Table of Contents
On February 25, 2026, NVIDIA reported record Q4 FY2026 revenue of $68.1 billion — and simultaneously, CNBC published an exclusive first look at Vera Rubin, NVIDIA’s next-generation AI platform. Built around six co-designed chips and assembled into the Vera Rubin NVL72 rack, this platform promises 10× more performance per watt and 10× lower inference token cost than Grace Blackwell. Here is everything you need to know, verified line-by-line against official sources.
Figure 1: Vera Rubin vs Grace Blackwell — verified specs. Sources: NVIDIA Newsroom (CES 2026), VideoCardz, Tom’s Hardware, StorageReview.
✅ EDITORIAL NOTE: Every statistic in this article has been individually verified against primary sources (NVIDIA Newsroom, CNBC, Tom’s Hardware, VideoCardz, StorageReview). Unconfirmed figures, analyst predictions, and estimates are clearly labeled as such.
What Is NVIDIA Vera Rubin NVL72?
NVIDIA CEO Jensen Huang announced Vera Rubin at CES 2026 (January 6, 2026), confirming that the chips had already entered full production — ahead of the original H2 2026 mass-deployment target. The Vera Rubin NVL72 is NVIDIA’s third-generation rack-scale AI platform, designed specifically for inference at scale and agentic AI workloads.
- Production status: Vera Rubin is ‘in full production’ as of CES 2026 (Jan 6, 2026). Partner availability: H2 2026. [Source: Jensen Huang CES 2026 keynote / Tom’s Hardware]
- Platform naming: Named after astronomer Vera Florence Cooper Rubin, who proved the existence of dark matter. [Source: NVIDIA / Network World]
- Architecture type: Six co-designed chips functioning as a unified system. NVIDIA calls this ‘extreme co-design.’ [Source: NVIDIA CES 2026 Keynote]
The six chips that form the platform, all verified from official sources:
- Rubin GPU — primary AI accelerator, 336 billion transistors, HBM4 memory (288 GB, 22 TB/s)
- Vera CPU — 88-core custom ARM ‘Olympus’ architecture, 227 billion transistors, 1.5 TB LPDDR5x
- NVLink 6 Switch — 3.6 TB/s per-GPU bandwidth, 260 TB/s total rack scale-up bandwidth
- ConnectX-9 SuperNIC — 1.6 Tb/s per-GPU bandwidth, GPU-direct RDMA networking
- BlueField-4 DPU — KV-cache storage platform for agentic AI workloads, offloads security from CPU/GPU
- Spectrum-6 Ethernet Switch — co-packaged optics (CPO), 5× better power efficiency vs prior gen, 102 TB/s capacity
Figure 2: The six co-designed chips of the Vera Rubin NVL72. All chip names and specs verified from NVIDIA CES 2026 Keynote, Tom’s Hardware, and StorageReview.
Vera Rubin vs Blackwell: Verified Specification Breakdown
The following comparison uses only official NVIDIA figures, cross-checked against Tom’s Hardware, VideoCardz, StorageReview, and The Register. No estimated or projected performance data is included in this table.
Inference Performance: Rubin GPU delivers 50 PFLOPS (NVFP4), 5× the performance of Blackwell B200’s 10 PFLOPS. Rack total: 3.6 EFLOPS. [Source: NVIDIA Newsroom, VideoCardz, Tom’s Hardware]
Training Performance: 35 PFLOPS (NVFP4) per GPU, 3.5× Blackwell. Rack total: 2.5 EFLOPS. [Source: NVIDIA Newsroom]
HBM4 Memory per GPU: 288 GB at 22 TB/s bandwidth — a 2.8× increase over Blackwell’s HBM3E (8 TB/s). [Source: VideoCardz, StorageReview, The Register]
Rack HBM4 Total: 20.7 TB of HBM4 at 1.6 PB/s bandwidth across all 72 GPUs. [Source: VideoCardz, Tom’s Hardware]
NVLink 6 Bandwidth: 3.6 TB/s per GPU (2× Blackwell NVLink 5), 260 TB/s total rack scale-up bandwidth. [Source: NetworkWorld, StorageReview]
LPDDR5x (CPU Memory): 54 TB total across 36 Vera CPUs — 2.5× the Blackwell NVL72. [Source: VideoCardz, ServeTheHome]
Token Cost Reduction: 10× lower inference token cost vs Blackwell. Benchmarked using the Kimi-K2-Thinking model (32K/8K ISL/OSL). [Source: NVIDIA official product page]
GPU Count for MoE Training: Requires 1/4 the GPUs of Blackwell to train an equivalent MoE model. [Source: NVIDIA Newsroom, VideoCardz]
Performance per Watt: 8× the inference compute per watt vs Blackwell NVL72. [Source: StorageReview — note: CNBC cited 10× total system efficiency, which includes cooling improvements]
⚠️ IMPORTANT CLARIFICATION: ‘8× perf/watt’ (GPU inference compute only) is confirmed by StorageReview from NVIDIA slides. ’10× performance per watt’ is the figure CNBC reported from NVIDIA’s broader system-level claim including cooling improvements. Both are from verified sources; they measure slightly different things.
Assembly Time: ~5 minutes for compute board assembly, vs ~2 hours for Blackwell. Jensen Huang confirmed this at CES 2026 keynote; cross-confirmed by Sherwood News analyst notes. [Source: Tom’s Hardware, Sherwood News]
Cooling: 100% liquid cooled (DLC — Direct Liquid Cooling). First fully liquid-cooled NVIDIA rack system. [Source: Tom’s Hardware, CNBC]
Confidential Computing: First rack-scale confidential computing from NVIDIA — extends across all CPUs, GPUs, and NVLink fabric. [Source: StorageReview, NetworkWorld]
NVIDIA’s $68.1B Quarter: Verified Earnings Data
Figure 3: NVIDIA FY2026 revenue by quarter. All figures from NVIDIA official earnings release (Feb 25, 2026) and NVIDIA Investor Relations.
- Q4 FY2026 Revenue: $68.1 billion — record quarterly revenue, up 73% YoY from $39.3B in Q4 FY2025. [Source: NVIDIA Newsroom press release, Feb 25, 2026]
- FY2026 Full-Year Revenue: $215.9 billion — up 65% from FY2025. [Source: NVIDIA Newsroom press release, Feb 25, 2026]
- Q4 Net Income (GAAP): $43 billion, up 94% year-over-year. [Source: CNBC earnings report, Feb 25, 2026]
- Data Center Revenue (Q4): $62.3 billion — 91%+ of total revenue. [Source: CNBC earnings report, Feb 25, 2026]
- Q1 FY2027 Guidance: $78 billion (±2%) — explicitly excludes China Data Center compute revenue. [Source: NVIDIA earnings call, Feb 25, 2026]
- FY2026 Quarterly Progression: Q1: $44.1B → Q2: $46.7B → Q3: $57.0B → Q4: $68.1B [Source: Fortune, CNBC, 247WallSt (Feb 26, 2026)]
Meanwhile, one important context note verified from multiple sources: NVIDIA’s Q1 FY2027 guidance of $78 billion explicitly excludes Data Center compute revenue from China, reflecting the ongoing impact of U.S. export controls on H20-class chips. This is not a speculative risk — it is explicitly stated in NVIDIA’s own earnings guidance.
Physical Facts About the Rack (Weight, Components, Cooling)
CNBC was granted an exclusive first look at Vera Rubin at NVIDIA’s headquarters in Santa Clara, California, on February 13, 2026. The following physical facts were reported directly from that visit and are confirmed by NVIDIA:
- Total rack components: 1.3 million components in the full NVL72 rack. [Source: CNBC exclusive, Feb 25, 2026]
- Total microchips per rack: ~1,300 microchips, compared with Grace Blackwell’s 864. [Source: CNBC exclusive, Feb 25, 2026]
- Superchip components: Each Vera Rubin superchip (2 Rubin GPUs + 1 Vera CPU) has ~17,000 components. [Source: CNBC exclusive, Feb 25, 2026]
- Rack weight: Nearly 2 tons. [Source: CNBC exclusive, Feb 25, 2026]
- Manufacturing locations: U.S., Taiwan (TSMC), and a new Foxconn plant in Mexico. [Source: CNBC exclusive / NVIDIA annual filing, Feb 2026]
- Power consumption: ~2× that of Grace Blackwell — exact figures not publicly released by NVIDIA. Estimated 120–130 kW per rack by passhulk.com technical analysis. [Source: CNBC (relative claim), technical estimate third-party]
- Cooling type: 100% liquid cooled (DLC). NVIDIA states this helps data centers consume ‘much less water’ than traditional evaporative cooling. [Source: CNBC exclusive, Feb 25, 2026]
📌 NOTE on power: The statement that Vera Rubin consumes ‘2× the power of Blackwell’ comes from CNBC’s report of NVIDIA’s own claim. The absolute figure (~120–130 kW) is a third-party engineering estimate and not an official NVIDIA specification. We label it as such.
5. Pricing: What Analysts Estimate (Not Official NVIDIA Pricing)
⚠️ IMPORTANT: NVIDIA does not publicly disclose rack pricing. The following figures are third-party analyst estimates only.
VERIFIED FACT — Analyst price estimate: Futurum Group (analyst: Daniel Newman) estimates $3.5M–$4M per rack — approximately 25% higher than Grace Blackwell. [Source: CNBC exclusive, Feb 25, 2026 / TipRanks, Feb 2026]
This estimate has been reported by both CNBC and TipRanks and is attributed to a named analyst at a named firm. It is not an official NVIDIA figure, and actual pricing through cloud providers will differ.
Step-by-Step Guide: Should You Build on Blackwell Now or Wait for Vera Rubin?
Step 1 — Assess Your Deployment Timeline
Vera Rubin partner availability is H2 2026. If you need production AI infrastructure before that window, the current Blackwell ecosystem (GB200, GB300) is in full production and broadly available today via AWS, Google Cloud, Microsoft Azure, CoreWeave, Lambda, Nebius, and Nscale — all confirmed Vera Rubin partners per CNBC.
✅ VERIFIED FACT — Production status: Blackwell is in full production now. GB300 systems also in full scale manufacturing as of CES 2026. [Source: Jensen Huang CES 2026 keynote]
Step 2 — Identify Your Workload Type
Vera Rubin’s most significant advantages are in inference-heavy workloads: chatbots, reasoning agents, long-context applications, and Mixture-of-Experts (MoE) model inference. For these workloads, the 10× token cost reduction is commercially significant. For batch training workloads on established architectures, Blackwell’s current economics may remain adequate through 2026.
Step 3 — Evaluate Agentic AI Requirements
The BlueField-4 DPU in Vera Rubin introduces a dedicated KV-cache storage platform — the NVIDIA Inference Context Memory Storage Platform. This is specifically designed for agentic AI workloads that require large context windows and state persistence across multi-step reasoning chains. This component does not exist in the current Blackwell architecture.
→ See our guide: MCP Agentic AI Systems: 2026 Production Architecture
Step 4 — Consider the Real Cost
Analyst estimates (Futurum Group / CNBC, Feb 25, 2026) place Vera Rubin NVL72 at approximately $3.5M–$4M per rack — ~25% higher than Grace Blackwell. Cloud providers will offer instance-based access, which reduces upfront capital requirement. Most developers will access Vera Rubin through cloud instances rather than direct rack procurement.
Step 5 — Plan the Transition
NVIDIA’s rack-scale design has been built for backward compatibility — the third-generation MGX design enables ‘seamless upgrades’ from prior generations, per NVIDIA’s official product page. Workloads running on Blackwell today can be migrated to Rubin instances when available, without requiring a full application rewrite.
What Vera Rubin Means for AI Developers & API Costs
The 10× inference token cost reduction is benchmarked using the Kimi-K2-Thinking model (32K input / 8K output sequence lengths) comparing Blackwell GB200 NVL72 and Rubin NVL72 — this is specified on NVIDIA’s official Vera Rubin NVL72 product page. Token costs through public APIs are set by cloud providers and AI model companies, not NVIDIA directly. The timeline for lower API pricing will depend on when cloud providers deploy Rubin capacity and how they price it.
📌 No verified source specifies a date when API costs from OpenAI, Anthropic, or Google will change. Any specific predicted date would be speculation. We do not speculate here.
What is factually confirmed from NVIDIA and verified third-party sources:
- Vera Rubin is in full production (confirmed CES 2026)
- Partner availability (AWS, Azure, GCP, CoreWeave, Lambda, Nebius, Nscale): H2 2026
- Meta has announced plans to use Vera Rubin in its data centers by 2027 (confirmed CNBC, Feb 25, 2026)
- OpenAI and Anthropic are on NVIDIA’s expected customer list (CNBC, Feb 25, 2026)
- The current Blackwell infrastructure (GB200, GB300) remains the production standard for all major cloud AI services
For developers building on AI APIs today:
Frequently Asked Questions
When will Vera Rubin be available?
NVIDIA confirmed partner availability for H2 2026 at CES 2026. This is the current official timeline. Consumer cloud instances will depend on individual provider deployment schedules.
How many chips are in the Vera Rubin NVL72 rack?
The NVL72 rack contains 72 Rubin GPUs and 36 Vera CPUs, along with NVLink 6 switches, ConnectX-9 SuperNICs, BlueField-4 DPUs, and Spectrum-6 switches. CNBC reported approximately 1,300 total microchips per rack and 1.3 million total components, verified by NVIDIA.
What is the performance improvement over Blackwell?
Per GPU: 5× inference performance (50 vs 10 PFLOPS NVFP4), 3.5× training performance (35 vs 10 PFLOPS NVFP4). Per rack: 5× inference (3.6 vs ~0.72 EFLOPS), 3.5× training (2.5 vs ~0.71 EFLOPS). All figures from NVIDIA official announcements.
What is the price of Vera Rubin?
NVIDIA does not publicly disclose rack pricing. Futurum Group analyst Daniel Newman estimated $3.5M–$4M per rack (approximately 25% higher than Grace Blackwell) in reporting cited by CNBC on February 25, 2026. This is an analyst estimate, not an official NVIDIA price.
How does Vera Rubin impact open-source AI?
The 4× reduction in GPUs needed to train MoE models reduces capital and energy cost for all training workloads, including open-source projects. This applies to models like DeepSeek-V3 and Llama 4 when training on Rubin infrastructure.
→ DeepSeek-V3 vs Llama 4 Benchmarks: 2026 Local AI Comparison
Is Vera Rubin relevant for AI security?
Yes. Vera Rubin introduces rack-scale confidential computing — hardware-level trust zones extending across all CPUs, GPUs, and NVLink fabric. This matters for enterprises in regulated industries running sensitive workloads on shared infrastructure.
→ AI Security Architecture 2026: Zero Trust & LLM Hardening
External Sources
- NVIDIA Newsroom — NVIDIA Announces Financial Results for Q4 and Fiscal 2026 (Feb 25, 2026)
- NVIDIA Official — Vera Rubin NVL72 Product Page
- NVIDIA Official — DGX Vera Rubin NVL72 Product Page
- CNBC — First look at Nvidia’s AI system Vera Rubin and how it beats Blackwell (Feb 25, 2026)
- CNBC — Nvidia (NVDA) earnings report Q4 2026 (Feb 25, 2026)
- Tom’s Hardware — Nvidia launches Vera Rubin NVL72 AI supercomputer at CES 2026 (Jan 5, 2026)
- VideoCardz — NVIDIA Vera Rubin NVL72 Detailed: 72 GPUs, 36 CPUs (Jan 6, 2026)
- StorageReview — NVIDIA Launches Vera Rubin Architecture at CES 2026 (Jan 7, 2026)
- ServeTheHome — NVIDIA Launches Next-Generation Rubin AI Compute Platform at CES 2026
- The Register — Nvidia unpacks Vera Rubin rack system at CES (Jan 6, 2026)
- NetworkWorld — At CES, Nvidia launches Vera Rubin platform for AI data centers (Jan 6, 2026)
- Fortune — Nvidia smashes Q4 2026 with $68 billion in revenue (Feb 25, 2026)
- Sherwood News — Wall Street’s key takeaways from Jensen Huang’s CES speech (Jan 6, 2026)
Have any thoughts?
Share your reaction or leave a quick response — we’d love to hear what you think!
-
0
-
0
-
0