The Annual Cadence: How Chip Roadmaps Are Built

The market leader now ships on a one-year clock. At the Computex trade show in Taipei in June 2024, NVIDIA's chief executive, Jensen Huang, said the company had "a one-year rhythm." He described the approach plainly: build at the scale of an entire data center, then sell the parts on an annual cycle. On the same stage he named the products behind the claim: Blackwell Ultra for 2025 and a new platform, Rubin, for 2026.

That sentence changed how the industry plans. Before it, data center accelerators from the market leader arrived roughly every two years. Hopper was announced at NVIDIA's annual developer conference in March 2022. Blackwell followed at the same event in March 2024. An annual cadence halves the interval between the moments a buyer's newest equipment becomes the previous generation.

Now look at how the promise was kept. Blackwell Ultra arrived in the second half of 2025, as announced. At the Consumer Electronics Show, or CES, in January 2026, NVIDIA launched the Rubin platform, which it described as six new chips: the Vera central processing unit, or CPU, the Rubin graphics processing unit, or GPU, a sixth-generation NVLink switch, a network interface card, a data processing unit and an Ethernet switch. NVIDIA said Rubin was in full production and that partner products would be available in the second half of 2026. NVIDIA's public roadmap then names Rubin Ultra for 2027, on the Kyber rack architecture, and a generation called Feynman for 2028.

Advanced Micro Devices, or AMD, made the same commitment. After launching its Instinct MI300X in December 2023, it moved to an annual cadence: the MI325X, then the MI350 series launched in June 2025, then the MI400 series in 2026 inside the Helios rack, with the MI500 named for 2027.

Why an annual cadence? The structural answer is that the product is no longer one chip. It is a stack: GPU, CPU, memory, links, switches, rack and software, each on its own development timeline. A one-year rhythm lets a vendor refresh some layers each year while holding others steady. Blackwell Ultra, for instance, kept Blackwell's system architecture while upgrading processors, memory and networking. Offsetting the layers is how a company ships every year without redesigning everything every year.

The consequence for everyone downstream is a mismatch of clocks. Chips now move on a one-year clock. Buildings move on a clock of several years, from land and power to commissioning. Power connections can take longer still. So a building designed around one generation's rack will often open to house a later one. That is why the previous module's question matters: a facility has to be built for density it cannot yet see in production.

Now the reading skill. A roadmap is a set of claims with different grades of certainty, and the vocabulary tells you which grade you are looking at. Learn these stages in order.

Announced means the vendor has named the product and some figures. Nothing has shipped.

Sampling means early units are with selected customers for testing.

In full production means manufacturing has started at volume. It does not mean the product is in data centers.

Partner availability means server makers and cloud providers can offer it. It is the first stage a typical customer can buy.

Deployed at scale means it is running in large numbers in real facilities. This is the stage that matters for power demand and for revenue.

Each step can take months. A headline saying a chip is in full production is a true statement about the factory and a weak statement about the data center.

Then treat dates by distance. The current generation's dates are commitments backed by production. The next generation's dates are plans. Anything two or more years out is a target. NVIDIA's own language respects this distinction, and press reports have at times questioned the timing of later rack generations. Faraday's read is that the far end of any roadmap should be read as a statement of direction and ambition, not as a delivery schedule.

Finally, watch what the roadmap says about the rack and the building, not only the chip. NVIDIA's plan to move to 800 volts direct current alongside Kyber is a roadmap item for electricians and switchgear makers as much as for chip buyers. The most useful lines on a chip roadmap are often the ones about power.