Meta Is Doubling Down, Not Pulling Back

Per a Reuters exclusive citing an internal memo: Meta plans to put its in-house AI chip, called Iris, into production in September. The chip is part of Meta's MTIA project, designed with Broadcom and manufactured by TSMC, aimed at reducing reliance on Nvidia and AMD GPUs. Meta plans to double its computing capacity to 14 gigawatts in 2027, and expects to spend up to $145 billion on AI infrastructure this year. The memo also locked in long-term supply deals — Samsung for memory, SanDisk for flash storage, Sumitomo Electric for fiber-optic equipment — and acknowledged that adopting the latest GPUs at Meta's scale "has been a heavy lift" and "has cost us time." Source: Reuters (Katie Paul, Max Cherney, Stephen Nellis)

Layer 2 of this framework is about AI capex growing faster than the revenue meant to justify it. This memo is that spending accelerating, not moderating. A week earlier, a report that Meta would sell "excess" compute was read by the market as a sign of capex discipline — hyperscalers finally slowing down. This memo says otherwise: you don't put your own chip into production, double your capacity, and commit $145 billion while sitting on a genuine surplus. The tell is buried in the memo itself — adopting the latest chips "has been a heavy lift" and "has cost us time." That is a company spending enormous sums into a buildout that is harder and slower than planned, with the payoff still unproven. It also broadens the picture: Meta building its own chip is one more large buyer designing around Nvidia, a slow erosion of the pricing power the whole GPU supply chain assumes.