GPU supply chain: big deals and big risks
NVIDIA signed a multi‑year supply agreement to furnish AWS with next‑gen GPUs and accelerators through 2027 — reportedly on the scale of ~1 million GPUs — while a DOJ indictment alleges Super Micro executives smuggled $2.5B of Nvidia servers to China, rattling markets and compliance teams. Together these stories highlight both locked-in cloud supply and acute geopolitical fragility in AI hardware sourcing, while HBM3E/HBM4 memory demand is driving a new 'AI memory supercycle.' (tradingpedia.com) (finance.yahoo.com) (markets.financialcontent.com)
Ian Buck, Nvidia’s VP for hyperscale and HPC, told Reuters that shipments tied to the AWS commitment will begin this year and run through the end of 2027. (finance.yahoo.com)) AWS’ own blog says Amazon will deploy “more than 1 million” Nvidia GPUs across regions starting in 2026, and listed specific integrations including Blackwell and Rubin architectures, new RTX PRO Blackwell server editions, and EC2 G7e instance support. (aws.amazon.com)) The Justice Department unsealed an indictment on March 19, 2026, naming Yih‑Shyan “Wally” Liaw, Ruei‑Tsang “Steven” Chang, and Ting‑Wei “Willy” Sun and alleging the use of false documents, staged dummy servers and complex transshipment schemes to divert U.S.‑assembled servers to China. (justice.gov)) Prosecutors say the operation involved roughly $2.5 billion in server purchases between 2024 and 2025, and the indictment alleges about $510 million in U.S.‑assembled systems were diverted to China between late April and mid‑May 2025. (convergedigest.com)) Micron’s fiscal Q2‑2026 results showed revenue of $23.86 billion, and company commentary released in mid‑March said advanced HBM3E and next‑generation HBM4 capacity is effectively booked through calendar 2026 as the firm ramps production and raises capital plans. (investingnews.com)) Nvidia’s acquisition/licensing of Groq’s inference technology (reported at about $20 billion) and Ian Buck’s remark that AWS will combine seven different Nvidia chip types for inference underscore why hyperscalers are locking multi‑architecture stacks while U.S. enforcement authorities — who said they worked with the Commerce Department’s Bureau of Industry and Security on the probe — are tightening export controls. (bloomberg.com))
