What's Inside
I've been watching the Chinese AI chip scene up close for years. When NVIDIA first unveiled the H20 in late 2023, I remember thinking “this feels like a compromise no one really wants.” But talk to cloud providers in Shanghai or Shenzhen today, and you'll hear a different story. The H20 isn't just tolerated — it's in high demand. Let me break down what's actually happening.
The Backstory: How H20 Was Born
The H20 is NVIDIA's answer to U.S. export controls that block the sale of high-end chips like the H100 and A100 to China. It's a trimmed-down version with reduced interconnect bandwidth and compute performance — roughly 15% of the H100's FP8 TFLOPS in some metrics. But crucially, it still uses the Hopper architecture and can handle AI inference and moderate training workloads.
Here's the part most analysis misses: Chinese companies don't see H20 as “crippled.” They see it as the most advanced chip they can legally get their hands on without risking supply chain disruptions. One hyperscaler procurement manager told me, “We'd rather have 80% of a Ferrari than 100% of a bicycle.” That mindset drives the current surge.
What's Driving the Demand? Three Forces
1. Inference workloads are exploding
Everyone talks about training giant models, but after DeepSeek and other players scaled up, the real bottleneck is inference — deploying those models in production. H20's memory bandwidth (almost 4 TB/s) and NVLink interconnect make it excellent for running large language models at scale. I visited a data center in Beijing where they replaced a cluster of older A800s with H20s and saw latency drop by 40%.
2. Domestic chip alternatives aren't ready
Huawei's Ascend 910B is the closest competitor, but software compatibility is still a nightmare. Developers have to port CUDA code to CANN (Huawei's platform), and it's not a smooth ride. I've talked to teams that spent months rewriting kernels and still got 30% worse performance. So even though H20 costs a premium — around $12,000–$15,000 per unit through gray channels — many companies consider it the safer bet.
3. Pre-buying before further restrictions
Industry insiders expect the U.S. to tighten rules again. Chinese firms are stockpiling H20s like toilet paper during a shortage. One logistics contact in Hong Kong told me container shipments tripled in the past quarter. They're not buying for immediate need — they're buying for future-proofing.
Quick numbers: Estimates suggest China will import roughly 800,000 to 1 million H20 units in the first full year of availability. Compare that to maybe 300,000 A800s in 2022. That's a 200%+ jump.
H20 vs. Domestic Alternatives: A Reality Check
Let's be honest: China's domestic AI chip industry isn't where headlines claim it is. Below is a comparison I've compiled from technical benchmarks and real deployments I've witnessed.
| Spec / Aspect | NVIDIA H20 | Huawei Ascend 910B | Cambricon MLU590 |
|---|---|---|---|
| FP16 TFLOPS (peak) | 148 | 256 | 96 |
| Memory Bandwidth | 4.0 TB/s | 1.4 TB/s | 0.8 TB/s |
| Software Ecosystem | CUDA (mature) | CANN (fragmented) | Proprietary (limited) |
| Power Efficiency | 5.4 petaflops/kW | 4.1 petaflops/kW | 2.9 petaflops/kW |
| Grey Market Price | ~$13,000 | ~$9,500 | ~$6,000 |
| Availability | Moderate (3–6 week lead) | Low (government priority) | Low (niche orders) |
See the pattern? H20 dominates in memory bandwidth and software maturity — the two things that matter most for AI inference clusters. Domestic chips often look better on paper (higher peak TFLOPS) but fail in real-world throughput because of bad memory or poor compiler.
Market Impact and Hidden Challenges
Supply chain kinks
NVIDIA can't simply flood China with H20s — each chip requires export licenses and customs scrutiny. I've heard stories of shipments held up at Singapore ports for weeks while customs verify end-user documentation. That pushes delivery times out, frustrating customers who need immediate capacity.
Software lock-in trap
Here's a non-consensus point: buying H20 now actually increases China's long-term dependence on NVIDIA. Companies that invest in CUDA-optimized codebases will find it even harder to switch to domestic chips later. It's like choosing iOS over Android — you get a smooth experience, but you're locked in.
A junior AI engineer at a startup told me, “I know we should diversify, but my boss says 'just get the H20s working now, worry about switching later.'” That short-term thinking is exactly what makes the H20 demand sticky.
What This Means for Investors and Stock Directions
If you're tracking stock directions, the H20 phenomenon has direct implications. NVIDIA's datacenter revenue from China — which dropped after export controls — is recovering thanks to H20. Meanwhile, Chinese chipmakers like HiSilicon (Huawei's chip arm) are losing mindshare. I'd argue that the H20 demand extends the “AI trade” narrative for NVIDIA, at least in the short term. But watch for a regulatory shoe to drop: if the U.S. bans H20 too, the demand could evaporate overnight, and stocks tied to Chinese AI infrastructure would wobble.
Personally, I think the smart money is watching the gray market premiums. They're a leading indicator. When H20 margins in unofficial channels start shrinking, it means supply has caught up – or demand is softening. Right now, premiums are still 15-20% above official pricing, signaling sustained hunger.
Reader Comments