NVIDIA 4Q26 Earnings Review: $68 Billion Was the Headline, $95 Billion Is the Story
The Second Engine: NVIDIA 4Q26 Earnings and the $95.2 Billion Forward Signal
NVIDIA's fiscal Q4 2026 delivered a textbook beat-and-raise quarter. Total revenue of $68.1 billion exceeded the company's own guidance by 4.8% and Bloomberg consensus by 3.5%, setting both a new all-time quarterly record and the largest sequential dollar increase in company history (QoQ +20%, YoY +73%). Non-GAAP EPS came in at $1.62, beating consensus by 5.9%. Next quarter's revenue guidance of $78.0 billion surpassed consensus by 7.6%. Full-year FY26 revenue totaled $215.9 billion, up 65% year-over-year, with $96.6 billion in free cash flow.
This marks the seventh consecutive beat-and-raise quarter since FY25 Q2. "They beat again" has long ceased to be news.
But beneath the numbers, this quarter carried three signals that matter far more than the size of the beat itself — signals about the structural evolution of NVIDIA's growth story, not merely its velocity.
NVIDIA Q4 FY26 Key Metrics: Beat/Miss vs. Consensus
source: Jason & Jarvis
Networking: The Quiet Rise of a Second Growth Engine
If you had to pick a single number to understand the upside surprise this quarter, it wouldn't be the $68.1 billion in total revenue, nor the $62.3 billion from Data Center. It would be Networking's $11.0 billion.
Up 264% year-over-year. 22% above consensus. The largest single-item surprise across every segment and sub-segment.
NVIDIA Q4 FY26 Data Center Networking Revenue
source: NVIDIA
A year ago, this number was under $3.0 billion.
Jensen Huang noted on the earnings call that NVIDIA entered the Ethernet switching market roughly two years ago and is now "probably the largest Ethernet networking company in the world." Full-year networking revenue exceeded $31.0 billion — a greater than 10x increase from FY2021, the year NVIDIA completed its acquisition of Mellanox.
This isn't a product line's natural growth curve. It's AI infrastructure hitting a structural inflection point. As AI clusters scale from dozens of servers to tens of thousands of GPUs, interconnect shifts from being a peripheral accessory to the bottleneck that determines an entire system's utilization and economics. NVLink extends individual compute nodes into rack-scale computers. As Jensen put it — "We don't ship nodes of computers, we ship racks of computers." Each NVL72 rack contains 9 switch nodes, each with 2 chips. NVLink handles scale-up; Spectrum-X and InfiniBand handle scale-out and cross-datacenter connectivity.
The math for customers is straightforward: in a $10 billion or $20 billion AI factory, a 10-20% improvement in network efficiency and utilization translates directly into hundreds of millions of dollars in revenue differential.
NVIDIA Q4 FY26 Data Center Mix (Compute vs. Networking)
source: Jason & Jarvis
Networking now accounts for 17.7% of Data Center revenue, up from under 10% a year ago. Both scale-up and scale-out demand hit record levels this quarter, each growing at double-digit rates sequentially. Jefferies stated plainly that Networking is rapidly emerging as "the critical second growth engine" after Compute.
We flagged this dynamic in our earlier analysis of AWS adopting NVLink Fusion — even competitors are buying into the standard. This quarter's $11.0 billion in Networking revenue is the hardest proof yet.
"Compute Directly Equals Revenue"
Bank of America's Vivek Arya raised a question on the earnings call that many investors have been quietly turning over: with global cloud CapEx approaching $700 billion and some customers showing cash flow strain, can NVIDIA still grow within that envelope?
Jensen's response deserves quoting in full:
"In this new world of AI, compute is revenues. Without compute, there's no way to generate tokens. Without tokens, there's no way to grow revenues."
In the traditional software era, IT CapEx bought "operating conditions" — servers, storage, and networking to run pre-compiled software, totaling roughly $300 to $400 billion per year globally. In the AI era, every inference run, every token generated, corresponds directly to billable output. CapEx has shifted from cost center to a direct function of revenue.
The old world ran on pre-recorded content: pre-compiled code, pre-built recommendation lists, pre-recorded video. Now everything is moving toward real-time generation — every query, every conversation, every coding assist produced on the fly based on the user, context, and intent. Jensen estimates that this mode of computing requires roughly 1,000 times more compute than the old approach.
"The amount of computation necessary is 1,000 times higher than the way we used to do computing."
This thesis found concrete validation over the past two to three months, as agentic AI reached an inflection point. Jensen repeatedly cited Anthropic's Claude Code and Claude Cowork, as well as OpenAI's Codex. These coding agents have spread rapidly among NVIDIA's own engineers, with individual sessions running from minutes to hours, each generating thousands to hundreds of thousands of tokens. Critically, agentic systems spawn multiple agents that collaborate as teams, driving geometric growth in token consumption.
Anthropic's numbers are particularly compelling: revenue grew 10x in a single year, and the company remains "severely capacity-constrained" — demand far outstripping available compute. Jensen described Claude Cowork's agent platform as "revolutionary," one that has "opened up floodgates for enterprise AI adoption."
Meta's experience illustrates the same logic from a different angle. Advances in its GEM models drove a 3.5x increase in Facebook ad click-through rates and more than 1% growth in Instagram conversations, producing meaningful revenue uplift. The same NVIDIA infrastructure that powers these workloads enables Meta Superintelligence Labs to train and deploy frontier agentic AI systems. GPUs produce tokens; tokens produce revenue; revenue justifies CapEx. This loop is closing.
For cloud service providers operating within power-constrained data centers, performance per watt translates directly into tokens per watt, which translates into dollars per watt.
"Choosing the right architecture, the one with the best performance per watt, is literally everything."
Architecture selection has graduated from a technical decision to a P&L decision.
$95.2 Billion in Commitments: Management Voting with Real Money
If the income statement captures "right now," purchase commitments capture "what management thinks about tomorrow."
NVIDIA Total Supply-Related Commitments ($bn)
source: Bernstein
This quarter, NVIDIA's total supply-related commitments surged to $95.2 billion. Last quarter: $50.3 billion. A year ago: $45.1 billion.
Nearly doubled.
This is an extraordinarily strong forward signal. Management's demand visibility now extends into calendar year 2027, and they've demonstrated sufficient conviction to lock in capacity and raw materials accordingly. Bernstein noted that NVIDIA's supply commitments approaching $100 billion "indicate the company's ability to meet growing demand with visibility through CY27."
On the balance sheet, inventories reached $21.4 billion at quarter-end (QoQ +8%), with days of inventory at roughly 115, down 4 days sequentially. A year-over-year doubling in inventory levels looks alarming on the surface, but context matters. NVIDIA is in the midst of a structural transition from shipping components to shipping rack-scale systems. A single GB200 NVL72 rack comprises 72 Blackwell GPUs, 36 Grace CPUs, liquid cooling infrastructure, and switches — system complexity and lead times have increased substantially. This inventory buildup is categorically different from the passive stockpiling of FY23's crypto downturn, when days of inventory spiked to 212.
Management confirmed that all component supply through CY2026 — including memory — has been secured. The company is now preparing its CY2027 supply chain. CFO Colette Kress expects CY2026 revenue to grow sequentially each quarter, exceeding the previously disclosed $500 billion Blackwell and Rubin revenue opportunity — a figure first revealed in last quarter's earnings, and now being surpassed.
Ecosystem Breadth
Several strategic developments from the quarter deserve a concise rundown.
Meta entered a multi-year strategic partnership with NVIDIA, deploying millions of Blackwell and Rubin GPUs, CPUs, and Spectrum-X networking across on-premises, cloud, and AI infrastructure. A notable detail emerged from Citi's management callback: all of NVIDIA's agreements with Meta are firm commitments, whereas competitors' deals with Meta include a 1GW guaranteed portion alongside optional components. The gap between commitment and option is, at its core, a temperature reading of platform trust. Meta also plans to deploy Vera CPUs in CY2027, with Vera Rubin NVL72 systems featuring confidential computing capabilities slated for applications like WhatsApp.
Strategic investments continued at scale. NVIDIA invested $10 billion in Anthropic this quarter and acquired a non-exclusive technology license from Groq for approximately $13 billion. Jensen compared Groq to the Mellanox acquisition — the core idea being to leverage Groq as an accelerator extending NVIDIA's architecture, with further details expected at GTC in March. (For a detailed analysis of the Groq deal's strategic logic — why pay nearly 3x the valuation to neutralize a potential rival — see our prior coverage.)
Sovereign AI continues to accelerate. FY2026 sovereign AI revenue grew over 3x year-over-year, exceeding $30 billion, with major clients in Canada, France, the Netherlands, Singapore, and the United Kingdom. Jensen drew an analogy to electricity and the internet — every nation will build and operate its own AI infrastructure, with long-term AI spending proportional to GDP.
Customer diversification is progressing steadily. The top 5 CSPs and hyperscalers still account for roughly 50% of total revenue, but non-hyperscaler customers — AI model makers, enterprises, supercomputing centers, sovereign clients — are growing faster. Management expects this trend to persist. For the full fiscal year, the single largest direct customer contributed 22% of total revenue, with the second-largest at 14%. Concentration remains high, but incremental customer cohorts are diluting it.
An unexpected bright spot came from Professional Visualization : Q4 revenue of $1.3 billion, up 159% year-over-year, beating consensus by 69%, crossing the $1 billion threshold for the first time. The surge was driven by strong demand for Blackwell-architecture professional GPUs — a sign that AI's penetration is extending from data centers into engineering, design, and creative workstations. Physical AI contributed over $6.0 billion to NVIDIA's FY2026 revenue. Gaming, by contrast, was one of the few soft spots — Q4 revenue of $3.7 billion missed consensus by 8.2%, primarily due to post-holiday channel inventory adjustments and memory supply constraints. Management expects memory supply limitations to remain a headwind over the coming quarters.
Gross Margins: Generational Leaps as the True Lever
Rising HBM prices, product-transition mix dilution, growing customer bargaining power — concerns about NVIDIA's gross margin have persisted throughout the Blackwell cycle. Q4 Non-GAAP gross margin of 75.2% once again provided an answer: 20 basis points above guidance, 16 basis points above consensus, up 170 basis points year-over-year.
NVIDIA Non-GAAP Gross Margin and Operating Margin: Historical Trend
source: Bernstein
Goldman Sachs noted that despite significant HBM price increases, strong gross margin performance reflected the company's substantial pre-purchase commitments for memory in CY2025, effectively locking in costs. But Jensen offered a more foundational framework on the earnings call:
"The single most important lever of our gross margins is actually delivering generational leaps to our customers."
The logic: if each product generation delivers performance-per-watt and performance-per-dollar improvements that substantially exceed the price increase of the system itself, margins hold. A deeper structural support lies in the fact that token demand is growing exponentially across multiple application inflection points — even GPUs that are six years old remain fully utilized in the cloud, with rising pricing. When legacy products remain profitable, pricing power for next-generation hardware rests on firmer ground. (For a detailed analysis of long-cycle GPU economics, see our earlier piece.)
Another frequently overlooked competitive advantage stems from architectural compatibility. All NVIDIA GPUs are backward-compatible — software optimizations developed for Blackwell simultaneously benefit Hopper and Ampere. This is why A100 remains relevant years after deployment. Jensen was blunt: NVIDIA's software works because the architecture is so good. The so-called "dielet tax" that competitors face — latency and power penalties from inter-chiplet communication — is avoided by NVIDIA's reticle-limited monolithic die design.
Management was also candid about ceilings: investors should not model gross margins persistently above 70%. The company's commitment to delivering best-in-class performance alongside the lowest total cost of ownership inherently caps profitability. Full-year Non-GAAP gross margin of 71.3% was dragged lower by Q1's $4.5 billion H10 charge, but Q3 and Q4 recovered to 73.5% and 75.2% respectively. Operating margin reached 67.7%, net margin 58.1%, and FCF margin approximately 51% — with full-year free cash flow of $96.6 billion and operating income of $137.3 billion, both all-time records.
This is high-quality growth.
China: The New Normal at 9%
China (including Hong Kong) accounted for just 9% of FY26 revenue — the lowest since FY16 and roughly a two-thirds decline from the FY22 peak of 26%.
China (Including Hong Kong) Revenue Share: Historical Trend
source: Bernstein
NVIDIA has shifted its revenue recognition methodology from point of invoicing to customer headquarters location. Under the new approach, approximately 76% of Data Center revenue attributed to Taiwan-headquartered entities is reassigned to end customers based in the United States and Europe. The US now accounts for 69% of FY26 total revenue.
A small number of H200 units destined for Chinese customers have received US government approval, but have yet to generate any revenue. China's direct impact on near-term earnings has diminished to very low levels. Management, however, did not dismiss the longer-term competitive undercurrent — "Our competitors in China, bolstered by recent IPOs, are making progress and have the potential to disrupt the structure of the global AI industry over the long term."
The Q1 FY27 guidance of $78.0 billion excludes any China-sourced Data Center Compute revenue. Any incremental contribution from China would be pure upside.
The Road Ahead: $78 Billion Guidance and the Next Catalysts
Q1 FY27 revenue guidance of $78.0 billion (±2%) topped consensus by 7.6%, extending the beat-and-raise cadence. Non-GAAP gross margin guidance of 75.0% (±50bps) held roughly flat versus the current quarter.
Two forward-looking details are worth flagging.
An accounting change for SBC. Starting in FY27, NVIDIA's Non-GAAP metrics will include stock-based compensation. Management explained this as an alignment with Magnificent Seven peers and a move favored by the SEC. Q1 SBC is expected at $1.9 billion, causing Non-GAAP OpEx to jump from Q4's $5.1 billion to approximately $7.5 billion. Strip out SBC, and underlying OpEx comes in at roughly $5.6 billion — up just 10% sequentially. The accounting change will create some short-term comparison noise, but over time delivers a more honest earnings picture. Full-year Non-GAAP OpEx is expected to grow in the low-40s percent year-over-year, though management emphasized that revenue growth has consistently outpaced operating expense growth over the past decade, and this year will be no exception.
Vera Rubin's delivery timeline. First engineering samples were shipped to customers this week, with volume production on track for the second half of the fiscal year. CFO Kress confirmed that "pretty much every single customer" is expected to purchase Vera Rubin — the only question is speed. The platform's six new chips — Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9, BlueField-4 DPU, and Spectrum-6 — promise inference token costs up to 10x lower than Blackwell. (Detailed specifications were covered in our CES Vera Rubin analysis; the platform has now entered its delivery countdown.)
Until Vera Rubin ramps, Blackwell and Blackwell Ultra remain the workhorses — Grace Blackwell systems accounted for roughly two-thirds of this quarter's Data Center revenue. CUDA software continues to compound: within four months, GB200 NVL72 performance improved by up to 5x. Blackwell Ultra, benchmarked by SemiAnalysis's InferenceX suite, demonstrated a 50x improvement in Agentic AI performance over Hopper and a 35x reduction in cost.
GTC in March stands as the next major catalyst: Groq integration details, Rubin deployment roadmaps, and potential Feynman architecture hints are all on the table.
Final Thoughts
$68.1 billion in quarterly revenue. $34.9 billion in free cash flow. Nearly $100 billion in forward supply commitments. Bernstein observed in their post-earnings note that the muted after-hours price reaction was "perhaps because investors were simply stunned by the sheer scale of the numbers."
That bewilderment is understandable. But it shouldn't substitute for judgment.
The quarter's core message is unambiguous: demand shows no sign of decelerating, the emergence of networking as a major franchise provides a new structural pillar for the growth narrative, the agentic AI inflection is injecting fresh acceleration into token economics, and $95.2 billion in supply commitments represent management's hardest possible forward vote on demand.
The fear that "the peak has passed" will never fully dissipate — nor should it. From the vantage point of our ongoing tracking of the AI infrastructure investment cycle (see our earlier analysis of macro-cycle risks), the current phase resembles the middle-leg acceleration of a long-distance race, not the final kick before the tape.
NVIDIA's greatest risk has never been insufficient demand.
It's expectations running too far ahead of reality.
GTC arrives in March. Next quarter's earnings follow in May. Together, they'll reveal whether the gap between expectations and reality is narrowing — or widening.
Earnings Call Recap
Open this more visual friendly version in a new tab/点击跳转查看原文,左上角切换中文
Financials
Open this more visual friendly version in a new tab/点击跳转查看原文,左上角切换中文
I will now share a more detailed version of this earnings behind the paywall:
This post is for subscribers only
Become a member to get access to all content
Subscribe now
Discussion in the ATmosphere