Anthropic is the fastest-growing startup in history, but staying there might prove to be bigger challenge.
Its numbers are genuinely extraordinary. Revenue went from $100 million in 2023 to $4.5 billion by mid-2025 — a 45x increase in eighteen months. Its current annualized run-rate is an astonishing $47 billion. Anthropic’s valuation has moved even faster, going from $61.5 billion in early 2025 to $965 billion by May 2026. That kind of trajectory tends to silence skeptics. But the speed of the ascent is precisely what makes it worth scrutinizing, because the assumptions baked into a $965 billion valuation are fragile in ways that aren’t getting enough airtime.
The 3-6 Month Problem
The central issue is one that the AI industry mostly avoids saying out loud: open-source models currently lag the frontier by approximately four months. That’s not a rumor or a pessimist’s estimate — it’s from Epoch AI’s analysis of the Epoch Capabilities Index, which tracks aggregate model performance across a broad range of tasks. DeepSeek itself, in its V4 technical report, acknowledged the same gap, describing its leading model as trailing state-of-the-art frontier models by approximately 3 to 6 months.
This creates a specific problem for Anthropic’s valuation. If AI capabilities plateau — or even just slow down — at any point, open-source would need only a few months to catch up to wherever the frontier stopped. At that point, the premium Anthropic charges for API access becomes very hard to defend. Models become a commodity. Pricing converges toward the raw cost of generating tokens, and the margin that justifies a near-trillion-dollar valuation evaporates.
The counterargument is that capabilities won’t plateau — that the frontier keeps moving fast enough to keep open-source perpetually behind. That’s possible. But it’s also circular: you’d be valuing Anthropic based on its ability to outrun a race it cannot afford to lose. And even if capabilities do keep improving indefinitely, there’s a subtler problem. The vast majority of tasks that humans do today don’t require the absolute frontier. They require something good enough, reliable, and cheap. Open-source models are already there for large swaths of real-world use cases. Anthropic’s premium pricing assumes enterprises need the best model available. Increasingly, they don’t — they need a model that works, at a price that makes the ROI math work.
Models Are Just Weights
There’s a structural issue that doesn’t get discussed enough: AI models, at their core, are files. Weights. A frozen set of parameters that can be copied, fine-tuned, and run by anyone with sufficient compute. That’s a fundamentally different kind of moat than, say, a proprietary drug formula with patent protection, or a network-effects business where switching is painful.
The talent dimension makes this more acute. The people who train frontier models are, to a significant degree, portable. Researchers leave labs. They start new labs. They take their intuitions, their architectural insights, and their training recipes with them — and they produce competing models. Anthropic itself is an illustration of this dynamic: it was founded by people who left OpenAI. There is no reason to believe the pattern stops there.
What’s more, AI is now accelerating the very process that threatens it. As models become capable enough to assist in their own training and improvement, the role of individual researchers diminishes. The barrier to standing up a new lab drops. When recursive self-improvement becomes a real workflow — not a sci-fi concept but a standard engineering practice — the notion that any single lab can maintain a durable, defensible advantage over a multi-year horizon becomes much harder to sustain. Competition compounds. Margins compress.
Anthropic’s Particular Vulnerability
OpenAI has built something genuinely hard to replicate: a consumer brand with nearly a billion daily users for ChatGPT. Google has Android, Search, and Gemini baked into products that people use whether they think about AI or not. Both companies have distribution that is, in a meaningful sense, structural.
Anthropic doesn’t have that. Its valuation is overwhelmingly a bet on enterprise API revenue — businesses paying for Claude access to power their products. That’s a legitimate business. But it’s also the most price-sensitive segment of the market, and the one with the most viable alternatives. Enterprise buyers have procurement teams. They benchmark. They switch.
There are already concrete signs of pressure. A startup CEO recently disclosed that his company switched entirely from Anthropic’s models to DeepSeek V4, reporting millions of dollars in savings and an actual improvement in performance on their core use cases. That’s a single data point, but it’s a meaningful one: the migration happened not because Claude was bad, but because the open alternative was good enough and dramatically cheaper. Uber has reportedly pushed back on API costs as well. When companies at that scale start doing the math, the math tends to win.
The Lindy case also reveals something about Anthropic’s pricing power. The CEO was explicit: inference was his company’s single largest cost line, more than payroll. For AI-native companies, this is common. When the difference between two models is 4x on cost and marginal on quality, the business case for paying the premium becomes genuinely hard to make. Crivello described the engineering migration as “100x more work than we thought” — and still concluded it was worth it. That’s a signal the premium is too high, not a sign of loyalty.
The On-Device Future
There’s a dimension of this that doesn’t fit neatly into the API pricing conversation, but may matter more in the long run: on-device models.
Google’s Gemma and a growing ecosystem of small, capable models are making it increasingly plausible that useful intelligence will simply live on a laptop or phone. No API call required. No inference cost. No dependency on a cloud provider. For a wide class of tasks — writing, summarization, code completion, basic reasoning — on-device models are already competitive with what frontier APIs offered two years ago, and improving rapidly.
If that trajectory continues, the addressable market for API-based AI shrinks. Not to zero — there will always be tasks that require the most capable models, running on the largest hardware. But the everyday intelligence layer, which is where most of the API call volume lives, could move to the edge. That’s a structural headwind for every cloud AI provider, and Anthropic, as the most API-dependent of the major labs, is the most exposed.
What Would Have to Be True
To justify $965 billion, Anthropic needs a set of things to remain true simultaneously. Capabilities must keep improving fast enough that open-source stays meaningfully behind. Enterprise buyers must remain willing to pay a significant premium for that gap. No well-funded competitor — from China, from within the US, or from a new entrant built on open-source infrastructure — can close that gap fast enough to force a price war. On-device AI must remain insufficient for enough use cases that API demand keeps growing. And Anthropic must build enough of a distribution moat — through integrations, trust, compliance certifications, enterprise relationships — that switching costs rise faster than the capability gap narrows.
None of those things are impossible. Some of them may even be likely. But all of them, simultaneously, over the multi-year horizon implied by a near-trillion-dollar valuation? That’s a heavy ask.
The skeptics who called Anthropic overvalued at $5 billion were spectacularly wrong. Keith Rabois, who in 2023 said he was “highly skeptical” the company would ever be worth $5 billion in ten years, has been proven wrong with a speed that’s almost embarrassing. But being wrong about $5 billion doesn’t mean the market is right about $965 billion. The business is real. The growth is real. The question is whether the durability of that growth is real — and there are more reasons to doubt it than the current valuation suggests.