Nvidia DMs TSMC: please sir can I have some more? The Chinese are starved for H200s

Steven

Nvidia DMs TSMC: please sir can I have some more? The Chinese are starved for H200s

Summary

Nvidia has reportedly approached TSMC to boost production of its H200 AI accelerators after Chinese firms placed orders totalling more than two million units, according to Reuters. That figure is a sharp rise from initial reports of 40,000–80,000 orders and far exceeds the roughly 700,000 H200s currently in stock.

The H200 is built on TSMC’s 4N process (an older node than Nvidia’s high-end Blackwell chips) and remains, for now, the most capable GPU Nvidia is permitted to sell into China after recent US export adjustments. Shipments are expected to begin in the second half of 2026, with 8‑GPU systems priced at about 1.5 million yuan each. Nvidia agreed to share around 25% of revenues with the US as part of the export arrangement.

Major Chinese hyperscalers — reportedly including ByteDance, which may spend roughly $14bn on H200s — are racing to secure capacity. However, Beijing has not yet approved imports, and Chinese regulators are simultaneously pushing domestic chip alternatives and restricting state-funded datacentres from using foreign AI accelerators amid security concerns Nvidia denies.

Key Points

  1. Chinese companies have reportedly ordered over 2 million Nvidia H200 GPUs, up sharply from earlier estimates of 40k–80k.
  2. Only ~700,000 H200s are currently in stock, prompting Nvidia to ask TSMC to ramp production on the 4N node.
  3. Shipments of H200s are expected in H2 2026; 8‑GPU systems are priced at about 1.5 million yuan (~US$215k).
  4. The US approved H200 exports to China under conditions including Nvidia sharing ~25% of revenues; Beijing’s approval is still required.
  5. ByteDance and other hyperscalers are major buyers, but Chinese policy pressures and domestic alternatives could block or limit adoption.
  6. The H200 significantly outperforms the export‑compliant H20: higher FP performance, more HBM3e and greater memory bandwidth.

Context and relevance

This story sits at the intersection of geopolitics, supply chains and the AI compute race. Availability of Nvidia accelerators influences which cloud providers and datacentres can train large models and deploy high‑performance AI services. The H200 scramble highlights Nvidia’s strong market leverage, TSMC’s critical role in chip supply, and how US and Chinese trade policies continue to shape global AI infrastructure procurement.

Why should I read this?

Because it’s the real‑world boil‑down of who’s going to get the horsepower for AI next year. If you buy cloud, build datacentres, run models, or follow the US‑China tech tug‑of‑war, this is the sort of supply‑chain drama that changes timelines and costs — and yes, it could affect availability and pricing for everyone.

Author’s take

Punchy: This isn’t just a sales spike — it’s a bellwether. Nvidia levering TSMC to meet a Chinese buying rush tells you where compute demand is headed and how fragile the global AI hardware market remains. Read the detail if you’re tracking AI infrastructure, vendor leverage or geopolitics; skipping it risks missing the next supply squeeze.

Source

Source: https://go.theregister.com/feed/www.theregister.com/2025/12/31/china_nvidia_h200/