NVIDIA Q2 FY2027 Earnings — The Beat Compresses Because Supply, Not Demand, Sets the Number
The Call
NVIDIA is supply-capped, not demand-capped, so the guide is a supply forecast rather than a demand forecast — and a supply forecast is far more accurate, which means the beat has to compress even though demand is not the constraint.
| Metric | Our Estimate | Company Guidance | Consensus | Actual | Error | vs Cons. |
|---|---|---|---|---|---|---|
| Revenuenon-GAAP Consensus is the yfinance-sourced figure in web/src/data/earnings-pages/NVDA.json as of 2026-08-07; a 41-analyst compilation put it at 92,070 and Visible Alpha at 92,200. Our call sits 3.3% above the guide midpoint and 2.3% above consensus. Built bottom-up: 68,300 compute ex-China + 1,300 China + 17,800 networking + 6,600 non-DC. | 94,000 USD M | 91,000 (+/- 2%, band 89,180-92,820) USD M | 91,846 USD M | 96,221 USD M | -2.31% | 2.0× |
| Gross profitnon-GAAP 94,000 x 74.9%. | 70,406 USD M | — | not separately published USD M | 72,166 USD M | -2.44% | — |
| Gross marginnon-GAAP Inside the guided band. See the gross-margin input note for why the memory-cost hit is deferred rather than absent. | 74.9% | 75.0 (+/- 50bps)% | 75.0% | 75.0% | -0.13% | 0.0× |
| Operating profitnon-GAAP 70,406 gross profit less 8,400 non-GAAP opex. | 62,006 USD M | — | not separately published USD M | 63,956 USD M | -3.05% | — |
| Operating marginnon-GAAP 62,006 / 94,000 = 65.96%. | 66.0% | — | not separately published% | 66.5% | -0.70% | — |
| Net incomenon-GAAP Pretax 62,706 (62,006 operating + 700 other income) less 17.0% tax of 10,660. GAAP net income is NOT forecast here: it is dominated by mark-to-market gains on strategic stakes, which ran $12.7B above non-GAAP in Q1 alone. | 52,046 USD M | — | not separately published USD M | 53,954 USD M | -3.54% | — |
| Diluted EPSnon-GAAP 52,046 / 24,300M shares = $2.1418. Implies a +2.9% beat over consensus, against an 8-quarter mean EPS surprise of +5.73% and a last-4 mean of +4.61%. The compression IS the call: a naive application of the base rate gives $2.18, and the 4-cent gap is what the supply-cap argument buys. If the beat lands at the historical rate instead, this note is wrong in the direction of too conservative. | 2.14 USD | — | 2.08 USD | 2.22 USD | -3.60% | 1.8× |
- → The supply cap binds harder than assumed and NVDA simply delivers its guide. Revenue $91-92B, EPS $2.05-2.08, and the entire 'beat compresses but persists' framing collapses toward the guide rather than above it.
- → Licensed China shipments landed fully inside the July quarter at Needham's $3B run rate rather than the $1.3B carried here. Revenue ~$95.7B, EPS ~$2.20. This single line carries the widest error bar in the note.
- → Memory cost inflation reaches COGS one quarter earlier than the long-term-agreement lag implies. Gross margin 74.3%, EPS ~$2.09 on unchanged revenue, and the margin call is wrong in the direction the whole industry is watching.
- → The opex anchor fails the way AMD's did and non-GAAP opex lands near $8.7B. EPS ~$2.13 — small on its own, but it is the failure mode that has already cost one run its accuracy.
- → Networking decelerates to the compute rate instead of holding a premium. Networking $16.7B rather than $17.8B, revenue ~$92.9B, EPS ~$2.11.
- → Honest EPS span across the defensible readings above: $2.05 to $2.20. The call is $2.14. The span is not a hedge — it is where the single judgment about supply-versus-demand can put the answer.
BEHAVIOUR FORECAST, NOT ARITHMETIC — held to a lower confidence than every line above, and scored separately. The mechanical construction is the FQ2 guided sequential step (+11.5% off FQ1 actual) applied to our $94.0B FQ2 call, which gives $104.8B. We are deliberately publishing ABOVE that mechanical step, at $106.0B, for one reason: this method has now under-predicted management twice in an up-cycle — STX FQ1 FY2027 at -9.3% and AMD Q3 2026 at -3.1% — because a step built from recent history cannot anticipate management choosing to signal acceleration. Rubin volume lands in exactly this quarter (H2 CY2026) and the CoWoS budget steps up, which is precisely the setup in which management signals. Consensus for the guide is $103.8B, so our call is above the street as well as above the mechanical step. If this comes in low a third time, the correct conclusion is that the step method is biased low in up-cycles and should be replaced, not re-tuned.
Evidence (9)
Positioning Suggestions
→ Options are cheap into this print relative to history: implied ~5.4-5.6% against a trailing 12-quarter realized average of 7.4%. That favours owning convexity over directional exposure, and it is the clearest asymmetry available here.
→ The memory complex is the higher-conviction expression of this quarter's mechanism than NVDA itself. MU and SK hynix collect the cost inflation that NVDA absorbs for at least one more quarter; consider that the cleaner long.
→ Rack integrators (SMCI, DELL) sit on the wrong side of the same fact — a >15% BOM increase they cannot reprice as fast. Consider trimming into strength rather than adding on an NVDA beat.
→ Do not size for a guide-driven move. Four consecutive declines on beats means the Q3 guide, not the Q2 print, is the tradeable event, and it arrives on the call at 17:00 ET rather than in the 16:20 release.
Predictions
10Q2 FY2027 revenue comes in at or above $93.0B
If revenue clears $93.0B but the stock still closes lower the next day, the growth-rate repricing thesis is confirmed and NVDA multiple compression should be treated as structural rather than sentiment.
Q2 FY2027 non-GAAP diluted EPS comes in at or above $2.12
Q2 FY2027 non-GAAP gross margin comes in below 75.0%, the guide midpoint
If gross margin lands below 74.5%, the memory cost has arrived a quarter early — add to MU/SK hynix and reduce rack integrators.
The Q3 FY2027 revenue guide midpoint is at or above $104.0B
NVDA closes lower on the next trading day (2026-08-27) than on 2026-08-26
A fifth consecutive post-print decline on a fifth consecutive beat argues for owning the memory suppliers rather than NVDA for the AI-capex exposure.
The absolute next-day move is smaller than 7.4%, the trailing 12-quarter average realized move
Data Center networking revenue comes in at or above $17.0B
Management quantifies a China revenue contribution for the July quarter on the call or in the CFO commentary
Management repeats supply-constrained framing for Rubin or Vera Rubin on the call
A new buyback authorization is announced alongside the print
Key Questions
- Is the constraint packages or demand? If Jensen repeats the 'supply-constrained through the entire life of Vera Rubin' framing from the Q1 call, the guide is a supply schedule and the beat should keep compressing every quarter until CoWoS steps up in 2027.
- When does memory cost actually reach COGS? The >15% price hike is priced for early-CY2027 shipments, so there is a window in which NVDA absorbs cost it has not yet repriced. Management naming that window is worth more than the FQ2 margin itself.
- Did licensed China shipments land inside the July quarter or slip to the October quarter? The Q1 guide carried China Data Center compute at zero, so anything disclosed is incremental and the timing decides roughly $1.7B of revenue.
- Does networking keep its growth premium over compute? Networking grew 35% against compute's 18% in Q1. If that premium holds, the value per rack is rising faster than the rack count and the ANET/MRVL competitive read changes.
- Does a fifth consecutive post-print decline happen on a fifth consecutive beat? Eight beats, six next-day declines, four in a row. At some point that stops being sentiment and starts being the market repricing the growth rate.
Monitoring Checklist
Post-Event Results
Actual Outcomes
- • Revenue 96,221 non-GAAP, +5.74% over the 91,000 guide midpoint and +4.76% over consensus. Non-GAAP diluted EPS 2.22 against 2.08 consensus. Gross margin landed exactly on the 75.0% guide for the third consecutive quarter.
- • The main line of the frozen forecast was wrong. It argued the beat would COMPRESS because a supply-capped guide is a supply forecast and therefore accurate. The beat EXPANDED, from +4.63% over guide in Q1 to +5.74% in Q2 — the second-largest of the last eight quarters.
- • NVIDIA guided a full fiscal year for the first time in its history: FY2028 revenue growth of approximately 70%, explicitly described as a supply-constrained outlook, against customer forecasts management said point to growth roughly doubling.
- • Management pre-announced a gross-margin trough: 74.0% guided for Q3, 71-72% in Q4, recovering to 72-73% across FY2028 as executed price increases take effect. The memory cost the frozen forecast said would arrive in FQ3/FQ4 is now confirmed by the issuer to arrive in exactly that window.
- • Q3 FY2027 revenue guided to 108,000 +/- 2%, above the 106,000 called here and well above the 103,800 consensus.
- • NVDA discontinued the Data Center compute/networking revenue split and replaced it with Hyperscale 48,710 (+13% QoQ) and ACIE 40,313 (+25% QoQ). One forecast line and one prediction card became permanently unscoreable as a result.
- • China is immaterial and dilutive: Hopper 200 shipments were under 1% of Data Center revenue, and management said current Hopper shipments dilute corporate margin.
- • Cash quality deteriorated sharply: operating cash flow fell to 24,077 from 50,300 despite higher earnings, as working capital consumed 30,707 — receivables alone took 22,346 and DSO went from 45 days to 60. Inventory rose to 31,600 from 25,800. The company issued 25,000 of senior unsecured notes.
- • Supply and capacity commitments rose from 119,000 to 279,000, described as primarily memory. Guarantees reached 108,500, of which 105,000 is credit support for SB Energy PORTS-Pike serving OpenAI's 20-year leases from FY2029.
- • Management addressed the financing question directly on the call: AI labs supported by NVIDIA's balance sheet are expected to be roughly a quarter of next year's business, and 'we know some will call this circular financing. We see it differently.'
Market Reaction
Thesis Updates Needed
Trade Recommendations
→ VOID — the gross-margin conditional did not fire. It required non-GAAP gross margin below 74.5% to add to MU/SK hynix and reduce rack integrators. Margin landed at 75.0%, exactly on guide, so the memory cost did not arrive a quarter early. Management has now told us when it does arrive: Q4 at a 71-72% trough. The trade is not wrong, it is early, and it now has a dated trigger instead of a guessed one.
→ RESOLVED against us. Both next-day conditionals failed on the 2026-08-27 close of 227.98 against 209.66, up 8.74%. There was no fifth consecutive post-print decline and the move was not contained: 8.74% realized against the 7.4% trailing 12-quarter average, so the convexity was cheap after all and we were on the wrong side of it. The reading that a supply-capped guide would compress the reaction was wrong at the day scale as well as the quarter scale.
→ The convexity suggestion was the right read on setup and is now spent. Implied ~5.4-5.6% against a trailing 12-quarter realized average of 7.4% was cheap into the print; that asymmetry does not survive the event.
→ No new NVDA position action is recommended from this print, because the target could not be settled. The re-rate cleared its arithmetic but not its convergence gate, and taking a directional action off a number two independent reviewers rejected would be worse than waiting one day.