Anthropic just admitted, in public, that it can't build compute fast enough to meet demand for its own best model.
On July 17, the company's official Claude account on X laid out the next stage of the rollout. It's already been delayed twice. Starting July 20, Claude Fable 5 becomes a standard part of every Max plan and every Team Premium plan, but capped at 50% of normal usage limits. Pro and Team Standard subscribers keep access too, though only through usage credits. Anthropic is handing them a one-time $100 credit to soften the transition. The company's own explanation, posted on X, was blunt: demand for Fable has been "challenging" to meet.
That's not marketing language. That's an infrastructure problem, spoken out loud.
Fable 5 first went live across Claude Platform, Claude.ai, Claude Code and Claude Cowork on July 1. The original plan gave Pro, Max, Team and select Enterprise customers full access at up to 50% of weekly limits, with a switch to metered usage credits set for July 7. That deadline came and went. Forbes reported the window got pushed to July 13, then again to July 19, as Anthropic scrambled to secure more capacity behind the scenes. July 20 is where the free run ends and the rationed version begins.
You don't extend a deadline three times because everything is running smoothly.
The capacity problem isn't a secret anymore. In May, Anthropic signed a deal to pay SpaceX close to $45 billion over three years, roughly $1.25 billion a month through 2029, for dedicated access to the Colossus 1 data center in Memphis, Tennessee. That single facility runs more than 220,000 Nvidia GPUs, a mix of H100, H200 and GB200 chips, drawing over 300 megawatts of power. Anthropic said at the time the extra compute would go straight into improving capacity for Claude Pro and Max subscribers.
It clearly wasn't enough. CNBC and the New York Times reported on July 17, the same day as the Fable 5 announcement, that Anthropic is now in early talks to lease a second chunk of compute from Meta. The deal could reach $10 billion over two years. The talks are preliminary and could collapse, according to that reporting. But the fact that Anthropic is shopping for a second multibillion-dollar compute deal within three months of signing the first tells you how tight the Nvidia GPU market really is. That's the real story here. Meta, for its part, is leaning into the role of landlord: the company is on pace to spend as much as $145 billion on capital expenditures in 2026, more than double the $72 billion it spent the year before, much of it aimed at exactly the kind of data center capacity Anthropic needs.
Frankly, this is the AI compute crunch playing out in real time, not in an analyst note but in a subscriber facing product change.
What it means if you're building on Claude #
If your startup runs on Claude Fable 5, plan around a hard ceiling now, not later. Half your normal usage limit on Max and Team Premium is a real constraint on any workflow that leans on the top tier model for coding, research or long context reasoning. Teams that blow through that 50% land in usage credit territory. Credits run $10 to $50 per million tokens depending on the tier, and that bill adds up fast for anyone running Fable 5 in production rather than for occasional queries. Build in a fallback. Route routine tasks to a lighter Claude model and save Fable 5 calls for the work that actually needs frontier level reasoning. Call it good architecture, not a hedge against Anthropic's execution - your own vendor has told you, in writing, that it's rationing the product.
Anthropic isn't alone in this squeeze, and it won't be the last lab to say so out loud. But it is, right now, the clearest public admission from a frontier AI company that demand for its best model has outrun its ability to serve it. Whether the Meta talks close, and how much compute they actually unlock, is still an open question. Founders building on Claude should plan for the current limits to stick around, not for them to lift on their own.
Also read: UK Safety Regulator Finds Jailbreaks That Turn GPT-5.6 Sol Into a Hacking Tool • Databricks Hits $188 Billion Valuation With New $3 Billion Coatue Round • Gemini 3.5 Pro's Delay Just Cost Alphabet Nearly $200 Billion in a Day