Last updated: 7 August 2026. Nvidia just closed a quarter with $81.6 billion in revenue, up 85% year on year, and guided to $91 billion for the next one. Its Vera Rubin platform has moved from samples to production. It released an autonomous-driving model that is genuinely open for commercial use. And the stock is down roughly 18% year to date, with Apple having overtaken it as the world’s most valuable company.
That combination — record operational results alongside deteriorating market sentiment — is the real Nvidia story in 2026. Below is what has actually been announced, with dates and figures, and a clear line between what Nvidia has confirmed and what is only reported.
The numbers: three quarters of Nvidia earnings
One thing trips up almost every summary of Nvidia’s results: its fiscal year is offset. Fiscal 2026 ended on 25 January 2026. So the quarter reported in February 2026 was Q4 FY2026, and the quarter reported in May 2026 was Q1 FY2027.
| Quarter (reported) | Revenue | Data Centre | GAAP EPS | Notes |
|---|---|---|---|---|
| Q3 FY2026 (19 Nov 2025) | $57.0B (+62% YoY) | $51.2B (+66%) | $1.30 | Gaming $4.3B. Huang: “cloud GPUs are sold out.” |
| Q4 FY2026 (25 Feb 2026) | $68.1B (+73% YoY) | $62.3B (+75%) | $1.76 | Beat the $65B guide. Gaming $3.7B. |
| Q1 FY2027 (20 May 2026) | $81.6B (+85% YoY) | $75.2B (+92%) | $2.39 | Networking alone $14.8B, up 199%. |
Full-year FY2026 revenue came in at $215.9 billion, up 65%, with GAAP net income of $120.1 billion. Guidance for Q2 FY2027 is $91.0 billion ±2% at a 75% non-GAAP gross margin. Nvidia also authorised a new $80 billion buyback and raised its dividend a cent to $0.25.
A reporting change that will confuse comparisons
From Q1 FY2027, Nvidia stopped separately disclosing Gaming, Professional Visualization and Automotive revenue. The new framework has two market platforms: Data Centre (split into Hyperscale and ACIE) and Edge Computing, which came in at $6.4 billion. If you see a “gaming revenue” figure for that quarter anywhere, it is an estimate — the last official standalone number is Q4 FY2026’s $3.7 billion.
Two other figures worth carrying: on the Q1 FY2027 call management expressed “full confidence in the $1 trillion in Blackwell and Rubin revenue we foresee from 2025 through calendar 2027” — raised from an earlier $500 billion figure — and said hyperscale customers accounted for roughly 50% of data centre revenue, about $38 billion.
Vera Rubin: what it is, and whether it has actually shipped
Announced at CES on 5 January 2026, Vera Rubin is a six-chip platform: the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch.
A naming note: Nvidia’s current product name is Vera Rubin NVL72 — 72 Rubin GPU packages paired with 36 Vera CPUs. Earlier material used “NVL144”, which counted the 144 compute dies (two per package). Same rack, different counting convention. Using NVL144 as the current name now reads as out of date.
| Specification | Vera Rubin NVL72 |
|---|---|
| NVFP4 inference | 3,600 PFLOPS (3.6 exaflops) |
| FP8 / FP6 training | 1,260 PFLOPS |
| HBM4 capacity | 20.7 TB |
| HBM4 bandwidth | 1,580 TB/s |
| Configuration | 72 Rubin GPUs, 36 Vera CPUs (3,168 Olympus cores) |
| NVLink 6 | 260 TB/s switch bandwidth |
| Per-GPU networking | ConnectX-9 at 1.6 Tb/s |
Per chip: each Rubin GPU delivers about 50 PFLOPS NVFP4 — roughly 3.3× Blackwell Ultra — with 288 GB of HBM4 across eight stacks at around 13 TB/s, built from two reticle-sized 3 nm tiles plus two I/O dies on CoWoS-L, at roughly 1.8 kW. Each Vera CPU carries 88 Armv9.2 Olympus cores and 176 threads with 1.2 TB/s of memory bandwidth, about 20% more than Grace.
Shipping status — confirmed
- 25 February 2026: CFO Colette Kress confirmed the first Vera Rubin samples shipped to customers that week.
- 20 May 2026: Nvidia said it was on track to begin production shipments in the second half of the year, starting in its fiscal Q3, with the ramp continuing into Q4.
- 15 July 2026: Jensen Huang, pushing back on delay reporting, said in Tokyo that “Vera Rubin is already in production. Giant amounts of production incoming.”
So: samples since February, production shipments beginning around now. No volume customer deployments have been confirmed as of early August 2026.
Reported but not confirmed: SemiAnalysis reported on 6 July 2026 that the Kyber rack for Rubin Ultra had slipped to 2028 over PCB midplane manufacturing problems. Nvidia’s only response was that its roadmap is intact. Nvidia’s public roadmap still shows Rubin Ultra in 2027 — around 100 PFLOPS FP4 per package, 1 TB of HBM4E, 576 GPUs per Kyber rack — and Feynman in 2028.
Blackwell is still the product paying the bills
Rubin gets the headlines, but essentially all of that $75.2 billion data centre quarter was Blackwell. The confirmed detail:
- Grace Blackwell systems were roughly two-thirds of data centre revenue in Q4 FY2026.
- Nearly nine gigawatts of Blackwell infrastructure is deployed and in use at the major cloud providers.
- GB300 NVL72 is described by Nvidia as the fastest product ramp in the company’s history, with claimed gains of up to 50× performance per watt and 35× lower cost per token versus Hopper.
- Partner data centres larger than 10 MW “nearly doubled in one year, now surpassing 80 sites.”
Alpamayo 2 Super: the autonomous-driving model that actually went commercial
This is the announcement most worth updating. The Alpamayo line has moved quickly and the naming has changed twice, so a lot of coverage is stale:
| Version | Date | Size | Licence |
|---|---|---|---|
| Alpamayo-R1 (research paper name) | Late 2025 | 10B | Research |
| Alpamayo 1 (renamed at CES) | 5 January 2026 | 10B | Weights non-commercial |
| Alpamayo 1.5 | March 2026 | 10B | — |
| Alpamayo 2 Super | 4 August 2026 | 34B | OpenMDW-1.1 weights, Apache 2.0 code — commercial use permitted |
Alpamayo 2 Super combines a 32-billion-parameter vision-language backbone built on Nvidia Cosmos 3 Super Reasoner with a 2.3-billion-parameter diffusion action decoder, reinforcement-learning post-trained. It was trained on roughly 115,000 hours of multi-camera driving video and about 3.7 million “chain-of-causation” traces.
On the LingoQA Lingo-Judge benchmark it scores 79.2, first out of roughly 40 models — ahead of Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 and GPT-4o by 23.2. For each scenario it outputs a trajectory, a causal explanation, a meta-action label, reasoning auto-labels and grounded visual question answering.
The licence change is the substantive news. Alpamayo 1’s weights were non-commercial; Alpamayo 2 Super explicitly permits fine-tuning, derivatives and commercial redistribution. Nvidia cites over 500,000 Hugging Face downloads across the family. Named ecosystem partners from the Alpamayo 1 launch included Lucid, JLR, Uber and Berkeley DeepDrive; we could not verify named commercial adopters of version 2 yet.
Physical AI and robotics
- Isaac GR00T N1.6 is the current shipping humanoid foundation model, integrating Cosmos Reason. GR00T N2 was announced at GTC on 16 March 2026 — a new world-action-model architecture that succeeds at unfamiliar tasks in unfamiliar environments more than twice as often as leading VLA models, and ranks first on MolmoSpaces and RoboArena. Availability is slated for the end of 2026, so N2 is announced, not released.
- Isaac GR00T reference humanoid robot (31 May 2026): a Unitree H2 Plus body, roughly six feet and 150 lb with 31 degrees of freedom, plus Sharpa Wave tactile hands and a Jetson AGX Thor T5000. Available late 2026 through Unitree; no price disclosed.
- Jetson AGX Thor / T5000 delivers 2,070 sparse FP4 TFLOPS with 128 GB of LPDDR5X at 273 GB/s in a 40–130 W envelope. The developer kit is $3,499. The newer Jetson T4000, announced at CES 2026, offers 1,200 FP4 TFLOPS and 64 GB at 70 W for $1,999 in 1,000-unit volume.
- Cosmos 3 (16 March 2026) unifies synthetic world generation, vision reasoning and action simulation, and is the backbone of Alpamayo 2 Super.
- Newton 1.0, the GPU-accelerated physics engine co-developed with Google DeepMind and Disney Research and governed by the Linux Foundation, reached 1.0 at GTC in March 2026.
Gaming: DLSS 4.5 now, DLSS 5 this autumn
DLSS 4.5 was announced at CES on 6 January 2026 and brings three things:
- A second-generation transformer model for Super Resolution, using five times the compute of the original transformer model. This works on all RTX GPUs, with FP8 optimisation on RTX 40 and 50 series and a heavier cost on RTX 20 and 30.
- Dynamic Multi Frame Generation, which varies the frame multiplier in real time to hold a target frame rate. Shipped 31 March 2026.
- 6X Multi Frame Generation — five generated frames per rendered frame, up from 4X. Dynamic MFG and 6X are RTX 50 series only.
As of January 2026 the counts were 400+ titles with DLSS Super Resolution and 250+ with Multi Frame Generation. If you are choosing hardware to take advantage of this, our RTX 5080 Super and RTX 5070 Ti Super breakdowns cover the current options, and our best 1440p GPU guide has the value picks.
DLSS 5 was announced at GTC on 17 March 2026: a real-time neural rendering model that adds photoreal lighting and materials rather than just upscaling. It ships in autumn 2026 with 15+ titles named. GPU requirements have not been disclosed.
Reported, not confirmed: The Information reported in February 2026 that an RTX 50 Super refresh had been cancelled and RTX 60 delayed, both blamed on the GDDR7 memory shortage. Nvidia has said nothing. What is confirmed is that Kress flagged gaming supply constraints as “a headwind to gaming in Q1 and beyond” on 25 February 2026 — which is at least consistent with the leaks.
Enterprise, industrial and the gigawatt deals
The Siemens Industrial AI Operating System partnership (6 January 2026) is the flagship industrial announcement: GPU acceleration across Siemens’ simulation portfolio, EDA workflows targeting 2–10× speed-ups in semiconductor design, and Omniverse, CUDA-X and PhysicsNeMo underneath. The first blueprint is the Siemens Electronics Factory at Erlangen, launching in 2026, with Foxconn, HD Hyundai, KION Group and PepsiCo evaluating.
On hardware, the RTX PRO 6000 Blackwell Server Edition ships with 24,064 CUDA cores, 96 GB of ECC GDDR7 on a 512-bit bus at 1,597 GB/s, 4 PFLOPS of FP4 and up to 600 W configurable power. Nvidia has not published adoption or unit numbers for it.
The infrastructure commitments are where the scale is:
| Partner | Scale | Nvidia involvement | Date |
|---|---|---|---|
| OpenAI | 10 GW of Nvidia systems | Intent to invest up to $100B progressively | September 2025 |
| Anthropic | Up to 1 GW of Grace Blackwell and Vera Rubin | Nvidia invests up to $10B; Microsoft up to $5B | 18 November 2025 |
| xAI / HUMAIN | 500 MW Saudi data centre | xAI is first customer of the Nvidia-backed JV | 19 November 2025 |
| NAVER + Brookfield | 55 MW growing to 200 MW by 2028 (~100,000 GPUs) | ~$10B structure; Nvidia $1B conditional | 24–25 July 2026 |
| SK Group | 2 GW AI cloud in Korea, first phase | Letters of intent worth over $500B — LOIs, not binding contracts | 24 July 2026 |
| Safe Superintelligence | “Order of magnitude” compute increase on Vera Rubin | Terms not disclosed — third-party “$5B” figures are unverified | 27 July 2026 |
China: approved, but generating no revenue
This is one of the most commonly mis-stated parts of the Nvidia story, so here is the sequence:
- 11 August 2025: Nvidia agrees to remit 15% of China chip-sale revenue to the US government in exchange for H20 export licences.
- 8 December 2025: the administration greenlights H200 sales to China with the US taking a 25% cut; formally cleared 14 January 2026.
- 25 February 2026: Kress confirms small amounts of H200 were approved but “we have yet to generate any revenue.”
- 20 May 2026 (most recent): Kress repeats that licences exist, “we have yet to generate any revenue, and we are uncertain whether any imports will be allowed into the country… we are not including any China Data Centre compute revenue in our outlook.”
The filings bear that out. In the quarter ended 26 April 2026, China including Hong Kong was $4.55 billion, just 6% of revenue, against 78% from the United States. International revenue fell to 22% of the total from 42% a year earlier. Two caveats worth flagging: we could not verify whether the original 15% arrangement still applies in 2026 or was superseded by the 25% H200 surcharge, and there is no 2026 confirmation that the China-specific B30A exists as a shipping product.
BioNeMo and life sciences
The January 2026 BioNeMo update added RNAPro for RNA structure prediction, ReaSyn v2 for synthesisability, nvMolKit for GPU cheminformatics and BioNeMo Recipes, with integrations across NeMo, NIM microservices and the open Clara models.
The adopter list is unusually concrete: Eli Lilly, which established a co-innovation lab with a commitment of up to $1 billion over five years in talent, infrastructure and compute; plus Thermo Fisher, Amgen, Roche, Basecamp Research, Chai Discovery, Natera, Dyno Therapeutics, OpenFold and Terray, alongside lab-robotics partners including Multiply Labs, Lila Sciences and Opentrons.
The risks Nvidia itself discloses
Record results have not translated into a rising share price, and the reasons are in the filings rather than the headlines.
- Customer concentration. In the quarter ended 26 April 2026, three direct customers accounted for 21%, 17% and 16% of total revenue — 54% combined. Three customers also represented 30%, 18% and 16% of accounts receivable, or 64% of the balance.
- Indirect exposure. Nvidia disclosed that “one AI research and deployment company contributed to a meaningful amount of our revenue by purchasing cloud services from our customers” — without quantifying it.
- Commitments. Purchase and supply commitments total $119 billion, of which $95 billion is payable within the rest of FY2027, plus $30 billion of cloud service agreements.
- Circular-financing concerns. On 27 July 2026 Nvidia fell about 5% after a report that it was weighing a ~$250 billion financing guarantee so OpenAI could lease SoftBank’s 10 GW Ohio campus. Credit-default-swap protection on Nvidia debt rose by record amounts, and Apple overtook Nvidia as the world’s most valuable public company. Those terms are reported and unsettled — for scale, Nvidia’s disclosed partner facility lease guarantees total $3.5 billion.
As of 6 August 2026, Nvidia traded around $218.99 with a market capitalisation of $5.30 trillion, on a trailing P/E of 33.6 and a forward P/E of 22.0 — down roughly 18% year to date despite trailing revenue of $253.5 billion and net income of $159.6 billion.
What to watch next
- 26 August 2026: Q2 FY2027 results. Guidance is $91.0 billion; the question is whether Vera Rubin revenue shows up in the mix.
- Vera Rubin volume deployments. Production has started; named customer installations have not been announced.
- Whether the Kyber / Rubin Ultra 2028 slip is real. Nvidia says the roadmap is intact.
- China. Licences exist and revenue is zero. Any actual shipment would be new information.
- Foundry diversification. Nvidia has invested $5 billion in Intel and agreed to co-develop x86 RTX SoCs, but reporting suggests its manufacturing engagement is limited to early packaging talks — see our Intel Foundry coverage for where that stands.
Frequently asked questions
Has Nvidia’s Vera Rubin platform shipped yet?
Partially. Nvidia confirmed the first customer samples shipped in late February 2026 and said production shipments would begin in its fiscal Q3, which runs from roughly August to October 2026. Jensen Huang said on 15 July 2026 that Vera Rubin was “already in production.” No volume customer deployments had been announced as of early August 2026.
What is Alpamayo, and can I use it commercially?
Alpamayo is Nvidia’s open vision-language-action model family for autonomous driving. The current release, Alpamayo 2 Super (4 August 2026), is a 34-billion-parameter model released under the permissive OpenMDW-1.1 licence for weights and Apache 2.0 for code — so yes, commercial use, fine-tuning and redistribution are permitted. The earlier 10-billion-parameter Alpamayo 1 had non-commercial weights.
Is DLSS 4.5 real, and do I need an RTX 50 card?
It is real, announced at CES on 6 January 2026. The second-generation transformer Super Resolution model works on all RTX GPUs. Dynamic Multi Frame Generation and 6X Multi Frame Generation are RTX 50 series exclusive and shipped on 31 March 2026. DLSS 5, a neural rendering model, was announced in March 2026 and ships in autumn 2026.
How much of Nvidia’s revenue comes from China?
About 6%, or $4.55 billion in the quarter ended 26 April 2026, and effectively none of it is data-centre compute. Although the US government approved H200 export licences in January 2026, Nvidia has stated it has yet to generate revenue from them and excludes China data-centre compute from its outlook entirely.
Why is Nvidia stock down when earnings keep breaking records?
Chiefly customer concentration and financing structure. Nvidia’s own filings show three direct customers accounted for 54% of revenue and 64% of receivables in the April 2026 quarter, alongside $119 billion of purchase commitments. Reports in July 2026 that Nvidia might guarantee roughly $250 billion of financing for an OpenAI data-centre lease intensified concerns about circular financing between Nvidia and its largest customers.
What is the difference between Vera Rubin NVL72 and NVL144?
They describe the same rack. NVL72 counts 72 Rubin GPU packages; NVL144 counted the 144 compute dies inside them, at two per package. Nvidia’s current official naming is Vera Rubin NVL72.
Is Blackwell being replaced by Rubin?
Not yet. Grace Blackwell systems made up roughly two-thirds of data centre revenue in the quarter ending January 2026, nearly nine gigawatts of Blackwell infrastructure is deployed at major cloud providers, and Nvidia describes the GB300 NVL72 ramp as the fastest in its history. Rubin is beginning production alongside it, not displacing it.
