Meta buys tens of millions of AWS Graviton 5 processor cores from Amazon

The scale here matters as much as the direction. Tens of millions of cores is not a pilot or a hedge — it suggests Meta is treating Graviton 5 as a primary compute path for at least some workload categories, which has real implications for how much of its AI spend stays off Nvidia's roadmap entirely.

The custom silicon bet fits a pattern that's been building across the coverage. When Cerebras filed for its IPO in mid-April (story [1]), the headline was the valuation, but the underlying signal was that specialized, non-Nvidia hardware is attracting serious institutional confidence. Meta's Graviton 5 commitment extends that same logic to hyperscaler chips: the industry is actively constructing alternatives to a single-vendor GPU dependency, and large buyers are now willing to commit at scale to make those alternatives viable. The RAM shortage story from The Verge adds a supply constraint dimension worth noting — if DRAM capacity stays tight through 2027, the economics of custom silicon optimized for memory efficiency become even more attractive.

Watch whether Google follows with a comparable public commitment to its own TPU v6 fleet at similar scale within the next two quarters. If it does, that confirms hyperscaler custom silicon has crossed from internal cost experiment to externally validated compute strategy.