China's AI models are closing the gap—and will continue to improve. However, this misses America's strategic compute advantage. In my new commentary, I argue that the TOTAL compute advantage is what export controls preserve and—if leveraged correctly—provides the real edge. 1/

Export control thresholds: &

Every year brings a new threshold and a new chip sitting right below it. This isn't a great chip, Huawei's chips have better specs. But price-performance and software ecosystem matter too. Nvidia's massive margins let them cut prices. They did the same with the H20 before.

Nvidia's B40: built to comply with export controls after the H20 ban. Probably: 148 TFLOP/s FP16, <1.4 TB/s memory bandwidth, max 300 GB/s networking. 15× fewer FLOP/s and 5.7× less bandwidth than the leading B200. But at $6.5-8K and with the software, it'll still sell.

RT @sj_manning: The window of opportunity is currently open. The PRC still lacks the compute necessary to widely export and embed their clo….

RT @sj_manning: Read the full piece: "Beyond Export Controls: How Strategic Promotion of American AI Abroad Can Protect the Homeland & Chec….

This is an important and good paper! As someone often seen as an advocate for training compute thresholds, I'm glad to see criticism focused on what matters most: the regulatory target. But some comments:. First, I shared similar reflections a few months ago after leaving.

RT @KonstantinPilz: We took a look at Chinese government efforts to accelerate AI. TL;DR: They are subsidizing at every layer of the stack….

RT @jacob_feldgoise: NVIDIA's latest chip architecture (Blackwell) achieves 10 TB/s die-to-die bandwidth—nearly 20x faster than the 910C's….

Here's the Substack: . Papers mentioned above: "AGI's Five Hard National Security Problems" ( and "Seeking Stability in the Competition for AI Advantage" (.

My colleagues just launched their new Geopolitics of AGI Substack! I recommend following. Expect relevant analysis from @RANDCorporation's experts, such as their recent "AGI's Five Hard National Security Problems" or their thoughtful take on "mutually assured AI malfunction.".

Plot twist: Replace "shipbuilding" with "chipbuilding" and swap China for that small island neighbor. Same story, different industry.

RT @GregHBurnham: The core idea is simple: building a broad and deep understanding, bit by bit, over time, is worthwhile. In particular, i….

me defending my most recent estimates.

TIL about the N-th Country Experiment.

We're planning to review the first batch of applications next week (Monday 6/23)! .Make sure to apply if you're interested in crunching numbers for better AI policy.

RT @tbpn: Morning. Here are our guest call-ins today:. - @BerntBornich (1X Tech).- @Emily_Sundberg (Feed Me).- @JoshuaSteinman (Galvanick)….

FWIW my production estimates are higher than 200k (and this calculation assumes no production ramp-up). The strategically relevant numbers are also different if this becomes a key priority. More analysis forthcoming on capacity and backfilling potential. 3/3.

What they can do: Strategic allocation of small shipments to build ecosystem lock-in for nations with smaller ambitions. That's the real concern—not massive backfilling, but targeted deployment and offering attractive package deals to draw countries into their ecosystem. 2/.

If China produces 200k Ascend 910Cs annually, filling the UAE's planned 5GW would require 15 years of production (3.1M chips) while delivering only ½ the performance. Or match performance with 7M chips consuming 2.2x more energy. China cannot backfill projects at this scale. 1/.

RT @KonstantinPilz: People keep asking me ‘Konstantin, where are all the data centers?’ Today, I can finally give you the answer. Explore….