Data as of 2026-09-14 (last US trading session: close of 2026-09-11).
After the close on 26 August, NVIDIA (NVDA) reported its second fiscal quarter of FY27 (quarter ended 2026-07-26): revenue of $96,221M, +106% y/y and +18% q/q; the Data Center segment at $89.0B, +117% y/y and +18% q/q. Against the sell-side consensus available before the print (revenue of roughly $92.2B on the AP/TIKR compilation basis; Data Center of $83-85B; non-GAAP EPS of $2.09-2.10), total revenue beat by about 4.4%, while Data Center ($89B against consensus of $83-85B) and the next-quarter guide beat by more: actual non-GAAP EPS came in at $2.22 and GAAP EPS at $2.46, and the next-quarter guide of $108.0B ± 2% sits above the pre-print compilation of $103-105B, with the explicit statement that it assumes no data center compute revenue from China. After hours the stock first fell about 1.3% on the gross margin warning, then rebounded 3.5-4.7% after the call to close at roughly $217 (the readings from WSJ, Yahoo and Kiplinger differ slightly). Over the following two weeks it worked its way back toward the high set in May: the 8 September close was $225.73, about 4.5% below that high (two secondary accounts put the high at $236.26 and $236.54, so the basis differs slightly); in the second week of September it fell 4.45% with the sector, closing at $218.29 on 11 September.
Our core judgement is that directional validation on the demand side is complete. Revenue doubling year over year, a customer mix split roughly in half (hyperscale at about 50%; AI clouds, industry enterprises and sovereign customers together about 50%), and the next-generation Vera Rubin platform in full production with the first wave of deployment customers named publicly — this growth line is not where the disagreement lies at present. The real disagreement is the shape of the risk: this cycle's soft spot has already moved from "will demand be falsified" to "how does the cost land on the statements." Q3 gross margin guidance is 74.0% ± 50bps, down from 75.0%, attributed to memory price increases. Supply commitments have risen to $279B and the ceiling on guarantee obligations is $108.5B (composition broken out below; 97% sits with a single project). Days sales outstanding has stretched from 45 days to 60 days, and the top five customers account for about 70% of receivables (disclosed directly in the 10-Q). A 26x forward P/E does not look expensive, conditional on delivery of the roughly 70% FY2028 growth outlook (company total-revenue basis, on a supply-constrained premise). That premise is itself the object of the quarter-by-quarter reconciliation that begins with the Q3 print in late November.
Revenue of $96.2B and Data Center of $89B: the demand radius widens from the big four clouds to sovereigns and enterprises
Start by laying out the demand-side evidence in full. Within Data Center's $89.0B, hyperscale accounts for about half, with the other half contributed by AI clouds, industry enterprises and sovereign customers combined (CFO basis). The sovereign and enterprise end produced named cases this quarter: Korea's national-level AI infrastructure with SK Telecom, NAVER and Brookfield; the Japanese government's national AI infrastructure; 35 AI HPC supercomputers in Europe; and Apple Private Cloud Compute adopting NVIDIA GPUs for confidential inference. The Vera Rubin platform has entered full production, with the press release naming CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud and Nebius as the first wave of deployers. On the inference side, the company reached a non-exclusive technology licence with Groq (the 10-Q cash flow statement shows roughly $2.944B of licence payments, with the note explicitly describing a non-exclusive license agreement), and the NVIDIA Groq 3 LPX inference accelerator entered full production alongside it — the in-house device stack is now complete, but the nature of the arrangement is a licence, not an acquisition. Demand is still being added to: on the day of the print, NVIDIA and AWS jointly announced a plan to deploy an additional 2 million GPUs across 2027-2028, spanning Blackwell Ultra, Rubin and Rubin Ultra (joint NVIDIA/AWS press release, 2026-08-26, also announced on the call). On the call, the CFO said Vera Rubin's first full shipping quarter (Q3) should account for about 20% of Data Center revenue, implying something on the order of $20B. Product economics are improving in the same direction: on the company's basis, revenue per gigawatt rises from roughly $18B in the Hopper generation to roughly $40B for Vera Rubin. Management gave an outlook on the call for roughly 70% growth in company total revenue in FY2028 (CFO verbatim: We expect to grow revenue by approximately 70% in fiscal 2028), and said supply is constrained with demand far exceeding supply.
What deserves attention is the evidentiary implication of the change in demand structure. The last bear script for "AI capex has peaked" depended on a single large customer cutting orders in order to falsify demand; once the customer radius widens to sovereigns, enterprises and AI clouds, that script becomes markedly harder to trigger. Alongside it sits an AI infrastructure financing consortium intended to mobilise over $500B (Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR; the official wording is mobilised over time, with final agreements still pending) — while three-year visibility on the capex pipeline rests on the FY2028 outlook and the $279B of supply commitments. One change in disclosure basis is worth recording in passing: from this fiscal year, Gaming, Automotive and Professional Visualization are no longer reported as separate lines and have been folded into two segments, Edge Computing ($7.2B, +27% y/y) and Data Center (the Q1 FY27 comparative column in the Q2 FY27 press release is already presented on the new basis), and networking revenue within Data Center is no longer broken out separately either. For an analyst this is bad news: segment comparability declines, and from here it can only be tracked indirectly via the call. One further point to hold in reserve: demand visibility and the vendor's own degree of participation in financing are two sides of the same coin, and the bear section returns to it.
Gross margin: the one hard negative signal, and the path is already written into the guidance
Q2 gross margin (GAAP) was 75.0%, up 2.6 percentage points y/y and 0.1 point q/q, still at a high level. The turn is in the guidance: Q3 steps down to 74.0% ± 50bps, with the CFO explicitly attributing it to memory price increases and rising component costs, and flagging that later quarters could move lower still. Management has itself laid out the path on the call: a trough around the fourth fiscal quarter of FY27 (roughly 71-72%), returning to the 72-73% range for the full year FY2028 (both CFO verbatim on the call, checkable in the public transcript). The supporting actions point the same way: a multi-year next-generation memory agreement signed with SK hynix, a net inventory build of $5.784B in the quarter (cash flow statement basis), and total inventory rising from $25.8B to $31.6B in preparation for Rubin production. That clarifies where the supply bottleneck sits: the rate limiter is HBM and CoWoS packaging capacity, not the GPU die itself.
Our judgement is that this looks more like deliberately trading cost for growth; weakening demand would show up first in separate leading markers (order and deployment cadence), and those did not appear this quarter. But gross margin directly determines the elasticity of EPS — the FY2028 EPS consensus has been revised up from $12.63 to $15.46 within 90 days (24/7 Wall St compilation), and if the gross margin assumption embedded in that moves below the 72% lower bound, the upward revisions rewind. Make the verification hook specific: where Q3 gross margin lands against 74.0% ± 50bps, and whether the Q4 guide digs the trough deeper.
The shape of the statements changes: $279B of supply commitments, $108.5B of guarantees, DSO from 45 to 60
The strongest set of bear evidence is not in the income statement but in the balance sheet and the commitments. Supply purchase commitments have risen to $279B, primarily memory volume locked in for Vera Rubin. The ceiling on guarantee obligations totals $108.5B, and the composition has to be taken apart: $105B is credit-support guarantees for SB Energy (on behalf of an OpenAI-related party, the 4.25GW PORTS campus in Ohio, with the first lease term beginning in FY2029), while the ceiling on land, power and facility guarantees for a number of AI cloud partners is only $3.5B — roughly 97% of the exposure sits with a single project and a single end customer. Days sales outstanding has stretched from 45 days to 60 days, and the top five customers account for about 70% of receivables (all of the above disclosed directly in the 10-Q). Viewed from this angle, the $500B financing consortium admits a second reading: the vendor is participating deeply in financing arrangements for its downstream customers, and part of demand's visibility is propped up by capital from the supply side. That supplies fuel to the "circular financing" objection, regardless of whether the objection holds at present. Cash flow itself remains strong: free cash flow of $21.34B in the quarter (24/7 Wall St basis), $26.0B of shareholder returns in the quarter, and $99.0B of remaining repurchase authorisation.
In our view, the correct reading of this set of disclosures is that NVIDIA is evolving from an asset-light chip supplier into a systems vendor deeply involved in customer financing and supply lock-in. On the way up, inventory, long-term contracts and guarantees are the moat, locking away capacity that competitors cannot get; at an inflection point, that same set of line items becomes the source of write-downs and losses — and on the guarantee side the load is highly concentrated, with $105B facing the same campus and the same end customer. As things stand it does not rise to a near-term risk signal, and the lengthening DSO also has a reasonable explanation in large-customer concentration and system-level delivery payment terms. What it changes is how the statements take the load if demand does decelerate. The qualitative call needs continuity: if Q3 DSO and inventory keep moving higher together while revenue growth starts to roll over, this section has to be upgraded from "shape observation" to a first-order variable.
China: revenue exposure near zero, inventory still bleeding, relaxation is an option with friction
H200 shipments into China in Q2 were less than 1% of Data Center revenue, and the Q3 guide contains no China data center compute revenue at all, consistent with the wording of the prior two quarters. The direction on the charge side has to be stated precisely: in the first fiscal quarter of FY26 there was a one-time $4.5B charge against H20; in the first half of this fiscal year a further roughly $0.4B was newly provisioned on soft H200 demand ($985M charged and $177M released in the quarter; $2.1B charged and $280M released in the first half, 10-Q basis) — China-related inventory is still generating new losses, but revenue exposure is already close to zero. The forward growth engine is now carried entirely by non-China demand, and transmission from China-related bad news to the fundamentals is close to nil: further tightening of controls is mainly a sentiment disturbance, while relaxation is an upside option, though a restart carries real friction — the 10-Q notes that even if the US side issues licences, purchase restrictions on the PRC side, a 25% import tariff (which cannot be passed through) and in-country US inspection procedures are all policy and cost obstacles, so it should not be read as pure upside. On the networking side, one industry variable is worth recording separately: the Spectrum-6 switch supports both traditional pluggable optical modules and co-packaged optics (CPO), which is a coexistence signal for both the optical module and CPO supply chains, and the substitution narrative does not hold, at least for this product generation. For share allocation among upstream components, the public criterion is the basis on which CPO's share of racks is disclosed in subsequent reporting.
Pricing: the "cheapness" of a 26x forward P/E is the price of a cyclical profile
As of 8 September ($225.73), the stock was up about 21% year to date in 2026 (as relayed by biggo); it then pulled back with the sector over the following week, closing at $218.29 on 11 September, roughly +17% year to date against the same start-of-year base. On multiples, the forward P/E is about 26x and PEG about 0.59 (24/7 Wall St, 2026-09-10, on a $223.67 share price basis). Set against AI software names that routinely trade at a hundred times forward earnings, the apparent cheapness of 26x reflects the market pricing this as a heavily cyclical stock, with peak-earnings risk discounted straight into the multiple. Management appeared at the Goldman Sachs Communacopia conference on 10 September, and the four things the market focused on (the Vera Rubin ramp, the CPU business, the memory bottleneck, and China controls) overlap closely with this note's verification hooks. That is not a coincidence; the contested points in the current price are concentrated in exactly those places.
We would put the real bull-bear divergence on the delivery path for the roughly 70% FY2028 growth outlook (company total-revenue basis) rather than on today's multiple. Three observable variables sit on that path: memory cost (which sets the lower bound on gross margin), the pace at which HBM and CoWoS supply eases (which sets the ceiling on revenue), and the diversion of incremental share to large customers' in-house silicon. On the third, the public evidence today supports complementarity rather than substitution: AWS is advancing Trainium while also announcing a plan to deploy 2 million NVIDIA GPUs. But this is a judgement that needs re-reading quarter by quarter, and it is too early to settle. Self-reported data gap: this note did not obtain a separate breakout of networking revenue within Data Center for Q2 (folded into the broader segment basis from this fiscal year), so that thread can only be tracked indirectly through the call and supply chain disclosures.
Verification calendar
| Timing | Event | How to read it |
|---|---|---|
| Late November | Q3 FY27 results | Revenue against $108B ± 2%; gross margin against 74.0% ± 50bps; the basis given for Data Center networking-related growth; whether DSO and inventory keep moving higher together |
| Same event | Q4 FY27 guidance | Whether gross margin is guided to a deeper trough below 74%; whether the roughly 70% FY2028 outlook (total-revenue basis) is reiterated |
| Ongoing | HBM/memory contract prices | The size of memory price increases determines the gross margin path and the credibility of upward FY2028 EPS revisions |
| Event-driven | China export controls | Relaxation = an upside option with friction (purchase restrictions/tariff/inspection procedures); tightening = a sentiment disturbance, with revenue exposure already close to zero |
| Ongoing | Large-customer deployment cadence | Progress on AWS's 2 million GPUs and the OpenAI deployment; the public production basis for hyperscalers' in-house silicon |
To draw it together: $96.2B and Data Center growth of +117% mean the demand side is not where the disagreement lies this quarter; the disagreement is concentrated entirely on the cost — the downward gross margin path starting from 75.0%, and the new shape of the statements formed by $279B of supply commitments, $108.5B of guarantees (of which $105B faces a single campus) and 60-day DSO. A 26x forward P/E is the price the market has put on that cyclical profile, and whether it is cheap depends on whether the roughly 70% FY2028 total-revenue growth outlook can be delivered without gross margin breaking the 72% lower bound.
The next observation point is the Q3 print in late November and the guidance given alongside it. In between, watch two lines: the monthly direction of memory contract prices, and public disclosure of large-customer deployment cadence. Whether the strength is delivered will answer itself when the time comes.
This note represents the research views of LunarTulip Research, compiled and analysed from publicly available information. It does not constitute investment advice. Markets carry risk; decisions must be made independently.