Meta’s infrastructure leadership told an OCP engineering session on Thursday that the company is now running production training workloads on AMD MI450 racks at scale. The deployment sits alongside the existing Blackwell Ultra and Vera Rubin fleet, which makes Meta the first publicly confirmed customer running mixed-vendor training at frontier scale rather than the usual mixed-vendor inference. Internal estimates floating around the session put the AMD share at roughly twelve percent of Meta’s training compute by year-end. AMD declined to comment on the share.
The interesting piece is that ROCm 7, which shipped to Meta in March under a private cadence and is rumored to land for the broader developer pool later this summer, is the actual enabling technology. The historical AMD pitch has been “comparable hardware, slightly cheaper price, miserable software stack.” ROCm 7 closes most of that gap. Pytorch parity, distributed-training primitives that no longer require a custom collective library, and an inference path that does not require porting CUDA kernels by hand. Meta’s release cycle suggests the parity holds for the workloads Meta cares about most, which are large dense and MoE transformer training.
For Nvidia, the read on this is more interesting than the share number. Vera Rubin Ultra gross-margin assumptions priced in something close to a monopoly on hyperscaler training. The Samsung HBM4 qualification last week removed one squeeze point on the memory side. A real AMD hyperscaler design win on the compute side removes the other. Neither story moves the Nvidia P&L this quarter. Both stories change the negotiating leverage at the procurement table for the next generation. The hyperscalers have spent eighteen months publicly demanding optionality. They have it now, on both sides of the bill of materials, for the first time since 2023.
The AMD strategic question for the back half of the year is whether this becomes a Meta-only story or a hyperscaler-wide story. The tells in the next ninety days will be Microsoft’s Azure capacity guidance and any quiet language change in Google’s compute disclosures. The customer that names AMD second decides whether this is one customer or a category.