AI demand outstrips capacity
Coverage shows AI scarcity is expanding beyond GPUs to cloud allocation and alternative silicon, with Alibaba opening an AI data centre using 10,000 domestic chips to reduce Nvidia dependence. At the same time some AWS customers are reportedly trying to buy out large swathes of capacity as providers push alternatives like Trainium. (archyde.com) (networkworld.com)
Artificial intelligence demand is now running ahead of the hardware needed to serve it, from Amazon Web Services cloud slots to Alibaba’s newest chip clusters in China. (aboutamazon.com) On April 9, Amazon chief executive Andy Jassy said two large Amazon Web Services customers had asked to buy all of the company’s 2026 Graviton capacity. He said Amazon refused because it still has to serve other customers. (aboutamazon.com) Jassy said Trainium2 capacity is effectively sold out, Trainium3 started shipping at the start of 2026 and is “nearly fully-subscribed,” and a “significant chunk” of Trainium4 capacity has already been reserved about 18 months before broad availability. Amazon said its chips business, including Graviton, Trainium and Nitro, is now running at more than $20 billion in annual revenue. (aboutamazon.com) A data center is a warehouse full of computers, and the scarce part in this cycle is the processor that trains and runs artificial intelligence models. For the past two years, Nvidia’s graphics processing units have been the default choice, but cloud providers are now steering customers to in-house chips when Nvidia supply is tight or too expensive. (networkworld.com) Alibaba moved the same week in a different direction: on April 8, Alibaba and China Telecom launched a data center in southern China built around 10,000 Alibaba-designed Zhenwu processors. CNBC reported the site is meant for both training and inferencing, the two main jobs of artificial intelligence systems. (cnbc.com) Alibaba’s project is also a response to export controls. CNBC reported Chinese companies have been accelerating domestic chip efforts as the United States has tried to restrict China’s access to advanced Nvidia processors and other key technologies. (cnbc.com) Amazon is making a different case to customers. Jassy said Trainium3 delivers 30% to 40% better price-performance than Trainium2, and he argued that owning more of the chip stack could save Amazon “tens of billions” of dollars in capital spending each year for inference, the step where a model answers a prompt. (aboutamazon.com) That leaves the market short on more than one thing at once. Companies are competing for Nvidia systems, for access to cloud regions that still have open capacity, and for alternatives from Amazon and Alibaba that can reduce dependence on a single supplier. (cnbc.com)
