TSMC’s Record June Revenue Signals a New AI Chip Pricing Supercycle - semivision

**Skyler Monroe** (0:10)
Hey everyone, welcome back to The AI Hardware Show. I'm your host Skyler Monroe. And if you're into chips, silicon, and the hardware actually powering the AI revolution, you are in exactly the right place. Huge thanks to our sponsors today, Limitless AI, helping businesses actually integrate AI into their real world workflows, and a Go Consulting that's, ah, Go your go-to for silicon development consulting. All right, we've got a loaded episode today. TSMC is breaking revenue records, and it might signal something much bigger, China is pushing hard on domestic chip tech, with some wild new architectures. The FBI is shopping for AI supercomputers, and Nvidia is quietly shipping chips to China again. Let's get into it, okay? First story TSMC posted record revenue for June, and analysts are starting to throw around a phrase that should get every investor and engineer's ears perked up, AI chip pricing supercycle. So let's break this down because there's a lot packed into that phrase.
TSMC, for anyone new here, is basically the world's most important chip manufacturer. If you have a cutting edge chip, whether it's from Nvidia, Apple, AMD, whoever, there's a very good chance TSMC made it. They are the foundry. The factory of factories, if you will, now revenue records from TSMC aren't just about TSMC, they're a signal about the entire semiconductor ecosystem. When TSMC is printing money, it means someone is ordering a lot of chips, and right now that someone is the AI industry.
The term supercycle is interesting here. In semiconductor history, you've had these periods where demand just explodes, usually tied to a major technology shift. The PC boom was one, the smartphone era was another, and now people are arguing we are in the early innings of an AI-driven supercycle. But here's what makes this one different. It's not just volume, it's pricing. AI chips are extraordinarily complex to manufacture. We're talking about the most advanced nodes, the most sophisticated packaging techniques, things like COOs packaging for stacking HBM memory on top of logic dies. That complexity commands premium pricing, and TSMC is essentially the only game in town for the leading edge.
So when hyperscalers, your Metis, your Googles, your Microsofts are all racing to build out AI infrastructure simultaneously, they're all knocking on TSMC's door at the same time, bidding up wafer prices.
That's the supercycle thesis in a nutshell.
Now, what does this mean going forward? It suggests that TSMC's revenue trajectory isn't a one-quarter blip. It could be a sustained multi-year run. For hardware folks, this is validation that AI infrastructure buildout is very real, very large and accelerating. Keep an eye on TSMC's numbers as a leading indicator for the health of the entire AI hardware space. When TSMC sneezes, the whole semiconductor world feels it.
Alright, story two, and this one is genuinely fascinating from a pure engineering standpoint. China's DFSX has unveiled what they're calling China's first 35D infinity chiplet architecture, along with 3D DRAM technology, and they've announced the DF1000 and AI accelerator built entirely from China's domestic supply chain.
Okay, so let's unpack all of that, because there are several different things going on here. First, why does China need to do this at all?
The short answer is export controls. The US has been tightening restrictions on what chips and chip making equipment can be sold to China, particularly around cutting-edge AI accelerators and high bandwidth memory, or HBM, which is the specialized memory that makes AI training and inference efficient. HBM is largely dominated by Korean companies like SK, Hynix and Samsung, and US export controls have made getting that into China increasingly difficult. So China is trying to build the whole stack domestically.
Now, the DF1000 itself, it's described as a software defined near in memory computing AI accelerator. Let's decode that.
Near in memory computing means you're doing computation very close to, or even inside the memory. Think about it this way. Traditionally, your processor and your memory are separate, and data has to travel between them constantly. That travel takes time and energy. Near memory computing shrinks that distance dramatically, which can be huge for AI workloads that are often memory bandwidth bound. Then there's the 3D DRAM angle. Instead of laying memory chips flat side by side, you stack them vertically, kind of like building up instead of out when you're running out of land in a city. This gives you much higher bandwidth in a smaller footprint. And the 35D chiplet packaging, this is their Infinity chiplet branding, sits between traditional 25D interposer-based designs and full 3D stacking, giving them more flexibility in how they assemble complex chips. Now, the really significant part here isn't just the technology, it's the domestic supply chain claim. If DFSX can actually deliver a competitive AI accelerator using only Chinese-made components and manufacturing, that would be a major milestone in China's push for semiconductor self-sufficiency. I'd be cautious though, domestic supply chain doesn't always mean leading edge, there's likely still a performance gap compared to what Nvidia or even AMD is shipping. But the direction of travel is clear, China is not waiting around. They are building their own path, and stories like this are going to keep coming.