Nvidia Rubin and Blackwell: GPU Product Mix and Ramp
Full research prompt
Analyze what Nvidia disclosed about the product mix across Blackwell, Blackwell Ultra, and the next generation (Rubin and Rubin-era systems), including networking, rack-scale systems such as NVL72, and any statements about the ramp timing. Assess how the transition could affect gross margin, pricing, and customer buying patterns. Include publicly estimated shipment and mix figures from analysts, and flag which are estimates rather than company-reported.
From Nvidia Earnings Q3 FY2027: Data Center Revenue, Guidance and the AI Cycle
Nvidia’s transition is not a pause between generations. In the October quarter, Vera Rubin added nearly $20 billion in its first production quarter while Blackwell Ultra kept growing and still improved year-over-year mix. The margin dip ahead is a memory-cost and early-ramp effect, not a demand air pocket: older GPUs remain fully utilized and are being re-priced higher in the rental market.
What Nvidia disclosed on the mix
The only architecture-level revenue number Nvidia has put on the current transition is for Vera Rubin, not a full Blackwell / Ultra / Rubin split. On the November 17, 2026 call for fiscal Q3 2027 (quarter ended October 25), interim CEO and CFO Colette Kress said the Rubin platform generated nearly $20 billion in its first quarter of production shipments and called it the fastest product ramp in company history. The investor presentation says the same figure was added quarter over quarter, with contribution across both Hyperscale and AI Clouds, Industrial & Enterprise (ACIE). Against official Data Center revenue of $107.6 billion, that is roughly high-teens of the segment—close to the “about 20% of data center revenue” Kress guided on the August call, before the print came in above the $108 billion total-revenue guide.
Production shipments started in August 2026. Three months later, racks were landing at every major hyperscaler and leading neoclouds. Kress said hyperscale partners brought Vera Rubin into production in about one month, half the time required for Grace Blackwell. Named sites with racks running: CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius, with SpaceXSI coming online. Oracle has made Vera Rubin NVL72 and Vera CPUs generally available; Google has opened Vera Rubin-powered A5X instances.
Blackwell was not displaced in the commentary. CFO commentary attributes Data Center growth to three things together: introduction of Rubin infrastructure, standalone Vera CPUs, and continued growth of Blackwell Ultra. The presentation’s line is “sustained strength for Blackwell, with Rubin off to an extraordinary start.” Nvidia still does not break out B200 vs. B300/GB300 vs. Rubin as a percentage of revenue. The last hard Ultra mix figure is older: in fiscal Q3 2026 (reported November 2025), GB300 had surpassed GB200 and accounted for roughly two-thirds of Blackwell revenue, and management called Blackwell Ultra the leading architecture across customer categories.
Roadmap timing that is company-stated, not inferred:
- Vera Rubin NVL72 is in volume shipment now.
- Vera Rubin Ultra is “on track to launch next year.”
- Feynman remains scheduled for 2028.
- CPU revenue is expected to more than double in fiscal 2028, with Vera adopted by AWS, OCI, CoreWeave, Meta, OpenAI, Nebius, Nscale, and SpaceXSI, including as a standalone processor, not only inside GPU racks.
Rack-scale systems and the networking attach
NVL72 is the commercial unit, not a chip SKU. Grace Blackwell NVL72 moved rack assembly out of the customer data center and into the supply chain; Vera Rubin ships in the same 72-GPU rack form (72 Rubin GPUs, 36 Vera CPUs). Kress’s next step is DSX POD, a modular AI-factory reference design that extends that idea from the rack to the cluster. Pods are meant to arrive ready to connect, cutting deployment from months to weeks. Nvidia plans first pods to U.S. and Central Europe customers early next year, with Dell, Foxconn, Schneider Electric, and others. The economic argument she attached: annual revenue from a one-gigawatt AI factory approaching $50 billion, so reaching first token one quarter earlier could be worth more than $10 billion to the operator. That is a company framing, not a booked number.
The content-per-watt story is how Nvidia explains why customers do not wait out a generation. On the August call, Jensen Huang’s revenue opportunity per gigawatt was about $18 billion for Hopper, about $25 billion for Grace Blackwell, and about $40 billion for Vera Rubin. The step-up is not only a faster GPU. The Rubin stack as described includes Vera CPU, Rubin GPU, NVLink scale-up, InfiniBand or Ethernet scale-out, and the Groq LPU. Selling more of the rack is the mechanism behind the higher dollars per constrained megawatt.
Networking is no longer a side line. Q3 networking revenue exceeded $21 billion. Co-packaged optics scale-out is in volume production. At Supercomputing 2026, Lambda announced the first production deployment of NVIDIA Photonics CPO switches across more than 10,000 GB300 GPUs, with Nvidia claiming 5x better power efficiency, 15x higher reliability, and 10% lower latency versus pluggable optics. Spectrum-X and BlueField-4 are part of the same attach motion, including BlueField-4 used for agent monitoring (NVIDIA Sentry).
Company performance claims, which are Nvidia’s figures rather than independent measurements: Vera Rubin NVL72 delivered up to 2x the performance per GPU of Blackwell Ultra in its MLPerf Training debut; on an MLPerf endpoints benchmark, Kress said Vera Rubin delivered 10x more tokens per watt versus GB300, with further gains expected from software still early in the lifecycle. Earlier company claims of 30x throughput per megawatt and 35x lower token cost versus Grace Blackwell Ultra should be treated as marketing benchmarks, not audited mix data.
Gross margin: Ultra helped, Rubin and memory hurt
The margin path is the clearest company statement on how the transition hits profitability.
| Period | GAAP gross margin | What management said |
|---|---|---|
| Q3 FY26 (Oct 2025) | 73.4% | Blackwell ramp, improved mix and cost vs. Hopper systems |
| Q2 FY27 (Jul 2026) | 75.0% | Blackwell Ultra in volume; memory costs already rising |
| Q3 FY27 (Oct 2026) | 74.8% (non-GAAP 74.9%) | YoY up on better Blackwell Ultra mix; sequential dip from the initial Rubin ramp. Came in above the 74% guide on mix |
| Q4 FY27 guide | 71–72% | Expected trough. Includes an extra week. No China Data Center compute assumed |
| FY2028 outlook | 72–73% | Recovery as price increases already executed take effect in fiscal Q1 2028 |
Two mechanisms are stacked, and they pull in opposite directions. Blackwell Ultra’s richer mix lifted margin versus a year ago. The initial Rubin ramp pulled it down slightly versus the prior quarter, which is the usual pattern when a new full-rack system is still absorbing yield, packaging, and systems cost. The larger guided drop into the January quarter is memory, not a Rubin demand problem. On the August call Kress described “extreme pricing conditions in memory,” with increases that had already exceeded prior expectations and were headed higher. Purchase commitments rose from $279 billion at the end of Q2 to $390 billion at the end of Q3, primarily multi-year memory procurement. Inventory rose to $39.7 billion to support the Rubin ramp; receivables rose on shipment linearity from the launch and extended terms on large multi-quarter agreements.
Q4 revenue is guided to $130 billion, plus or minus 3%, with sequential growth from AI infrastructure demand and the continued Vera Rubin ramp. Fiscal 2028 revenue growth was raised on the November call to about 75% year over year, from about 70% in August, off a higher fiscal 2027 base. Management still frames supply, not orders, as the constraint through at least the end of fiscal 2028.
Pricing and how customers are buying through the transition
Nvidia has not published a price list. What it has said is that price increases have already been executed and are expected to show up in fiscal Q1 2028 margins, which is why the 71–72% Q4 trough is described as the bottom rather than a new run-rate. Separately, Bloomberg reported on August 22, 2026 that contract server builders told large customers—including those supplying Microsoft, Google, and Oracle—to expect increases of more than 15% on many Grace Blackwell and Vera Rubin configurations shipping in early 2027, with the size varying by generation and memory configuration. The Information’s figure, cited by multiple outlets, was around 17%, and one estimate attached at least $5 billion of extra hardware cost to a 1 GW build at that rate. Nvidia has not confirmed those percentages in an SEC filing or on the earnings call. Treat them as reported customer communications, not company guidance.
The rental market is doing the opposite of a clearance sale. Kress said token prices have fallen nearly 60% since May, which accelerated usage rather than freeing capacity. Hopper GPUs have been recontracted at up to twice their original price under multi-year agreements. B300 rental rates are up more than 60% since introduction. Compute is fully utilized across every hyperscaler and neocloud Nvidia serves. That is the buying-pattern tell: customers are not skipping Ultra to wait for Rubin, and they are not dumping Hopper to fund Rubin. They are running all three, and scarcity is repricing the installed base upward.
Customer mix is broadening at the same time the architecture turns over:
- Hyperscale: $60.1 billion, up 23% sequentially and 71% year over year, driven by the Rubin ramp plus sustained Grace Blackwell.
- ACIE: $47.5 billion, up 18% sequentially and 195% year over year, with sequential growth from Rubin shipments to neoclouds and frontier labs starting to build their own data centers.
- Cloud industry backlog cited on the call: $2.5 trillion, up from $2 trillion three months earlier.
- Frontier labs expected to be about 15% of Data Center revenue this year.
- China is not in the forward compute outlook. Q3 shipments of licensed Hopper Data Center products to China were $2.4 billion, and Nvidia said it is no longer shipping those products there.
The practical pattern is a fast cutover at the hyperscalers (about a month to production) layered on top of a long tail. GB300 stays in the fleet because it is sold out and rental rates are rising. Hopper still clears at a premium. Rubin is additive capacity, not a replacement that creates a quarter of idle orders.
Analyst shipment and mix figures — estimates, not company numbers
Nvidia does not report GPU units, rack counts, or a Blackwell/Ultra/Rubin unit mix. Everything below is third-party and the sources do not agree.
Rack counts (estimates). Bernstein, as summarized on November 14, 2026, forecasts about 66,000 Nvidia rack shipments in 2026 and about 85,000 in 2027, with Vera Rubin about 60% of 2027 rack shipments, GB300 most of the rest, and Rubin Ultra only about 2,000 NVL72 racks in the fourth quarter of 2027. Morgan Stanley checks circulated in August 2026 pointed to nearly 7 million Rubin and Rubin Ultra units in 2027 and Rubin NVL72 rack shipments that could reach about 90,000 in 2027; an October 2026 Morgan Stanley note put 2027 racks at 90,400 versus an 82,000 rack forecast for 2026. Those two houses are not on the same 2026 base (66,000 vs. 82,000). Do not average them into a false precision.
Rack price / BOM (estimates). Morgan Stanley’s May 2026 bill-of-materials work, widely cited since, put a Vera Rubin VR200 NVL72 at about $7.8 million for hyperscalers versus about $4.0 million (one write-up uses $3.99 million) for a GB300 NVL72, with memory content near $2.0 million versus about $374,000—a roughly 435% increase—and memory’s share of the rack rising sharply. A Bernstein figure cited in secondary coverage is about $9.1 million per rack and about $47 billion per gigawatt all-in. That $47 billion is not the same concept as Nvidia’s $40 billion per gigawatt, which is Nvidia’s own revenue opportunity, not the customer’s full data-center cost. An HSBC estimate put about 70% of a roughly $4 million VR200 NVL72 BOM into the Rubin GPUs excluding HBM. The $4 million versus $7.8–9.1 million spread is large enough that these should be read as different scopes (GPU-centric BOM vs. ODM purchase price vs. all-in), not as a consensus ASP.
Unit mix and timing (estimates, partly overtaken by the print). SemiAnalysis, in a November 13, 2026 update reported by secondary outlets, moved about 100,000 Rubin GPUs from calendar Q3 to Q4 2026 and kept full-year 2026 Vera Rubin chip shipments near 980,000. It raised Q3 GB300 by about 130,000 units to fill the gap, then cut Q4 GB300 by about 260,000 and H200 by about 150,000. Its model implied about $109.1 billion of Nvidia revenue in calendar Q3 2026 and $125.3 billion in Q4. Actual fiscal Q3 revenue was $115.8 billion, so that near-term model was already low before the call. The same note cut Nvidia’s 2027–2030 compute-revenue shipment forecasts by 22%, 39%, 56%, and 66%. Those out-year cuts are a scenario, not a shipment report.
TrendForce’s September 2026 summary said 2026 high-end GPU shipments would stay Blackwell-led—one Chinese-language summary put Blackwell near 80%—with Rubin’s share reduced by HBM4 yield and readiness issues and a mainstream ramp only after the first half of 2027. That was written before Nvidia reported nearly $20 billion of Rubin revenue in the October quarter. The right reading is that Rubin is already a large revenue contributor in the second half of 2026, while unit share can still be a minority if ASP per rack is much higher than GB300. TrendForce also flagged a possible shift from 12-high to 8-high HBM4/HBM4e to improve yield. That is an estimate of a spec still under evaluation, not a confirmed Nvidia change.
A separate July 2026 SemiAnalysis report said the Kyber NVL144 rack for Rubin Ultra had slipped into 2028 on orthogonal-backplane manufacturing problems, and that a stopgap wider scale-up design had been dropped. Nvidia’s response at the time was “our roadmap is intact,” without confirming or denying the cabinet date. On the November call the company reiterated Vera Rubin Ultra for next year and did not discuss Kyber. Treat Kyber timing as unresolved.
What this means if you are buying, competing, or modeling the ramp
The transition raises Nvidia’s dollars per constrained megawatt and pulls customers forward rather than making them wait. A buyer who delays for a cheaper GB300 quarter is competing with rental rates that are already up more than 60% on B300 and with a platform Nvidia says produces about $40 billion of its revenue per gigawatt versus about $25 billion on Grace Blackwell. The cost of waiting is utilization and time-to-token, not just ASP.
For gross margin models, the company path is specific: Ultra mix was the YoY support, the Rubin start was a small sequential drag, and the guided trough to 71–72% in the January quarter is memory cost plus early systems ramp, with a stated recovery to 72–73% in fiscal 2028 if the already-executed price increases land in the April quarter. A model that marks Rubin as structurally lower-margin than Ultra is ahead of what Nvidia has said. A model that ignores HBM and the $390 billion commitment stack will miss the trough.
For unit models, do not back into “Rubin is 20% of units” from the revenue share. At Morgan Stanley’s roughly 2x rack price versus GB300, high-teens revenue share can be a much smaller unit share. Bernstein’s 60% of 2027 racks and Morgan Stanley’s ~90,000 Rubin racks are the published estimates to stress-test, and they disagree on the 2026 base. SemiAnalysis’s 980,000 Rubin chips for calendar 2026 is an estimate whose associated revenue forecast already undershot the October quarter.
For anyone trying to displace Nvidia at the rack, the attach has widened. Networking alone was more than $21 billion in the quarter, CPO is in volume on GB300 clusters, Vera CPU is being bought standalone and is guided to more than double next fiscal year, and the next packaging step (DSX POD) moves cluster integration into the supply chain the way NVL72 moved rack integration. Matching a GPU on FLOPS does not match the purchase order, which is now a liquid-cooled 72-GPU domain plus CPU, scale-up fabric, scale-out optics, and software that hyperscalers are qualifying in about a month.
Recent Findings Supplement (November 2026)
Nvidia’s Q3 FY2027 earnings (reported November 17, 2026, for the quarter ended October 25, 2026) provided the primary new disclosures on product mix, ramps, and related dynamics after mid-May 2026.[1][2]
Vera Rubin NVL72 (the rack-scale successor to Grace Blackwell NVL72/GB300 variants) entered volume production shipments in August 2026 and generated nearly $20 billion in revenue in its first full quarter of shipments. Nvidia called this the fastest product ramp in company history, with production racks already running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius (and coming online soon at SpaceXSI). Hyperscalers brought Rubin systems into production in roughly one month—about half the time required for Grace Blackwell.[1][3]
Data Center revenue reached $107.6 billion (up 21% sequentially and 110% YoY), split between Hyperscale (~$60.1 billion) and AI Clouds/Industrial & Enterprise ($47.5 billion). Rubin contributed across both segments with strong, diverse demand. Blackwell (including Ultra variants) showed sustained strength, while Rubin added incremental volume. Networking exceeded $21 billion.[4][5]
Gross margin commentary directly tied the mix to profitability. YoY margin expansion (to 74.8% GAAP / 74.9% non-GAAP) was driven by an improved Blackwell Ultra mix. The slight sequential dip reflected the initial Rubin ramp (higher memory and system costs early in production). Q4 FY2027 guidance (which includes an extra week) calls for revenue of $130 billion ±3% (no China Data Center compute assumed) and gross margins of 71–72%, described as the trough before recovery to the 72–73% range in FY2028.[6][7]
Roadmap and next-generation updates: Vera Rubin Ultra remains on track for 2027; Feynman is slated for 2028. Vera CPUs are ramping alongside GPUs, with adoption by hyperscalers, neoclouds, and AI labs (e.g., Meta, OpenAI, CoreWeave) for agentic workloads.[8]
Performance and system-level details (new benchmarks): In its MLPerf Training v6.1 debut, Vera Rubin NVL72 delivered up to 2x performance per GPU versus an optimized GB300 NVL72 (Blackwell Ultra) baseline across multiple models, with examples including 2x on DeepSeek-V3 671B MoE at 1,024 GPUs. Rack-level claims include ~2.3–2.5x dense FP4/FP8 compute and 2.4x HBM bandwidth versus GB300 NVL72, with inference cost reductions projected at up to 10x lower per million tokens. Each Rubin GPU features 288 GB HBM4.[9][10]
Analyst estimates (explicitly flagged as such, not company-reported): Bernstein (cited November 14, 2026) projects ~66K Nvidia rack shipments in 2026 and ~85K in 2027, with Vera Rubin accounting for ~60% of 2027 rack shipments (GB300 supplying most of the rest) and Rubin Ultra contributing ~2K NVL72 racks in 4Q27. Earlier TrendForce analysis (September 2026) noted Blackwell dominance (~80% of high-end GPU shipments in 2026) with Rubin ramping to mainstream status in 2027 amid HBM4 yield constraints; some supply-chain commentary pointed to potential short-term delays into early 2027 before the earnings update confirmed the August start.[11][12]
Implications for gross margins, pricing, and buying patterns: The transition to rack-scale systems (NVL72) and fuller platforms (GPUs + Vera CPUs + networking) is shifting revenue mix toward higher-ASP integrated solutions, supporting Nvidia’s cited revenue opportunity of ~$40B per gigawatt for Rubin versus ~$25B for Blackwell. Initial margin pressure stems from ramp costs and mix shifts but is expected to normalize as volumes scale and software optimizations improve efficiency. Customers are accelerating adoption of full-rack deployments for both training and inference; cheaper tokens (down ~60% since May in some reports) are driving higher usage and agentic workloads, which in turn boost demand for CPUs and sustained GPU utilization. Hyperscalers and neoclouds are qualifying and deploying faster than prior generations, with broader sovereign/enterprise layering expected later.[13][5]
No major new regulatory or policy updates specific to product mix appeared in the latest sources. Supply commitments rose sharply to $390 billion (primarily multi-year memory deals). All figures above derive from the November 17, 2026, earnings release, accompanying materials, or contemporaneous analyst commentary; earlier 2026 references (e.g., August shipment start) are contextual only where they inform the most recent updates.