NVIDIA Corporation — Bull & Bear Memo
SECTION ONEBusiness in one line
NVIDIA sells rack-scale AI data-centre systems — GPUs, Grace CPUs, NVLink and networking, and the CUDA software stack — architected as one “extreme co-designed” unit and manufactured entirely by others: TSMC and Samsung make the wafers, SK hynix, Micron and Samsung supply the memory, Hon Hai, Wistron and Fabrinet assemble [FY2026 10-K]. The economics are the spread between a very high system price and a bought-in bill of materials, widened by software given away to make the hardware indispensable. Data Center was $89.0bn of $96.2bn of Q2 FY2027 revenue — 92% of the company [Q2 FY2027 10-Q].
Type: fast grower resting on a cyclical. Revenue rose 106% year on year last quarter, but the demand behind it is a capital-expenditure budget set by a handful of buyers, not recurring consumption. Remaining performance obligations beyond one year are $3.2bn against a $96bn quarter — there is no backlog to cushion a pause [Q2 FY2027 10-Q]. It must be underwritten as a cyclical wearing a fast grower’s clothes.
SECTIONS TWO & THREEBull case, and the attack on it
Four load-bearing assumptions, each paired with the bear argument that attacks it. Use the toggle to read one side alone; click any card to open it.
The simple reason this stock could work is that the world has decided to build a new kind of computer, NVIDIA sells the whole computer rather than a component of it, and at a 75.0% gross margin and 66.2% operating margin it generated $74.4bn of operating cash flow in the last six months alone, while the buildout is still early.
The most likely way I lose money is by owning a company that has quietly become the credit support for its own customers, and finding out that the demand I was paying a growth multiple for was partly demand NVIDIA financed.
Why the market might be missing it: the reported multiple is calculated on earnings that include $23.7bn of mostly unrealised H1 equity gains, which flatters the “E” and disguises how much cheaper the stock looks against operating income — while simultaneously hiding that those same marks are correlated with the revenue.
SECTION FOURSignals to monitor
Every signal traces to a named assumption or bear risk. A signal that tracks nothing in the memo is noise.
SECTION FIVEBottom line
NVIDIA could work because it sells the entire AI computer rather than a part of it, and an annual full-stack cadence keeps resetting the comparison against anyone building a substitute, at margins that fund the next reset.
For that to hold, the rack must remain the unit of competition, gross margin must survive memory repricing, and demand must broaden to buyers who fund themselves.
The thesis most likely breaks through the financing loop: $47.9bn of equity stakes in customers, $36bn of AI-cloud capacity NVIDIA has agreed to buy back, $63.1bn of receivables on extended terms and $108.5bn of guarantees mean a customer funding failure hits revenue, earnings and asset values in the same quarter.
What would change my mind, in either direction, is the relationship between the commitment book and revenue — if commitments and equity additions decelerate while ACIE keeps compounding, the demand is real and independent; if they must keep rising to hold the growth rate, it is not.
EVIDENCESources
Form 10-Q, Q2 FY2027 — quarter ended 26 Jul 2026, filed 26 Aug 2026 primary
- Condensed income statement — revenue $96,221m; gross margin 75.0%; operating income $63,734m (66.2%); net income $59,688m; H1 operating margin 66.0%
- Note 5 / Note 6 — investments — non-marketable equity $22,251m → $47,898m, net additions $31,005m in H1; marketable equity $42,783m; cumulative gross unrealised gains $9.1bn
- Note 7 — other income — gains from equity securities $23,707m in H1; other income $24,140m
- Note 10 — commitments — supply and capacity $279bn (from $119bn), $92bn due in remainder of FY2027; AI-cloud agreements $36bn; equity investment commitments $25bn ($18bn within FY2027); guarantees $108.5bn including $105bn SB Energy / OpenAI
- Note 13 — segments & customers — Data Center $89,023m; Hyperscale $48,710m (+102%); ACIE $40,313m (+138%); Edge $7,198m; one direct customer 16% of Q2 revenue; three at 16% / 15% / 13% of H1
- MD&A — inventory provisions $985m gross, $177m released, 0.8pt gross-margin drag; receivables $63,059m; operating cash flow $74,421m; buybacks $39.8bn H1; RPO beyond one year $3.2bn
Form 10-K, FY2026 — year ended 25 Jan 2026 primary
- Item 1 — business — extreme co-design; supply chain (TSMC, Samsung, SK hynix, Micron, Hon Hai, Wistron, Fabrinet); $76.7bn cumulative R&D; 7.5m CUDA developers; ~42,000 employees, 31,000 in R&D
- Item 1 — competition — Alphabet, Amazon, Microsoft, Alibaba, Baidu and Huawei named as competitors with internal accelerator teams
- Item 1A — risk factors — “most of our sales are made on a purchase order basis”; customers “can generally cancel, change, or delay product purchase commitments with little notice and without penalty”; one direct customer 22% and another 14% of FY2026 revenue
- Financials — revenue $215,938m (+65%); operating income $130,387m; net income $120,067m; operating cash flow $102,718m; capex $6,042m
Form 10-Q, Q1 FY2027 comparatives
- Prior-quarter values for the signals table: supply commitments $119bn; Data Center $75,246m; Hyperscale $43,050m; ACIE $32,196m; gross margin 74.9%
Not used excluded
- External market data, sell-side estimates and press commentary — outside the source folder and outside the scope of this memo
- Earnings-call transcripts and press-release exhibits — not present in the source folder, so no management commentary is quoted anywhere above
- The prior Coverage File and Business Overview in the research folder — prior work, used only for orientation, never as evidence