Nvidia, taken apart

By Dylan Siegel

This is my understanding and opinion, for education. Not personal financial advice.

New to this? Read first: How companies make money

Nvidia sells the machines the AI boom runs on. Here's what it makes, who builds it, who pays for it, and what it earns.

What Nvidia actually sells

People say "GPUs", but customers increasingly buy a whole AI supercomputer: GPUs, Nvidia's own CPUs, the networking that ties thousands of chips together, and the software to run it. The current flagship rack, Vera Rubin NVL72, joins 72 GPUs and 36 CPUs into one system.

Who builds it

Nvidia designs; others manufacture. That chain is the real bottleneck.

Nvidia doesn't control TSMC. It's one of TSMC's biggest customers, and TSMC controls the scarcest thing in the industry: leading-edge capacity. TSMC's own numbers confirm the demand: in the quarter that ended September 2026 it made $46.7 billion in sales, up 41% from a year earlier, at a 66.4% gross margin.

More TSMC capacity cuts both ways. Nvidia can ship more. But the same factories also build chips for AMD, Google, Amazon and others, so scarcity, and Nvidia's pricing power with it, can fade.

Who pays for it

The buyers are the few companies that can spend tens of billions before an investment pays off:

On its August earnings call, Nvidia said the five biggest cloud companies are expected to spend close to $800 billion on capital projects in 2026 and $1.3 trillion in 2027. That's Nvidia's estimate, and it covers buildings, power and networking, not just Nvidia chips.

What it earns

Yearly revenue went from $60.9 billion (fiscal 2024) to $130.5 billion (fiscal 2025) to $215.9 billion (fiscal 2026). Management guided next quarter's revenue without any data-center sales to China.

Three things to watch inside those numbers:

• Commitments. Nvidia has promised suppliers $279 billion for future manufacturing and memory capacity, up from $119 billion six months earlier. That's a bet on demand, and exposure if demand slips.
• Investments. Part of Nvidia's profit now comes from gains on its stakes in AI companies, not from selling chips.
• It's financing its customers. Nvidia invested $30 billion in OpenAI's $110 billion round, and it guaranteed up to $105 billion of lease and power payments for an OpenAI data-center campus in Ohio. That's covered in the next post.

China

China is a market Nvidia is losing and a competitor it's creating. In the first half of 2026, Nvidia's share of China's data-center AI chip market fell to 34%, according to IDC. It shipped 678,000 units, down 44%, while Huawei roughly doubled to 572,000 (28% of China's overall accelerator market). For the first time, more non-GPU AI chips shipped in China than GPUs (1.02 million vs. 0.99 million).

Licensed sales to China are now a rounding error in Nvidia's revenue. The bigger risk is the opposite: export limits pushing China to build a full rival ecosystem, and, worst case, a conflict that disrupts Taiwan, where the chips are made.

Common questions

Does Nvidia make its own chips? No. It designs them; TSMC manufactures them.

Is Nvidia's lead just hardware? No. The software (CUDA), the networking and the full systems are a big part of why customers stay.

Who are its competitors? AMD, plus the big cloud companies' own chips (Google's TPUs, Amazon's Trainium, Microsoft's Maia), plus Huawei and others in China.