Nvidia Steps Back Into China Market With First H200 Chip Sales
THE SO WHAT
Nvidia trickling H200s into China under export caps reinforces a bifurcated compute world—Western hyperscalers get the bleeding edge at scale, Chinese buyers get constrained volume and performance. If your supply chain or customer base touches China, assume uneven model capability and latency across regions and architect for asymmetric deployments.
READ THE SOURCE
MORE FROM THE WIRE
Deep & Emerging TechKioxia Plans New Chip Facility in Japan to Meet AI Memory Demand
New Kioxia fabs in Iwate aimed at AI storage are a reminder that memory, not just compute, is a binding constraint on model scale and inference economics. Infra buyers should be modeling NAND and HBM availability and pricing into their 3–5 year AI roadmap, not just GPU counts.
Deep & Emerging TechChina sets new airborne wind turbine record at 4,000m — this 747-sized helium blimp generates megawatt power from high-altitude winds
Megawatt-scale power from a 4,000m airborne turbine is a bet that high-altitude wind can complement ground-based renewables where land or permitting are constrained. Data center and heavy-compute operators should clock this as a potential future option in geographies where grid expansion is slow.
Deep & Emerging TechNvidia forecasts Q3 revenue of $108B, above a $104B estimate, and expects FY 2028 revenue to grow by about 70%; NVDA jumps 4%+ after hours
Nvidia guiding to ~$108B in Q3 and ~70% growth through FY 2028 says hyperscaler and AI infra capex is not slowing—your GPU scarcity assumptions may be wrong on both price and availability. For builders, this buys a bit more confidence to plan multi-year GPU-dependent roadmaps, but also raises the bar on showing real ROI against a very visible cost line.
Deep & Emerging TechDataBank’s Incoming CEO on AI Buildout, Nvidia Earnings
A 70+ data center operator publicly framing power availability as the constraint tells you where the bottleneck has moved—chips are only useful if you can light them. If AI is material to your roadmap, you should be mapping your workloads to specific metros and power commitments, not just cloud SKUs.