What reportedly happened
Meta had been buying Gemini capacity at scale alongside its own model development. Google, facing its own internal and customer demand, throttled the account. Both companies declined to comment on specifics, but neither disputed the underlying tension: there is not enough accelerator capacity for everyone who is willing to pay.
The signal
For two years the industry debated which lab had the smartest model. That race continues, but the binding constraint has shifted to who can serve their model at scale. Data centers, power contracts and chip allocations are the new moats, which is exactly what TSMC’s record quarter shows from the supply side.
What it means for founders
If hyperscalers ration each other, smaller customers should assume they are rationed first. Concretely: negotiate committed capacity if AI is core to your product, build graceful degradation into features that depend on third-party models, and keep a tested fallback provider. Availability is now a product-reliability issue, not a procurement detail.