GPU supply crunch intensifies

The AI boom is now running into the oldest problem in hardware. There are not enough parts. Not enough finished systems. Not enough memory stacked close enough to the GPU to keep the whole machine fed. That shortage is showing up in two places at once. Prices are climbing for the accelerators companies want right now, and the chips Nvidia wants to sell next are already being constrained by the memory they will need. The immediate squeeze is on Blackwell. Nvidia’s current data center line has become the default engine for training and serving large models, and the company’s own DGX B200 packs eight Blackwell GPUs into one box with 1,440 GB of total GPU memory. Nvidia says that system delivers 3x the training performance and 15x the inference performance of the previous generation. That kind of jump does not just attract buyers. It concentrates demand on a very small set of components, especially high-bandwidth memory, or HBM, which sits beside the GPU and moves data far faster than ordinary server DRAM. (nvidia.com) That memory has become the choke point. Blackwell-class GPUs use HBM3e today, and Rubin-class parts are moving to HBM4. SK hynix used CES 2026 to show a 16-layer, 48 GB HBM4 stack, while Samsung said at GTC 2026 that its HBM4 is in mass production for the Vera Rubin platform and can run at 11.7 Gbps, above the 8 Gbps industry baseline it cited. Nvidia, meanwhile, says Rubin is already in full production and positions the platform as the next rack-scale step after Blackwell. The implication is plain enough: Nvidia is trying to ramp a new GPU generation while the hardest part of the bill of materials is still scarce, specialized, and controlled by a tiny number of suppliers. (news.skhynix.com) That is why the supply story is no longer just about Nvidia’s fabs or TSMC’s packaging lines. It is about who can deliver the memory stacks, at what speed, and in what volume. Parameter reported in March that Samsung and SK hynix had secured the HBM4 supply positions for Rubin, with Micron left out of that flagship program. Even if that specific supplier split changes, the broader fact does not. The pipeline has narrowed to a handful of companies making a component that every top-end AI GPU now depends on. When one part of the stack becomes that concentrated, launches stop being about chip design and start being about allocation. (parameter.io) Vendors are responding by moving the scarce hardware into more expensive, more tightly integrated systems. ServeTheHome’s tour of MSI’s GTC 2026 booth showed exactly where the market is heading: not loose accelerator cards for hobbyists or even ordinary workstation buyers, but GB300 stations and liquid-cooled PCIe servers aimed at offices, labs, and enterprise deployments. Nvidia is doing the same at the high end with rack-scale products like Vera Rubin NVL72, which combines 72 Rubin GPUs and 36 Vera CPUs in a single system. The form factor tells the story. When supply is tight and every GPU is precious, vendors do not sell cheap abundance. They sell complete infrastructure. (servethehome.com) That shift also explains the price signals now coming out of the market. AI Tool Discovery’s Blackwell overview pegs a B200 GPU module at roughly $30,000 to $40,000 and notes 192 GB of HBM3e per GPU. Those figures are not surprising anymore. The surprising part is how normal they have become. A top-end AI GPU is no longer just a processor. It is a bundle of advanced packaging, bleeding-edge memory, power delivery, interconnect, and cooling, all competing for the same industrial bottlenecks. MSI’s WS300, a GB300 workstation-class system that can run from a standard outlet, is the compact version of that reality. (aitooldiscovery.com)