Nvidia Plans Small-Batch Shipments of Custom AI Chips to China Within This Year
Nashnova编辑部
Nvidia plans to begin shipping a custom LPU variant to Chinese customers by year-end — a chip that meets export controls without any performance cuts — but whether Beijing will allow the orders remains the biggest unknown.
What is this chip, and why does it skip the redesign?
Nvidia modified the software around its Language Processing Unit — an LPU, a chip built specifically for AI inference — so it can work with processors already available in China.
This means → the LPU's raw compute already sits below the export-control threshold; no hardware cuts were needed, only a software adaptation.
In plain terms = the chip is compliant by design. Nvidia's job was simply to make it run on Chinese hardware stacks.
According to *The Information*, several Chinese customers have placed orders. Reuters cited the report but said it could not independently verify the details.
Will Beijing allow it?
The biggest uncertainty is not in Washington but in Beijing. Previously, Beijing blocked or discouraged Chinese firms from buying Nvidia chips that already had U.S. export clearance, including the H20.
This means → even a fully compliant chip can be stopped on the Chinese side for industrial-policy reasons. Compliant ≠ sellable.
A recent signal of easing: Chinese regulators told firms they plan to permit some H200 purchases, and small batches have started arriving.
How severe is China's inference-chip shortage?
Moonshot AI suspended new subscriptions after launching Kimi K3 due to insufficient compute. DeepSeek sharply raised developer-access pricing for its latest V4 model this month.
Major Chinese tech firms have locked up most of domestic chip designers' capacity for this year and have begun booking 2027 orders.
This reflects a supply-demand gap far from closed. Nvidia's LPU variant targets exactly this window — re-establishing a market presence before Chinese customers fully pivot to domestic chips.
Who is it competing against?
In China's inference-chip market, the LPU variant will go head-to-head with Huawei's Ascend 950DT, also designed for inference and expected to ship in volume later this year.
This means → Nvidia and Huawei will clash in the same time window, chasing the same customers. First mover gains the edge.
What do the grey channels reveal?
Three people with direct knowledge say some Chinese firms have obtained H200s and Nvidia gaming chips — both banned by Beijing — through resellers and intermediaries.
In plain terms = the blockade is porous. The more acute the compute shortage, the harder demand is to suppress, and the more active grey channels become.
On the supply side, chip packager Amkor Technology completed packaging of over one million H200 units in Q2, primarily for Chinese customers, with core components supplied continuously by TSMC.
What determines whether this gambit works?
Whether Nvidia can lock in enough Chinese customers with the LPU variant before Beijing makes a final call is the decisive test.
This means → this is a race against policy. Chip compliance is just the entry ticket — speed and customer lock-in are what decide the outcome.
Content is for reference only, not financial advice.