When your supplier is also your rival: the five forces for AI applications
On 3 June 2025 Windsurf, a popular AI coding tool, said that Anthropic had sharply reduced its direct access to Claude models, giving, in the words of its chief executive Varun Mohan, little notice. The cut came shortly after reports that OpenAI was acquiring Windsurf. Anthropic's co-founder Jared Kaplan said it would be odd for the company to be selling Claude to OpenAI, and cited limited computing capacity as a factor too. Two months later, on 1 August, Anthropic revoked OpenAI's own API access to Claude, saying OpenAI had violated commercial terms that bar customers from using the service to build competing products or to train competing models. OpenAI replied that evaluating other systems is standard industry practice.
Whatever one thinks of the merits of either decision, they illustrate something about the structure of this industry. The party that supplies the model can also be a competitor, can decide for strategic reasons who may buy, and can change terms with little notice. For a company that builds applications on someone else's model, those facts belong in the strategy as much as any feature does.
Michael Porter made a general argument about technology twenty-five years ago that applies. In his 2001 article on the internet he claimed that the new technology tended to weaken industry profitability, because it lowered barriers to entry, increased the bargaining power of buyers and made rivals' offerings easier to compare, and that advantages gained from it were quickly competed away. His framework of five forces, set out in 1979 and revised in 2008, asks where the profit goes in an industry: to suppliers, buyers, new entrants, makers of substitutes or rivals, according to who has the most bargaining power. The technology does not decide the answer. The structure does.
I want to apply the five forces to a company that builds products on top of foundation models, and argue that the result is less comfortable than the enthusiasm around AI applications suggests, though not hopeless.
Supplier power, with a twist
In a conventional analysis a supplier is powerful when there are few of them, when their input is important to the buyer, when switching is costly, and when they can threaten to integrate forward into the buyer's business. All four conditions are partly met.
Concentration is high. Menlo Ventures' mid-2025 survey of 150 technical leaders, as reported in the press, put Anthropic at 32% of enterprise LLM API usage, OpenAI at 25%, Google at 20%, Meta's Llama at 9% and DeepSeek at 1%. Anthropic's share of coding workloads was reported at 42%, against 21% for OpenAI. The figures come from a survey and have limits, and they describe a market of a few suppliers where the leader changed in two years.
The input is critical, since the model is the product's engine. Switching costs are real, if lower than in older software: prompts, tool definitions and evaluation baselines are tuned to a model, and moving takes weeks of work.
Forward integration is the twist. Model providers build their own applications, from coding agents to research assistants to enterprise search, and sell them directly to the customers of the companies that buy their models. The supplier is also a competitor, with the advantage of knowing what its customers are building through its API and the ability to ship a similar feature with first access to its own next model. The Windsurf and OpenAI episodes show the further point that a supplier can withdraw supply when it judges a customer to be a strategic threat, and that capacity limits give it a reason beyond strategy.
The other forces
Buyer power depends on the customer's alternatives. A buyer of a horizontal AI assistant has plenty, including going straight to the model provider's own product, and can compare outputs easily. A buyer whose workflow is embedded in an application, with the integrations, history and trained users that come with it, has higher switching costs. The strength of the buyer's position therefore depends on how deeply the product is woven into how the customer works.
The threat of new entrants is high where the product is a thin layer. A small team can build a working application in weeks, and the cost of the underlying capability is falling. Epoch AI estimates that the price of reaching a fixed level of performance has been falling by roughly 40 times a year, with rates that vary from 9 times to 900 times by task. Lower input costs help every entrant equally. Barriers survive where the product depends on things that cannot be assembled quickly: distribution, accumulated data on outcomes, regulatory approval, integration into systems of record, and a reputation for trustworthiness in a domain where errors are costly.
Substitutes are an unusual force here, because the main substitute for a specialized AI product is a general one. A lawyer who might have bought a contract-review tool can paste a contract into a general assistant. A manager who might have bought a meeting summarizer can use the one built into the video-call software. The specialized product has to be demonstrably better at something the general tool does not do well enough, and the gap has to survive the next release of the general tool.
Rivalry is intense, because entry is easy and products are similar in capability. The competition tends to move to price, packaging and features that the next model release makes obsolete.
| Force | What it looks like for an AI application | A move that helps |
|---|---|---|
| Supplier power | A few labs, which can integrate forward and restrict access | Keep a tested alternative model and your own evaluation suite |
| Buyer power | Customers can go direct, compare easily, and switch where integration is shallow | Embed in the workflow and the system of record |
| New entrants | Small teams ship quickly, and input costs fall for all | Build on assets that take time: outcome data, approvals, distribution |
| Substitutes | General assistants cover much of what specialized tools do | Specialize where errors are costly and the general tool is not accountable |
| Rivalry | Similar capabilities, price competition, rapid obsolescence | Compete on outcomes you can measure, not on model capability |
An old pattern on a new platform
Software has seen this before. Platform owners learn what third parties build on top of them, and when a category proves popular they ship their own version. The industry has a word for it, from 2002, when Apple added to its Sherlock search tool the main features of a third-party product called Watson: being "sherlocked." Social platforms and app stores have changed API terms in ways that ended whole categories of dependent products. The lesson is not that building on a platform is a mistake. It is that the platform owner has an option the dependent company does not, and that the dependent company's strategy has to include what it will do on the day the option is exercised.
Andy Grove added a sixth force to Porter's five in "Only the Paranoid Survive": the complementors, whose products make yours more valuable. The model labs are complementors as well as suppliers, since a better model makes the applications on top of it more valuable, and the relationship runs in both directions at once. Where the lab's interests and the application's align, as when a new model lifts quality across the board, the lab is an ally. Where they diverge, as when the lab enters the same market, it is a competitor. A company should know, for each product, which of the two the lab is likely to be in the next eighteen months.
An illustration, invented for the purpose. Consider a company that sells an AI assistant for collections teams at Brazilian banks. Its supplier risk is moderate: it runs on one hosted model with a tested alternative. Its buyer power is mixed, with banks large enough to build in-house but unwilling to take on the regulatory and operational burden of doing so. Substitutes are the general assistants that bank staff already use, which cannot access the bank's accounts or follow its collections rules. Entrants face the barrier of integration with core banking systems and of the compliance reviews that take banks months. Rivalry is moderate and rising. Its defensible positions are the integrations, the compliance record and the data on which scripts and offers lead to payments in Brazilian consumer credit. Its exposed position is the language-model layer, which any competitor can match.
Falling input prices are not a gift
It is tempting to treat cheaper inference as pure good news for an application company, and it is good news, but it is also available to every competitor and every customer. When all rivals' costs fall together, competition passes the saving to buyers in the form of lower prices or more generous usage. The company that merely resells capability gains little, since the capability is what gets cheaper.
The margin evidence points the same way. Reports from venture investors put the gross margins of AI-native products well below the 75% to 85% typical of software: an ICONIQ snapshot cited in secondary coverage puts the 2026 average near 52%, and Bessemer's 2025 analysis of fast-growing AI startups described a group of explosive growers at around 25% gross margin alongside more efficient ones near 60%. I have not seen the primary reports, and the categories blur, so the figures should be read as indications. The mechanism behind them is solid. Inference is a cost of goods sold that scales with use, unlike the near-zero marginal cost of conventional software, and agentic features consume many times the tokens of a chat. A seat priced as flat-rate software can have a heavy user whose inference alone exceeds the price.
Here is a simple version. A seat sells for $30 a month, and the average user costs $9 a month in inference, leaving a 70% gross margin. A heavy user who runs agents all day costs $45. Ten percent of users like that add $3.60 to the average cost per seat, which takes it to $12.60, or a 58% margin, and the company pays more to serve its best customers than they pay it. If a competitor then cuts price to $25, the margin at the same cost falls to 50%. Cheaper models restore some margin, and they also invite the next price cut.
Where the profit can sit
Porter's logic gives a direct instruction: position the company where the forces are weakest, and work on the strongest one. For an application company, a few positions have held up in the analysis above.
The first is the system of record or the workflow itself. If the product is where the customer's work is done, with history, integrations and trained users, buyer power falls and switching costs rise. The model underneath can change without the customer noticing, which also weakens supplier power.
The second is outcome data and the feedback loop around it. A product that learns from the results of its decisions in a particular domain gains something that neither a general assistant nor a new entrant can copy quickly.
The third is accountability in a regulated or high-stakes domain. A general assistant carries no liability for its advice, and a specialized provider that stands behind its work, with audit trails, certifications and insurance, sells something the substitute does not. The value is the trust, which is expensive to produce and valuable when mistakes cost money.
The fourth is distribution. A company with a customer base and a sales channel in a niche that the labs do not serve directly has a durable advantage in a market of small segments, where it is not worth a lab's attention to build a bespoke product.
On the supplier side, the moves are the ordinary ones, executed seriously: keep more than one model qualified, so no single provider can set terms; own the evaluation suite and the prompts, so that migration is a matter of days; avoid features that exist only at one vendor in the core of the product; read the terms about competing products closely, since the clause that cost OpenAI its access applies to every customer.
Run the table once a year
The five-force table is cheap to produce and easy to get wrong in the direction of optimism. Fill it in for each product line, write one sentence of evidence for each cell, and mark which force is rising. The forces in this industry move quickly: a lab's new feature can create a substitute overnight, and a price cut can erode a rival's margin by the next quarter. A yearly review is a minimum, and the review should end in a decision about the weakest cell and what to do about it before the next one.
Sources
- Porter, M. E. (1979). How competitive forces shape strategy. Harvard Business Review, March-April 1979; Porter, M. E. (2008). The five competitive forces that shape strategy. Harvard Business Review, January 2008.
- Porter, M. E. (2001). Strategy and the Internet. Harvard Business Review, March 2001.
- TechCrunch (3 June 2025), Windsurf says Anthropic is limiting its direct access to Claude AI models; The Decoder (June 2025), coverage of Anthropic's statements on Windsurf.
- Coverage of Anthropic's revocation of OpenAI's access to Claude (1 August 2025) by TechRepublic, Slashdot and others.
- Menlo Ventures (July 2025). 2025 Mid-Year LLM Market Update, as reported by TechCrunch (31 July 2025) and BigDATAwire.
- Epoch AI. LLM inference price trends. epoch.ai/data-insights/llm-inference-price-trends.
- ICONIQ and Bessemer Venture Partners, State of AI reports (2025-2026), as reported in secondary coverage; figures indicative.
- Illustrative seat economics: $30 seat price, $9 average inference cost, a heavy user at $45, assumptions stated in the text.
- Grove, A. S. (1996). Only the Paranoid Survive. Currency/Doubleday.