The U.S.’s efforts to stymie Chinese chip production are not watertight

Tencent president Martin Lau reportedly said in an analyst call today that the tech giant would be forced to look for domestic alternatives to the likes of Nvidia’s H100 and H800 processors to train its AI models. “We will have to figure out ways to make the usage of our AI chips more efficient,” he said, according to Reuters. “And we will also try to look for domestic source[s] for these training chips.”

Tencent’s Hunyuan AI model will be able to go through “at least a couple more generations” of development using the company’s stockpiled Nvidia chips, Lau added, but the constraints will limit its ability to offer those processors’ capabilities to others through its cloud services.

Local alternatives do exist, with Baidu having just ordered 1,600 Huawei Ascend 910B chips. And a high-level U.S. Commission said yesterday that Baidu’s growing success in providing such alternatives suggests American efforts to stymie Chinese chip production aren’t watertight.

The Biden Administration last year banned the export to China of chipmaking equipment that could enable production processes at anything at or below the now-antiquated 14-nanometer scale. (In case you ever wonder what these numbers refer to, it used to mean the size of individual transistors, but these days it’s really just a marketing term in which smaller numbers denote greater component density and better performance. TSMC’s 3nm manufacturing process, which is responsible for Apple’s newest chips, is the latest and greatest.)

However, Baidu contractor SMIC is churning out chips that use 7nm processes, for AI training and for mobile phones such as the new Huawei Mate 60 Pro.

“Importers are often able to purchase the equipment if they claim it is being used on an older production line, and with limited capacity for end-use inspections it is difficult to verify the equipment is not being used to produce more advanced chips,” the U.S.-China Economic and Security Review Commission griped in its annual report to Congress.

There’s a “proliferation of new firms acting on behalf of” the Chinese military, and finding the data to track them all may not be possible, the commission added. The U.S. Doesn’t have “sufficient control over the supply chain to introduce effective controls unilaterally,” and “multilateral coordination is difficult.”

The report did not provide any recommendations for how to fix this situation. So as things stand, everyone will just have to accept that the export controls’ effectiveness is only temporary and—while really meaningful—not absolute.

More news below.

David Meyer

Want to send thoughts or suggestions to Data Sheet? Drop a line here.

NEWSWORTHY

ON OUR FEED

“We’ll require creators to disclose when they've created altered or synthetic content that is realistic, including using AI tools…For example, this could be an AI-generated video that realistically depicts an event that never happened, or content showing someone saying or doing something they didn't actually do.”

—YouTube cracks down on deepfakes. As the Guardian reports, YouTube is also giving record labels and distributors a tool to let them flag for removal songs that feature AI-generated simulacra of their artists’ voices.

IN CASE YOU MISSED IT

Former Googler Adrian Aoun raises $100 million for walk-in AI healthcare pods, by Alexei Oreskovic

Google’s lawyer cringes in court when expert reveals confidential revenue figure paid to Apple over search advertising, by Bloomberg

OpenAI’s next big bet: Custom GPTs for everyone, by Sage Lazzaro

AI will live up to the hype and more unicorns will lose their horns: 10 health care predictions for 2024 from top investors, by Bob Cocher and Bryan Roberts (Commentary)

Tesla admits Cybertruck won’t pull it out of its current slowdown as hopes rest on new low-cost car: ‘We’re between two major growth waves’, by Christiaan Hetzner

Elon Musk and Satya Nadella are among the CEOs lining up for face time with China’s Xi, APEC’s biggest business kingmaker, by Bloomberg

BEFORE YOU GO

AI weather forecasting. Researchers at Google DeepMind said in a Science paper that their GraphCast AI model is the first that can seriously outperform traditional methods of forecasting the weather. The model was trained on historical weather data, and the researchers said it “significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets.” What’s more, GraphCast seems to be much better at predicting severe weather events such as extreme temperatures, and tracking tropical cyclones.

Matthew Chantry, machine-learning chief at the outperformed European Centre for Medium-range Weather Forecasts, told the Financial Times that AI weather prediction systems had evolved “far sooner and more impressively than we expected even two years ago.”

However—and this is a big however—the DeepMind team acknowledges that GraphCast is a proof of concept, rather than something that can imminently revolutionize weather prediction in its own right. "Our approach should not be regarded as a replacement for traditional weather forecasting methods, which have been developed for decades, rigorously tested in many real-world contexts, and offer many features we have not yet explored,” they wrote.