1/ Trump is meeting Xi this week for China trade talks.
Congress is worried Trump may offer downgraded Blackwell AI chips as a concession.
If this happens, it could effectively mean the end of US chip restrictions.
Thread with highlights from our new 7,000-word report.
Congress is worried Trump may offer downgraded Blackwell AI chips as a concession.
If this happens, it could effectively mean the end of US chip restrictions.
Thread with highlights from our new 7,000-word report.
2/ First - the reported chip specs:
The “B30A” is rumored to be half of NVIDIA’s flagship B300: half the processing performance, half the memory bandwidth, and half the price.
This means the B30A's performance per $ is similar to the B300.
The “B30A” is rumored to be half of NVIDIA’s flagship B300: half the processing performance, half the memory bandwidth, and half the price.
This means the B30A's performance per $ is similar to the B300.
3/ The B30A would:
- Be far better than any Chinese AI chip
- Have >12x the processing performance of the H20, a chip requiring an export license that has been approved for export in only limited quantities.
- Exceed current export control thresholds by >18x
- Be far better than any Chinese AI chip
- Have >12x the processing performance of the H20, a chip requiring an export license that has been approved for export in only limited quantities.
- Exceed current export control thresholds by >18x
4/ At a system level, a B30A-cluster would cost only ~20% more than a B300-cluster, a cost China can subsidize.
Chinese AI labs would have access to supercomputers for AI training as powerful as those available to US AI labs.
Chinese AI labs would have access to supercomputers for AI training as powerful as those available to US AI labs.
5/ Approving a chip with these specs would also be a serious departure from the Trump admin’s export control strategy, which seeks to deny powerful AI compute to strategic rivals.
6/ Another implication: exporting B30As = fewer AI chips for the rest of the world.
2 main scenarios:
- Net demand for chips exceeds TSMC’s manufacturing capacity
- Chinese companies using B30As could also take market share from US cloud companies (in China or elsewhere)
2 main scenarios:
- Net demand for chips exceeds TSMC’s manufacturing capacity
- Chinese companies using B30As could also take market share from US cloud companies (in China or elsewhere)
7/ And most importantly, the total US AI compute advantage over China would shrink dramatically.
Going from a 10-30x advantage with no exports, to China in the lead if the admin allows an aggressive export strategy
Going from a 10-30x advantage with no exports, to China in the lead if the admin allows an aggressive export strategy
8/ But won’t US chip restrictions cause Huawei to backfill with its own AI chips?
No, for both supply and demand reasons.
No, for both supply and demand reasons.
9/ On the supply side, China faces bottlenecks due to US/allied chipmaking tool controls.
AI chips require two components: processor dies and high-bandwidth memory (HBM).
US capacity for processors is 35-38x of China’s (or adjusting for China’s higher mfg errors, 160-170x).
AI chips require two components: processor dies and high-bandwidth memory (HBM).
US capacity for processors is 35-38x of China’s (or adjusting for China’s higher mfg errors, 160-170x).
10/ China fares even worse on HBM, making virtually none this year.
Even next year, the US advantage will be 70x.
Even next year, the US advantage will be 70x.
11/ As a result, five different analysts find Huawei makes an extremely small number of AI chips.
They’ll be at 1-4% of US AI chips this year, and 1-2% in 2026 as the US ramps and Huawei stalls.
They’ll be at 1-4% of US AI chips this year, and 1-2% in 2026 as the US ramps and Huawei stalls.
12/ On the demand side, China will likely create artificial demand for inferior Huawei chips.
So B30A sales to China will have minimal effect on Huawei market expansion.
Instead, sales would supercharge China’s frontier AI & arm Chinese cloud to compete globally with US cloud.
So B30A sales to China will have minimal effect on Huawei market expansion.
Instead, sales would supercharge China’s frontier AI & arm Chinese cloud to compete globally with US cloud.
13/ The more effective policy for US AI chips to beat Chinese chips is to tighten:
1) country-wide controls on chipmaking tools including DUV immersion tools
2) enforcement of HBM controls
This is consistent with recs from the China Select Committee: selectcommitteeontheccp.house.gov/media/reports/…
1) country-wide controls on chipmaking tools including DUV immersion tools
2) enforcement of HBM controls
This is consistent with recs from the China Select Committee: selectcommitteeontheccp.house.gov/media/reports/…
14/ This decision is an inflection point.
Whether we export the B30A or other downgraded Blackwell chips will define the future of US-China AI competition.
Read the full report here from @GeorgiaCAdamson, @taoburr, @fiiiiiist and me: ifp.org/the-b30a-decis…
Whether we export the B30A or other downgraded Blackwell chips will define the future of US-China AI competition.
Read the full report here from @GeorgiaCAdamson, @taoburr, @fiiiiiist and me: ifp.org/the-b30a-decis…
• • •
Missing some Tweet in this thread? You can try to
force a refresh
