The AI Policy Podcast

Host: Center for Strategic and International Studies (CSIS)
Episode: What Selling Nvidia's Blackwell Chips to China Would Mean for the AI Race
Date: November 5, 2025
Guests:

• Gregory C. Allen (Host), CSIS
• Georgia Adamson, Technology Fellow, Institute for Progress (IFP)
• Saif Khan, Distinguished Technology Fellow, IFP; Former National Security Council and Department of Commerce

Episode Overview

This episode examines the ongoing debate over whether the United States should permit Nvidia to export its latest Blackwell generation B30A chips to China. Prompted by breaking news that the Trump administration had decided not to discuss sales of the most advanced Blackwell chips to China, the discussion focuses on the economic, technological, and geopolitical ramifications. The guests—co-authors of a recent policy analysis on the same subject—provide data-rich insights into China’s AI capabilities, U.S.-China tech competition, and the strategic stakes of export policy.

Key Discussion Points & Insights

1. Background on Export Controls and Nvidia’s China Strategy

[03:25–08:40]

• The Trump administration previously tightened, then partially reversed, restrictions on exporting Nvidia’s H20 chips to China.
• The B30A (a downgraded Blackwell chip) enters the scene as Nvidia’s pitch to sell a slightly less powerful version to China, amid regulatory uncertainty and intense lobbying.
  • Georgia Adamson explains: The B30A was designed with half the high bandwidth memory (HBM) and processing dies of the flagship B300, making it about 50% less powerful but more affordable.
• The latest trade summit saw President Trump decide not to discuss Blackwell chip exports after being advised by key officials (see [00:10] and WSJ reporting).

Quote:

“Take 30% to 50% off of it. This is after reports that Jensen Huang [Nvidia CEO] had been sort of lobbying the Trump administration to export B30A.” — Georgia Adamson [04:16]

2. B30A Performance: How Good Is It?

[11:52–17:14]

• Comparison Metrics:
  • B300 (“flagship”): 60,000 TPP (Total Processing Performance, a Commerce Dept. compliance metric)
  • B30A: ~30,000 TPP (half the performance)
  • H20 (older chip): ~2,300 TPP
  • Huawei Ascend 910C (China’s top domestic chip): ~15,000 TPP
• Price-Performance:
  • B30A is 12–13x more powerful than an H20 and outpaces Huawei’s chip on a performance-per-dollar basis (~3x better).
  • Despite technical progress, Chinese chips have poor reliability and a weak software ecosystem, with data center utilization below 30% compared to up to 100% for legacy Google TPUs.
• Reliability:
  • Nvidia’s Hopper chips can run ~6 years before failure; Huawei’s are far less reliable, limiting their scalability in clusters.

Quote:

“Even though it is better in TPP … Huawei Ascend 910C is better than the H20. But there are other metrics—memory bandwidth, software reliability—where Nvidia’s chips are still far ahead.” — Greg Allen [17:14]

3. Market Dynamics and China’s Persistent Preference for Nvidia

[23:26–25:06]

• Chinese firms (Tencent, Baidu, Deep Seek) have stockpiled billions in H20 chips, despite impending bans, due to superior Nvidia reliability and software compatibility.
• China’s AI development bottleneck remains access to high-end chips.
• Revealed Preference: Most state-of-the-art Chinese AI models are still trained on U.S. chips, not domestic ones.

4. Debate: Should the US Allow B30A Sales to China?

[25:57–30:29]

• Steelman argument for sales (presented by Saif Khan):
  • Keeping China reliant on U.S. chips benefits the U.S. tech ecosystem and could stymie Huawei’s ambitions.
  • Major Chinese bans on lesser chips (H20) may be bluff or a negotiating gambit to secure approval for more powerful, desirable chips.
• Evidence from smuggling and enforcement ambiguity—Chinese authorities may tacitly enable smuggling of advanced chips while publicly banning lesser ones.
• Host counterpoint: China’s long-term strategy is self-reliance and indigenization; efforts to use access as leverage are undermined by China’s determination to reduce dependency.

5. Major Findings of the Guests’ Policy Paper

Four Key Takeaways from Their Analysis

[31:29–44:17]

1. Allowing B30A Exports Would Break From U.S. Policy Aims

• The B30A is 12.6x faster than the H20 (the previously allowed export), so enabling its export would meaningfully weaken U.S. export controls.
• Direct contradiction of Trump administration’s AI Action Plan, which prioritizes denying advanced compute to U.S. adversaries.

Quote:

“Denying our foreign adversaries access to this resource then is a matter of both geostrategic competition and national security.” — Trump admin AI Action Plan, quoted by Georgia Adamson [32:32]

2. US Companies Would Have Fewer Chips

• With supply bottlenecks, increasing Chinese access would divert chips from U.S. customers, raising concerns around domestic innovation and strategic autonomy.
• Nvidia's commercial incentives for customer diversification mean sales to China will increase prices globally and diminish U.S. leverage over buyers.

Quote:

“If you give B30As to Chinese companies … they will actually decrease the demand that is available to U.S. cloud providers to serve those same customers.” — Saif Khan [37:56]

3. Chinese AI Labs Could Build Comparable Supercomputers

• Even with a nominal 50% performance degradation, Chinese data centers could double up B30As to match U.S. B300s at only a 20% increased cost, erasing much of the intended gap.

Quote:

"They can take two B30As and put them together and get something that is almost as good as a B300 ... only a 20% performance penalty in a performance per dollar basis." — Greg Allen [41:16]

4. The U.S. Compute Advantage Would Nosedive

• Policy modeling:
  • No exports, heavy smuggling: U.S. still maintains at least an 11.7x advantage; with no smuggling, potentially 31x.
  • If B30A exports are allowed and China is free to buy: U.S. advantage collapses—possibly giving China parity or even a slight lead in new compute by 2026.

Quote:

"The difference between selling B30As and not selling B30As could be the difference between an America that has 31 times as much AI compute ... and a different world in which China actually gets more AI compute than us." — Greg Allen [49:32]

6. Chinese Chipmaking Constraints and the Importance of Equipment Controls

[49:32–54:57]

• China’s leading chipmaker, SMIC, struggles with poor yield rates (~20% vs. 60–80% at TSMC/Nvidia), severely capping China’s domestic AI hardware capability.
• Export controls on equipment are working, but the guests recommend tightening further—“The more effective way to limit Huawei’s influence… is not to sell chips to China, but to restrict equipment access.” — Saif Khan [61:24]

Quote (on SMIC yields):

"Their yields are still dreadful ... not making progress that TSMC was making six years ago. That tells you a lot is riding on these export controls." — Greg Allen [53:22]

Notable Quotes & Memorable Moments

• "This is Nvidia's pitch to the U.S. government. They know that this chip is currently illegal under existing export regulations ... So this is Nvidia coming up with a product and saying this is what we think you should allow us to sell." — Greg Allen [10:53]
• "The software ecosystem is quite immature ... once you aggregate this many GPUs into a cluster, the chances you'll have one or two GPUs going offline at a given time and causing instability increases quite a bit." — Saif Khan [21:07]
• "If you think about the total cost of ownership of our chips ... even if our competitors’ chips were free, it would not be cheap enough for them to be competitive with us." — Jensen Huang (quoted by Greg Allen) [44:17]
• "China is utterly committed to indigenization ... do we want to sell them everything they need to build a bridge to their decoupled future?" — Greg Allen [62:03]

Timestamps for Major Segments

• 00:10 — Breaking news: U.S. administration decision on Blackwell chips
• 03:25 — Introduction to the B30A export issue and the guests' policy paper
• 08:40 — Details on B30A technical specs and U.S./China policy context
• 11:52 — Deep-dive into hardware performance comparisons (TPP, HBM, price-perf.)
• 21:07 — Software stack, reliability, and real-world data center utilization
• 25:57 — Steelman case for allowing B30A sales; counterarguments
• 31:29 — Paper's four key findings and policy implications
• 39:28 — How Chinese labs could match U.S. supercomputing power
• 45:14 — Modeling of the U.S. compute advantage vs. China under different scenarios
• 49:32 — Chinese chip production limits, SMIC’s struggles
• 61:24 — Final recommendations and summary

Conclusion

The episode makes a strong case—grounded in data, technical analysis, and policy modeling—that exporting downgraded but still cutting-edge B30A Nvidia chips to China would be a “game changer,” severely eroding the U.S. strategic lead in AI compute. The hosts and guests agree that China's persistent goal of technological independence, together with its capacity to “bridge the gap” via smuggling, stockpiling, and massive government investment, means that U.S. policymakers must be extremely cautious about relaxing existing export controls. Further tightening, especially in manufacturing equipment, is also recommended.

Final Words:

“The numbers … like half of a B30 can be misleading as to just how enormous the stakes are here.” — Greg Allen [62:03]