Nvidia stages overlapping cycles
Nvidia appears to be overlapping product generations to sustain demand: reports say Rubin R100 entered mass production while memory (HBM4) and sample timing issues are keeping attention on Blackwell-class systems. That mix—earlier production on a next-gen part plus constrained memory—means customers should expect continuous generation overlap and that roadmap clarity will become a key commercial promise. (markets.financialcontent.com) (coincentral.com)
Nvidia is trying to sell tomorrow’s chips before today’s chips are fully out of the spotlight. On March 16, 2026, Nvidia said its Vera Rubin platform had “seven new chips in full production,” even though Blackwell systems are still the boxes most customers are deploying now. (nvidia.com) That is unusual only if you think chip launches happen one at a time. In data centers, they overlap like car model years, because cloud companies place orders months before racks are installed and years before those machines are retired. (nvidia.com) The reason memory keeps showing up in this story is simple: an artificial intelligence chip is useless without the ultra-fast memory stacks bolted next to it. Rubin is built around High Bandwidth Memory 4, which is the next version of that stacked memory and is meant to move more than 2 terabytes of data per second in SK hynix’s first 12-layer samples. (news.skhynix.com) Those memory chips are not flowing at full industrial volume yet. SK hynix said in March 2025 that it had shipped 12-layer High Bandwidth Memory 4 samples early, but it was still targeting mass production in the second half of the year, which tells you how long the gap can be between a sample and a real supply ramp. (news.skhynix.com) That gap helps explain why Blackwell does not disappear just because Rubin is “in full production.” Blackwell already has qualified systems, existing customer designs, and memory supply that is further along, so buyers who need capacity in 2026 can keep taking Blackwell while Nvidia lines up enough High Bandwidth Memory 4 for Rubin-class machines. (nvidia.com) (news.skhynix.com) Nvidia also changed the rhythm of its roadmap before this week’s headlines. At Computex 2024, Jensen Huang said the company had moved to a one-year cadence for its data-center artificial intelligence accelerators, which makes overlap less of a glitch and more of the plan. (nvidianews.nvidia.com) Once you move to annual launches, every delay in memory, packaging, power delivery, or server qualification spills into the next cycle. The result is a market where “current generation” means the gear you can install now, while “next generation” means the gear vendors want you budgeting for before the current build-out is finished. (nvidia.com) (news.skhynix.com) Rubin itself is not a small step. Nvidia said the Vera Rubin NVL144 system links 144 graphics processing units and 144 Vera central processing units in one rack-scale design, which is the kind of machine sold to hyperscalers planning giant training clusters and agent-style inference farms. (nvidia.com) That scale raises the cost of uncertainty. If a cloud company is reserving power, buildings, liquid cooling, and tens of thousands of memory stacks, it needs to know whether Nvidia’s next rack ships in calendar 2026, slips into 2027, or arrives in limited volume first. (nvidia.com) (news.skhynix.com) So the real product Nvidia is selling alongside Blackwell and Rubin is scheduling confidence. In an annual-release market with memory bottlenecks, customers are buying a place in line as much as they are buying a chip. (nvidia.com) (news.skhynix.com)
