Chip Race Gets Messy
Hyperscalers and labs are widening their hardware bets: Amazon is reportedly weighing selling its in‑house AI chips to other companies, while Anthropic is exploring designing its own silicon — all amid rising reliance on Taiwan fabs and packaging chokepoints. At the same time, reports surfaced of banned Nvidia server shipments disclosed by a Chinese AI firm and tighter U.S. export controls, highlighting how supply‑chain and policy fragility now shape system design choices ( ).
The artificial intelligence boom runs on a very physical stack: a chip gets designed on software, made in a factory, wrapped in advanced packaging, dropped into a server, and then cleared through export rules before anyone trains a model. This week, trouble showed up at every layer at once. (bloomberg.com, cnbc.com, benzinga.com, bloomberg.com) Amazon said on April 9 that it is considering selling its in-house chips to other companies, not just using them inside Amazon Web Services. Chief executive Andy Jassy said Amazon’s chip business is already running at more than $20 billion a year. (bloomberg.com, reuters.com) Those chips include Trainium for training large models and Inferentia for serving them after training, which is the part users feel when they type into a chatbot and wait for an answer. Amazon says its newer Trainium2 systems offer 30% to 40% better price performance than current graphics processing unit instances on its cloud. (aws.amazon.com, aboutamazon.com) On the same week, Reuters reported that Anthropic is exploring designing its own chips instead of relying only on outside suppliers. The twist is that Anthropic already has a deep partnership with Amazon and agreed in 2023 to use Amazon Web Services as its primary cloud provider for critical workloads. (cnbc.com, anthropic.com, aboutamazon.com) That sounds contradictory until you look at the bottleneck. The scarce part is no longer just the chip itself; it is also advanced packaging, the step that bolts multiple pieces of silicon and high-bandwidth memory together like wiring an engine into a car before it can leave the factory. (benzinga.com, cnbc.com) Taiwan Semiconductor Manufacturing Company is building more in Arizona, but even chips made in the United States have still been getting sent back to Taiwan for final packaging, according to recent reporting. Taiwan Semiconductor says its Arizona plans now include two advanced packaging facilities, which shows how central that missing step has become. (benzinga.com, tsmc.com) Nvidia sits in the middle of this squeeze because its systems are still the default choice for many model builders, and CNBC reported on April 8 that Nvidia has reserved the majority of Taiwan Semiconductor’s most advanced packaging capacity. When one buyer locks up the assembly line, everyone else starts looking for a different engine. (cnbc.com) Then policy hit the same supply chain. Bloomberg reported on April 10 that a Shenzhen company disclosed about $92 million of banned Nvidia server purchases to Beijing, tied to Super Micro systems containing high-end Nvidia chips that have required U.S. permission for China since 2022. (bloomberg.com, bis.gov) Those U.S. controls were first rolled out on October 7, 2022 and expanded again on October 17, 2023 by the Bureau of Industry and Security, which widened restrictions on advanced computing chips and semiconductor tools going to China. A chip company now has to plan not just for speed and cost, but for whether a finished server can legally cross a border. (bis.gov, cset.georgetown.edu) That is why Amazon wants to turn its internal chips into a business, and why Anthropic is at least considering becoming a chip designer itself. In 2026, the artificial intelligence race is no longer just about better models; it is about who can secure silicon, packaging, servers, and export clearance at the same time. (bloomberg.com, cnbc.com, benzinga.com, bis.gov)
