OpenAI Anthropic AI commoditisation margins open source AI

Margin Call: Why AI's Biggest Players Are Building on Sand

19 May 2026

Leading AI companies like Anthropic and OpenAI are currently unprofitable, losing money even on premium subscriptions, and are under growing pressure to raise revenue while facing inevitable commoditisation of their models. Cheaper open-weight models from China and API proxy networks are eroding the pricing power of US frontier labs, with Chinese models expected to match current leaders by end of 2026. The likely long-term winners are platform gatekeepers like Apple, Google, and Microsoft, who control software distribution, while pure-play AI labs face shrinking margins and an increasingly difficult path to profitability.

Jeff Bezos once said your margin is my opportunity. He was talking about retail, but the observation applies just as well to anyone currently charging frontier prices for AI inference.

Anthropicand OpenAI are both burning cash at a spectacular rate. Their pitch to investors essentially boils down to 'trust us, the numbers will eventually work.' Some investors keep writing cheques. Others are quietly doing the maths.

The maths is not flattering. Claude Code subscribers paying $200 a month can apparently consume around $5,000 worth of tokens if they push hard enough. OpenAI is losing money on subscriptions too. Both companies have responded by nudging customers toward metered pricing, which at least means heavy users pay proportionally more. But it doesn't solve the underlying problem, which is that the cost of serving these models remains punishing and the competition is accelerating.

Benedict Evans, one of the more clear-eyed analysts watching this space, recently updated his 'AI eats the world' presentation with a fairly blunt assessment. The supply and demand imbalance that has let frontier labs charge premium prices won't last. Models will become commodity infrastructure. Pricing power will have to move up the stack or it won't exist at all.

You can see Anthropic already trying to get ahead of this. Claude Code, the CLI, the desktop app, Claude Cowork, Claude Design, assorted productivity wrappers. The goal is obvious: if the model itself becomes interchangeable, lock developers into your tooling and your workflow integrations instead. Whether that works is another question.

The more uncomfortable pressure is coming from China. Researcher Zilan Qian at the Oxford China Policy Lab has documented how Chinese developers are accessing US models through API proxy networks, often at a fraction of the official price. Every new control measure the labs introduce, geoblocking, phone verification, biometric KYC, gets matched by a new layer of workaround infrastructure. The logs from all that usage may themselves be getting traded for model training and other purposes.

It's a mess, and probably not sustainable in its current form. But it illustrates how difficult it is to maintain exclusivity when your product is, fundamentally, text going in and text coming out.

Open weight models are closing the gap faster than the labs would like to admit. GLM-5.1, Kimi K2.6, DeepSeek V4-Pro, Qwen3-Coder-Next. These are already good enough for a significant chunk of software development work. Qwen3.6-27B runs reasonably well on local hardware if you've provisioned it properly. US firms are estimated to be around seven months ahead of their Chinese counterparts right now. That's not nothing, but it's also not a moat.

Current projections suggest open weight models will reach rough parity with today's frontier leaders by the end of 2026. At that point, better benchmarks will still get announced and celebrated, but for most enterprise use cases it probably won't matter. Good enough will be good enough.

And the market for 'good enough' AI is the market that actually exists. According to Andreessen Horowitz, enterprises are spending around $3 billion annually on AI for coding. Every other category, legal, support, healthcare, sits well below $500 million. Evans's figures show that outside of the tech industry, daily AI usage barely registers. Consumer adoption looks impressive until you notice that only around five percent of ChatGPT's 900 million weekly users pay anything at all.

So Anthropic and OpenAI need prices to go up and adoption to broaden significantly. Both of those things are running in the wrong direction. Meanwhile, the companies best positioned to benefit from widespread AI adoption are the ones that already control how software reaches users. Apple, Google, Microsoft on the OS and device side. Amazon, Google, Microsoft again on cloud infrastructure. Notice how none of those are primarily AI labs.

The frontier model companies will cut distribution deals with these incumbents because they won't have much choice, especially given the constraints of mobile hardware. Those deals will extract a toll.

High margins attract competition. That's not a controversial statement. When you combine high margins with billions in accumulated losses, the window for turning things around gets narrower with every passing quarter.

Evans puts it well: sometimes software eats the world, and sometimes it only nibbles. The question for Anthropic and OpenAI is whether they get to be at the table, or whether they end up on the menu.