Blackwell Ultra supply ramp clears the CoWoS packaging bottleneck
The correct forward model stops forecasting from accelerator demand and starts forecasting from grid interconnection and cooling capacity. Revenue visibility improves and the set of variables that can delay it changes.
Two quarters of accelerator shipments were bounded by advanced packaging allocation. That constraint has cleared and the resulting revenue step-up is visible in both sequential acceleration and a guidance raise at both ends of the range. The forward question is whether rack-level delivery, not die supply, becomes the next binding constraint, and the evidence suggests it already has.
- Data center revenue reached $54.2 billion, up 62 percent year over year
- Full-year guidance floor raised from $295 billion to $310 billion
- Networking revenue $8.9 billion, up 41 percent, as full-stack attach rises
- Rubin generation draws roughly three times current rack power, moving the constraint to facility supply
All Signals (3)
NVIDIA data center revenue $54.2B, up 62% YoY, as Blackwell Ultra ramp clears supply bottleneck
NVIDIA signs 11-year sovereign compute agreement valued in the 'tens of billions' with Gulf state fund
CUDA 13.2 release adds unified memory tiering and deprecates two legacy APIs used by 8% of sampled workflows
Related Themes
Cluster Debate
The cluster documents a supply constraint clearing rather than a demand inflection: data center revenue reached $54.2 billion, up 62 percent, and full-year guidance was raised to $310-315 billion from $295-305 billion in the same quarter.
- Data center revenue $54.2 billion, up 62 percent year over year, with full-year guidance raised to $310-315 billion from $295-305 billion
- Gross margin declined 200 basis points to 74.8 percent on a higher proportion of complete systems rather than standalone silicon
- CUDA 13.2 added unified memory tiering with a two-release compatibility shim for APIs used by approximately 8 percent of sampled workflows
- The sovereign agreement carries an eleven-year term with delivery extending beyond the current hardware generation
Every signal in this cluster is a capacity statement dressed as a demand statement. They can ship racks, they can tier memory, and they signed a country for eleven years, which tells you the constraint they are actually managing is how fast they can spend.
- Guidance floor raised rather than the midpoint, which is a supply statement
- Gross margin decline explained by systems mix rather than pricing
- Rubin generation roughly triples rack power and ships with a retrofit reference design rather than a greenfield spec
- Sovereign contract valued only as 'tens of billions' over an eleven-year term
- The sequential acceleration is attributable to supply clearing rather than a demand shock
- The systems and networking mix is rising structurally across the product line
- The next binding constraint sits outside the chip supply chain entirely
An eleven-year term extends demand visibility well beyond the sector's two-to-three-quarter cycle, which is a structural change in forecastability regardless of the value disclosure.
A contract with no number is a press release. The term is long because sovereign funds think in decades, and the value is unstated because a figure would become the comp for every government deal that follows.
- Eleven-year term with phased delivery beginning in the current fiscal year
- Value described only as in the tens of billions of dollars over the term
A two-release shim covering 8 percent of sampled production workflows is a managed migration, which is standard practice for a platform this deeply embedded.
CUDA had no short deprecation window before this release. The deadline exists because the new memory tiering is load-bearing for the next rack generation, and the ecosystem has to be on it before the hardware ships.
- Compatibility shim provided for two release cycles
- Approximately 8 percent of sampled production workflows depend on a deprecated call
Both agents agree on the mechanism: the packaging constraint cleared and the bottleneck moved to the building. The unresolved item is whether the sovereign agreement and the CUDA migration window describe confidence or a schedule being managed. The delivery schedule for the first sovereign phase would settle it: on-time delivery validates the multi-generation demand read, while a slip that aligns with the Rubin launch would recast the agreement as roadmap coordination rather than demand visibility.