DEEP DIVE №1 · WEEK 32, 2026
NVIDIA NVDA
The tollbooth of the intelligence age — and, increasingly, the banker of its own customers.
BY SARVESH PATEL · AUGUST 2026
IF YOU READ NOTHING ELSE
- ▸ NVIDIA sells the machines that make AI — and the software (CUDA) that locks the industry to those machines.
- ▸ Quarterly revenue grew $13.5B → $81.6B in under three years (+85% YoY, +43% vs. the prior reported quarter) — all from filed 10-Qs.
- ▸ Its balance sheet is a fortress ($195B equity, $7.5B long-term debt), but its income increasingly includes marks on investments in its own customers.
- ▸ The bull case is compute demand compounding; the bear case is circular financing, customers building their own chips, and efficiency gains shrinking chip needs.
What it actually sells
Three things, in descending order of fame: accelerators (the GPU systems — now full racks, not chips — that train and run frontier models), networking (the interconnect that makes a hundred thousand GPUs behave like one computer), and CUDA — the software layer two decades of AI code is written against. The third is the quiet moat: switching hardware vendors is a purchase decision, but leaving CUDA is a rewrite.
Where it sits in the chain
Directly between the foundry and everyone else. NVIDIA designs; TSMC manufactures on capacity the two negotiate years ahead; SK Hynix and Micron supply the high-bandwidth memory that is routinely the binding constraint. Downstream, its customer list is effectively the rest of this website: the ~$725B of 2026 hyperscaler capex is, to a large degree, a queue for NVIDIA systems (Statista), and the AI-native clouds — CoreWeave, Nebius, Lambda — were built around allocations of its chips.
The move that defines 2026: funding its own demand
This year NVIDIA stopped being only a supplier and became one of the largest investors in AI itself: $30B into OpenAI's $110B roundat a $730B valuation, alongside Amazon's $50B and SoftBank's $30B (Cloud Wars), plus equity stakes in the GPU clouds that buy its hardware. Bloomberg mapped the resulting loop — the circular-deals economy: NVIDIA funds customers, who buy NVIDIA systems, often through clouds NVIDIA part-owns. It is vendor financing at a scale tech has never seen — not automatically sinister, but it means demand, revenue, and valuation now partially reference each other.
What it depends on
Four dependencies worth naming: TSMC (essentially sole manufacturer — the deepest single-point risk in the global economy), HBM supply (memory, not logic, has repeatedly been the bottleneck), power (its chips are worthless without the gigawatts tracked on our indicator board), and export policy — the China market has been repeatedly resized by Washington, and a parallel Huawei-led stack is being built in response (see the China stack).
How it got here — a 30-year overnight success
1993Founded by Jensen Huang, Chris Malachowsky, and Curtis Priem to build graphics chips for gaming.
1999Ships the GeForce 256 — marketed as the world's first 'GPU' — and IPOs on Nasdaq.
2006The company-defining bet: CUDA, a programming layer that let scientists use gaming chips for general math. Wall Street saw wasted R&D for years.
2012AlexNet — the neural network that ignited modern deep learning — is trained on two consumer NVIDIA GPUs. AI researchers standardize on CUDA.
2019Buys Mellanox for ~$6.9B — the networking company whose InfiniBand now stitches AI clusters together. In hindsight, one of tech's great acquisitions.
2022ChatGPT launches; the H100 becomes the scarcest industrial commodity on earth.
2024–25Blackwell generation ships as full rack-scale systems; NVIDIA becomes, at points, the most valuable company in the world.
2026Turns banker: $30B into OpenAI, stakes across the GPU clouds — funding the demand for its own chips.
The product line, in plain English
Accelerators (the famous part): Hopper (H100/H200) and now Blackwell (B200, GB200) — sold less as chips than as full racks where 72 GPUs act as one machine. This is what the hyperscalers' capex queues are for.
Networking (the underrated part): from the Mellanox deal — InfiniBand and Spectrum-X Ethernet, plus NVLink inside the rack. Roughly speaking: the chips think, the network lets a hundred thousand of them think together. Competitors can match a chip; matching the cluster is far harder.
CUDA + software (the moat): the programming platform, libraries, and now inference microservices (NIM). Two decades of AI code targets CUDA; leaving means rewriting. This is why customers who resent the prices keep paying them.
Grace CPU, DGX systems, automotive (DRIVE), robotics (Jetson/Isaac): smaller today, but robotics is the declared next act — the physical-AI bet.
Who runs it, who buys it
Jensen Huang — co-founder and CEO for the entire 30+ years, the longest-tenured founder-CEO in large-cap tech and its most recognizable figure (the leather jacket is a brand asset). CFO Colette Kresshas run the numbers since 2013. The customer base is extraordinarily concentrated: the hyperscalers and AI-native clouds dominate, and NVIDIA's own filings routinely disclose individual customers each exceeding 10% of revenue — meaning a handful of buyers' capex decisions move the whole income statement. That concentration is the mirror image of the moat.
Competition, mapped honestly
Merchant rivals: AMD's MI series is the credible alternative and wins real deployments; Intel's accelerator effort has repeatedly stumbled. Neither has CUDA's gravity.
The customers themselves (the real threat): Google's TPU (mature, trains Gemini), Amazon's Trainium (trains for Anthropic), Microsoft's Maia, Meta's MTIA, and the OpenAI–Broadcom custom-chip program. Every major buyer is spending billions to need NVIDIA less.
Inference specialists: Groq, Cerebras, SambaNova attack the serving side, where CUDA lock-in is weakest and cost-per-token is everything.
China: export controls capped NVIDIA's China business and midwifed Huawei's Ascend line — a subsidized rival with a guaranteed home market (see the China stack).
Scoreboard today: in frontier training, NVIDIA remains overwhelmingly dominant; in inference, the field is genuinely contested. The bear case runs through inference and custom silicon, not through a head-on rival.
The numbers — from the latest SEC filing
Straight from the 10-Q for the quarter ended April 26, 2026 (fiscal Q1 2027), via SEC EDGAR's XBRL data:
| Revenue (quarter) | $81.6B |
| Gross profit (quarter) | $61.2B (~75% gross margin) |
| Net income (quarter) | $58.3B — flattered by non-operating gains, incl. marks on its AI investment book |
| Total assets | $259.5B |
| Total liabilities | $64.0B |
| Stockholders' equity | $195.5B |
| Cash & equivalents | $13.2B |
| Long-term debt | $7.5B |
The balance-sheet read: a fortress — equity three times total liabilities, debt trivial against one quarter's profit. The asset side is the thing to watch: as the investment book (OpenAI at a $730B valuation, GPU-cloud stakes) grows, more of "assets" and even "net income" reflects marks on private AI valuations rather than chip sales — the circularity theme showing up inside the 10-Q itself. Source: NVIDIA filings on SEC EDGAR.
The growth picture — quarter by quarter
The fastest way to understand this business is to watch the filed numbers move. Nine reported quarters of revenue and net income, straight from the XBRL data behind NVIDIA's SEC filings (January quarter-ends omitted where the data reports annual figures only):
| GROWTH READ | REVENUE | NET INCOME |
|---|---|---|
| Year over year (Apr-26 vs Apr-25) | +85% | +210% |
| Vs. prior reported quarter (Apr-26 vs Oct-25) | +43% | +83% |
| Three-year multiple (vs Jul-23) | ~6.0× | ~9.4× |
Two teaching notes on reading these. First, net income growing fasterthan revenue means margins are expanding — the classic signature of pricing power. Second, the Apr-26 net-income spike overstates the operating business: as flagged above, it includes non-operating gains on the investment book. When income jumps faster than gross profit, always ask what's underneath. (We don't chart the stock price here — filings don't contain it, and any brokerage app does; this page is about the business — but for orientation, TradingView's free widget below shows the market's reaction to everything described here.)
The bull case, honestly
If the timelines this desk tracks are even half right, compute demand compounds for years: task horizons doubling every ~4 months, labs writing 70–100% of their code with AI, and $725B of annual capex still growing 77% year over year. In that world, the company selling the scarce input with a software moat keeps pricing power — and its investment book (OpenAI at $730B, the GPU clouds) compounds on top.
The bear case, honestly
Three real ones. Circularity: if AI revenue disappoints, the vendor-financing loop unwinds in both directions at once — sales fall and the investment book marks down together. Custom silicon: every major customer (Google TPU, Amazon Trainium, Microsoft, Meta, OpenAI-Broadcom) is designing its own chips precisely to escape the tollbooth. Efficiency: DeepSeek-class algorithmic gains — the 5–40×/yr cost collapse on our board — could mean the world eventually needs fewer chips per unit of intelligence than the capex queue assumes. The counterargument to all three so far: demand has outrun every one of them.
What we're watching
The three tells for this name on our board: hyperscaler capex guidance (confirmation or the first honest stall signal), HBM supply announcements, and any customer shifting a flagship model's training run to in-house silicon. Each is tracked weekly.
Glossary — the six terms that unlock everything above
GPU / accelerator: a chip that does thousands of simple calculations at once — the shape of work neural networks need.
CUDA: NVIDIA's programming platform; the reason AI software runs on NVIDIA by default.
HBM: high-bandwidth memory stacked next to the GPU; the physical bottleneck that decides how many accelerators exist each quarter.
Foundry: a factory that manufactures chips others design — for leading-edge AI chips, effectively just TSMC.
Training vs. inference: training is teaching the model (huge one-time compute); inference is running it for users (forever compute — the larger long-run market).
Hyperscaler: the giant clouds (Microsoft, Amazon, Google, Meta, Oracle) whose capex is the AI buildout.
INFORMATIONAL ONLY — not investment advice, no recommendation to buy or sell any security. Facts are sourced inline as of August 2026 and may change. See the full disclaimer.