GPU supply still tight
Demand for Hopper- and Blackwell-class GPUs remains strong and supply is constrained by advanced packaging and high‑bandwidth memory bottlenecks, leaving enterprises with long lead times. Market reports note hyperscaler demand has absorbed much of the available capacity and some transitions (like Vera Rubin) could delay parts of the next HBM cycle (ibtimes.com.au) (stocktitan.net) (cloudnews.tech).
Companies still trying to buy Nvidia’s top artificial intelligence chips in April 2026 are running into the same problem as last year: demand is outrunning supply. (nvidianews.nvidia.com) Nvidia reported $62.3 billion in data center revenue for the quarter ended January 25, 2026, up 75% from a year earlier, and Chief Executive Officer Jensen Huang said customers are “racing to invest in AI compute.” (nvidianews.nvidia.com) The constraint is no longer just the graphics processor itself. High-bandwidth memory, which stacks memory chips like a vertical sandwich next to the processor, and advanced packaging, which wires those pieces together, have become the choke points. (news.skhynix.com, investors.micron.com) Micron told investors on March 20, 2025 that its high-bandwidth memory output for calendar 2025 was already sold out and that strong 2026 demand was forcing customer negotiations a year ahead. The company also said the artificial intelligence ramp was tightening leading-edge dynamic random-access memory supply outside high-bandwidth memory. (investors.micron.com) Taiwan Semiconductor Manufacturing Co. is still expanding the packaging step known as Chip-on-Wafer-on-Substrate, or CoWoS, but the build-out has not erased the shortage. TSMC’s April 2026 quarterly materials showed first-quarter revenue of $35.9 billion and guided the next quarter even higher, reflecting continued demand across advanced chip programs. (investor.tsmc.com, trendforce.com) Memory makers are pushing the next generation, but that transition also ties up supply. SK hynix said on September 12, 2025 that it had completed HBM4 development and prepared mass production, then showed a 16-layer 48-gigabyte HBM4 part at CES in January 2026. (news.skhynix.com, news.skhynix.com) That matters because Nvidia’s newest systems use more of the scarce parts per deployment. Blackwell Ultra systems announced in March 2025 pack up to 288 gigabytes of HBM3E memory per graphics processor, and Nvidia said Equinix would offer preconfigured Blackwell-ready facilities in 45 markets. (nvidianews.nvidia.com, nvidianews.nvidia.com, developer.nvidia.com) Market trackers now expect Blackwell to dominate Nvidia’s high-end shipment mix this year even with supply-chain friction. TrendForce said on April 8, 2026 that Blackwell’s share of Nvidia’s high-end graphics processor shipments could rise to 71% in 2026, while Rubin’s share was cut to 22% because of delays tied to HBM4 certification, networking upgrades and higher power and cooling demands. (trendforce.com.tw) The result for enterprises is simple: the biggest cloud companies and model builders are absorbing much of the available capacity first, while everyone else waits for memory stacks, packaging slots and complete racks to line up. Nvidia’s own results show the spending wave is still accelerating, not easing. (nvidianews.nvidia.com, investors.micron.com, trendforce.com.tw)
