Compute Chain Tightens

Nvidia is pushing Blackwell beyond giant artificial intelligence training clusters just as it tightens scrutiny over who gets its chips. (nvidia.com) (digitimes.com) On April 22, Nvidia said its RTX PRO 4500 Blackwell Server Edition pairs with vGPU 20 software to run virtual desktops, graphics, video and lightweight AI development in the same data center stack. Nvidia said the setup delivers nearly 1.9 times the graphics acceleration of prior architectures in virtualized environments. (developer.nvidia.com) The card is a 165-watt, single-slot server GPU with 32 gigabytes of GDDR7 memory, aimed at denser enterprise deployments than the larger Blackwell parts sold for hyperscale model training. Nvidia’s documentation says the 4500 supports Multi-Instance GPU, which slices one chip into isolated partitions for separate virtual machines. (nvidia.com) (docs.nvidia.com) A virtual GPU works like carving one physical chip into rented apartments for different users, instead of assigning the whole building to one tenant. Nvidia said the RTX PRO 4500 Blackwell Server Edition supports MIG-backed vGPU, letting enterprises split one accelerator across multiple workloads with hardware isolation. (nvidia.com) (docs.nvidia.com) That product push landed a month after federal prosecutors unsealed charges against Supermicro co-founder Yih-Shyan “Wally” Liaw and two others on March 19, alleging a conspiracy to divert U.S.-made high-performance servers with advanced Nvidia technology to China. The Justice Department said the scheme involved export-control violations, smuggling and fraud. (justice.gov) DigiTimes reported on April 23 that Nvidia has since upgraded global supply-chain monitoring, with closer audits of shipments and customer lists after the indictment. PCMag reported in March that Nvidia and Supermicro had already flagged suspicious orders and canceled them. (digitimes.com) (pcmag.com) Supermicro said on March 19 that it is not named as a defendant in the indictment, called the alleged conduct a violation of company policy, and said it was cooperating with investigators. The company also said it put two employees on administrative leave and ended its relationship with a contractor tied to the case. (ir.supermicro.com) (sec.gov) At the same time, Google is building a parallel route for AI compute with its own Tensor Processing Units, or TPUs, which are custom chips tuned for AI math the way a calculator is tuned for arithmetic. Google said this week that Ironwood is live at 4.6 petaFLOPS per chip, and that its eighth-generation TPU family will split into TPU 8t for training and TPU 8i for inference. (blog.google) (cloud.google.com) (thenextweb.com) Google’s design shift means the biggest cloud companies are no longer waiting for one merchant chip road map to cover every workload. Nvidia is adding lower-power Blackwell options for enterprise fleets, while Google is separating training and inference into different in-house silicon lines. (developer.nvidia.com) (blog.google) (cloud.google.com) The result is a tighter compute chain: more ways to deploy AI chips inside corporate data centers, and more checks on how those chips move across borders. Nvidia is trying to widen access to Blackwell inside approved channels while narrowing the paths around them. (digitimes.com) (nvidia.com)