The Physical Stack of AI · Chips and accelerators
You can name Nvidia's three-stop accelerator cadence (Hopper → Blackwell → Rubin), explain what changes between them, and place upcoming Rubin Ultra and Feynman generations on the timeline.
Nvidia ships a new accelerator architecture roughly every two years, and every frontier lab's training schedule, every datacenter buildout, every inference price cut is keyed to that cadence. If you want to know what AI infrastructure 2027 will look like, you look at what Nvidia is launching in 2026.
The four lessons in this chapter walk the staircase: from Hopper (the H100, the workhorse of the GPT-4 era) to Blackwell (B200 / GB200 / GB300, in volume production through 2026), to Rubin (announced as a full-production platform in 2026), and then to what's announced but not yet broadly shipping: Rubin Ultra in 2H 2027, the specialized Rubin CPX inference chip, and the next-next-gen "Feynman" architecture expected in 2028.
By the end you'll be able to read a sentence like "we're building out 450,000 GB200s under a 15-year Oracle lease" and know which hardware generation that is, how it compares to what's next, and why the timeline matters.
Type: multi-choice
Prompt: > In May 2026, which Nvidia AI accelerator generation is shipping in volume to customers?
Choices: - (a) Rubin (R200) — full production - (b) Blackwell (B200 / GB200) — in volume since early 2026; GB300 ramping ✓ - (c) Hopper (H100) — still the current generation - (d) Feynman — the post-Rubin architecture
Chapter contains 4 lessons.