AI chip race shifts
Commentators say the AI competition is moving from pure hardware leadership toward who can compute cheaply at scale, with the U.S. still ahead on advanced chips but China possibly winning on cost and token production. A South China Morning Post opinion framed the gap as ‘US controls chips, China controls the scoreboard,’ and sources cite Nvidia concerns that models like DeepSeek V4 running on Huawei Ascend chips could hurt U.S. interests. Separately, reporting notes Amazon’s in‑house AI‑chip business is growing quickly, adding another competitor to the market. (scmp.com) (huaweicentral.com) (fool.com)
The artificial intelligence chip race is tilting away from who has the fastest hardware and toward who can generate tokens cheapest at massive scale. (scmp.com) A token is a small unit of text an artificial intelligence model reads or writes, and serving billions of them is what turns a model into a business. South China Morning Post columnist Jeffrey Wu wrote on April 15 that “whoever produces tokens cheaply, at scale” holds the advantage in the artificial intelligence economy. (scmp.com) The United States still leads on advanced artificial intelligence chips, especially through Nvidia, whose graphics processing units remain the standard for training large models. Poynter reported in February that the U.S. lead rests on both chip quality and chip supply, even as Chinese rivals narrow the gap. (poynter.org) The China side of the argument is now less about beating Nvidia at the very top end and more about building a cheaper domestic stack. Reuters reported on April 3, via The Information, that DeepSeek’s upcoming V4 model will run on Huawei chips instead of American ones. (msn.com) That matters because export controls hit China’s access to the best American accelerators, so Chinese companies have spent the last two years adapting models and software to homegrown chips. TrendForce said on April 7 that Huawei’s Ascend line is benefiting as curbs limit Nvidia’s H200 in China and push customers toward domestic alternatives. (trendforce.com) The DeepSeek story is still partly a report about plans, not a finished launch. Huawei Central said last week that V4 would likely arrive in the coming weeks, while Reuters attributed the claim to The Information and neither DeepSeek nor Huawei had publicly confirmed the shift. (huaweicentral.com) (msn.com) A second pressure point for Nvidia is coming from inside the United States, where cloud companies are designing their own chips to cut costs. Amazon Web Services says its Trainium2 chip delivers up to four times the performance of first-generation Trainium and offers 30 percent to 40 percent better price performance than certain graphics processing unit-based EC2 instances. (aws.amazon.com) Bloomberg reported on April 9 that Andy Jassy said Amazon’s in-house chip effort would have a $50 billion annual run rate if it were a standalone business selling semiconductors to Amazon Web Services customers and outside buyers. Motley Fool repeated that figure on April 15 in arguing Amazon has become a more serious rival in artificial intelligence infrastructure. (bloomberg.com) (fool.com) That leaves the market splitting into at least three contests at once: Nvidia at the high end, Chinese domestic systems built around Huawei, and cloud giants like Amazon trying to make artificial intelligence compute cheaper inside their own platforms. The scoreboard Wu described is no longer just about the best chip; it is about who can keep the meters running at the lowest cost. (scmp.com) (aws.amazon.com)
