NVIDIA (NVDA): what it does, how it makes money, and why the AI build-out runs on it

The standard $NVDA story is that NVIDIA makes graphics cards, and graphics cards happen to be good at AI. That was true ten years ago. It is a poor description of the company today.

NVIDIA now sells complete AI computers: the GPU, the CPU next to it, the memory package, the switches that wire thousands of chips together, and the software that makes all of it programmable. In the quarter ended 2026-07-26, data centers brought in $89.0B of a $96.2B total, about 93% of revenue. Gaming, the business most people still picture, no longer has its own line in the results. This piece walks through what NVIDIA does, how it makes money, where it sits in QA's AI compute bubble, and the risks that sit under a $5.65T market cap. Market data is as of 2026-10-02 unless noted.

The full video teardown, plus live price, ETF holdings and bubble correlation, is on /stocks/nvda.

What does NVIDIA do?

NVIDIA designs chips and systems that do a very large number of simple calculations at once. That is exactly what training and running AI models requires: multiplying huge grids of numbers, billions of times per second. A GPU (graphics processing unit) was built to do this for pixels; NVIDIA turned it into the general-purpose engine for AI.

NVIDIA does not manufacture its chips. It designs them, and Taiwan Semiconductor ($TSM) fabricates them. The memory stacked next to each GPU, high-bandwidth memory or HBM, comes from a small group of Korean and US suppliers. The supply side is its own story: see why NVIDIA's shipments depend on the HBM fabs.

Three layers sit on top of the silicon:

1. The accelerators. Product generations ship roughly once a year: Hopper (H100, H200), then Blackwell (B200, GB200), now Vera Rubin, which the company says is in full production. The cadence itself is part of the moat; the Blackwell to Rubin roadmap covers why.
2. The networking. A single GPU is useless for frontier AI; tens of thousands have to act like one machine. NVLink, InfiniBand and Spectrum-X Ethernet, much of it from the 2020 Mellanox acquisition, do that wiring. This is the second product the market under-discusses.
3. The software. CUDA, NVIDIA's programming platform, has nearly two decades of libraries, tools and developer habit behind it. Switching away means rewriting and re-tuning code, which is why the CUDA moat outlasts any single chip generation.

How NVIDIA makes money

Starting with fiscal 2027, NVIDIA reports two platforms. NVIDIA's fiscal year runs about eleven months ahead of the calendar: fiscal 2027 roughly covers calendar 2026.

• Data Center, split into Hyperscale (public clouds and the largest consumer internet companies) and ACIE (AI clouds, industrial and enterprise: neoclouds, sovereign AI projects, corporate AI factories).
• Edge Computing, which absorbed Gaming and now also holds PCs, consoles, workstations, robotics, automotive and AI RAN base stations.

The latest reported quarter, Q2 FY27 (ended 2026-07-26, per NVIDIA's press release filed with the SEC):

Post-earnings coverage of the call reported Hyperscale revenue more than doubling year on year and ACIE up about 138%; the press release itself gives only the Data Center total.

The concentration point is the one to hold onto. QA's desk lists Microsoft ($MSFT) Azure, Amazon ($AMZN) Web Services, Google Cloud ($GOOGL), Meta ($META), OpenAI and xAI as anchor customers, plus a neocloud bloc (CoreWeave, Lambda, Crusoe). The top handful of buyers carry the majority of Data Center revenue. The detail, and what happens if one of them builds its own chips, is in NVIDIA's customer concentration.

The cash goes back out, too. NVIDIA returned about $26.0B to shareholders in Q2 FY27 through buybacks and dividends, with roughly $99.0B of repurchase authorization left. Shares outstanding were about 24.15B as of 2026-10-02, and the desk notes the share count has been shrinking, the opposite of most AI growth names, which dilute.

Where it sits in the AI compute bubble

NVIDIA is the anchor of QA's AI Compute Accelerators bubble, with full weight; the desk's one-line rationale is "the AI compute monopoly." It also sits, at a 0.6 weight, in the Robotics / Automation bubble through its Isaac and Jetson robotics platforms, and it maps onto the AI Hardware and Robot Foundation Models themes.

On the AI stack, NVIDIA owns the compute layer and a growing share of the networking layer, and it depends on two layers it does not own: foundry (TSMC) and memory (HBM). Its tightest correlations in the QA universe reflect exactly that chain. As of 2026-10-02:

• TSMC (TSM): 0.65, the foundry that makes its chips
• Broadcom ($AVGO): 0.54, networking silicon and the custom-chip designer for hyperscalers
• ASML ($ASML): 0.51, the lithography machines behind every leading-edge fab
• Lam Research ($LRCX): 0.50, wafer fab equipment

$NVDA trades with the AI capex chain, not with consumer tech. When hyperscaler spending guidance moves, the whole bloc tends to move together.

The main competitive pressure comes from two directions. AMD ($AMD) sells merchant GPUs against NVIDIA directly. The bigger structural question is the hyperscalers' own chips: Google's TPU, Amazon's Trainium, Meta's MTIA, mostly aimed at inference. NVIDIA vs custom ASICs walks through what each one actually does.

The numbers

Two things stand out. First, a net margin near 64% on more than $300B of trailing revenue is unusual for a hardware company; it is closer to a software margin profile, and it reflects how much pricing power the full-system offering carries. Second, the trailing P/E of 29.5 is lower than many readers expect for the largest company in the US market, because earnings have grown about as fast as the price. When QA's teardown video went up on 2026-08-09, the market cap was about $5.4T; it was $5.65T on 2026-10-02.

The Street mean target is an average of analyst estimates, not a QA view and not a promise. On a company this size, it mostly measures how much of the next two years of hyperscaler spending the consensus is willing to count.

The bull case

• Scale and growth at once: $96.2B in a quarter, up 106% year on year, with guidance for $108.0B (plus or minus 2%) in Q3 FY27.
• Full-stack lock-in: GPU, CPU, networking and CUDA are sold as one system, which makes a partial switch to a competitor costly.
• An annual product cadence (Blackwell, then Rubin) that keeps the performance gap moving faster than rivals can close it.
• Demand broadening beyond the hyperscalers: ACIE (neoclouds, sovereign AI, enterprise) was the faster-growing Data Center sub-market in Q2 FY27, per call coverage.
• A net-cash balance sheet and heavy buybacks: about $26.0B returned in one quarter.

The bear case

• Customer concentration. A handful of hyperscalers and AI labs carry most of Data Center revenue; a single quarter of capex deceleration at one of them shows up directly.
• Custom silicon. Google, Amazon and Meta are all building their own accelerators, with Broadcom as a key design partner; every workload that moves to a TPU or Trainium is one NVIDIA does not ship.
• Supply chain dependence. One foundry (TSMC, in Taiwan) and a small set of HBM suppliers cap how much NVIDIA can ship, and add geopolitical exposure.
• China. Q3 FY27 guidance assumes no Data Center compute revenue from China at all. Export rules have shifted several times; see the China export loophole.
• Size. At $5.65T, the stock is already priced on sustained AI spending; a normalization of that spending, even at a high base, would test the multiple.

How to access

NVIDIA trades on NASDAQ as $NVDA. To trade it from a US-retail account, see /stack/ibkr.

Most investors already own it indirectly. As of 2026-10-02, NVDA is about 8.1% of the S&P 500 funds ($SPY, $VOO), 14.4% of the Technology Select Sector SPDR ($XLK), 17.7% of Vanguard Information Technology ($VGT) and 22.6% of the VanEck Semiconductor ETF (SMH, $SMH). It is also 9.6% of Global X Robotics & AI ($BOTZ) and 4.9% of ARK Autonomous Technology & Robotics ($ARKQ). Anyone holding a broad US index fund has a meaningful NVIDIA position without having bought a single share.

Subscribing on /stocks/nvda sends a free Ticker Teardown Dossier, and bubble-correlation shifts and rule-based alerts on $NVDA are part of /pro.

What to watch

• Next earnings: Q3 FY27 results, expected around 2026-11-18 (calendar estimates; NVIDIA had not confirmed the date as of writing). Guidance is $108.0B revenue and a 74.0% gross margin.
• Hyperscaler capex guidance from Microsoft, Amazon, Alphabet and Meta in their own late-October prints: the leading indicator for NVIDIA's Data Center line.
• The Vera Rubin ramp, and whether gross margin holds near the mid-70s through the product transition.
• Any China policy change, since current guidance assumes zero Data Center compute revenue there.
• Observable reference levels from the trailing 52-week range ($164.27 to $237.96): the auto-computed retracements sit at $209.81 (0.382), $201.12 (0.5) and $192.42 (0.618), all below the 2026-10-02 close. They are reference points recomputed daily, not trade levels or targets.
• Bubble-level shift: if NVDA's correlation with TSMC and the equipment names breaks down, the "one AI capex trade" read of the semiconductor bloc changes.

Live data on this ticker: /stocks/nvda. Price, ETF holdings, bubble correlation, curated levels, bot positions.

Bubble context: /bubbles/semiconductors. The cluster this name belongs to and how it's moving.

QuantAbundance is educational research. Nothing here is investment advice. See /disclosures.