Anthropic traded competitive distance for survival. Demand for Claude has outstripped its hardware, forcing the AI lab into a deal with a rival: leasing the Colossus 1 data center controlled by Elon Musk’s SpaceX and xAI. On FYI, Frank Downing explained the 300-megawatt facility’s mixed chip architecture is ideal for inference, allowing Anthropic to run existing models while hoarding its own Nvidia GPUs for research.
"This marks a sharp pivot. Anthropic previously blocked Musk’s xAI from accessing its models. Now, the need for inference capacity has overruled strategic insulation."
- Frank Downing, FYI - For Your Innovation
The pivot reveals a staggering price of entry. Building a one-gigawatt AI data center costs $60 billion upfront. On FYI, Downing noted that if a model builder like Anthropic uses the capacity itself, it can generate $30 billion annually - a two-year payback period that explains the current frenzy. SpaceX aims to bypass terrestrial limits by launching compute clusters into orbit, where Brett Winton expects the shift to begin around 2028.
Meanwhile, Anthropic’s restrictive stock policies have spawned a chaotic $200 billion shadow market. On Bankless, Dio Casares detailed a landscape where layered special purpose vehicles (SPVs) extract 10% fees plus carry, and roughly 10-20% of proposed deals involve fraud or misrepresentation. Brokers sell forward contracts on employee shares that vanish if the employee is fired or sued.
"It’s a gold rush where the middlemen are out-earning the primary investors. The further an investor moves from the cap table, the more they are buying a promise rather than a share."
- Dio Casares, Bankless
Anthropic is now explicitly warning investors. On Bitcoin And, host David Bennett reported the company declared any stock transfer via SPVs or crypto tokens without board approval is void, labeling platforms that imply a $1.5 trillion valuation as potential fraud. This fight for cap table control clashes with a secondary market that Casares predicts will trigger a "logistical nightmare" at IPO, with years of litigation to untangle nested SPV distributions.
The company’s compute crisis is degrading its product. On Nerd Snipe, Theo argued Anthropic routes public users onto inferior Google TPUs and Amazon Trainium chips to reserve Nvidia stacks for researchers. An audit by AMD’s head of AI showed that between January and March, input tokens for similar tasks in Claude Code increased 170x, pushing one project’s estimated monthly costs from $26 to over $42,000.
"The data is staggering: the input tokens required for similar tasks increased by 170x, while output tokens grew 64x. This inefficiency translates to a massive cost explosion."
- Theo, Nerd Snipe
The pressure is structural. On The AI Daily Brief, Nathaniel Whittemore noted Claude Code’s annualized revenue jumped from $1 billion to $2.5 billion in two months, demonstrating the market’s shift from chatbots to autonomous builders. Yet Anthropic’s clash with the Pentagon over military use led the Defense Secretary to designate it a supply chain risk - a fracture that OpenAI capitalized on by signing an agreement with the Department of War. For Anthropic, the race is now threefold: secure enough compute to serve growth, police a fraudulent secondary market, and maintain a product that hasn’t been lobotomized by its own infrastructure choices.
