Porter's Five Forces Assume Scarcity. Intelligence Is Now Abundant

In brief

Michael Porter's five forces explain why some industries earn more than others, and they still do. What has aged is how strategy teams score them. The scores assume that analysis, writing and expert judgment are scarce and priced by the hour. That assumption is failing. Stanford's AI Index 2025 found the price of GPT-3.5-level model answers fell more than 280-fold in under two years, and by February 2025 the best open models trailed the best closed ones by under 2 percent. Cheap intelligence lowers entry barriers, makes substitutes good enough, and weakens any single model supplier's hold, while buyers have begun negotiating through software. The gains are uneven, as a Harvard and BCG field experiment with 758 consultants showed. This piece re-scores each force, gives a five-question test of how exposed an industry is, and states the strongest objection: Porter himself said technology acts through the forces, not around them.

Key Takeaways

→GPT-3.5-level answers got more than 280 times cheaper between November 2022 and October 2024.
→The five forces still explain industry profits. The scoring has aged: it assumes thinking work is scarce and expensive.
→Entry, substitutes and supplier power move fastest when intelligence gets cheap; buyer power and rivalry follow.
→The gains are jagged: in the Harvard and BCG field experiment, AI helped consultants on some tasks and cut correct answers by 19 percentage points on another.
→Score your industry on five questions before your next strategy review. A high score means the old Five Forces slide is out of date.

Between November 2022 and October 2024, the price of getting an answer from a model as capable as GPT-3.5 fell from $20 to 7 cents per million tokens, according to Stanford's AI Index 2025. That is a fall of more than 280-fold in under two years. I can't think of an input to a Five Forces analysis that has ever moved that fast.

My argument is narrow. The five forces still describe how an industry divides its profits. But most strategy teams still score them as if thinking work were scarce and expensive. At today's prices for machine intelligence, that calibration rates entry barriers, substitutes and supplier lock-in as safer than they now are. If your Five Forces slide looks the same as it did in 2022, it probably overstates any protection that rests on expert headcount and billable hours.

I call the new condition infinite leverage. It is what happens when the cost of a unit of cognitive work falls far faster than the cost of the people who used to do it, so that headcount stops being the limit on how much expert work a firm can produce. The marginal cost never reaches zero, and the word "infinite" is a direction rather than a measurement.

This piece is for the chief strategy officer or business-unit head who owns next year's strategy review, in a business where most of the cost base is people doing thinking work. It re-scores each force, gives a five-question test of exposure, and ends with the strongest case against my argument.

What Porter's Forces Assumed

Porter's 1979 Harvard Business Review article opens with the line "The essence of strategy formulation is coping with competition." He named five forces that set an industry's profit potential: the threat of new entrants, the bargaining power of suppliers, the bargaining power of buyers, the threat of substitute products, and rivalry among existing competitors. His 2008 restatement keeps the five and adds a section called "Factors, Not Forces", which warns against mistaking visible features of an industry, technology among them, for its structure.

The framework never said that intelligence was scarce. The scores did. When a strategy team rates the threat of entry as low, it usually means a newcomer would need years and hundreds of skilled people to match what the incumbent does. When it rates substitutes as weak, it usually means the alternatives are worse or no cheaper. When it rates supplier power as moderate, it is thinking of scarce specialists and a few large vendors. Each judgment carries a hidden price for expertise, and for as long as the framework has been taught it was safe to assume that expertise cost what experts cost.

That is the assumption I think has broken. The forces still operate. The inputs most teams feed them no longer match the market.

Intelligence Got Cheap, Unevenly

Other measures agree. Epoch AI measured the same kind of fall across several benchmarks and found prices dropping between 9 and 900 times a year, depending on the task, with a median of 50 times a year. Epoch adds a caution: the fastest drops came in the most recent year, and they may not persist.

Capability also stopped being the property of a few vendors. The AI Index reports that in January 2024 the leading closed-weight model beat the best open-weight model by 8.0 percent on the Chatbot Arena leaderboard. By February 2025 the gap was 1.7 percent.

Cheap doesn't mean uniformly useful. The best evidence on that is the field experiment run by researchers from Harvard Business School and others with 758 Boston Consulting Group consultants, first circulated in September 2023 and published in Organization Science in 2026. On tasks inside what the authors called the jagged technological frontier, consultants with AI completed 12.2 percent more tasks, finished 25.1 percent faster, and produced work rated more than 40 percent higher in quality. On a task chosen to sit outside the frontier, consultants using AI were 19 percentage points less likely to reach the correct answer than those without it.

Both halves matter for strategy. Intelligence is now abundant enough to change the economics of whole categories of work. It is also jagged enough that a competitor who deploys it carelessly can get worse. The forces shift where the frontier covers an industry's core work, and they shift little where it does not.

Force by Force: What Changes

Each force, scored under two assumptions

Force	Scored for scarce intelligence	Scored for abundant intelligence
Threat of entry	Capital and expert headcount deter newcomers.	Small teams can reach scale fast.
Substitutes	A cheaper substitute is usually a worse one.	Cheaper and often good enough, unevenly.
Supplier power	Few experts and few vendors set terms.	Open models cap any one model vendor.
Buyer power	Relationships slow down switching.	Buyers negotiate through software.
Rivalry	Rivals come from inside the industry.	Rivals can come from any industry.

Source: Author's analysis; the evidence for each row is in the text below

Threat of entry

Entry barriers built on headcount are the first to weaken. Stripe's 2024 annual letter reports that the top 100 AI companies on its platform reached $5 million in annualized revenue in a median of 24 months, against 37 months for the top 100 software-as-a-service companies of 2018. Cursor, the AI code editor, said in June 2025 that it had passed $500 million in annual recurring revenue, and in November 2025 that it had crossed $1 billion in annualized revenue with a team of more than 300. Barriers that rest on proprietary data, distribution, licences and regulation hold up much better, and in regulated markets the rules themselves remain one of the stronger ones.

Threat of substitutes

Porter's test for a substitute is its price and performance against yours. For work done by people answering questions or drafting documents, the substitute now often wins on price and comes close on performance. Klarna said in February 2024 that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, doing the equivalent work of 700 full-time agents. The same case shows the limit. In May 2025 its chief executive told Bloomberg that cost had been "a too predominant evaluation factor" and the result was "lower quality", as Fortune reported, and Klarna said it would recruit people again so customers could always reach a human. Substitutes now arrive faster and cheaper, and they still have to be good enough for the customer.

Supplier power

For any firm that buys intelligence, the model vendor is a new supplier, and an early worry was that a few closed-model providers would hold all the power. Open releases have capped that. Meta released LLaMA in February 2023 under a research licence. The Technology Innovation Institute in Abu Dhabi made Falcon 40B royalty-free under Apache 2.0 in May 2023, Mistral AI released Mistral 7B under the same licence that September, and DeepSeek released R1 under the MIT License in January 2025. With the open-closed gap at 1.7 percent by February 2025, a buyer who keeps the ability to switch models holds real bargaining power. The supplier risk moves elsewhere, to whoever controls the firm's data, its compute contracts and its integrations.

Buyer power

Buyers are starting to negotiate through software. In a pilot described in Harvard Business Review in 2022, Walmart International invited about 100 of its tail-end suppliers to negotiate with a chatbot built by Pactum. Eighty-nine took part, the chatbot reached agreement with 64 percent of them, and Walmart gained on average 1.5 percent savings and payment terms extended to an average of 35 days. When a large buyer can run hundreds of negotiations at once, the supplier's account manager and the information gap the relationship depended on count for less. Pricing power moves to products that are hard to replace because they sit deep in the customer's workflow.

Rivalry among existing competitors

Porter drew rivalry inside an industry's boundaries. Cheap intelligence makes those boundaries porous, because a firm that has learned to sell its expertise as software in one market can offer it in another at little extra cost. A legal-services firm's next rival may be a software company, and a bank's may be a payments platform. The question for rivalry becomes how fast each competitor learns from its own customers, which favours whoever runs the most useful work through its systems and improves on what comes back.

Is Your Industry Exposed? A Five-Question Test

Not every industry will feel this at the same speed. The useful question is how much of your value depends on thinking work that machines now do cheaply. In McKinsey's 2023 estimate, current generative AI and other technologies could automate work activities that absorb 60 to 70 percent of employees' time. That is potential, not adoption, and it varies widely by occupation, which is why the test below asks about your own work rather than the economy's.

Score your industry from 1 to 5 on each question, then add the scores.

Five-Question Exposure Test

Score each question from 1 (low) to 5 (high), then add the five scores

Can the core product or service be delivered as information, without physically changing anything? Advice, analysis, software, media and most financial services score 4 or 5. Building, growing and moving physical things score 1 or 2.

Estimate the share of operating cost that goes to people analysing, writing, deciding and advising. Use your own payroll, not an industry average. The higher the share, the more of your cost base a cheaper substitute can reach.

If each new customer needs proportionally more expert hours, your marginal cost is exactly what cheap intelligence attacks. A business that bills for hours scores high; a business whose next customer costs almost nothing already scores low.

Test your core tasks, not adjacent ones, against current models. If they handle most of the work well, score high. If they fail on the judgment that matters, as AI did on the out-of-frontier task in the Harvard and BCG experiment, score low and test again next year.

Look at who has been funded in the last eighteen months to do what you do, and at which suppliers or adjacent firms have started offering it. Startups aimed at your main revenue line mean a high score.

I set these bands by judgment; nobody has measured them. A total of 20 or more means the shift is already under way, and the Five Forces analysis should be redone with this year's prices for intelligence before the next budget is set. Between 12 and 19, the forces are moving but not yet decisive, and the useful work is finding which of your core tasks sit inside the frontier. Below 12, the change is further out, and the right step is to watch questions 4 and 5 once a year.

The Strongest Objection

The best case against this piece is Porter's own. In the 2008 article he writes: "Advanced technology or innovations are not by themselves enough to make an industry structurally attractive (or unattractive)." With James Heppelmann he applied the forces to an earlier wave in 2014, writing that "industry structure changes when new technology, customer needs, or other factors shift these five forces." On that view, AI is a factor, the forces absorb it, and nothing needs rewriting. He had made a related case about the internet in 2001: a technology every competitor can buy tends to weaken industry profitability without giving anyone a proprietary edge, which makes distinctive strategy matter more.

I agree with most of that, and this piece sits in the line of the 2001 argument. My claim is about calibration, not the framework. When intelligence is cheap for everyone, Porter's logic says the profits flow to whatever stays scarce: proprietary data, trust, distribution, regulatory standing, and someone accountable for the result. Incumbents that already own those assets may come out of this stronger. The ones exposed are those whose protection was the cost of hiring and training expert staff. The objection does catch one thing. The Klarna reversal and the jagged frontier both show that cheap intelligence is not free advantage, and a firm that treats it as one can damage the quality its customers pay for.

A second objection is selection. Stripe's figures describe the top 100 AI companies on its platform, and Cursor is among the fastest-growing software companies of its time. Most AI start-ups will not grow like that, and many will fail. I accept the point for any single entrant. That doesn't rescue the incumbent, because in Porter's terms the threat of entry turns on what it costs to make a credible attempt, and the price data above say that cost has fallen for every attempt, including the ones that fail. Cheap intelligence also cuts the incumbent's own costs, so part of the gain goes to buyers as lower prices, which is why entry and buyer power move together.

What to Do Before the Next Strategy Review

Rerun the Five Forces with three changes. Price a finished, reviewed piece of expert work at today's model prices, counting the review time and the tasks that fall outside the frontier, instead of at what your people cost. Name rivals by the customer need they serve, wherever they come from. And list what stays scarce in your business once intelligence is cheap, including who signs the work and answers for it when it is wrong, because that list is where your next advantage will come from. Part 2 of this series puts that list through two tests to find which items can become moats.

To test a single AI initiative against these forces, the AI Use Case Canvas gives a twelve-block evaluation. For the regulatory barriers that still hold, the Regulation Tracker shows where AI law stands, country by country.

Sources and Changes

Michael E. Porter, "How Competitive Forces Shape Strategy", Harvard Business Review, March–April 1979.
Michael E. Porter, "The Five Competitive Forces That Shape Strategy", Harvard Business Review, January 2008.
Michael E. Porter, "Strategy and the Internet", Harvard Business Review, March 2001.
Michael E. Porter and James E. Heppelmann, "How Smart, Connected Products Are Transforming Competition", Harvard Business Review, November 2014.
Stanford HAI, AI Index Report 2025, 7 April 2025 (inference cost, open and closed model gap).
Ben Cottier and others, Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks", 12 March 2025.
Fabrizio Dell'Acqua and others, "Navigating the Jagged Technological Frontier", Harvard Business School Working Paper 24-013, 22 September 2023; published in Organization Science, 11 March 2026.
Stripe, 2024 annual letter, 27 February 2025.
Cursor, "Series C", 6 June 2025, and "Series D", 13 November 2025.
Klarna, press release on its AI assistant, 27 February 2024; Bloomberg, 8 May 2025, as reported by Fortune, 9 May 2025.
Remko Van Hoek, Michael DeWitt, Mary Lacity and Travis Johnson, "How Walmart Automated Supplier Negotiations", Harvard Business Review, 8 November 2022.
Meta AI, LLaMA release post, 24 February 2023; Technology Innovation Institute, Falcon 40B royalty-free release, 31 May 2023; Mistral AI, "Announcing Mistral 7B", 27 September 2023; DeepSeek, DeepSeek-R1 release note, 20 January 2025.
McKinsey & Company, "The economic potential of generative AI: The next productivity frontier", 14 June 2023.

All sources checked on 6 October 2026.

Changes on 6 October 2026. New headline, summary and takeaways, and the piece now stands on its own; Part 2 remains a separate article. The opening strategy-offsite story is gone: it was a composed scene, not a single engagement I can document, and its figures went with it. The training-cost chart is replaced, because everything after 2022 on it was a projection drawn as data; the Ark Invest training-cost figures are replaced by Stanford's and Epoch's newer measures of inference prices. Charts that scored industries and forces without a source are removed. The Harvard and BCG finding now reads 19 percentage points, as the paper states, and cites the published version. New evidence on entry, substitutes, suppliers and buyers is sourced above.