AI infra: Boost Run validated
Boost Run announced it achieved NVIDIA Exemplar Cloud validation on NVIDIA's Blackwell architecture, claiming reproducible performance on real AI training workloads. Reports from the same window show Blackwell GPU rental rates spiking and vendors like CoreWeave and Premio securing preferential access or validations — signalling tighter supply and rising infrastructure costs for large-scale AI projects. (prnewswire.com) (alltoc.com) (intellectia.ai) (natlawreview.com)
Boost Run said on April 13 it earned NVIDIA Exemplar Cloud validation on Blackwell, a certification NVIDIA reserves for cloud platforms that hit its performance targets on AI workloads. (prnewswire.com) The company said providers must land within 5% of NVIDIA reference performance on real training jobs, production software stacks, and large multi-GPU setups. Boost Run said that result puts it alongside CoreWeave, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius in the current Exemplar Cloud group. (prnewswire.com) NVIDIA launched Exemplar Clouds in May 2025 to give customers a standardized way to compare AI cloud performance, security, and resiliency instead of relying on vendor marketing or narrow chip tests. NVIDIA says the program uses benchmark recipes, reference hardware and software, and workload-level measurements such as throughput and time-to-train. (nvidianews.nvidia.com) (nvidia.com) That matters because Blackwell is the newest NVIDIA data center architecture, and cloud access has become a gating factor for companies training large language models. NVIDIA said in February 2025 that CoreWeave was the first cloud provider to make Blackwell generally available in the cloud with GB200 NVL72 instances. (blogs.nvidia.com) CoreWeave is now advertising NVIDIA HGX B300 capacity and taking reservations for Blackwell Ultra, with 270 gigabytes of HBM3e memory per GPU on HGX B300 and GB300 NVL72 rack systems built for larger reasoning models. Those product pages show how quickly the market has moved from first access to a race for validated, reserved capacity. (coreweave.com) The supply squeeze is showing up in rental markets. Public B200 listings compiled in March and April 2026 showed single-GPU Blackwell rentals around $5.98 to $6.08 per hour on RunPod and Lambda, while an eight-GPU CoreWeave cluster was listed at $68.80 per hour. (deploybase.ai) (getdeploying.com) Other vendors are also moving to lock in Blackwell credentials and inventory. Premio said on April 13 that it had validated support for NVIDIA RTX Pro Blackwell graphics processing units across several industrial and on-premises systems, extending Blackwell branding beyond hyperscale cloud into edge and enterprise hardware. (natlawreview.com) Boost Run is also using the certification as a public-market milestone. The company said it has a definitive merger agreement with Willow Lane Acquisition Corp. and expects to trade on Nasdaq under the ticker BRUN after the deal closes. (prnewswire.com) The immediate question is no longer whether Blackwell systems exist in the cloud. It is which providers can prove performance on real workloads, and how much that proof will cost customers as the newest NVIDIA capacity gets spoken for. (nvidia.com) (prnewswire.com)