Source item

From a supply-chain perspective, here’s a quick look at OpenAI’s plan to start deploying 1 GW of AMD MI450 systems in 2H26—and what it could mean for Nvidia. 1. There have been no significant changes to AMD’s 2026 CoWoS orders over the past two weeks. 2. 1 GW MI450 deployment translates to roughly 50K CoWoS-L wafers. Current estimates put AMD’s 2026 CoWoS orders at 60K–80K wafers (80–90% for MI400 series). This means AMD's current CoWoS order plan is sufficient to support the 1 GW build-out, under both optimistic and conservative scenarios. 3. A quick look suggests that the main beneficiaries of OpenAI’s AMD order are suppliers tied to HBM and UALink. The MI450’s HBM4 is mainly supplied by Samsung. Although the full UALink spec won’t reach mass production until the MI500 series in 2027, the strong order visibility should be reflected early in the stock prices of related suppliers, such as Astera Labs. 4. Judging from Nvidia’s painful experience with rack-level deployments, rolling out 1 GW of MI450 systems is unlikely to be a smooth ride either. What Nvidia can do—and clearly has been doing—is raise the competitive landscape before AMD’s rack-level servers start shipping in volume, thereby securing its next phase of advantage. As long as the overall AI compute market keeps expanding, OpenAI’s partnership with AMD should have limited impact on Nvidia.