Inside Anthropic's $100 Billion AI Compute Commitment | CFO Krishna Rao

Key points

• **Anthropic financial trajectory disclosure (CFO-direct, not press-leaked).** Q1 2026 went from ~$9B run-rate to >$30B by end of quarter. **Last month closed >$100B in compute deals:** Google + Broadcom for 5 GW of TPU capacity starting 2027, Amazon Trainium up to 5 GW additional. **NDR (net dollar retention) >500% annualised.** Has now sold to **9 of the Fortune 10.** Anecdote: 'signed two double-digit-million dollar commits in an Uber ride on the way here.' **Anthropic is no longer in pilot territory at the enterprise — buyers are committing 8-figures per deal at first sale.**
• **Three-platform compute strategy = the multi-year flexibility moat.** TPU (Google), Trainium (Amazon Annapurna Labs), and Nvidia GPU. **Only model on all three clouds; only LLM lab using all three chip platforms.** Built a custom orchestration layer so compute is fungible across model dev / internal acceleration / customer serving. 'It took us years to do this — when we started on TPU v3, people said we were crazy.' Compilers + bare-metal optimisation built in-house. **Krishna spends 30-40% of his time as CFO on compute alone.** **The architectural moat at the inference/compute layer that other labs don't have.**
• **'Cone of uncertainty' as the compute-budgeting framework.** Small movements in monthly/weekly growth compound very differently over 12-24 months on an exponential. 'Humans think linearly; I had to break that paradigm for myself.' Plan compute to the *top* of the range (so you can serve at the exponential), then dynamically reallocate when reality is between scenarios. 'If you buy too much you go out of business; if you buy too little you can't serve customers and you fall off the frontier — they're the same thing.'
• **'Returns to frontier intelligence are extremely high — especially in enterprise.'** Krishna's central operating thesis. Intelligence is multidimensional (not just IQ): also long-horizon task ability, tool use, agentic action, computer use. **Each model generation has both capability gain AND efficiency gain (5x more capability + 3x cheaper to serve simultaneously is now normal).** Customer behaviour: 'when a new Opus drops, even consumers immediately switch — they want the best.' Direct counter to 'use the 6-month-old cheaper model' framing.
• **Compute floor for model dev is non-negotiable.** 'There's a level of compute for model development that we will not go below — even if it means it's harder to serve customers or we have to do unnatural things.' Internal allocation conversation is dynamic but the floor is sacred. **Sub-text: the customer-facing rate-limits we keep hearing about in [Issue 03 / 04](/issues/2026-05-10) are not capacity bugs — they're Krishna actively choosing to preserve research-team compute over customer-serving compute.**
• **Pricing-stability disclosure — the Opus 4.5 cut as deliberate Jevons paradox.** Pricing stable across Haiku/Sonnet/Opus tiers since launch (~3 years). **Only material price change: dropping Opus price at 4.5 launch.** Reason: 'customers were fitting Opus problems into Sonnet workloads because Opus felt too expensive — they were getting suboptimal results.' Efficiency gains made it serveable at lower price. Result: **consumption went up far more than the price cut warranted — Jevons paradox playing out exactly as theory predicts.** Confirms that pricing power is real but they're deliberately *not* using it to maximise short-term margin.
• **90%+ of Anthropic's code is written by Claude Code.** Direct disclosure: 'a lot of Claude Code's code is written by Claude Code.' Cross-references this week's [Toby Lütke 'River' agentic engineering thesis](/issues/2026-05-10) and Foroughi's 80%+ from [Issue 04](/issues/2026-05-10). **This is now the third operator-class disclosure of >80% AI-coded production codebases in 2 weeks. The 'AI-native engineering' transition appears to be functionally complete at the frontier.**
• **Internal use of compute as accelerant.** 'If we said employees can't use our models, we could serve billions of dollars more revenue with that internal compute.' But it pays off via faster R&D, better evals, faster model development. **Recursive self-improvement is happening at the operational level (not just model-architecture level) — they're trading near-term revenue for compounding model-quality velocity.** Krishna: 'scaling laws are alive and well; we see that even with Mythos.' Founding team includes original scaling-law-paper authors.
• **Two-week Dario all-hands as the cultural cadence.** Dario writes a short doc with 3-4 topics and takes open questions (not planted). Not a decision-making forum — a window into leadership thinking. **All 7 co-founders still at the company.** Vast majority of the first 20-30 employees still there. Lost only 2 people to Meta during this year's compensation poaching war (other labs lost dozens). **'Talent density beats talent mass'** as the explicit hiring frame.
• **Cyber-only release framework for Mythos.** Mythos's cyber capability 'spiked' relative to other dimensions — they decided to release it differently from prior models. **Phased release to a defensive-focused user group, then expanding over time.** Open-source-code-base test: prior model found 22 vulnerabilities, Mythos found 250. **'That's a template that could be used for the future' — high-cyber-capability releases get gated phased deployment to defenders first.**
• **'Virtual collaborator' as the next product frontier.** Has organisational context, tool access (homegrown + procured), memory, ability to learn from mistakes (its own + the company's), works over very long time horizon — not just a task but an idea. **Cowork is growing faster than Claude Code did at the same point.** 'Even our product development isn't 1 PM + 2 engineers shipping over 3 months any more — it's daily shipping with a fleet of agents.' **Cross-references Chesky's 'AI founder mode' Project Hawaii from [Issue 04](/issues/2026-05-10).**
• **Government partnership posture.** Anthropic 'America First in approach.' Worked closely with administration on the Mythos release process. Reads the regulatory-overhang situation as 'an honest conversation about implications' — actively in dialogue, not lobbying against. **Implicit: doesn't expect aggressive pre-release-approval regimes given the constructive working relationship.** Cross-references [Issue 03's 'No federal AI regulation within 12 months' prediction](/issues/2026-05-03).
• **Personal reflection — Tom Brown 2024 walk anecdote.** Joined Anthropic at ~$250M run-rate. Chief Compute Officer Tom Brown took him on a 2.5-hour walk around the Mission and described his vision for compute scale + model capability over 2-3 years. Krishna: 'sounded crazy; came home and told my wife — if even 10% is true this is going to bend every paradigm. Turns out most of it has come true.' **Useful as a base-rate for evaluating any 'this sounds crazy' AI-trajectory claim from a frontier-lab insider.**

Notable quotes

Themes

• Anthropic CFO-direct confirmation of $30B+ Q1 ARR and $100B+ compute commitment
• Three-platform compute strategy as the multi-year flexibility moat
• 90%+ of Anthropic code written by Claude Code — frontier-lab AI-native confirmed
• Pricing-stability + deliberate Jevons-paradox via Opus 4.5 price cut
• Talent density beats talent mass — lost only 2 to Meta in poaching war