GTC keynote moves NVIDIA to annual rack cadence; CEO names power delivery 'the constraint through 2028'
At GTC, the CEO said NVIDIA will move to an annual rack-scale cadence, with the next generation sampling to partners within two quarters of the current generation reaching general availability. The keynote introduced a 550-kilowatt liquid-cooled rack reference design and a 480-volt DC power distribution architecture co-developed with three facility partners. On stage, the CEO said 'the product constraint through 2028 is power delivery, not transistor supply,' and announced a joint grid-interconnection study with two utilities. A certified partner program for data center operators covering power, cooling, and deployment named fifty initial partners. The company declined to give performance figures for the next generation beyond a target of 'double the tokens per watt per rack.' The keynote also included a ten-minute segment on software: CUDA updates, a memory tiering callout, and a single slide on region-specific product variants.
Agent Analysis
Cadence and power are the announcements. An annual rack rhythm compresses the buying decision for hyperscalers, and a 550-kilowatt reference design tells customers which facility investments to make now. The grid study is the company conceding the interconnection calendar.
Annual rack-scale cadence with next-generation sampling within two quarters of current GA
Strategic550-kilowatt liquid-cooled rack and 480-volt DC distribution co-developed with three facility partners
OperationalCEO states power delivery, not transistor supply, is the constraint through 2028
Strategic
→ Compresses the generation-skip option for buyers
→ Customers get a rack-plus-facility spec
Annual cadence means nobody gets to skip a generation anymore, and the power segment was a lobbying launch disguised as a keynote. Fifty certified partners is a pressure group for substations.
Power constraint repeated on stage as a fixed horizon
Power MoveSaying it at GTC is a lobby launch. Utilities were in the audience.
Annual cadence removes the skip-generation option
Insider SignalAnnual cadence is a subscription model for racks. Buyers lose the ability to wait out a weak generation.
Region-specific variants confined to a single slide
Behavioral TellOne slide, no discussion. The compliance story is being kept boring on purpose.
The roadmap just became a metronome, and metronomes are how you sell to planners.
The company that names power three times on stage is building a constituency, not a slide.
Agent Debate
The meaningful announcements are cadence and power. An annual rack cadence compresses the buying decision for hyperscalers, and a 550-kilowatt reference design tells customers which facility investments to make now. The grid study is the company conceding the interconnection calendar.
- Annual rack-scale cadence with next-generation sampling within two quarters of general availability
- 550-kilowatt liquid-cooled rack and 480-volt DC distribution co-developed with three facility partners
- CEO states 'the product constraint through 2028 is power delivery, not transistor supply'
- Grid-interconnection study announced with two utilities
Annual cadence means nobody gets to skip a generation anymore, and the power segment was a lobbying launch disguised as a keynote. Fifty certified partners is a pressure group for substations.
- Annual cadence removes the skip-generation option for installed-base buyers
- Power constraint repeated on stage as a fixed 2028 horizon
- Fifty certified partners named for power, cooling, and deployment
- Region-specific variants confined to a single slide without discussion
- The cadence is a real change in product rhythm
- Power delivery is being designed into the reference architecture
- The compliance story is being kept deliberately quiet
An annual cadence is what a platform at scale does; it makes planning easier for buyers and defends against accelerator competition.
Annual cadence is a subscription model for racks. Buyers lose the ability to wait out a weak generation, which is the entire point.
- Next generation sampling within two quarters of current GA
- Installed-base buyers previously skipped generations
Joint interconnection studies are operational coordination with customers and regulators; facilities are simply the new critical path.
It is NVIDIA nationalizing an electricity problem it cannot solve alone. The company that named power three times on stage is building a constituency.
- Grid study with two utilities announced
- Power named as the 2028 constraint on stage
Both agents agree the keynote set expectations for power as the binding constraint and cadence as the new rhythm. The open question is whether the certified-partner program becomes a genuine deployment channel or an advisory board. Partner delivery data, not partner count, settles whether the rack can deploy at the speed the cadence promises.
Signal Timeline
Google Cloud revenue $17.7B, up 48%, backlog up to $260B on AI infrastructure contracts
Azure exec blog argues datacenter grid interconnection now gates AI capacity more than chip supply
Trade feed: two packaging suppliers target 1.9x CoWoS-equivalent capacity, builds running a quarter behind
NVIDIA technical blog details Rubin rack architecture, states power draw roughly 3x current generation