Anthropic's IPO risk isn't doom: it's the free model next door

What if the scariest sentence in Anthropic's IPO filing isn't about existential risk, but about customers deciding they don't need the smartest model in the room?

The filing has produced plenty of dramatic coverage: warnings about catastrophic harm, China using Claude to hack targets. But as the Australian Financial Review points out, Anthropic also slipped in a far more mundane threat. The growing popularity of open-weight models, built by companies other than Anthropic and OpenAI, means most businesses may never buy the frontier capabilities Anthropic is spending billions to invent. And the corporate market is exactly where the AI giants and their backers expect the bulk of profits to end up.

The stakes are big. CNBC reports that Anthropic was valued at $965 billion in May and is reportedly looking for an IPO valuation between $1.5 trillion and $2 trillion. That's not a price for a good software company. It's a price for a durable, monopoly-like position.

The moat is a treadmill, not a wall

The bullish story goes like this: Anthropic trains the best models, enterprises pay a premium, and the premium funds the next, better model. It works only as long as "best" is worth paying for. Open-weight models attack that premise from below. They don't need to beat Claude. They need to be good enough for most corporate tasks, the ones that don't require frontier reasoning, and then be free to run, tune and keep in-house.

Look at what small open models are doing. Liquid AI just released LFM3, six open-weight models built on a 1.7-billion-parameter backbone, downloadable immediately and fine-tunable on a single consumer GPU. At 4-bit precision the text model fits in roughly 1 GB. The company says it beats a model more than five times its size on long browser and business tasks, though that comparison is self-reported and unverified. Treat the claim with caution. But the direction matters more than any single benchmark: capability that used to require a data center keeps sliding toward a laptop or a phone.

In our view, that is the real problem with a valuation built on model quality. Quality is a depreciating asset. Every frontier gain becomes, with a lag, a commodity someone can download.

Cost structure makes it worse

Frontier labs fight this treadmill with enormous fixed costs. Forbes describes the filing as revealing soaring revenue, massive losses and about $500 billion in infrastructure commitments. A tracker account on social media attributes to Michael Burry even starker figures on compute obligations and slowing growth. We can't verify those numbers, so we won't lean on them. But the logic holds without them: if you carry huge commitments to build the best model, you need pricing power. Open weights are what erode pricing power.

This is also why the China angle in the filing matters. SCMP reports that the popularity of Chinese AI models has emerged as a key risk factor ahead of the listing, according to a US research firm. Cheap, capable models from abroad aren't just a geopolitical talking point; they are a price anchor.

Price the trust and the tooling, not the benchmark

Does that mean Anthropic is doomed? No, and we'd be wary of anyone who says so. It means the defensible part of the business is probably different from the one in the pitch.

What can't be downloaded for free? Reliability a legal department can sign off on. Security posture and accountability when something breaks. Integrations, agent tooling, support, and the boring work of making a model behave inside a bank or a hospital. A company can grab open weights; it still has to staff, secure, evaluate and update them. Plenty will happily pay to avoid that. We think Anthropic's safety-first image could be an asset here, not just a marketing line.

So the question for investors isn't "will Claude stay ahead on benchmarks?" It probably will, for a while. The question is how much of revenue comes from customers who would leave the moment a cheaper model is nearly as good, and how much comes from customers locked in by workflow, compliance and trust. The filing's disclosure of this risk is welcome, but the market will need numbers, not warnings.

Our verdict

If the IPO prices at anything near the reported $1.5 trillion to $2 trillion range, buyers are paying for a future where frontier intelligence stays scarce. Open-weight releases like LFM3 suggest scarcity is the shakiest assumption in the story. We'd value Anthropic as a trusted enterprise platform with a great research lab attached, not as a lab that happens to sell software. Those two companies deserve very different multiples, and the reported valuation looks to us like it prices the second one. Retail investors eyeing the debut should read the risk-factor section on open-weight competition before the one on existential doom. It's less cinematic, and likelier to hit the stock.