Why Compute Is the Lifeline of AI: Anthropic CFO Reveals the Industry’s Harsh Truth

Developers are shocked by Anthropic’s June 15 pricing overhaul that separates programmatic Claude usage from the standard chat subscription. A former $20 Pro plan that covered chat, code, debugging, and agents now grants only $20 worth of credits per month for Pro users, $100 for Max 5× users, and $200 for Max 20× users, after which usage stops unless additional money is added.

This change hurts both individual developers and enterprises: credits are allocated per user and cannot be shared across teams, making it impossible to create a unified budget pool for shared automation tasks. A single runaway agent or poorly designed prompt can quickly exhaust a day’s worth of credits, and CI pipelines that retry or long‑context agents can cause unpredictable cost spikes.

In a recent interview, Anthropic CFO Krishna Rao emphasized that “compute is the lifeline of our business,” noting he still spends 30‑40 % of his time on compute issues. He described the core dilemma: buying too little compute prevents serving customers, while buying too much drags the company into unsustainable costs.

Rao detailed Anthropic’s massive compute spending: exclusive leasing of SpaceX’s Colossus 1 super‑computer, a ten‑year commitment of over $100 billion for 5 GW of AWS Trainium chips, a 3.5 GW TPU deal with Google and Broadcom slated for delivery from 2027, and a $50 billion investment to build its own U.S. AI infrastructure. Anthropic is the only language‑model lab using three chip platforms—NVIDIA GPUs, Amazon Trainium, and Google TPUs—layered as a “compute cake” that dynamically selects the most cost‑effective platform at any time.

Rao claims Anthropic’s $1 of compute generates more value than anywhere else, yet admits the company still feels “the money never stretches far enough.” Internally, over 90 % of Anthropic’s code is now written by Claude Code, and the internal compute is being used to train the next generation of models rather than being sold to customers.

The rising subscription fees are not merely a profit‑driven move; they reflect the underlying cost of electricity, chips, and data‑center GPUs, which have only risen over time. After Anthropic dramatically cut Opus model prices last year, usage surged—a classic Jevons paradox where cheaper, more useful tools lead to higher overall consumption.

Rao concludes that the AI industry’s ultimate paradox is not that AI is too expensive for users, but that its increasing usefulness makes it inevitably more costly. Anthropic bets that sufficiently powerful models will keep users willing to pay, even as developer backlash can outpace model iteration, echoing Microsoft’s recent decision to phase out Claude Code licenses in favor of its own Copilot CLI because the cost accounting simply didn’t add up.