Nvidia asks TSMC to ramp H200 production: China orders surge

Nvidia has asked TSMC to raise production of its H200 data-center AI GPUs. Chinese buyers placed orders for more than 2 million units, Reuters and The Register report. Nvidia currently holds about 700,000 H200s in inventory and is seeking additional wafer capacity to meet demand. Shipments from any expanded run are expected to start in early 2026.

The H200 is Nvidia's newest AI accelerator for large language model training and inference. Limited wafer supply and strong China demand have created acute allocation pressure. Multiple outlets report Nvidia asked TSMC to boost manufacturing throughput rather than reroute existing stock. Separately, Tom's Hardware says ByteDance may buy roughly $14 billion of Nvidia GPUs in 2026 (single-source report).

If TSMC approves a ramp, lead times will fall only after extra wafer and packaging capacity is online, so procurement should expect rolling deliveries. Procurement teams should confirm allocation schedules and prioritize critical workloads. Cloud and data-center operators should verify allocation terms and plan multi-vendor capacity to reduce risk.

Why It Matters

• Expect staggered deliveries; verify allocation schedules with vendors and adjust procurement timelines rather than assuming immediate fulfillment.
• Prioritize critical training and inference workloads for confirmed capacity, and plan fallbacks to cloud or alternative vendors to avoid delays.
• Capacity gains from a TSMC ramp require wafer and packaging cycles—improvements should appear starting in early 2026, not immediately.