The argument over whether OpenAI bought too much compute is already obsolete — the question now is whether Anthropic bought enough.
The Summary
- OpenAI locked in massive compute deals early while Anthropic CEO Dario Amodei warned rivals were "YOLOing" — but Anthropic now faces service degradations while OpenAI ships.
- Internal OpenAI tension reportedly exists between Altman's capacity bets and CFO Sarah Friar's budget concerns, though both deny any real rift.
- Anthropic signed its own mega-compute deals at least six months behind OpenAI, a delay that's showing up as downtime and reliability issues.
The Signal
The AI infrastructure race has clarified into a simple truth: you can't reason your way out of insufficient compute. Anthropic is experiencing "genuine downtime and really degraded service" according to engineers managing production AI systems, while OpenAI ships reasoning models and multimodal tools at pace. The technical sophistication of Claude 3.5 Sonnet matters less when your users hit rate limits or 503 errors.
Altman's "compute is destiny" thesis looked expensive in early 2024. The reported internal friction with CFO Sarah Friar suggests real financial strain from those commitments. But the strain appears manageable compared to the alternative: being capacity-constrained while your competitor scales.
"OpenAI has been clearly ahead of the curve on compute" — Peter Gostev, Arena.ai
Anthropic waited at least six months after OpenAI to sign comparable infrastructure deals. That timing gap is the entire story. Six months in AI infrastructure procurement translates to:
- Later access to new chip generations
- Higher prices as hyperscaler capacity fills
- Missed compounding advantages in model iteration speed
- Service reliability issues that damage enterprise trust
The irony is sharp. Amodei positioned Anthropic as the thoughtful, sustainable AI lab while accusing unnamed rivals of reckless "YOLOing" on compute. That framing worked when both companies had similar capacity constraints. It breaks when one company ships consistently and the other apologizes for downtime.
This isn't about better models or smarter research. Lawrence Jones from Incident.io works with Netflix and Etsy on production reliability. When he says Anthropic's service is "bad right now," he's reporting what paying enterprise customers experience. Reliability determines which model gets embedded in actual products.
The Implication
The compute race isn't separate from the AI race — it is the AI race. Model quality matters, but only if you can deliver it consistently at scale. Anthropic's technical prowess won't overcome infrastructure gaps that manifest as user-facing failures.
For companies building on AI: diversify your model providers, but watch reliability metrics more than benchmarks. An 85% solution that ships beats a 95% solution that times out. For investors: compute commitments that looked aggressive in 2024 now look prescient. The companies that hesitated are scrambling to catch up at worse economics.