NVIDIA closed its fourth consecutive quarter of accelerating growth with $96.0 billion in revenue and, for the first time in company history, issued a full fiscal-year-ahead guide of approximately 70% revenue growth for FY2028. What makes this call consequential is not the number itself but what management chose to say around it: Jensen Huang was explicit that underlying customer demand is running closer to a doubling, meaning the 70% figure is a supply ceiling rather than a demand forecast. In the same breath, the company reset gross margin guidance meaningfully lower on a memory-cost shock it admits it did not see coming, and it used the words “circular financing” about its own frontier-lab funding arrangements before any analyst raised the term.
First published: August 30th, 2026.
The tension running through this call is simple to state and harder to resolve: an unprecedented full-year growth guide, delivered in the same breath as an admitted margin forecasting miss. Everything below works through both sides of that ledger.
Participants & Roles
Every question below was fielded by Huang personally, with one supplementary answer from Kress — a stylistic pattern worth noting, though not itself a red flag.
| Analyst | Firm | Focus of Question |
|---|---|---|
| Joe Moore | Morgan Stanley | Credibility of the 70% FY28 guide vs. ~100% demand signal |
| CJ Muse | Cantor Fitzgerald | Inference market share, agentic workload mix, ACIE/Groq contribution |
| Stacy Rasgon | Bernstein | Bridge of the ~$200B guidance uptick; unconstrained demand and pricing |
| Vivek Arya | BofA Securities | Scope/cash-timing of ~$500B ecosystem commitments; custom-silicon risk |
| Timothy Arcuri | UBS | Open- vs. closed-model economics |
| Ben Reitzes | Melius Research | RSI / AGI progress as a demand catalyst |
| James Schneider | Goldman Sachs | Rank-order of supply constraints (power, DRAM, foundry) |
| Aaron Rakers | Wells Fargo | Long-run $/gigawatt trajectory and linearity of capacity scaling |
Key Investment Themes
A supply-constrained growth story, by management’s own framing
Huang repeatedly distinguished ~100% underlying demand growth from the ~70% FY28 guide, explicitly calling the latter a function of what the supply chain can deliver. Notably, he declined to rank-order the actual bottleneck when asked directly.
Rising revenue-per-gigawatt as the master economic metric
$18B/GW (Hopper) → $25B/GW (Grace Blackwell) → $40B/GW (Vera Rubin), with networking, CPU and Groq LPU content layered in each generation. The most repeated data point on the call — and entirely company-sourced.
ACIE — the “invisible” half of the business
Sovereigns, regional neoclouds, enterprise and AI-native start-ups are said to already represent ~half of data-center revenue, growing ~100%/yr, with management asserting this segment could eventually exceed hyperscale in size. It is described as structurally invisible to conventional channel-check data because customers here buy an entire factory platform rather than discrete, countable chip units — one reason ACIE’s scale and growth rate are consistently underappreciated relative to the more closely tracked hyperscaler segment.
From chip vendor to financier of its own demand base
~$50B already invested directly in frontier labs; new financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR targeting $500B+ of third-party capital; a 4.25GW SoftBank Energy/Portsmouth deployment for OpenAI; and a neocloud revenue-share structure paying NVIDIA once on hardware and again on rental upside.
A margin reset arriving alongside the growth upgrade
Gross margin guided down to a 71–72% trough in Q4 FY27 on a DRAM price shock management says “exceeded our prior expectations” — an admitted forecasting miss on cost, delivered in the same breath as an unprecedented full-year revenue guide.
The full-stack moat, tested against customer-built silicon
Pressed on OpenAI’s and Anthropic’s in-house chip efforts (e.g., “Jalapeño”), Huang’s defense rested on platform fungibility and lifecycle breadth rather than any single point of technical superiority — a thesis that will be tested as top customers diversify silicon roadmaps.
An expanded AWS partnership (2M additional GPUs through Q2 FY29, Vera CPU/Rubin GPU mix, Nemotron on Bedrock/SageMaker, full physical-AI stack for warehouse robotics); Groq 3 LPX in full production (Nebius first volume customer) positioned as a bolt-on for high-interactivity workloads rather than core volume; and a formal removal of all China data-center compute revenue from the forward outlook on geopolitical grounds.
Forward-Looking Financial Signals
Formal, numeric guidance given on the call:
| Metric | Period | Guide |
|---|---|---|
| Total revenue | Q3 FY27 | $108B ± 2% |
| Gross margin (GAAP / non-GAAP) | Q3 FY27 | 74% ± 50bps |
| Vera Rubin % of DC revenue | Q3 FY27 | ~20% |
| Opex (GAAP / non-GAAP) | Q3 FY27 | ~$9.2B / $9.0B |
| Gross margin | Q4 FY27 | 71–72% (trough) |
| Opex growth, full year | FY27 | Low-50s % |
| Tax rate | FY27 | 16–18% |
| Revenue growth | FY28 | ~70% YoY |
| Gross margin | FY28 | 72–73% (recovery) |
Hyperscale growth to “reaccelerate” in Q4 FY27 / FY28 as Vera Rubin supply grows; the supply-demand gap “expected to remain a bottleneck at least through the end of fiscal year 2028”; CPU (Vera) revenue “preliminary expectation… to more than double” in FY28; AI-lab-linked demand expected at roughly a quarter of next year’s business; memory pricing “headed even higher into next year.” Management twice declined, when directly asked, to quantify the gap between the ~70% supply-based guide and the ~100% demand signal.
The widely referenced “~$200B uptick” and “$500B aggregate commitments” figures originated from analysts’ own back-of-envelope math against CFO commentary (Rasgon, Arya) — Kress did not confirm the specific totals, she addressed only the timing of commitments (front-loaded in years 1–3). Treat these as analyst-derived, not company-confirmed.
Management Q&A — Key Exchanges
Huang cited agentic compute intensity (15–100x a human user), ACIE’s ~100%/yr growth as an underappreciated demand pool, and multi-year land/power/shell lead times as the real constraint. This is the first time NVIDIA has guided a full fiscal year ahead.
Huang described a four-phase AI lifecycle (data prep, pretraining, post-training, agentic inference) and framed NVLink-72 rack-scale architecture as the one fungible platform spanning all four — with Groq positioned narrowly for high-interactivity, lower-throughput niches.
Huang confirmed unconstrained demand would be “a lot higher” than 70% without giving a number, attributing the uplift to both hyperscale and ACIE growth plus rising $/GW economics ($18B → $25B → $40B across generations).
Huang argued NVIDIA’s full-stack, cloud-agnostic platform differs structurally from single-purpose inference chips being built in-house by OpenAI and Anthropic, and expects those labs to remain long-term customers regardless. Kress clarified that the bulk of supply commitments are front-loaded in the first three years.
Huang said both open and closed models are surging and both run predominantly on NVIDIA due to footprint and fungibility, citing autonomous-cybersecurity defense as an open-model-enabled use case — “delighted by any model succeeding.”
Huang framed demand as set to “inflect further” as agent counts scale from thousands to millions per company, describing a “coarse-grained” self-improvement loop already occurring via per-run skill-file updates, while downplaying AGI-milestone framing as largely semantic.
Huang declined to single out one bottleneck, describing the entire supply chain as “challenged” and running flat out, with capacity and yield improving incrementally through the year rather than in a single step.
Huang described the directional goal as maximizing compute density per gigawatt of land/power/shell — “the perfect answer is actually infinity per gigawatt” — and cited an anecdotal claim of sub-one-year ROI on $50B-scale data centers as evidence supporting continued reinvestment.
Semantic Analysis of Management Language
Beyond what management said, a word-level pass through the transcript — comparing Colette Kress’s scripted remarks against Jensen Huang’s live Q&A — surfaces a few patterns in the specific financial and strategic vocabulary that carry their own signal, independent of the headline numbers.
“Margin” appears 8 times across the entire call — 6 of them in Kress’s scripted remarks, where the Q4 FY27 trough and the memory-driven reset are disclosed, and only 2 in the full Q&A session. Neither of those 2 live mentions is even about NVIDIA’s own margin — both come from Huang describing how profitable NVIDIA’s compute is for hyperscaler and frontier-lab customers. Management put its own margin reset on the record once, in the safest possible setting — a prepared, single-pass statement — and no analyst pulled it back into live conversation afterward.
For a call disclosing ~$50B of direct frontier-lab investment, $500B+ of facilitated third-party financing, and a named “circular financing” arrangement, the word “risk” itself appears exactly once in 90 minutes — in the script, proactively, in a single reassuring line: “our risk is limited.” It is never used again by management or by any analyst, including during Arya’s direct question on the scope of ecosystem commitments and competitive exposure to customers’ in-house silicon. The substance of risk is discussed at length; the word itself is almost entirely avoided.
The same two words do double duty across the call. Early on, they build the competitive moat: “our architecture is the most fungible… one platform, fungible for every model and workload, durable for the entire life cycle of AI.” Later, when defending the scale of frontier-lab financing, management reaches for near-identical phrasing to argue the exposure is safe: “the NVIDIA Compute platform is fungible and durable and can be redeployed to support other customers.” The vocabulary built to describe commercial strength is reused, largely unchanged, to describe financial safety — worth noting as a rhetorical shortcut rather than a separately reasoned risk argument.
“Approximately” appears 8 times in the call — every instance in Kress’s prepared remarks, and every instance attached to a hard figure (approximately 70% growth, approximately $500 billion in third-party capital, approximately 12 gigawatts, approximately a quarter of next year’s business). In live Q&A, the vaguer magnitude phrase “a lot” appears 18 times — never once attached to a number (“we just need a lot more,” “the unconstrained would be a lot higher,” “we’ve got lots and lots of time”). The same asymmetry shows up in “unconstrained” demand, used 3 times to describe the gap above the 70% guide, and never quantified even when asked directly twice (Rasgon, Schneider). When a figure was ready, it was written into the script in advance; when one was requested live and withheld, an imprecise intensifier filled the gap.
“Demand” is mentioned more in the script (15) than in Q&A (11) — management led with the demand narrative on its own terms. “Supply” runs the other way: 8 mentions in the script versus 17 in Q&A, more than double, largely because analysts (Moore, Rasgon, Schneider, Rakers) kept steering the conversation there. Read alongside the margin finding above, live questioning extracted materially more detail on supply mechanics than it did on the one negative revision — margin — disclosed the same day.
Arya’s question on OpenAI’s and Anthropic’s in-house chip efforts is the single most competitively pointed exchange on the call — and also where language repetition is tightest. Huang says “I’m delighted” four times in that one answer, and some variant of “100% confidence” three times (“I have 100% confidence that our technology…,” “I have every confident…,” “I have 100% confidence that, through quite a long period of time…”). Reaching for the same reassurance phrase repeatedly, rather than building a new argument each time, reads as a prepared talking point deployed under strategic pressure rather than a freshly reasoned rebuttal.
Methodology note: counts are drawn from a full-text pass of the corrected transcript, splitting Colette Kress’s prepared remarks (through “we will now transition to Q&A”) from the live Q&A that followed. Figures are indicative of pattern and direction, not a substitute for reading the underlying exchanges in Section 04.
Red Flags & Areas of Concern
Financing“Circular financing,” in management’s own words — with the counterparty undisclosed. Credit enhancement extended to a frontier lab that has also secured capacity independently is acknowledged by name as “circular financing.” The rebuttal offered (fungibility, redeployability, “our risk is limited”) is qualitative, with no disclosed LTV, collateral haircut, or loss-given-default framework — and the ~2GW recipient itself is not identified, limiting independent credit assessment.
Balance sheetScale of vendor-financing exposure. ~$50B invested directly, $500B+ in facilitated third-party capital, and take-or-pay commitments under the neocloud revenue-share model are growing quickly; management calls this “a small fraction” of expected free cash flow but discloses no explicit ratio or stress-test figure on the call.
GuidanceFirst-ever full-year guide, alongside an admitted forecasting miss. Management states it has “never forecasted or guided to a year in advance” before this call — a change in disclosure posture that coincides with a memory-cost forecast the CFO admits “exceeded our prior expectations.”
Credit qualityExtended payment terms, rising DSO. Days-sales-outstanding rose to 60 days, explicitly tied to extended terms for large, multi-quarter shipments to “certain investment-grade customers” — worth cross-checking against customer-concentration disclosure once filed.
Open Questions for Further Research
From primary CFO Commentary / 8-K disclosures — separate balance-sheet capital from facilitated third-party capital and contingent guarantees. This distinction determines how much of the ecosystem build-out is actually NVIDIA’s own balance-sheet risk versus capital raised and held by outside institutions.
Receiving ~2GW of credit enhancement — cross-reference recent 8-K filings and lab funding announcements. Given the scale is roughly half of the disclosed OpenAI/Portsmouth deployment, naming the counterparty is a prerequisite for assessing credit quality rather than taking management’s risk characterization at face value.
Behind the DSO increase, via the 10-Q’s credit-loss allowance and concentration footnotes. A concentrated base of extended-term customers would turn the reported DSO increase into a credit question, not just a working-capital one.
The call confirms an acceleration narrative — a fourth straight quarter of accelerating growth, record capital returns, and the first full-year-ahead guide in company history — built on rising per-gigawatt revenue capture, an under-recognized non-hyperscaler growth engine, and an increasingly active role financing its own frontier-lab customer base. What actually changed versus prior expectations sits on the cost side, not the volume side: gross margin guidance was reset meaningfully lower on a memory-cost shock management admits it underestimated, even as revenue guidance was both raised and extended further out than ever before.
The key swing factors going forward are whether the supply chain — deliberately not rank-ordered by management on this call — closes the stated 70%-versus-~100% demand gap without further slippage; the credit quality and eventual cash conversion of the rapidly scaling frontier-lab financing book, given management’s own use of the term “circular financing”; and execution of the Q1 FY28 price increases needed to restore gross margin to the 72–73% range. None of these are demonstrated impairments today — each is a disclosed, falsifiable checkpoint to track against primary filings rather than take at face value from the call’s narrative framing.