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.