TL;DR: Multiple reports say Nvidia is developing a China-compliant Blackwell-based accelerator dubbed B30A. It’s positioned as the H20’s successor, with far higher performance (reports suggest up to ~6× vs H20) and a price roughly double H20—i.e., around $20k–$25k per GPU, pending export approval. Sampling is rumored for September 2025, with commercial availability late 2025 to early 2026 if regulators give the green light. Treat all details as provisional until Nvidia makes it official. (Reuters, Tom's Hardware)
Is Nvidia really developing a “B30A” for China?
Yes—according to Reuters and industry press, Nvidia is preparing a new, China-specific AI chip, B30A, built on its latest Blackwell architecture. It is intended to outperform the H20 (a Hopper-era, export-compliant part) while still fitting within U.S. export rules. Nvidia hasn’t formally launched the product; reports cite unnamed industry sources and Nvidia commentary that any go-ahead depends on U.S. government approvals.
How much will it cost? (And is it really pricier than H20?)
Pricing chatter centers on $20,000–$25,000 per GPU—roughly double the H20’s $10,000–$12,000 street range in China. Despite the hike, major Chinese buyers reportedly see it as a good deal because of the expected performance leap. Some outlets quote buyers prepared to pay ~$24,000. Keep in mind: these are market whispers, not official list prices.
When will B30A be available?
Timelines are rumored, not confirmed:
● Sampling: as early as September 2025, contingent on U.S. export licensing.
● Volume availability: late 2025 or early 2026 is the prevailing industry expectation if approvals proceed smoothly.
The schedule is tightly coupled to evolving export policy. Nvidia has only recently been able to resume H20 sales under new licensing terms, illustrating how fluid the regulatory landscape is.
What will the B30A look like (technically)?
While Nvidia hasn’t published specs, consistent leaks and reporting sketch a plausible configuration:
● Architecture: Blackwell, a generation beyond Hopper (H20’s lineage).
● Compute scope vs. B300: Roughly half the compute of flagship Blackwell B300, by design, to meet export thresholds.
● Package & memory: A single Blackwell compute chiplet paired with four HBM3E stacks, totaling ~144 GB, likely on TSMC CoWoS-S packaging (more economical than the B300’s top-end packaging).
● Throughput (rumored): up to ~7.5 PFLOPS (FP4); take this as indicative rather than final.
● Interconnect: Expected NVLink support for multi-GPU scaling, though cluster size/topology may be constrained by compliance rules.
Bottom line: B30A aims to substantially outperform H20—some sources say up to 6×—while staying below U.S. performance ceilings.
What is it for?
Primary role: AI training and large-scale inference in data centers. Like H20, B30A would target Chinese hyperscalers and enterprise buyers building or expanding AI clusters (LLM training, fine-tuning, retrieval-augmented generation, multi-modal workloads). Reports specifically position B30A as the training-class part, with an additional RTX 6000-series “D/Pro”-style product expected for inference and professional visualization under China-compliant specs.
Why would buyers pay more than H20?
Throughput per dollar: Even at 2× the price, a ≥4–6× performance uplift improves time-to-train, cluster utilization, and TCO (fewer nodes to hit a training target).
Ecosystem lock-in: Nvidia’s CUDA + NVLink + NCCL + software stack maturity reduces integration risk and engineering cost compared with switching platforms. (Inference from market behavior; buyers’ continued pursuit of Nvidia parts supports this.)
Export-compliant headroom: If B30A pushes right up to allowed limits, buyers get maximum legal performance in China while preserving compatibility with the broader Nvidia ecosystem.
How does policy shape B30A?
Recent months have seen stop-start export policy shifts. H20 production and sales were curtailed, then resumed under licensing, and reports now discuss revenue-share conditions on certain China sales and backlogs in approvals. Against that backdrop, Nvidia is said to be preparing B30A subject to U.S. approval—hence the cautious language from the company. Expect timelines and specs to track evolving rules.
Should you plan around B30A now?
If you operate in China and are eyeing major AI capacity in 2025–2026, it’s reasonable to:
● Scenario-plan two tracks: (a) H20/RTX Pro–based expansions if approvals lag, and (b) B30A adoption if sampling/shipments proceed.
● Design clusters that can scale horizontally (rack-level power/cooling ready for higher-TDP parts) and network-wise (NVLink domain sizing vs. Ethernet/IB spine), so you can slot in B30A later without re-architecting fabric. (General best practice; consistent with rumored NVLink support.)
● Budget at $20k–$25k per GPU for planning purposes, with the understanding that allocation, bundles, and channels may move real-world pricing.
Key takeaways
● In development? Multiple reports say yes—Nvidia is preparing B30A for China, a Blackwell-based successor to H20, pending U.S. approval.
● B30A Price vs H20: About 2×, i.e., $20k–$25k vs $10k–$12k for H20, justified by a large performance jump.
● ETA: Sampling rumored September 2025; broader availability late 2025 / early 2026 if permits arrive.
● What it’s for: AI training/inference in data centers—China-compliant, high-throughput compute with NVLink and HBM3E.
● Caveat: All details subject to change until Nvidia announces official specs, pricing, and ship dates.
Related reading:
NVIDIA H100 AI GPUs: Revolutionizing Machine Learning and Data Processing
Inside NVIDIA DGX B200: A Deep Dive into NVIDIA’s Next-Gen AI Supercomputing Beast
NVIDIA GB200: Inside the Blackwell Superchip Powering AI at Scale
The NVIDIA A100 GPU: The AI Workhorse That Changed Everything
