Three Companies Will Decide What You Can Think With
The Quiet Consolidation of Intelligence Into a Handful of Frontier Labs, and Why That Should Worry You More Than Any Single AI
The fear sold to the public is a single rogue machine, one runaway intelligence that outsmarts us all. What is actually arriving is quieter and harder to put on film. A handful of private companies are becoming the gatekeepers of how a civilization reasons, decides, and remembers. The danger isn’t a mind we can’t control. It’s a small group of owners we can’t route around.
Frontier-scale AI is collapsing into three or four labs that set the cost of compute, build the models, and write the terms of access. The system growing too smart was always the cinematic worry. The real one is plainer: too few people own the system you increasingly think with, and almost no one is watching the ownership.
Three companies, 88% of enterprise LLM usage
Share of enterprise large-language-model API usage, 2025
Published by Kymata Labs · Independent Research Institution.
Count the AIs you actually use. The list is short on purpose.
Think about the email you cleaned up this morning, the code your engineers shipped, the brief your analyst summarised, the answer you trusted at 11pm. How many distinct companies stood behind all of it? Almost certainly three or fewer. The narrowness wasn’t your decision. The market made it for you, and it chose a very small handful.
According to one 2025 enterprise survey, three companies account for 88% of enterprise LLM API usage: Anthropic, OpenAI, and Google, with everyone else splitting the remaining 12%.1 Calling that a market with a few big players undersells it. At 88%, the few big players more or less arethe market. And the product they’re selling is the raw material of thought.
The choice of model was made for you, and it came down to three names.
“A decade of fearing one all-powerful machine left almost no one watching the more ordinary question of who came to own the machines.”
Kymata LabsNone of this is a forecast. It’s the current scoreboard.
Four independent measures point at the same small set of companies: market share, the cost of building a frontier model, who funds that cost, and who is now investigating it. The convergence is the argument.
The market: three names, 88% of the usage
In Menlo Ventures’ 2025 enterprise study, three companies account for 88% of enterprise LLM API usage: Anthropic at 40%, OpenAI at 27%, and Google at 21%. The remaining 12% is spread across Meta’s Llama, Cohere, Mistral, and a long tail of everyone else. Sit with that distribution for a moment. The entire rest of the industry, from open-weights projects to well-funded challengers to national champions, is competing for the scraps left on the table after three firms have eaten. Treat the exact figures as a market estimate from a survey rather than an audited census; the shape is the point, and the shape is an oligopoly.
Menlo Ventures, “2025: The State of Generative AI in the Enterprise” · survey-based market estimate.The cost moat: the price of a frontier model is now a barricade
You cannot simply decide to compete here. Stanford’s AI Index estimated the training compute behind GPT-4 at roughly $78 million and Gemini Ultra at roughly $191 million. Epoch AI finds frontier training cost has grown about 2.4× every year since 2016, a curve that compounds the barriers faster than any challenger can clear them. Anthropic’s CEO, Dario Amodei, projected $1 billion training runs in 2024–25 and around $10 billion by 2026. At ten-billion-dollar entry tickets, the list of who can even attempt a frontier model stays short, and it stays short because of the economics, not by accident.
Stanford HAI, AI Index 2024 (estimated training-compute cost); Epoch AI, 2024; Amodei via Business Insider, 2024.The choke points: the labs differ, the landlords don’t
Follow the capital and the data centres and the picture narrows further. Microsoft has put roughly $13 billion into OpenAI. Amazon has committed $8 billion to Anthropic. Google has committed $3 billion-plus to Anthropic for a stake of around 14%, which means Anthropic is backed by both Amazon and Google. The tidy framing of one cloud capturing one lab turns out to understate things. What the money actually shows is the more uncomfortable version: the frontier labs are capital-and-compute-dependent on the same tiny set of hyperscalers, namely Microsoft, Amazon, and Google. Two layers of concentration sit stacked on each other, and the visible competition between models rests on a shared three-company foundation underneath.
CNBC (Microsoft–OpenAI ~$13B); Amazon (Amazon–Anthropic $8B); The New York Times (Google–Anthropic $3B+, ~14% stake).The regulators: watching closely, mostly not blocking
This has not escaped notice. In January 2024 the FTC issued 6(b) orders to Microsoft, OpenAI, Amazon, Anthropic, and Alphabet, and in January 2025 published a staff report examining whether the cloud–AI partnerships “risk distorting innovation and undermining fair competition.” In the UK and EU, the CMA and European regulators looked at the same deals and largely declined to escalate. The scrutiny has been intense and the intervention slight. The structure got studied and flagged, then, for now, left exactly where it stood.
U.S. FTC, 6(b) inquiry (Jan 2024) and staff report (Jan 2025); UK CMA and EU reviews.
The labs differ. The landlords don’t.
Who funds and hosts the frontier — two layers of concentration, stacked
The barricade built itself, one training run at a time.
None of this required a conspiracy. It required a cost curve. When the price of a competitive model climbs 2.4× a year, every passing season quietly prices out another would-be rival.3 A startup that could have built a near-frontier system for tens of millions in 2023 now faces a multi-billion-dollar bar by 2026. Concentration is the natural sediment left behind by an economics this steep.
The price of a frontier model is now a barricade
Estimated cost of a single training run — the bar that keeps the room small
The capital made things worse, though depending on where you sit it also made them possible at all. To afford the compute, the labs took the hyperscalers’ money: ~$13B from Microsoft into OpenAI, $8B from Amazon and $3B+ from Google into Anthropic.5, 6, 7 Each cheque bought compute today and entrenched dependence tomorrow. The clouds that train the models also rent them back to the world, so the same three companies end up sitting at both the top and the bottom of the stack.
The leaderboard kept reshuffling, and that motion fooled a lot of people into reading the market as open. Between Menlo’s 2024 and 2025 surveys, OpenAI’s enterprise share fell from roughly 50% to 27% while Anthropic rose to 40%.1, 5That is a dramatic upset, and it happened entirely among the same three names. The competition is real. It simply takes place inside a very small room.
A cost curve, compounding quietly, did the work no villain had to.
We are losing the argument by fearing the wrong thing.
Two stories compete for our attention, and the louder one is winning. The cinematic version stars a single superintelligence, self-aware and unaligned, slipping its leash; it is vivid, it is eminently fundable, and it keeps our eyes fixed on the machine. The quieter version is made of paperwork instead: a quarterly invoice, a terms-of-service update, a board seat, an equity stake. That second story is dull to look at, which is part of why it is the one actually reshaping who controls cognition.
One story turns on “what if the AI decides?” The more pressing one turns on “who decides what the AI does?”, and the answer there is a handful of private companies setting price, capability, and access for the systems a civilization increasingly reasons with. A single rogue mind is, in principle, something you could switch off. A few permanent gatekeepers over how everyone else thinks is something you have to live inside instead. We have been staring so hard at the monster that no one stopped to read the lease.
The same scoreboard, read by three different readers.
Concentration is not yet destiny. The same data that diagnoses the problem also points to where the leverage still sits, and what you do with that leverage depends a great deal on who you are.
For individuals
Three vendors shouldn’t end up owning your only way to think.
- Spread your dependence on purpose: use more than one provider, and keep capable open-weights options in the mix so no single lease owns your workflow.
- Keep the hard cognitive work yours, and treat the model as a tool you direct rather than the place your reasoning lives.
- Read the terms. Because only three sellers set them, price, access, and capability can all shift under you with little warning.
For employers
A workflow wired to one frontier lab is a single point of failure.
- Avoid lock-in: build for portability across providers so a price hike or a policy shift can’t hold your operations hostage.
- Remember the stack underneath, where a lab and its hyperscaler can amount to the same exposure, and map your real concentration rather than just your vendor list.
- Fund and pilot the long tail of open weights and smaller labs to keep a credible alternative alive on your own balance sheet.
For policymakers
The structure was studied and left standing. That was a choice, and it can be revisited.
- The FTC’s own 2025 staff report asked whether these partnerships “risk distorting innovation and undermining fair competition.” Answer the question with teeth, not just another report.
- Look at the compute layer rather than only the model layer, since concentration among hyperscalers is the foundation the whole oligopoly stands on.
- Fund public and independent compute so the frontier stays reachable by more than the handful who already own it.
FAQ
Often it is, and concentration by itself isn't the alarm. Search consolidated. So did operating systems and cloud infrastructure. What's different here is the layer doing the consolidating. The market being captured isn't for a product you use; it's for the faculty you reason with. When three firms set the price, the capabilities, and the terms of access for the systems people increasingly use to write, decide, and remember, the concentration sits on top of cognition itself. A worse search engine you can route around. A narrower set of minds doing the thinking for a civilization is harder to leave.
It's the strongest counterweight we have, and it genuinely matters. But notice the shape of the Menlo estimate. Meta's Llama, Cohere, Mistral, and the entire long tail divide the remaining 12% of enterprise API usage against the three leaders' 88%. Open weights also run into the deeper choke point: training a competitive frontier model still demands compute at a scale only a few hyperscalers can supply. So open-source has widened the set of people who can deploy a strong model. It has not yet widened the set who can build one at the frontier.
Because the membership stays stable even when the ranking doesn't. Between Menlo's 2024 and 2025 surveys, OpenAI's enterprise share fell from roughly 50% to 27% while Anthropic rose to 40%. The leaderboard reshuffled, yet the set of names on it didn't change. An oligopoly was never defined by a permanent number one; it's defined by the same small group capturing the market year after year while everyone else stays pinned to the margins. That is what the data shows.
They vary, and we've flagged that inline throughout. The 88% figure is a market estimate from a 2025 enterprise survey rather than an audited census, so read it as directional, not exact. The training-cost figures are published estimates from Stanford HAI and Epoch AI, plus a CEO's projection from Amodei; they are not invoices. The investment and stake figures come from company announcements and reporting. What makes the pattern hard to wave away is the convergence: market share, compute economics, capital structure, and regulatory attention all four point at the same names from four different directions.
Watch the compute rather than the chatbots. The visible competition plays out at the model layer, where names rise and fall and the headlines follow. The durable power sits one level down, with whoever owns the data centres, the chips, and the capital that funds each training run. Anthropic is backed by both Amazon and Google; OpenAI by Microsoft. You can swap which lab you admire this quarter and still find the same three hyperscalers underneath it. Follow the infrastructure and the concentration the leaderboard keeps hiding comes into view.
The danger was never a single mind grown too powerful. It’s a circle of owners grown too small.
None of this is an argument against frontier AI. The models are remarkable, and the labs building them are doing real work. It is an argument for noticing the ownership before it hardens, and for keeping more than three doors open while opening them is still possible. Concentration over cognition is the kind of power that sets its own terms once it settles, so the moment to widen the room is now, while you can still make out the walls.
Whatever else you track, track who owns the machines.
Sources
Every figure in this paper is drawn from the primary sources below. Where a number is a market estimate, a projection, or a survey result rather than an audited fact, we have said so in the text.
- Menlo Ventures (2025). "2025: The State of Generative AI in the Enterprise." Market-share figures are estimates from an enterprise survey of LLM API usage. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
- Stanford HAI (2024). "The AI Index Report 2024": estimated training-compute costs (GPT-4 ~$78M; Gemini Ultra ~$191M). https://hai.stanford.edu/news/ai-index-state-ai-13-charts
- Cottier, B. et al. / Epoch AI (2024). "The rising costs of training frontier AI models." arXiv:2405.21015 (~2.4× annual growth in training cost since 2016). https://arxiv.org/abs/2405.21015
- Business Insider (2024). Dario Amodei, Anthropic CEO, on $1B training runs in 2024–25 and ~$10B by 2026. https://www.businessinsider.com/anthropic-ceo-cost-10-billion-train-ai-years-language-model-2024-4
- CNBC (2024). "How Microsoft's $13 billion bet on OpenAI…" https://www.cnbc.com/2024/08/10/rise-of-openai-microsofts-13-billion-artificial-intelligence-bet.html
- Amazon (2024). "Amazon and Anthropic deepen their shared commitment": $8B total investment. https://www.aboutamazon.com/news/company-news/amazon-invests-additional-5-billion-anthropic-ai
- The New York Times (2025). "Google Invests Another $1 Billion in Anthropic": $3B+ committed, ~14% stake. https://www.nytimes.com/2025/03/11/technology/google-investment-anthropic.html
- U.S. Federal Trade Commission (2024). "FTC Launches Inquiry into Generative AI Investments and Partnerships": 6(b) orders to Microsoft, OpenAI, Amazon, Anthropic, Alphabet; staff report Jan 2025. https://www.ftc.gov/news-events/news/press-releases/2024/01/ftc-launches-inquiry-generative-ai-investments-partnerships
Read it as a PDF
The complete argument, the cross-domain evidence, and all primary sources, typeset as a white paper. Free, no signup.
↓ Download the PDF