How Claude went from $9 billion to $45 billion in one year | CFO explains

Krishna spends 30 to 40% of his time on compute decisions—the most consequential and hardest decisions in the company. The stakes are binary: buy too much compute and you go out of business from capital inefficiency; buy too little and you can't serve customers or stay at the frontier, which amounts to the same outcome. The company operates within what Krishna calls a «cone of uncertainty,» where small movements in weekly or monthly growth rates compound into vastly different outcomes over 12 to 24 months.

Anthropic uses three different chip platforms—Amazon's Trainium, Google's TPUs, and Nvidia's GPUs—fungibly across model development, internal acceleration, and customer inference. This flexibility didn't happen overnight; it required years of investment to become what Krishna believes are the most efficient users of compute among frontier labs. The orchestration layer they've built allows them to dynamically allocate compute across workloads, running inference on a chip in the morning and model development on the same hardware in the afternoon.

Flexibility extends to procurement strategy as well. The company builds optionality into deals, balancing near-term compute (like the Memphis Colossus partnership with SpaceX) with massive long-term commitments (5 gigawatt deals with Google/Broadcom and Amazon totaling over $100 billion, with $50 billion still to be deployed). They assess each opportunity through a consistent framework: price-performance, duration, location, chip type, and how efficiently they can run it—whether it lands next month or in 2027.

For Anthropic, scaling laws are alive and well—progress is accelerating, not slowing. Over 90% of the company's code is written by Claude Code, and much of Claude Code itself is written by Claude Code. This recursive self-improvement loop means that the models are helping to build the next generation of models. Talent remains critical—setting direction, conducting experiments, identifying new areas of discovery—but the models accentuate and accelerate that talent in ways that break prior paradigms.

Internally, the company evaluates models at various points during pre-training runs, comparing loss curves and capability snapshots to prior generations. They also rely heavily on customer feedback: pain points become training targets. Customers tell them, «I wish the model were better at X,» and the response is, «Build your product for that capability—we're going to improve it.» This connected loop between R&D and deployment creates a flywheel where better models unlock more TAM, which funds more compute, which trains better models.

The implications are profound. Tom Brown, the Chief Compute Officer, described a vision on a walk with Krishna in early 2024 that sounded like science fiction—compute scale, model capabilities, and timelines that seemed impossible. Much of what Tom predicted has already come to fruition, and there's more beyond the current frontier. Krishna's takeaway from that conversation: everything is going to happen much quicker than we think, and the capabilities will bend all prior paradigms of what's possible in enterprise software.

The release of Mythos marked an inflection point—the first time Krishna heard friends who carefully watch AI say, «This one kind of makes me scared.» Mythos is an incredibly capable model across many dimensions, but it spiked particularly in cyber capabilities. An earlier model found 22 security vulnerabilities in an open-source codebase; Mythos found 250. That statistic is both promising (for defensive patching) and concerning (for offensive potential).

For the first time, Anthropic decided on a phased release. Rather than never releasing the model or releasing it broadly, they created an expanding access tier focused on ensuring the cyber capabilities are used defensively—to patch code bases, not exploit them. This approach reflects the company's cultural commitment to AI safety and responsible deployment. The decision wasn't to suppress capability, but to shape how it enters the world. Krishna sees this as a template that could be used for future releases, acknowledging that with great capability comes the need for thoughtful stewardship.