optimizing for the previous war. The margin story compounds this: [F] a 33 percent gross margin and 3.2 billion euros of trailing free cash flow fund experiments rivals cannot match sustainably.

For labels and artists’ teams. [CI] Reserved is a preview of leverage migration. If Spotify’s opaque algorithm decides which fans get early tickets, then artist teams’ direct channels, mailing lists, fan clubs, presale codes, lose their function as the superfan gateway, and with it their first-party data. The implication for any artist-side organization: the cost of not being on Spotify’s preferred terms now extends beyond streams into the live business, which is where most artist income lives.

For ticketing incumbents. [SI] The Live Nation partnership reads as cooperative today, but the structure, Spotify owning fan identification while the ticketer owns inventory, positions Spotify to commoditize its partner over time. Single-source in the sense that it rests on the design of one just-launched product. Watch whether Spotify extends Reserved to venues and promoters outside the Live Nation system.

Layer 4: Actions and the watch list

For a rival streaming service’s strategy team, three recommendations, each with an owner and a clock:

Reframe the competitive assessment (strategy team, this quarter). Retire content-parity dashboards as the primary Spotify tracker; replace with a fan-relationship scorecard: ticketing moves, engagement-gated perks, personalization shipping velocity.
Decide the live-music posture (corp dev plus product, 90 days). Either secure your own ticketing or presale partnership before exclusivity norms harden, or make an explicit, written decision not to compete on live access. The worst position is drift.
Re-test the thesis (CI, next two earnings cycles). Two promotion tests: does a paid superfan tier finally launch, and in what form, which would confirm or kill the “bundle dismantled” inference; and do premium net adds hold through the price increase in full-year data, which stress-tests the pricing-power inference.

Open speculation, labeled as such. [SP] Spotify’s endgame may be an advertising and commerce identity layer for music fandom, monetizing who loves which artist rather than access to audio. Little direct evidence today beyond the segment reporting reshuffle and the engagement-gating pattern. It stays on the watch list with one indicator: any move to sell fan-targeting or fan-data products to artist teams or brands.

What this teardown demonstrates

Every claim above traces down to a dated, public source or wears an explicit inference label, and every recommendation traces up from the evidence through a stated insight. That is the whole system working at once: the Triage Matrix decided Spotify deserved the hours, the Ladder kept the evidence honest, and the Pyramid forced the climb from 761 million users to three decisions a competitor could accept or reject in one meeting.

The raw material was available to anyone: filings, press releases, product launches, two wire reports. The difference between this and a news summary is not access. It is structure.

Views are my own and do not represent my employer.

The Signal Triage Matrix: deciding what deserves analysis before you analyze anything

Thu, 28 May 2026 00:00:00 GMT

Every CI function drowns in the same way. Alerts, news, rumors, filings, posts, and forwarded articles arrive faster than any team can analyze them, so the team defaults to one of two failure modes: analyze everything shallowly, or analyze whatever arrived most recently. Both feel busy. Neither is triage.

The Signal Triage Matrix is a decision tool you apply to a signal before spending analytical effort on it. Two questions, one placement, one pre-committed response. It takes under a minute per signal, and that minute is the highest-leverage minute in the entire CI workflow, because it decides where all the other hours go.

The two axes:

Impact if true. If this signal turns out to be real, how much does it move our revenue, margin, strategy, or competitive position? Note the phrasing: if true. Triage is not verification. You assess impact assuming the signal is real, precisely so that low-confidence but high-stakes signals do not get dismissed before anyone checks them.

Urgency, meaning time-to-relevance. How soon does this start affecting decisions we have to make? Not how new the signal is, and not how loudly it is being discussed. A signal can be breaking news and still be slow, or quiet and already late.

Two axes, four quadrants, four pre-committed responses.

Quadrant 1: High impact, high urgency. Escalate and verify now. This signal, if true, changes decisions that are being made this quarter. It jumps the queue. The response is a same-day effort: verify what can be verified, label what cannot, and get a short brief to decision makers with a clear fact vs. inference split. Speed matters more than completeness; a two-paragraph note today beats a deck next week.

Quadrant 2: High impact, low urgency. Schedule a deep dive. The strategic heavyweights live here: structural shifts that will matter enormously but not this month. The danger is that urgent noise perpetually crowds them out, so the response is to put the analysis on the calendar with an owner and a date, and to define tripwires, specific observable events that would move this signal into quadrant 1.

Quadrant 3: Low impact, high urgency. Brief note, no project. These are signals that executives will ask about tomorrow because they are loud, even though they change little. The response is a short, pre-emptive note: what happened, why it matters less than the headlines suggest, one labeled inference. The trap in this quadrant is analysis inflation, spending a week on something that deserved an afternoon because the noise made it feel important.

Quadrant 4: Low impact, low urgency. Log and move on. Most signals live here. The response is a one-line entry in a watch log, nothing more. Logging is not the same as ignoring; patterns in quadrant 4 entries are often the raw material for a future quadrant 2 insight. But no individual quadrant 4 signal earns analytical hours.

The matrix applied: an AI-chip CI desk in 2023-2024

Imagine you ran competitive intelligence at a semiconductor company competing in AI accelerators during 2023 and 2024, watching the environment around Nvidia. Four real signal types from that period, one per quadrant.

Quadrant 1: US export controls on advanced AI chips to China. In October 2023 the US government tightened restrictions covering Nvidia’s China-market accelerators. Impact if true: enormous; the rules redrew who could sell what into a major market, for Nvidia and for every rival. Urgency: immediate; customers, pricing, and compliance postures were shifting within days. Correct response: same-day brief, verified against the published rules themselves rather than press summaries, with clearly labeled inferences about how competitors might reposition China-specific products.

Quadrant 2: Hyperscalers building custom AI silicon. Through this period, Google’s TPUs, Amazon’s Trainium, and Microsoft’s Maia all signaled the same structural shift: the largest chip buyers were becoming chip designers. Impact if true: very high; it threatens the long-term shape of the merchant accelerator market. Urgency: low in any given week; no single announcement changed that quarter’s decisions. Correct response: a scheduled deep dive on hyperscaler silicon roadmaps, plus tripwires such as announced production volumes or a hyperscaler shifting a flagship workload off merchant chips.

Quadrant 3: Order-cut and demand-wobble headlines. The AI trade produced periodic rumor cycles, reports of shifted orders, supply reallocations, or demand pauses, that dominated a news cycle and then faded. Impact if true: usually modest and often unverifiable at the moment of the headline. Urgency: high in one narrow sense; leadership would be asking about it by the next morning. Correct response: a half-page note within a day, heavy on rung labeling from the Fact vs. Inference Ladder, explicitly stating what was a reported claim versus a verified fact. Not a project. Not a deck.

Quadrant 4: Startup accelerator announcements. The period produced a steady stream of AI-chip startup claims, benchmark wins on narrow tests, funding rounds, architectural manifestos, mostly from companies with no shipped volume. Impact if true: low for now. Urgency: none. Correct response: one line each in the watch log. If the log later shows a cluster, for example several startups converging on the same architecture or the same customer segment, that cluster becomes a quadrant 2 deep-dive candidate.

The discipline the example illustrates: the loudest signal of any given week (quadrant 3) received the least analysis, and the quietest structural signal (quadrant 2) received the most. That inversion is the entire point of triage. Volume of coverage is not a proxy for either axis.

Placement rules that keep the matrix honest

Triage the signal, not the source. A weak source can carry a high-impact signal. Confidence gets handled at the verification stage using the Ladder; the matrix deliberately excludes it so that inconvenient signals cannot be quietly filed as unimportant just because they are unconfirmed.

Re-triage on new evidence, not on news volume. Signals move between quadrants when facts change: a tripwire fires, a rumor gets confirmed, a timeline compresses. They do not move because coverage got louder.

Pre-commit the responses. The quadrant responses (same-day brief, scheduled deep dive, short note, log line) should be defined before signals arrive. Triage decided in the moment collapses back into recency bias, which is the disease it exists to cure.

Time-box quadrant 3 ruthlessly. If a “brief note” is entering its second day, either the signal was misplaced and belongs in quadrant 1, or the analysis has inflated and needs to stop.

The complete system

This matrix closes a three-framework system for CI work, and the three tools map cleanly onto the lifecycle of a signal.

The Signal Triage Matrix decides whether and when a signal deserves analysis. The Fact vs. Inference Ladder governs how honestly each claim inside that analysis is labeled. The CI Report Pyramid structures what the output must contain to drive a decision: data, insight, implication, action.

Triage the input. Label the evidence. Build the pyramid. A CI function running all three is doing intelligence. A CI function running none of them is doing expensive news summarization, on whatever topic happened to arrive last.

Views are my own and do not represent my employer.

The Fact vs. Inference Ladder: the honesty system behind credible CI

Tue, 19 May 2026 00:00:00 GMT

Competitive intelligence has one non-negotiable asset: credibility. Lose it once, by presenting a guess as a fact, and every future report you write gets read with a discount applied. The Fact vs. Inference Ladder is a simple labeling system that protects that asset. It forces every claim in a report to declare what it is, so the reader always knows how much weight it can bear.

The ladder has five rungs. Top rungs carry decisions. Bottom rungs carry questions. Nothing is banned from a report, not even speculation; what is banned is a claim wearing the wrong label.

Rung 1: Verified Fact. Directly observable, documented, and checkable by anyone. A published price list. A regulatory filing. A job posting with a date. The test: could a skeptical reader verify this in ten minutes without trusting you?

Rung 2: Reported Claim. Someone else asserts it, and you are passing it on with attribution. A journalist citing “people familiar with the matter.” An executive’s statement about their own strategy. A supplier’s comment at a trade show. It might be true, but you did not verify it; the source did the asserting, and your report should say so. The most common labeling error in CI is promoting rung 2 to rung 1 by dropping the attribution.

Rung 3: Corroborated Inference. Your own conclusion, supported by multiple independent data points that all point the same direction. This is the workhorse rung of good CI; it is where insight actually lives. The word “independent” is doing the heavy lifting: three news articles all citing the same original report is one data point wearing three costumes.

Rung 4: Single-Source Inference. Your own conclusion resting on one data point or one source. Legitimate to include, often valuable as an early warning, but fragile. The honest move is to state it, flag it, and name what evidence would promote it to rung 3.

Rung 5: Speculation. A hypothesis with little or no direct evidence, included because it is worth watching. Speculation is not a sin; unlabeled speculation is. The best CI teams keep a visible “watch list” of rung 5 items precisely so they can be tested rather than forgotten.

The ladder applied: Tesla’s early-2023 price cuts

Consider how one well-known episode sorts across the rungs. In January 2023, Tesla cut prices on its main models in the US and several other markets, in some cases by double-digit percentages, and continued adjusting prices downward through the year.

Rung 1, Verified Fact. The price changes themselves. They appeared on Tesla’s own published pricing, were dated, and were checkable by anyone with a browser. Also rung 1: Tesla’s reported automotive gross margins declined through 2023, because that figure sits in quarterly filings.

Rung 2, Reported Claim. Elon Musk’s own framing that Tesla was choosing volume over margin and could push price aggressively because of its cost position. That is the company asserting its own motive. Competitors’ statements about how they would respond, reported in the press, also sit here. All of it is attributable, none of it is self-verified.

Rung 3, Corroborated Inference. “The cuts were a deliberate offensive to pressure higher-cost EV rivals, not a distress reaction.” An analyst could support that with independent points: Tesla’s filed margins remained above most EV competitors even after cutting; production was scaling at new factories, consistent with a volume strategy; and rivals publicly wrestled with matching cuts they could not afford. Several independent data streams, one direction.

Rung 4, Single-Source Inference. Early in the episode, a claim like “these cuts signal softening demand in China” rested largely on one type of evidence, such as reported order intake data from a single research house. Worth stating; worth flagging; worth naming the promotion test, for example registration data over the following two quarters.

Rung 5, Speculation. “Tesla intends to cut prices until several legacy EV programs become unviable and are cancelled.” In early 2023 that was a hypothesis, not a finding. The correct treatment was to write it down, label it, and attach indicators, such as announced program delays or cancellations at rivals. Notably, some of those indicators did later materialize across the industry, which is exactly why speculation deserves a labeled place on the page rather than deletion.

The point of the example is not the automotive story. It is that the same episode contains all five rungs simultaneously, and a report that mashes them into one undifferentiated narrative forces the reader to guess which sentences are load-bearing.

Why analysts skip the labels

Three reasons, all human.

First, labeling feels like weakness. “We infer” reads as less authoritative than “Tesla is.” The opposite is true over time: readers learn that your unlabeled claims are always rung 1, which means your rung 1 claims get accepted without friction. Precision compounds into authority.

Second, deadlines compress rigor. Under time pressure, rung 4 claims quietly get written in rung 3 language because hunting for corroboration is the step that takes hours. The fix is not more hours; it is honest labeling of what the hours did not cover.

Third, narrative momentum. A clean story wants every sentence at the same confidence level. Real intelligence is jagged: solid facts next to fragile inferences next to open questions. A report that reads too smoothly has usually been sanded down at the expense of the labels.

Using the ladder in practice

Three habits make the ladder operational rather than decorative.

Label at the claim level, not the report level. A confidence disclaimer on page one covers nothing. Each analytical sentence should be identifiable as fact, attributed claim, or inference from its own wording.

Name the promotion test for anything on rungs 4 and 5. One sentence: what evidence, observable within a defined window, would move this claim up a rung or kill it? This converts weak claims from liabilities into a monitoring plan.

Audit the mix. A finished report that is 90 percent rung 1 is a news summary. A report that is mostly rungs 4 and 5 is an opinion column. Strong CI typically carries its weight at rung 3: independent facts, assembled into inferences the reader could not have made alone, sitting on top of a visible factual base.

The ladder pairs directly with the CI Report Pyramid. The Pyramid tells you a report must climb from data to action; the Ladder tells you how honestly each step of that climb is labeled. Data lives on rungs 1 and 2. Insight lives on rungs 3 and 4. And the discipline of both frameworks is the same: make your judgment visible, because visible judgment is what readers learn to trust.

Views are my own and do not represent my employer.

The CI Report Pyramid: why most competitive intelligence dies on page one

Sat, 09 May 2026 00:00:00 GMT

Most CI reports are read once, filed, and never acted on. Not because the research was weak, but because the report stopped one or two levels short of being useful. The analyst gathered facts, maybe spotted a pattern, and then handed the “so what” over to the reader. Readers do not do that work. They skim, nod, and move on.

The CI Report Pyramid is a simple test for whether a report will drive a decision or decorate a shared drive. It has four layers, and every layer has to be earned from the one below it.

Data > Insight > Implication > Action.

Each layer answers a different question. If you cannot answer the question, you have not reached that layer, no matter what your slide title says.

A four-layer pyramid. From the bottom: data, asking what happened; insight, asking what pattern connects it; implication, asking what this means for us; and action, asking who does what, by when. An arrow runs up the stack, because each layer is earned from the one below it.

The CI Report Pyramid. Each layer carries its own test question, and has to be earned from the layer beneath it.

Layer 1: Data. “What happened?”

Data is the raw, verifiable material: filings, job postings, pricing pages, import records, press releases, customer reviews, conference remarks. It is the only layer where the standard is factual accuracy, and the standard is absolute. Every item here should be sourceable, dated, and labeled as fact.

This is also where most CI reports live and die. A 30-page competitor profile that lists everything a rival did last quarter is a data dump wearing a report’s clothing. It feels rigorous because it is dense. It is actually unfinished work.

Running example. You cover the industrial pumps market. Over six weeks you log the following about a competitor, Veltrax Flow Systems: Veltrax posted 14 openings for field service technicians across the US Gulf Coast; it announced a distribution agreement with a regional MRO supplier in Texas; its latest price list shows a 6 percent reduction on mid-range centrifugal pumps; and its CEO mentioned “aftermarket revenue” four times in the last earnings call, up from zero mentions a year ago.

Four facts. All checkable. None of them, on their own, tell your leadership anything worth a meeting.

Layer 2: Insight. “What changed, and what pattern connects it?”

Insight is where separate facts become one finding. The test: can you state, in one sentence, a pattern that a reasonable reader could not have seen from any single data point alone?

This is also where intellectual honesty starts to matter. An insight is an inference, not a fact, and it should be labeled as one. The discipline of writing “we infer” instead of “Veltrax is” keeps you credible when you are wrong, and you will sometimes be wrong.

Running example. The four facts converge: We infer Veltrax is shifting from selling pumps to selling uptime, building a Gulf Coast service footprint and using price cuts on new units to seed an installed base it can monetize through aftermarket contracts.

That is a real finding. Hiring alone could mean growth. Price cuts alone could mean weakness. Together, with the earnings language, they point at a strategy change. Notice the sentence is falsifiable; future evidence can confirm or kill it. That is what separates insight from narrative.

Layer 3: Implication. “What does this mean for us, specifically?”

An insight about a competitor is still about the competitor. The implication layer translates it into consequences for your company: your accounts, your margins, your roadmap, your pricing corridor. The test: does the sentence contain your company as the subject or object? If it could appear unchanged in a rival’s report, it is not yet an implication.

This layer is where analysts get timid, because implications require judgment and expose you to disagreement. Push through. A report without implications forces every reader to derive their own, and ten readers will derive ten different ones, most of them casually.

Running example. If Veltrax converts even a quarter of its new installed base to service contracts, our Gulf Coast aftermarket revenue, roughly 30 percent of regional margin, comes under direct attack within 18 months. Our three largest regional accounts all have Veltrax units on site already, making them natural first targets for bundled service offers.

Now the reader is awake. The competitor’s move has a location, a timeline, and a number attached to your own P&L. Label your assumptions here too; the “quarter of the installed base” figure is a scenario, not a forecast, and saying so costs you nothing.

Layer 4: Action. “What do we do, who does it, and by when?”

The top of the pyramid is a recommendation concrete enough to be accepted or rejected in the meeting where it is presented. The test is brutal and simple: does it name an owner, a move, and a deadline? “Monitor the situation” fails. “Consider strategic options” fails. Both are the analyst handing the work back.

Actions do not need to be grand. Often the strongest recommendation is a cheap, fast, reversible move that buys information.

Running example. Recommend: (1) Commercial team offers multi-year service agreements to our three exposed Gulf Coast accounts before end of quarter, before Veltrax’s field team is fully staffed. (2) Product team prices a service-plus-hardware bundle for mid-range pumps within 60 days. (3) CI re-tests the Veltrax thesis in 90 days against two indicators: technician headcount actually onboarded, and any published service contract wins.

Point 3 matters. Building a feedback loop into the action layer is what makes CI a system rather than a sequence of one-off memos.

The two ways reports fail the pyramid

Failure mode 1: The truncated pyramid. The report stops at data or insight. It is accurate, thorough, and inert. You can diagnose it by scanning for the words “we recommend”; if they never appear, the pyramid was never finished. This failure is comfortable because it carries no risk of being wrong. It also carries no chance of being useful.

Failure mode 2: The floating peak. The opposite disease: bold recommendations resting on thin or unlabeled evidence. The action layer exists, but you cannot trace it down through implication and insight to specific facts. These reports feel decisive and age badly. When one confident-but-unsupported call goes wrong, every future report from that analyst gets discounted.

The pyramid’s real function is forcing traceability in both directions. Every action should trace down to named facts. Every fact worth including should trace up toward a possible action. Facts that trace to nothing are scope creep; cut them or move them to an appendix.

The 30-second self-audit

Take the last CI report you wrote or received and ask four questions:

Can every factual claim be sourced and dated? (Data)
Is there at least one sentence that connects multiple facts into a labeled inference? (Insight)
Does the report name consequences with your own company as the subject? (Implication)
Is there a recommendation with an owner and a deadline, plus a date to re-test the thesis? (Action)

Four yeses: the report can drive a decision. Three or fewer: you know exactly which layer to build next.

The pyramid will not make your judgment better. It will make it visible, and visible judgment is what separates intelligence from expensive news summarization.

Views are my own and do not represent my employer.