When ordinary people hear “NVIDIA’s GPU,” the picture in their heads is: a chip that NVIDIA made from start to finish. But the reality is that NVIDIA doesn’t manufacture a single chip itself.
The H100 or B200 you have in your hands today is a product created through the extremely refined collaboration of three companies — NVIDIA does the design, TSMC does the manufacturing, and SK Hynix (or Micron, or Samsung) provides the HBM. Each of the three companies takes a slice of the profit along this value chain.
Understanding their division of labor means understanding the profit distribution structure of the semiconductor industry.
First, let’s look at what NVIDIA does.
It does GPU architecture design, the CUDA software stack, and the sales network. The circuit diagram, transistor layout, and programming interface for the entire chip all come from NVIDIA, but it doesn’t have a single wafer fab line.
NVIDIA hands the design over to TSMC, and TSMC builds according to the blueprint, turning silicon wafers into chips. This model is called Fabless (no-fab design), and NVIDIA is the world’s most typical Fabless company. Its gross margin reached about 75% in fiscal 2026, which is almost unimaginable in the hardware industry as a whole — it captures the profit from the “design and ecosystem” segment, with almost no manufacturing cost.
TSMC is the true manufacturer of this chip.
It does wafer foundry work, and after receiving design files from NVIDIA, AMD, and Apple, it turns them into physical products in its own factories.
TSMC’s moat lies in advanced process nodes and advanced packaging — 3nm, 2nm, CoWoS. No one else in the world can do these processes. Its pricing power is reflected in wafer prices: a 2nm wafer sells for about $30,000, and 3nm wafers will rise another 15% in the second half of 2026. TSMC’s gross margin is also around 55%. Although that is not as high as NVIDIA’s, it is the only link in the entire AI industry chain that can “hold up capacity” — if it doesn’t approve, NVIDIA can’t expand output either.
SK Hynix does HBM, which I’ve already discussed in the previous article.
Its position is to “stack another layer of scarcity on top of NVIDIA’s shipment pace,” and its gross margin keeps being lifted as HBM3E and HBM4 iterate forward. It is not the core manufacturer of the GPU, but it is the only memory partner that can make the GPU truly “run.”
Once you see this triangular division of labor clearly, you’ll understand one thing:
The profit from AI GPUs is not eaten up by NVIDIA alone; it is distributed along the chain of “NVIDIA — TSMC — SK Hynix.” Every Blackwell that rolls off the line, NVIDIA takes about 60% to 70% of the value, TSMC takes about 10% to 20%, and the HBM vendors take a little over the remaining 10%. Each of the three companies has a globally dominant position in its own segment, which together supports this industry.
This also explains why the stock prices of these three companies have moved together over the past two years. They are not in a competitive relationship; they are in a symbiotic relationship. The more NVIDIA sells, the tighter TSMC’s CoWoS becomes, and the more expensive SK Hynix’s HBM gets. If you only buy NVIDIA, you only capture one segment; but if you understand the whole chain, you can see which segment has the thickest profits and is the hardest to replace.
This is the next layer of the “sell shovels” law — the shovel itself is not made by one company either, but by an entire industry chain working together.