All research gathered — EDGAR financials through FY2026, Q1 FY2027 actuals, China/H200 status, ownership, peer multiples, and the Rubin roadmap. Here is the complete report.

NVDA — Research Report (2026-07-13)

TL;DR

• What it is: NVIDIA designs the GPUs, networking, and CUDA software that run essentially all serious AI training — the world's most valuable company at ~$5.1T market cap.
• My take: the best business model in tech, now at a historically reasonable price (~21–24× forward earnings) — the risk is no longer valuation froth, it's whether AI capex keeps compounding.
• $215.9B FY26 revenue, +65% (hyper-growth at a scale nobody has ever grown from) — FY26 results
• ~$97B FY26 free cash flow (the profits are real cash, not accounting)
• ~$1T in Blackwell+Rubin orders through 2027 (a huge slice of future revenue already contracted) — CNBC
• 36% of revenue from just 2 customers (its fate is tied to a handful of buyers' budgets)
• Strongest bull point: Rubin already in full production + contracted backlog ≈ growth essentially locked through 2027.
• Strongest bear point: its biggest customers are simultaneously its biggest threat — they're building their own chips and control the capex tap.
• Watch next: Q2 FY27 earnings (late Aug 2026) — $91B guided, with zero China revenue baked in, so any China restart is pure upside.

1. Business overview

So what: NVIDIA stopped being a "chip company" — it sells entire AI factories, and ~90% of revenue now comes from data centers.

NVIDIA designs (but does not manufacture) the processors that train and run AI models, plus the networking that wires thousands of them together, plus CUDA — the software layer that 20 years of AI code is written against. It sells these increasingly as full rack-scale systems (a GB200/GB300 "NVL72" rack integrates 72 GPUs, CPUs, and networking into one ~$3M unit) rather than individual chips.

Customers are hyperscalers (the giant cloud companies — Microsoft, Amazon, Google, Meta), AI labs (OpenAI, Anthropic, xAI), "neoclouds" like CoreWeave, and now sovereign governments. Concentration is extreme: in FY26, one direct customer was 22% of revenue and another 14% — 36% from two buyers (FY26 10-K).

Note on dates: NVIDIA's fiscal year ends in late January, so "FY2026" ≈ calendar 2025, and we're now in FY2027.

2. Financials — 5 years

So what: revenue 8×'d in four years, and unlike most growth stories, cash flow kept pace with reported profit.

All figures from SEC EDGAR company facts unless noted; capex/FCF cross-checked against Macrotrends and FinanceCharts.

*One-offs flagged: FY26 gross margin was dragged down by a $4.5B write-off of China-only H20 chips after export bans (Q1 FY26); the exit rate was back to 75.0% by Q4. FY23 was the crypto-mining hangover — worth remembering that this company has seen demand evaporate before.

Balance sheet: $62.6B cash + marketable securities vs. ~$8.5B total debt → roughly $54B net cash (FY26 10-K). Capital returns FY26: $40.1B buybacks + ~$1.0B dividends; shares outstanding shrank from 24.64B (FY24) to 24.30B (FY26).

Current run-rate (Q1 FY27, quarter ended 2026-04-26): revenue $81.6B, +85% y/y; Data Center $75.2B, +92%; FCF $48.5B in one quarter; dividend raised 25× ($0.01→$0.25/qtr) plus a new $80B buyback authorization. Caution: Q1 GAAP net income ($58.3B) was inflated by $15.9B of paper gains on its equity stakes — the cleaner non-GAAP figure is ~$45B. Q2 FY27 guidance: $91.0B ±2%, assuming zero China data-center revenue.

3. Key metrics — and why each matters here

So what: on almost every metric, NVDA looks like a software company wearing a chip company's clothes — the exception is customer concentration.

4. Unit economics — the one number this business runs on

So what: NVIDIA earns ~75 cents of gross profit on every AI-compute dollar, and its revenue is effectively a tax on global AI data-center capex.

The atomic unit is a GPU (roughly $30–40K each, ~$3M per integrated rack), sold at ~75% gross margin. Volume is the multiplier: management expects ~20M Blackwell-generation chips sold vs. 4M for all of Hopper — 5× units at higher prices and higher margin per unit.

The upstream driver to watch is therefore not NVDA's own reporting but hyperscaler capex budgets — NVDA's Data Center revenue is a share of that spending. As long as its customers' AI capex grows and NVDA holds share, revenue growth is nearly mechanical.

5. Value-chain position — bottleneck or commodity?

So what: NVDA owns today's bottleneck (the compute layer + CUDA), but sits between two suppliers with real pricing power and customers actively engineering their escape.

• Who has pricing power over NVDA: TSMC — the only fab that can make these chips and package them (CoWoS advanced packaging), and it has been raising prices; and the HBM memory oligopoly (SK Hynix/Samsung/Micron) — HBM4 supply is the physical ceiling on Rubin output.
• NVDA's moat over customers: CUDA software lock-in (rewriting AI infrastructure for another chip is expensive and slow), systems integration nobody else does at rack scale, and a brutal annual product cadence — Vera Rubin hit full production in ~18 months after Blackwell, at 3.5× training / 5× inference performance.
• The counterattack: Google's TPUs, Amazon's Trainium, and Broadcom's custom-chip business (AI revenue >$30B in FY26) are the customers' lever. So far they've grown alongside NVDA, not instead of it.
• China is currently a zero: H20 was banned (the $4.5B charge), then the H200 was approved for China in Jan 2026 with a 25% surcharge and per-customer caps — but Beijing told its companies not to buy, and nothing has shipped. Guidance assumes zero China, so this is optionality, not risk.

M&A track record: excellent but small-ball — Mellanox (2019, $6.9B) built today's networking segment and looks like one of tech's best acquisitions; the Arm bid was blocked (2022). The current strategy is minority stakes instead: $18.6B invested in one quarter into AI labs and neoclouds — which is also a yellow flag (see bear case: it partially funds its own customers).

6. Ownership, management, insider signals

So what: founder-led with massive skin in the game, but the insider tape in mid-2026 leans one way — out.

• Jensen Huang (founder-CEO, 30+ years): owns ~870.6M shares, ~3.6% — worth ~$180B. His incentives are overwhelmingly equity; no control-share structure, big institutions (Vanguard, BlackRock) top the register.
• Insider selling: Huang sells via pre-scheduled 10b5-1 plans (plans set in advance to avoid insider-trading issues), including an ~$865M plan. Routine in isolation — but in June 2026 NVIDIA's five most senior officers sold on the same day, and insider buying across AI hardware has nearly vanished. A soft signal, not proof — sized against his stake, Huang's sales are trivial.
• Shareholder-return posture turned real in 2026: dividend up 25×, $80B fresh buyback, and a stated target of returning ~50% of FCF (company statements; still leaves ~$50B/yr for investments).

7. Valuation view

So what: the market already prices in sharp deceleration — NVDA trades below the average chip stock on forward earnings despite growing 3–4× faster.

Arithmetic (estimated): ~$5.1T market cap − $54B net cash ≈ $5.05T enterprise value. Consensus FY27 EPS ~$9.34 implies ~$225B net income; at ~$210/share that's ~22× forward earnings and roughly ~20× my estimated FY27 EV/EBITDA (whole-business price vs. operating cash proxy — here EBITDA ≈ operating income since capex/depreciation are tiny).

What a reasonable buyer might pay: nobody acquires a $5T company, so the question is what return the price implies. Paying ~22× forward earnings for contracted-backlog growth, 75% gross margins, and net cash is cheaper than the market average multiple for a business growing several times faster — Motley Fool notes it hasn't been this cheap in ~7 years. A buyer underwriting FY28 EPS of ~$11–12 (post-Rubin, estimated) at a held 22–25× gets $250–300/share; the same buyer assuming an AI-capex digestion year (EPS flat, multiple derates to a cyclical 15×) gets $140. The market price ($210) sits between those — i.e., it prices real cyclical risk, not perfection.

8. Bull / base / bear — and the kill switch

So what: the next 12 months are largely contracted; the 2028+ question is whether AI spending is infrastructure (like power grids) or a capex bubble (like 2000 telecom fiber).

• Bull: ~$1T Blackwell+Rubin orders through 2027, Rubin already in full production, inference demand (running AI, recurring) replacing one-off training demand, China as free upside → FY28 EPS $12+, multiple re-rates toward peers, $280–320.
• Base: backlog converts, growth decelerates from ~85% toward 30–40% by FY28, margins hold ~73–75%, multiple stays ~20–22× → stock compounds roughly with earnings, ~$230–260 over 12 months.
• Bear: hyperscaler ROI on AI disappoints → capex digestion in 2027; custom silicon takes the marginal dollar; concentration bites (one customer pausing = a visible revenue hole); EPS stalls ~$9 and the multiple derates to 15× → ~$135–150.
• Catalysts: Q2 FY27 print (late Aug 2026, $91B guide); Rubin volume ramp in 2H26 (watch HBM4 supply); hyperscaler capex guidance each quarter; any China H200 shipment (currently zero); OpenAI deal revenue (explicitly excluded from the backlog figure).
• What would prove the thesis wrong: (1) hyperscalers guiding capex down — the single cleanest kill signal; (2) Data Center revenue growing slower than customer capex (= share loss to TPU/Trainium/Broadcom); (3) gross margin sustained below ~70% (= pricing power broken); (4) Rubin orders being deferred rather than grown; (5) a widening gap between NVDA's investments into customers and those customers' independent ability to pay — the circularity risk ($18.6B/quarter invested into the ecosystem that buys its chips).

9. PE / quality lens

So what: a buyer would love the cash machine and the lock-in; the fear is that peak margins, peak share, and peak customer-capex could all be the same peak.

Love: 75% gross margin with falling capital intensity, net cash, CUDA switching costs, annual cadence competitors can't match, founder-operator. Fear: two-customer concentration, demand that is itself venture-funded in part, a supplier (TSMC) it cannot substitute, geopolitics on both ends (US export rules + Beijing procurement bans), and FY23 as proof this company can hit an air pocket fast.

📚 What this company teaches

1. Sell the bottleneck, price like software. NVDA's margins come from owning the scarce layer (compute + CUDA) in a value chain where everyone else — fabs, memory, clouds — must route through it. When analyzing any industry: find who must be paid regardless of which end-product wins (the HBM-bonder / HALEU logic from this vault's discovery playbook).
2. Backlog vs. market cap is the most honest growth metric. Guidance is a promise; a $1T order book is a contract. When a fifth of a company's entire valuation is already contracted revenue, the near-term debate is settled — only the terminal question remains.
3. Watch the gap between profit and cash — in both directions. NVDA's FCF ≈ 80% of net income says earnings are real. But Q1 FY27 also shows the reverse trick: GAAP profit inflated by $15.9B of paper gains on investments in its own customers. Cash flow statements expose both.

Self-test: NVIDIA trades at ~22× forward earnings while Broadcom trades at ~41×, yet NVIDIA grows much faster. List two rational reasons the market might still be right to award Broadcom the higher multiple — then decide which reason applies to NVDA's 2028.

Sources: SEC EDGAR company facts / FY26 10-K (primary), Q4+FY26 results, Q1 FY27 results, Q1 FY27 8-K, CNBC (GTC 2026), CNBC (H200 China), Digitimes ($500B pipeline), Macrotrends, GuruFocus, RevenueMemo (ownership), Kresmion (insider brief), HeyGoTrade (peer multiples). Estimates labeled in text; FY27/FY28 EPS and EV/EBITDA figures are consensus-based or my arithmetic, not reported.