Nvidia News

[Adam Kay]

Abridged earnings call highlights and points to note.

Record financial performance: Revenue reached $68 billion in Q4 (+73% year-on-year), with record operating income and free cash flow. Full-year data centre revenue was $194 billion (+68%). Free cash flow totalled $97 billion for fiscal 2026 (period ending jan 2026)

Data centre dominance: Q4 data centre revenue was $62 billion (+75% YoY, +22% sequentially), driven by strong demand for the Blackwell architecture and Blackwell Ultra ramp. Nearly 9 gigawatts of Blackwell infrastructure are deployed. Note AMD crowing about 6GW over 5 years with Meta. Nvidia will deploy > 20GW this year alone.

Sustained growth outlook: Q1 revenue is expected to be $78 billion ±2%, mainly from data centres. (I think I stated 74 early but it's +$10B without china-aligns with our expectation of +$10B each Q). Sequential revenue growth is expected throughout calendar 2026, with visibility into 2027. No China data centre compute revenue is assumed. Looks like our 80/90/100/110 prediction aligns. And could be low. Translates TTM fwd of $380M and earnings of $220B over the next 12 months. Soon approaching twice the earnings of the worlds second most profitable company. Think about that!

Networking surge: Networking generated $11 billion in Q4 (+3.5x YoY which is 250%). Nvidia is now the biggest network company in the world. Full-year networking revenue exceeded $31 billion (over 10x since the Mellanox acquisition). They paid buttons for Mellanox-what an acquisition. NVLink, Spectrum-X Ethernet and InfiniBand adoption hit record levels.

Performance leadership**: GB300 NVL72 delivers up to 50x performance per watt and 35x lower cost per token versus Hopper.** Continuous CUDA optimisation improved GB200 NVL72 performance by up to 5x in four months. NVIDIA positions itself as delivering the lowest cost per token. Right there is why they are Nr1. Monetising AI is ALL about token generation at lowest cost. AMD can't even match Hopper today! Asics, TPU are not even close.

AI demand inflection: Agentic AI has reached a turning point. Inference equals revenue, as token generation drives customer monetisation. Hyperscaler 2026 CapEx expectations are approaching $700 billion, up nearly $120 billion since the start of the year.

Sovereign AI expansion: Sovereign AI revenue more than tripled YoY to over $30 billion, led by Canada, France, the Netherlands, Singapore and the UK. China revenue remains uncertain despite limited H200 approvals. NB- nice to have it back but irrelevant for the next year or so. Ignore the noise. I do think it will get resolved. I don't buy the 'China might catch up' They could have very competitive models but who is going to buy it. Not the West. Would you rent a Chinese AI agent? Will any S&P companies use Chinese AI. So imo it's an irrelevance. China will consume it, which is fine. And it would appear Deepseek is using Blackwell anyway so no china sales is not the case. The Wall Street Journal claims Deepseek is a fraud-all their work is stolen (distilled) from OpenAI et al, big surprise.

New platform – Rubin: Unveiled at CES, Rubin includes six new chips (Vera CPU, Rubin GPU and new networking components). It will train models using one-quarter the GPUs and reduce inference token costs by up to 10x versus Blackwell. Production shipments begin in H2. The math says Rubin will be 500X lower token cost than Hopper. Just incredible.

Gaming and other segments:
Gaming revenue: $3.7 billion (+47% YoY), though supply constraints are expected to weigh on Q1 and beyond.

Professional Visualisation: $1.3 billion (+159% YoY).
Automotive: $604 million (+6% YoY), driven by self-driving demand. Physical AI contributed over $6 billion in FY2026.

Gross margins remain strong: Q4 GAAP gross margin was 75%. Full-year margins are expected in the mid-70s. Management argues sustainability depends on delivering generational leaps in performance per watt and per dollar. And precisely what they are doing. Every year, a new leap fwd.

Capital allocation: Returned $41 billion (43% of free cash flow) to shareholders via buybacks and dividends, while increasing inventory and purchase commitments to secure future supply.
Strategic ecosystem expansion: Deepened partnerships with OpenAI, Meta and Anthropic (including a $10 billion investment in Anthropic). NVIDIA emphasises CUDA ecosystem breadth, full-stack AI infrastructure, and extreme co-design as competitive advantages.

Long-term thesis: Computing has shifted to AI token generation. “Compute equals revenue.” Agentic AI is the current growth driver; physical AI (robotics, manufacturing, autonomous systems) is the next major wave. NB-we know this is going to be huge. Data centre CapEx could reach multi-trillion USD levels by 2030 if AI adoption continues accelerating. The question is now, when not if Nvidia generate $1T in a fiscal year.

[Adam Kay]

Validation Milestone
Microsoft Azure became the first major cloud provider to power on and begin validating a Vera Rubin NVL72 rack (announced by Satya Nadella on 13 March 2026). This is a significant engineering win: the full rack (72 Rubin GPUs + 36 Vera CPUs, NVLink-6 fabric, liquid cooling) is integrated and undergoing qualification in Azure datacentres. It positions Microsoft ahead for early deployments, with broad availability still guided for the second half of 2026 (H2 2026, i.e., July–December).Rack Cost Estimate
A Vera Rubin NVL72 rack is likely priced in the $3.5 million to $5 million range (most analyst estimates cluster around $3–4 million, with some supply-chain views up to $5–5.7 million). This represents a premium over Blackwell GB200/GB300 NVL72 racks (around $3 million).
The uplift stems from advanced components: HBM4 memory, denser NVLink-6, Vera CPUs, and enhanced liquid cooling (cooling alone rises from ~$50,000 on Blackwell to ~$55–56,000 on Rubin).
NVIDIA doesn't publish official prices, but the economics favour rapid payback through vastly higher efficiency.

Performance Improvements
Rubin delivers massive leaps, especially for inference (the dominant AI workload now): Vs. Blackwell (GB200/GB300 NVL72): Up to 5x higher inference performance per rack (e.g., 3.6 exaFLOPS FP4 vs. ~0.7–0.8 exaFLOPS equivalents). Per-GPU gains include ~50 PFLOPS NVFP4 inference (5x vs. Blackwell), plus better power efficiency and features for agentic/long-context models. Training MoE models needs ~4x fewer GPUs.

Cost per Token Shrinking
This is where Rubin crushes economics—driving the cost of intelligence off a cliff for inference-heavy workloads (e.g., agentic AI, reasoning, MoE models): Vs. Blackwell: NVIDIA states 10x lower cost per million tokens (official claim on specific MoE/reasoning benchmarks like Kimi-K2-Thinking).
Vs. Hopper: Blackwell cut costs by up to 10x (real deployments saw drops from $0.20/million tokens to $0.05 or lower with NVFP4). Rubin stacks another 10x reduction → potentially 100x lower effective cost per token over Hopper in optimised cases.
Providers already realised 4x–10x drops moving Hopper → Blackwell (e.g., 20¢ → 5¢/million tokens for MoE). Rubin positions sub-1¢/million at scale once volumes ramp in H2 2026.
The upfront rack cost ($4M average) is offset by far more useful compute per dollar, lower power/token, and fewer units needed—making massive AI scaling dramatically cheaper.

In short: Validation is a big early win for Microsoft/NVIDIA, racks cost a hefty $3.5–5M each (premium justified), performance jumps 5x over Blackwell (20–25x over Hopper), and token costs plummet another 10x vs. Blackwell—paving the way for agentic AI at unprecedented scale and affordability.