Hook
More breakout videos from this creator.
NVIDIA Q2 earnings 👀 NVIDIA is now earning over $1 billion a day. But the craziest part is they don't make chips. Apple makes the iPhone. If you want to know who exactly is making these AI chips, here it is in under three minutes. There's three things we need to set the stage. The first is where we are within the AI ecosystem. We're in layer two, the chips. Second is when people talk about chips, they distinguish between memory and logic chips. The poster child for logic chips is NVIDIA. But NVIDIA is only one chip maker. And the reality is they don't make chips anymore than Apple makes the iPhone. They outsource the manufacturing. So to orient ourselves with the chip manufacturing process, we're going to use the SEMI ETF SMH. SMH includes ~25 of the largest U.S.-listed companies involved in semiconductor production, design, and equipment. The four companies listed are NVIDIA, AMD, and Broadcom, but backing up a second, infrastructure players. What actually happens when an infrastructure player building a data center places a chip order with NVIDIA, AMD, or Broadcom? The first thing is the chipmaker will have to use some kind of special software called EDA software to create a chip design. The two major companies that sell the software are Cadence and Synopsys. For my fellow engineers, think of the software like CAD. NVIDIA or whoever creates the design and that is sent to the fabrication facility or fab, TSMC, which stands for Taiwan Semiconductor Manufacturing Company. It produces 90% of the cutting edge logic chips. This is why Taiwan has become so important to you politically. So what actually happens inside these fabs, and more importantly, which companies are involved? It starts with polished silicon wafers. Two suppliers, Shin-Etsu and Sumco, make up about half the world's supply. Both are in Tokyo, which is why you won't find them in SMH. The next step is really two in one. First is deposition where a thin film gets laid down on the wafer. That film is the actual material the circuit is built from. Then a second coating goes on top, a light sensitive chemical called photo resist, which acts as a stencil for what comes next. The deposition machines are from Applied Materials and Lam Research. The photo resist machine comes from Tokyo Electron. And the photo resist itself from Shin-Etsu, who you've seen before, and JSR, a Japanese company taken private by Japanese state-backed fund in 2024. With the photo resist on, the wafer goes into a machine where light bounces off a series of mirrors and prints the circuit pattern into the coating. The light marks up resist survive and therefore where the circuit ends up. This is called lithography. Only one company makes the machine for leading edge chips: ASML, based in the Netherlands. China however just built its first machine, one tier down now. The whole process repeats as the wafer goes through the same sets of machines around 80 times. The deposited film accumulates into patterned layers, building transistors at the bottom and the wiring at the top. Once every layer is done, the wafer gets cut into individual chips on a machine made by a Japanese company, which is where we get our GPU. And then gets mounted side by side with the memory, specifically HBM, which we covered in a prior video. But as a reminder, Micron and SK Hynix are two of the major players in HBM. The whole thing then sits on an ABF substrate, a film made by Ajinomoto, a Japanese company best known for inventing MSG. The last step is testing, though really it's been happening all along. KLA makes the inspection tools that check every layer as it's built. And the final electrical test comes down to two companies: Atest and Teradyne. The reality is NVIDIA doesn't sell chips, they sell racks. TSMC ships the finished product to a manufacturer like Foxconn or Wistron. They mount the chips onto NVIDIA's board, add switch trays, liquid cooling, and the whole thing becomes a rack that runs $6 million. That's what ships to the customer.