Blackwell is the de facto stack

Anthropic did not just buy more chips this week. On April 10, 2026, it signed a multiyear cloud deal with CoreWeave so new computing capacity for Claude starts coming online later in 2026. (coreweave.com) That sounds like a normal cloud contract until you look at who CoreWeave is. CoreWeave built its business around renting out giant clusters of NVIDIA graphics processors for artificial intelligence work, and NVIDIA made Blackwell generally available in the cloud through CoreWeave first. (coreweave.com) (blogs.nvidia.com) A modern artificial intelligence buyer is not really choosing one chip anymore. The buyer is choosing a whole stack: the model company, the cloud operator, the networking setup, the software tools, and the benchmark recipe that says the system performs the same way every time. (nvidia.com) (developer.nvidia.com) NVIDIA has been turning that stack into a product of its own with something called Exemplar Cloud. The program gives cloud providers NVIDIA benchmark recipes and checks whether they hit NVIDIA’s reference performance, security, and reliability targets on real workloads such as inference, fine-tuning, and pretraining. (developer.nvidia.com) (nvidia.com) That is why a smaller provider like Vultr mattered this week. On April 7, 2026, Vultr said it was among the first to become an NVIDIA Exemplar Cloud after meeting NVIDIA Blackwell reference-design performance targets. (newswire.telecomramblings.com) Vultr said that validation came from tests on a 512-node NVIDIA HGX B200 cluster. In plain English, NVIDIA is no longer only selling the engine; it is helping certify which rental-car fleets can prove they drive the engine at full speed. (markets.financialcontent.com) (developer.nvidia.com) The bigger clouds are lining up the same way. NVIDIA says Oracle Cloud Infrastructure has Exemplar Cloud validation on Blackwell too, while Microsoft, CoreWeave, and Oracle Cloud Infrastructure are deploying Blackwell Ultra systems at scale for reasoning models and artificial intelligence agents. (nvidia.com) (blogs.nvidia.com) Anthropic is also spreading its bets across multiple infrastructure partners instead of picking one landlord. Anthropic said on April 6, 2026, that it expanded its partnership with Google and Broadcom for multiple gigawatts of next-generation compute, and four days later CoreWeave joined that infrastructure roster for Claude. (anthropic.com) (coreweave.com) That combination changes how the market gets sold. If a chief information officer can buy Claude on a cloud that already has Blackwell hardware, NVIDIA software, and NVIDIA-approved benchmark results, the purchasing decision starts to look less like picking parts and more like buying a finished factory line. (nvidia.com) (developer.nvidia.com) CoreWeave’s stock jumped 10.8% on April 10 after the Anthropic deal was announced, which is a clue to what investors think is being bought here. They are not pricing a one-off rental of graphics processors; they are pricing a position inside the default route from model maker to cloud to customer. (siliconangle.com) If Blackwell keeps becoming the standard test, the standard rack, and the standard cloud badge, competitors will have to beat an ecosystem instead of a chip. That is a much harder job, because ecosystems lock in buyers long before the next benchmark chart shows up. (nvidia.com) (developer.nvidia.com)