This just scratches the surface of what’s discussed in the article. Check out the full article here for all the fun details: scout.eto.tech?id=3978.

A few thoughts:
- Emphasizes the importance of thinking about training vs. inference when designing export control policies
- Underscores the point that while stockpiles of chips are available, export controls will have a delayed impact (as I and many others have argued)

The key for China moving forward (according to the article)? Making better use of existing compute. Strategies include improving software efficiency, offering cloud services that boost chip utilization, and optimizing model architectures.

Plus, the article notes potential future export controls on advanced AI chips could contribute to China’s long-term shortage in diffusing AI broadly.

Long term: The demand for inference compute is expected to increase as deployment of AI generally becomes more important, potentially leading to a "long-term shortage."

Many small cities and towns are launching intelligent computing center projects (check out the map). As of October 2024, local governments invested at least $3.8 billion in compute infrastructure. www.datawrapper.de/_/qX6Zx/

Short term: existing compute meets basic business needs (and there are surpluses in some regions). Large tech companies (e.g., Alibaba, Tencent) & telecom operators (e.g., China Mobile) are investing heavily in compute but smaller companies are selling idle (and stockpiled) compute.

Does China have a compute surplus or shortage? Another interesting article flagged by @emergingtechobs.bsky.social describes the complex picture as “short-term sufficiency, long-term shortage” and includes a list of local government compute projects.

Hearing Tomorrow🏛️

CSET's @hannadohmen.bsky.social will testify at the U.S.-China Commission's hearing, "Made in China 2025—Who Is Winning?"

Tune in at 11:20 am (ET) for Panel II: The Next Decade of U.S.-China Tech Competition. www.uscc.gov/hearings/mad...

A few thoughts from me and the @csetgeorgetown.bsky.social hivemind on developments over the last week (Stargate, DeepSeek, etc.). These events aren't necessarily surprising from an analytical lens, but they certainly seem to have caught markets and the media off guard.

This is partially by design: US export controls aim to slow down PRC leading-edge compute access, but incentivizing lower compute innovations in China was always a risk. See @hannadohmen.bsky.social and @jacob-feldgoise.bsky.social in @foreignaffairs.com www.foreignaffairs.com/china/limits...

These days it's hard to keep up with AI & emerging tech policy. My colleagues and I @csetgeorgetown.bsky.social are committed to providing data-driven analysis for policymakers and the public as we all confront many difficult tech policy questions.

Follow our team here:
bsky.app/starter-pack...

@hannadohmen.bsky.social and I wrote about the Ascend 910B’s architectural changes, which adds additional context to the die shots in this report

cset.georgetown.edu/publication/...

It appears that the Biden admin is set to announce major new restrictions on China's access to cutting-edge AI chips on Monday. The move aims to slow China's ability to build powerful AI models. 1/

www.wired.com/story/memory...

7 of the 25 announced investment awards have been finalized. That's 53% (~$19 bn) of the total announced investments thus far.

How is M&A changing the AI landscape?

Our new report explores how acquisitions of AI companies have grown over the last decade and how non-AI companies are using acquisitions to adopt the technology.

Explore the report 👉 bit.ly/4g7p6ro

The CHIPS office finalized its award to Intel this morning! www.nist.gov/news-events/...

go.bsky.app/QhUsZTG My global tech and economic security starter pack is only a week old, and I'm sure it's already badly out of date. If there are people who ought to be there, let me know!