GPU supply keeps Blackwell centre-stage
Analysis continues to put NVIDIA’s Blackwell generation at the heart of current data‑centre deployments, while reports suggest Rubin delays tied to HBM4 issues might extend Blackwell’s dominance. Vendors and cloud partners are positioning around available capacity—Vultr was named an NVIDIA Exemplar Cloud after Blackwell testing—while AMD is still pushing to gain ground. (www.fool.com, parameter.io, itbrief.co.uk)
NVIDIA’s Blackwell chips are still the main event in artificial-intelligence data centers, and supply questions around the next Rubin generation are keeping them there. (nvidia.com) Blackwell is NVIDIA’s current data-center graphics processing unit family, and cloud providers are building around it now rather than waiting for Rubin systems promised for the second half of 2026. NVIDIA said in January that Rubin-based DGX systems are coming next, while its current Mission Control software already runs Blackwell-based DGX deployments. (nvidia.com, nvidia.com) The supply question centers on HBM4, short for high-bandwidth memory 4, the stacked memory that sits next to an artificial-intelligence chip and feeds it data at very high speed. Tom’s Hardware reported in January that TrendForce expected HBM4 volume production no earlier than the end of the first quarter of 2026, though NVIDIA said production remained on schedule. (tomshardware.com) That timing matters because the market is buying complete systems, not just future roadmaps. Vultr said on April 7 that it became one of the first NVIDIA Exemplar Clouds after meeting NVIDIA reference-design performance targets on Blackwell hardware. (vultr.com, finance.yahoo.com) Vultr’s testing used a 512-GPU Blackwell cluster across 11 training models, including Llama 3.1 405B, Grok-1 314B and DeepSeek-v3 671B. The company said the program measured throughput, latency and scaling against NVIDIA’s benchmarking recipes. (vultr.com) Rubin is not absent from the market conversation; it is just not the hardware most customers can deploy today. NVIDIA launched the Rubin platform at Consumer Electronics Show 2026, and Vultr said three weeks ago that Rubin cloud instances are expected in the second half of 2026. (nvidianews.nvidia.com, vultr.com) Reports on Rubin timing are mixed. Parameter wrote last month that Samsung and SK Hynix planned to start HBM4 manufacturing in March 2026 and that Rubin remained on track for commercial availability in the second half, while other reports have focused on possible memory bottlenecks. (parameter.io, tomshardware.com) AMD is still trying to use that window to gain share. AMD said its Instinct MI350 series is part of an annual accelerator roadmap, following the MI325X and preceding the MI400 series, as it pushes more memory and higher inference performance into the same data-center buying cycle. (ir.amd.com) For now, the practical choice for many cloud builders is the chip generation they can test, rack and sell this year. As long as Blackwell systems keep clearing performance benchmarks and Rubin timing stays tied to memory supply, Blackwell remains the center of deployment plans. (vultr.com, tomshardware.com, nvidia.com)
