Nvidia's B30A chip turns export controls into paywall

💡 TL;DR - The 30 Seconds Version

👉 Nvidia develops B30A chip for China delivering 50% of flagship B300 performance while paying U.S. government 15% of revenue from sales.

📊 Single-die Blackwell design targets September sample delivery, maintaining high-bandwidth memory and NVLink features of current H20 chips.

💰 Revenue-sharing arrangement transforms export controls from binary restrictions into monetized access model for advanced semiconductor technology.

🏭 Chinese firms continue buying Nvidia despite Beijing warnings, with CUDA ecosystem creating switching costs beyond hardware performance gaps.

🌍 Washington discovers middle path between complete bans and unrestricted access, potentially extending model to other strategic technology sectors.

🚀 Software dependencies now anchor market power as much as hardware specs, bending policy outcomes toward managed diffusion strategies.

Washington is monetizing controls as Beijing manages chip sovereignty.

Nvidia is building a China-specific AI accelerator that’s faster than today’s H20 but below its flagship, a move documented in Reuters’ report on Nvidia’s B30A chip plan. The single-die “B30A,” based on the Blackwell architecture, is designed to deliver roughly half the computing power of the dual-die B300 while retaining high-bandwidth memory and NVLink. Samples could reach Chinese customers for testing as soon as next month, pending specifications and U.S. approval.

The policy backdrop is shifting from simple prohibition to conditional access. After an April halt, Washington allowed H20 sales to resume in July, even as it pressed for tighter guardrails. The latest turn is financial: under newly announced license conditions, Nvidia and AMD will share 15% of revenue from certain advanced chip sales in China with the U.S. government. President Donald Trump also floated allowing “30%–50%” compute reductions on next-gen parts for China while blocking full-fat models. The direction is clear. Access now carries a toll.

This is novel industrial policy. Export controls are becoming dials, not switches.

The ecosystem variable

Hardware is only half the story. Nvidia argues that keeping Chinese developers on its stack matters as much as chip ceilings. CUDA, NVLink networking, and mature software tooling create switching costs that outlast a single product cycle. That’s why a throttled Blackwell can still be valuable to Chinese buyers who have accumulated CUDA-native models, libraries, and ops know-how. Momentum compounds. So do dependencies.

Beijing knows this. Which is why guidance has been calibrated, not absolute.

Beijing’s calibrated pressure

In recent weeks, Chinese regulators questioned major platforms about their H20 purchases and raised “information risk” concerns tied to the export-license paperwork. Companies were asked why they weren’t using domestic options and were cautioned about government-related deployments. Crucially, firms were not ordered to stop buying H20s. That keeps political pressure on, while giving labs latitude to hit training deadlines.

It’s a familiar playbook: nudge toward localization, avoid derailing near-term AI builds.

Engineering to the rulebook

The B30A’s single-die design is an engineering compromise aimed squarely at compliance thresholds and manufacturability. One die means fewer yield headaches and a cleaner story for export screening, even if raw throughput lands at ~50% of B300. High-bandwidth memory and NVLink stay, preserving multi-GPU scaling and a software-compatible path for teams already on Hopper-class systems.

Nvidia is also prepping the RTX6000D—another China-tuned Blackwell part aimed at inference. It uses conventional GDDR, with bandwidth capped at 1,398 GB/s to sit just under the 1.4 TB/s limit introduced in the spring. Initial deliveries are slated for September. Built-to-spec is now a product category.

What changes now

First, Washington has discovered a middle lane. Instead of a binary “yes/no,” it is trialing “yes, but pay—and accept reduced capability.” That keeps visibility on shipments, keeps leverage over feature sets, and taxes the very activity controls were meant to slow. Whether that improves national security or simply normalizes a rent is the open question.

Second, Nvidia is showing how to defend share inside a ring-fenced market. It’s offering a ladder: H20 today, RTX6000D for cheap inference, B30A for larger training jobs. All on CUDA. That de-risks a hard pivot to domestic platforms, which is precisely what U.S. policymakers say they fear. Paradoxically, the accommodation could entrench the incumbent.

Limits and unknowns

Three caveats deserve emphasis. Regulatory approval for B30A is not assured; a single interagency objection can freeze samples at the dock. Beijing’s “advice” can harden into mandates with a memo, especially for government-adjacent workloads. And economics bite: revenue-sharing, capped specs, and packaging constraints (from HBM to CoWoS slots) may squeeze margins enough to test Nvidia’s appetite for bespoke China SKUs at volume.

Yet both capitals seem to prefer friction to a clean break. That keeps the door ajar.

Why this matters:

Export controls are turning into adjustable, revenue-generating levers, reshaping how tech power is exercised across borders.
Software ecosystems now anchor market power; CUDA lock-in can outweigh hardware gaps, bending policy outcomes toward managed diffusion.

❓ Frequently Asked Questions

Q: Where does the 15% revenue sharing money actually go?

A: The funds go directly to the U.S. Treasury, with no specific earmarking for particular programs. This represents the first time export controls have been explicitly monetized, creating a precedent for other strategic technology sectors.

Q: How does B30A performance compare to existing chips?

A: B30A delivers roughly 50% of B300's computing power but outperforms the current H20. For context, Nvidia sold 1 million H20 units in China during 2024—five times Huawei's comparable chip sales.

Q: Why don't Chinese companies just switch to Huawei's chips?

A: Switching from Nvidia's CUDA platform to Huawei's CANN requires retraining development teams, rewriting existing AI models, and accepting performance trade-offs. Industry sources report Huawei chips have higher defect rates and efficiency challenges.

Q: What's the timeline for B30A approval and delivery?

A: Nvidia hopes to deliver samples for testing as early as next month, though chip specifications aren't finalized. U.S. regulatory approval remains uncertain, with any single interagency objection potentially blocking shipments.

Q: What does CUDA "ecosystem lock-in" mean practically?

A: CUDA is Nvidia's programming interface that connects chips to software. Chinese AI labs have built models, libraries, and operations expertise around CUDA, making hardware switches costly even when alternatives exist.

Q: How is this different from previous export control approaches?

A: Traditional controls were binary—either allowed or banned. This creates a middle path: controlled access with performance limits and revenue sharing. It's managed diffusion rather than wholesale restriction.

Q: Why use single-die instead of dual-die design for B30A?

A: Single-die designs have fewer manufacturing yield problems and simpler export compliance stories. It prioritizes production reliability over maximum performance—fitting for export-restricted markets with capability ceilings.

Q: What happens if Chinese demand for Nvidia chips drops significantly?

A: China generated 13% of Nvidia's revenue in fiscal 2024. Revenue-sharing arrangements and specialized manufacturing for China-specific chips create margin pressures that could test Nvidia's appetite for bespoke products at scale.