External Publication

NVIDIA 1Q FY27 Earnings Review: Reality Stepped Up, Visibility Stretched Forward

Jason with his AI analysts May 26, 2026

NVIDIA 1Q FY27 Earnings Review: $81B Is the Headline, $1T Is the Story

Three months ago, at the close of my 4Q FY26 review "$68B Was the Headline, $95B Is the Story" , I left a thought: NVIDIA's real risk has never been insufficient demand. It's expectations running ahead of reality. GTC March together with the 1Q FY27 print would tell us whether that gap was closing or widening.

The answer this call delivered in late May 2026 turned out to be two things happening at once. Reality stepped up — Revenue $81.6bn beat consensus by 3.07%, Non-GAAP EPS $1.87 beat by 5.06%, Q2 guidance of $91.0bn implies +94.7% YoY re-acceleration, FCF $48.6bn beat by 21.5%, and operating margin lifted to 65.9%. The numbers themselves are clean enough that you have to look hard to find a flaw. At the same time, Jensen extended the runway by another stretch: Blackwell + Rubin forward visibility, framed at GTC as $500bn through CY25-CY26, was restated as $1tn through CY25-CY27, with a hint that CY26 alone has overtaken the original two-year envelope; a new $200bn Vera CPU TAM was formally added to the product matrix; and the reporting framework was rebuilt into a Hyperscale + ACIE + Edge three-segment structure that better matches the actual shape of the business.

Tucked alongside it: a fresh $80bn buyback authorization and a 25x quarterly dividend hike, with full-year cash return committed at ~50% of FCF — a direct answer to the "NVDA returns too little capital" critique that's been building over the past year.

The core question this quarter is no longer "is NVIDIA still delivering". The question is whether the valuation frameworks investors are using can keep up with the simultaneous extension of both reality and visibility.

The Numbers Themselves: A Structurally Clean Beat

Let me walk the account first, because every piece of framing later anchors on these facts.

Revenue of $81.6bn. Under the new reporting structure, the Data Center platform contributed $75.2bn (YoY +92.3%, QoQ +20.7%), and Edge Computing contributed $6.4bn (YoY +29.3%). Inside Data Center, the two new sub-segments — Hyperscale and ACIE (AI Cloud + Industrial + Enterprise) — produced $37.9bn and $37.4bn respectively, almost a clean 50/50. Networking, still broken out under the old framework, ran $14.8bn (YoY +198.6%, beating consensus by 16.14%). That's the strongest outlier of the quarter and the main fuel behind Q2 guidance's implied +94.7% YoY re-acceleration.

GAAP P&L Sankey: Hyperscale + ACIE + Edge → Net Income $58.3B

Source: Author's own chart based on NVIDIA Q1 FY27 reported financials. The purple node represents the $15.9bn equity securities fair-value gain — already excluded from Non-GAAP.

Non-GAAP gross margin of 75.0% matched guidance to the decimal and ran 0.07pts off consensus. Worth noting because management had previously described 75% as the implied gross margin ceiling — this quarter ran right against that ceiling. Q2 guidance is again 75.0% (+/- 50bps), and full-year commentary is "stay in the mid-70s". Against persistent price increases for HBM and CoWoS, the print itself answers one strand of the sell-side preview debate — namely whether FY27 H2 GM might slide below 75%.

Non-GAAP operating income of $53.8bn , YoY +146.7%, with operating margin lifting to 65.9%. This is the cleanest line of the quarter: revenue +85.2% YoY, OpEx +49.2% YoY, and operating leverage carries OI into triple-digit growth. None of this relies on accounting magic — it's what scale and pricing power look like when they show up together.

Operating leverage visualized: OpEx +51% << Rev +85% << OI +147%

Source: Author's own chart based on NVIDIA Q1 FY27 reported financials. The orange dashed line is the Revenue YoY 85% baseline.

Non-GAAP diluted EPS of $1.87 , beating consensus by 5.06%. One methodological note: starting FY27, NVIDIA no longer excludes SBC from Non-GAAP, and historical data has been restated under the new convention. So this is the first "complete" SBC-inclusive quarter, and a few sell-side models hadn't fully switched conventions — which means the apparent beat magnitude varies by which model you read. The cleaner like-for-like comparison uses ex-SBC EPS of ~$1.79, which still beats consensus by 4.5%. Folding SBC into Non-GAAP is an improvement in accounting transparency, but it introduces a small comparability noise in the short term.

GAAP EPS of $2.39 looks much higher than Non-GAAP — and the $0.52 gap isn't operating in nature. It comes from NVIDIA's ecosystem investments in CoreWeave, OpenAI, Anthropic, Wayve, and Mistral, with $15.9bn of mark-to-market fair-value uplift this quarter. The valuation anchor needs to be Non-GAAP. One related fact worth noting: NVIDIA's marketable + non-marketable securities pool ballooned from $35.1bn at the start of the quarter to $73.6bn, a $38.5bn QoQ jump — of which $18.6bn represents new non-public investment outlay this quarter. Two competing narratives will pull at this figure; I've already laid out the facts-and-logic supporting each in "From Selling Engines to Building Cars" — won't rehearse them here, just recording the direction.

Free cash flow of $48.6bn , with FCF / Non-GAAP NI conversion at 107%, and capex at just 2.15% of revenue. This is the structural difference between NVIDIA and all its hyperscaler customers: their capex intensity runs 70%+, while NVIDIA stays asset-light, so the same revenue scale converts into far more free cash flow. That fact is the physical basis for the capital-return step-function I'll get to below.

That's the accounting story. It's neat, has almost no fine print, and every segment beat/miss is within noise level. If the call had ended here, this would be a "very good" quarter — but not one that produces much incremental insight. What Jensen and Colette Kress did over the next 90 minutes of prepared remarks and Q&A was expand the call from "quarterly results" into something closer to a formal disclosure upgrade about where the company's actual business boundary now sits.

The Real News: Three Disclosure Upgrades

If you string together everything Jensen and Colette said about future scale and business structure during prepared remarks and Q&A, this quarter delivered three simultaneous disclosure upgrades. They're better understood not as redrawing the map, but as formally accounting for business that was already running across the past several quarters — under a more accurate reporting structure than the company had used before.

Upgrade #1: Reporting Framework Rewrite

Starting this quarter, NVIDIA replaced the old five-segment structure (Data Center / Gaming / Pro Viz / Auto / OEM) with two market platforms: Data Center (subdivided into Hyperscale and ACIE) and Edge Computing (PC, workstation, game console, robotics, automotive, AI-RAN, and so on).

When Joseph Moore asked about the philosophy behind the change, Jensen gave an answer worth reading slowly. He framed the reason as letting the market "better understand our business", noting that AI and computing themselves carry diversity along four axes: content (language models; 3D graphics for robotics; proteins/materials for life sciences; physics for science), use case (enterprise, energy, manufacturing), deployment location (hyperscale cloud / AI native cloud / enterprise on-prem / sovereign / factory / edge), and governance (public cloud / regulated cloud / confidential computing / national security).

In Q&A he condensed Data Center's "simplest factoring" into three segments: hyperscale (only 5-6 players; the simplest go-to-market), AI natives + enterprise + industrial + sovereign (corresponding to roughly 250,000 enterprises globally), and robotic edge (NVIDIA today the practically sole supplier). All three carry distinct stacks, operating systems, and go-to-market models. ACIE in the new framework is the formal naming of "segment two".

New framework revenue mix: Hyperscale and ACIE roughly even, Edge at ~8%

Source: Author's own chart based on NVIDIA Q1 FY27 reclassified reporting. ACIE = AI Cloud + Industrial + Enterprise.

ACIE ran $37.4bn this quarter, YoY +74%, QoQ +31% — already in the same league as Hyperscale. Inside that line: AI cloud (neoclouds like CoreWeave, Crusoe, Lambda) YoY >3x, sovereign revenue YoY +80%, partner data centers above 10MW "roughly doubled in one year" to 80+ sites, and NVIDIA AI infrastructure now deployed across ~40 countries covering $50tn of GDP.

The most important read-through: one of the most common investor concerns over the past few years has been NVIDIA's heavy reliance on 4-5 hyperscalers. The top-3 direct customers this quarter still combined for 54% of revenue (21% / 17% / 16%), so customer concentration hasn't dropped meaningfully on the headline metric. But ACIE at nearly 50% of DC and growing 74% YoY means customer diversification is progressing faster and more structurally than the market had assumed. That's hard evidence of de-risking, not narrative.

A bit of honest layering is required. The AI cloud (neoclouds) inside ACIE remains, to some degree, derived demand from hyperscalers — CoreWeave and its peers route significant compute to Anthropic and OpenAI, who are themselves tightly coupled to hyperscalers. But even setting that overlap aside, the sovereign + enterprise on-prem + industrial on-prem layer inside ACIE shows meaningful YoY growth — and is a slice the company couldn't cleanly quantify at the 4Q26 review. Bringing it onto the formal reporting menu is itself a useful disclosure.

Upgrade #2: $500bn → $1tn Blackwell + Rubin Visibility

At GTC, the forward visibility for Blackwell + Rubin was $500bn covering CY25-CY26. This quarter, Colette restated it: "We have full confidence in the $1 trillion in Blackwell and Rubin revenue we foresee from 2025 through calendar 2027" — the time window extended by a year, the total doubled. She added that CY26 revenue is expected to grow sequentially each quarter and is on track to exceed the prior $500bn Blackwell + Rubin revenue opportunity.

The two sentences need to be parsed separately. Extending the window is partly an accounting maneuver — doubling $500bn relies on adding a year. But the "CY26 alone will exceed the prior $500bn revenue opportunity" comment is more substantive. It effectively says a single CY26 has now absorbed the original two-year $500bn envelope. Read carefully, that implies CY27 itself can be nudged higher.

Barclays quantified the gap in its post-print review: this $1tn figure sits $40bn above the cumulative Data Center revenue implied by Street models. Goldman Sachs added that the $1tn framework does not include contributions from Vera CPU, LPX, or Rubin-CPX — all of which are pure incremental.

The reason management can put this number on the table so directly hinges on one background fact: NVIDIA's total supply this quarter (inventory + purchase commitments + prepayments) climbed to $145bn, up from $95bn at Q4 FY26 — a $50bn QoQ rise, +53%. Per Morgan Stanley's read, $95bn of the $145bn is "to be consumed in the next three quarters of production". The implication: $1tn isn't a castle in the air. The first 18-24 months of it are already locked in through supply commitments. That's the most credible "reality anchor" of the quarter.

Upgrade #3: $200bn New CPU TAM

Vera CPU is positioned as "the world's first CPU purpose-built for agentic AI". What makes the announcement distinctive is the two numbers management gave alongside it: CY2026 standalone CPU revenue visibility of $20bn, and a 2030 CPU TAM of $200bn — a market NVIDIA has historically not addressed.

Jensen's framing in response to Vivek Arya is worth keeping intact: the CPU economics of the past were "dollars per core". In the agent era, the economics shift to "tokens per dollar / dollar per token" — agents don't rent cores; they need work done fast. Vera is purpose-built for this "agentic CPU" economics. He elaborated with a use-case picture: the world today has roughly a billion human users; tomorrow it will have billions of agents, each using tools — much as humans use PCs today. Every agent will spin off sub-agents, and every spin-off triggers inference. As he put it: "all of the thinking happens on GPUs, all of the orchestration runs on CPUs".

Sell-side opinion on $20bn's plausibility splits — but not evenly. J.P. Morgan decomposes it as ASP $5-8k per CPU, implying 2.5-4M CPU units and 6-8% server CPU unit share — "reasonably achievable". BofA Global Research accepts $200bn TAM (revised up from the prior $125bn) and notes that memory/wafer pre-bookings are in place on the demand side. Citi folds $20bn into its 2026 sales model and flags the 2030 TAM at $200bn as "separate from and not double-counted against the prior $1tn framework". Morgan Stanley is neutral, awaiting June 1 Computex for more detail.

The one publicly aired strong pushback comes from Barclays : the $20bn target is "difficult to justify through back-of-envelope math; our calculation is closer to a double-digit billion range". That's the most distinctive cautionary signal of this sell-side review cycle and shouldn't be dismissed — but with a 4 vs 1 vs 1 sell-side distribution, $20bn is broadly accepted as plausible, with Barclays the lone distinctive outlier.

Jensen offered his own defense in Q&A: $20bn refers specifically to standalone CPU, separated from the Vera CPU that ships as the companion CPU inside Vera Rubin systems. Vera has four use cases: (1) inside Vera Rubin at a 2 Rubin GPUs to 1 Vera CPU bundling ratio; (2) standalone CPU; (3) Vera + CX9 with the storage software stack; (4) Vera + CX9 with security / confidential computing. It's a product matrix with internal structure, not a single SKU chasing a single number.

Taken together, the three disclosure upgrades don't conjure new demand out of thin air — the demand was already there. They take pieces of the business that had been scattered across line items, or weren't being clearly seen by the market, and bring them onto the formal reporting menu. From an analyst-modeling angle, this means forward estimates and valuation frameworks both need recalibration — not that NVDA's fundamentals have changed.

The Capital Return Step-Function: Lifting Cash Returns to ~50% of FCF

If TAM expansion has no credible cash-conversion path, it's just a forecast. Jensen and Colette evidently know this.

On May 18, 2026, the Board approved a new $80bn buyback authorization. Combined with the $38.5bn remaining at Q1 end, total unexecuted authorization stands at $118.5bn. The quarterly cash dividend was raised from $0.01 to $0.25 — a 25x step-up. Capital return for the quarter alone was $19.6bn (buybacks $19.3bn + dividend $0.24bn), YoY +36.4%, a single-quarter company record. Full-year 2026 commitment is to return ~50% of free cash flow — roughly $97bn annualized based on Q1 FCF.

Capital return step-function visualized

Source: Author's own chart based on NVIDIA Q1 FY27 financials and company announcements.

This policy directly answers the critique BofA Global Research had laid out in its preview: across CY22-25, NVIDIA returned only 47% of FCF — well below the 80% peer benchmark. BofA's 1Q27 review explicitly flagged the $0.25 quarterly dividend as "1 quarter earlier than expected" and framed the cumulative buyback authorization as "approximately $120bn" — a meaningful upward revision, clearly running ahead of their own prior expectations.

A subtler implication is worth one extra sentence. When a company is simultaneously expanding OpEx (full-year OpEx growth guidance moved from "low 40s" to "upper 40s"), pushing non-public equity investments by $18.6bn QoQ, and lifting capital return to 50% of FCF — it is saying three things in parallel: it has enough cash generation to do all three at once; it has full confidence in long-term FCF durability; and it is deliberately using share repurchase to offset whatever valuation discount the "circular financing" perception is creating.

That's a precise piece of capital-allocation triangulation. It doesn't require the market to believe any specific TAM number — as long as the cash-return reality stays credible, the valuation floor is well anchored.

A Few Open Questions Worth Tracking

A bit of fine print is needed — there isn't much, but three open questions are worth watching over the next 1-2 quarters.

First, the most visible non-numerical surprise this quarter was full-year OpEx growth shifting from "low 40s" to "upper 40s". Q2 Non-GAAP OpEx guidance of $8.3bn runs 8.18% above consensus $7.67bn. My read splits into two layers. One, operating leverage remains strikingly intact — Q1 Non-GAAP OI YoY +147% against OpEx YoY +49% — so the OpEx step-up does not dilute earnings quality. Two, the acceleration tells you that Vera Rubin production prep, Vera CPU, BlueField-4 STX, NVIDIA Dynamo, NemoClaw, OpenShell, and Agent Toolkit are all ramping in parallel — management is actively choosing platform completeness over near-term operating efficiency. That's a reasonable trade, but it does trim near-term EPS upside elasticity somewhat.

Q2 FY27 guidance tracker: Revenue +4.2% beat / OpEx +8.2% miss

Source: Author's own chart based on NVIDIA Q2 FY27 guidance and sell-side consensus. Purple = guidance range; blue circles = Q1 actual; grey diamonds = consensus.

Second, Rubin production cadence carries a light tension. Colette, in response to Joshua Buchalter, confirmed that the company delivered the first Vera Rubin samples to customers this quarter , with Q3 starting the initial phase and Q4 entering sustained ramp. But Barclays' Asia supply-chain checks read a signal that "Rubin may slip to later this year on shipment timing", while J.P. Morgan's global team's supply-chain checks "still show Vera Rubin ramping strongly in late CY2026 and early CY2027" — JPM and Barclays disagree directionally on the supply-chain read. My base case treats H2 Rubin ramp as the working assumption, but supply-chain signals heading into Q3 earnings will be the key incremental data point.

Third, China DC compute remains a "could go either way". The U.S. government has approved H200 export licenses to Chinese customers, but NVIDIA reports "zero revenue to date" and "uncertainty whether any imports will be allowed into the country". Q2 guidance continues to assume $0 China DC compute contribution — meaning any actual ramp would be pure upside rather than already baked in. Citi notes that "any commentary on H200 delivery timing or China re-entry will be a major variable for the stock" — making this the next standalone catalyst.

DC sub-segment beat/miss: Networking +16% outlier, Compute -1.1% miss

Source: Author's own chart based on NVIDIA Q1 FY27 reported financials and old-framework sell-side consensus.

(The -1.1% Compute miss is the only sub-segment miss of the quarter under the old framework, driven by below-consensus China DC compute contribution, mix-shift crowd-out from Networking's rising share, and partial reclassification of compute revenue under system-level sales — three plausible explanations, none of which point to softening core demand.)

How the Story Goes from Here

Back to the "verdict quarter" question I posed at the open. Did the gap between expectations and reality narrow or widen in 1Q FY27?

Near-term (next 2-3 quarters): reality is closing in on expectations, while visibility has been pushed forward another stretch. The $1tn Blackwell + Rubin window (CY25-CY27), the $200bn Vera CPU TAM, the $80bn new buyback authorization, the 25x dividend step-up, the $145bn total supply commitments — these are all hard signals arriving together and reinforcing one another. With Q2 guidance implying +94.7% YoY re-acceleration, the thesis won't be broken inside Q2.

Medium-term (next 4-8 quarters) : the falsification conditions are reasonably clear — whether AI infrastructure annual spend can reach the $3-4tn range by 2030 (Colette reiterated; BofA already raised AI TAM to $3tn+); whether Vera Rubin will be supply-constrained across its full life cycle (Jensen's explicit call); whether Vera CPU can grow from standalone $20bn to the $200bn 2030 TAM (the lone Barclays pushback); and whether physical AI can scale from today's LTM $9bn into the next growth engine. Of the four, I lean toward the first two being base-case true — the token-economics and capex-sustainability arguments are developed in "Multi-Trillion AI Capex Bull Case" and "Can a GPU Really Be Profitable for Six Years?". The latter two remain open, but with 4 of 6 sell-side desks accepting the Vera CPU number and cumulative $1tn already supply-locked, the base-case probability distribution leans positive.

Longer-term : Colette gave a framing I expect to return to repeatedly — "Customers do not buy GPUs, they build AI factories." The right economic yardstick isn't the GPU purchase price; it's the full-life-cycle cost of producing intelligence inside an AI factory: token per watt, tokens per dollar, uptime, utilization, time to production, software durability, asset life. If customers make their decisions with this scorecard, NVIDIA's platform durability gains multi-cycle stability. Morgan Stanley distilled this thesis well: "Lower-cost chips have not been proven to deliver a lower cost-per-token, and best-in-class capability is an enormous advantage for customers pursuing the longest asset life". Not a final answer — but the evidence today leans in support.

Closing

I closed the 4Q FY26 review with that thought — expectations running faster than reality. The answer this late-May 2026 call delivered isn't reality catching up to expectations, and isn't expectations stepping down. It's reality and visibility moving forward together — on one side, a clean beat with record operating leverage, FCF margin of 59.5%, capex at 2.15% of revenue, fundamentals tidy enough that they don't need any narrative dressing; on the other, management lifting forward visibility from $500bn to $1tn, formally adding a $200bn Vera CPU TAM to disclosure, and anchoring cash return at 50% of FCF.

The real impact of this "simultaneous forward move" on investors isn't in the fundamental thesis — that thesis remains very solid. It's in the valuation framework itself. Bringing ACIE, Edge, and Vera CPU — previously scattered across line items — onto formal consolidated disclosure means the SOTP and TAM-anchored DCF templates that sell-side and buy-side are running both need rebuilding. This will take a quarter or two to absorb — but it shouldn't be read as the company actively manipulating perception. It looks more like a business whose structure has outgrown its old reporting framework.

The next standalone catalyst is the June 1 Computex / GTC Taipei — Vera CPU details, further Rubin ramp signals, more concrete physical AI and robotics partnership disclosures. That will tell us how much of this disclosure upgrade has already been walked on the ground.

My current base case: the ground actually walked will outrun what's been put on paper.

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

I will now share a more detailed analysis on this earnings result behind the paywall:

Upgrade your subscribtion?

This post is for subscribers only

Become a member to get access to all content

Subscribe now