GTC 2026, the Structural Evolution of Semiconductors, and Alsemy

GTC 2026, the Structural Evolution of Semiconductors, and Alsemy

GTC always gets me excited. It takes me back to the days when I used to tinker with graphics cards as a kid, or when I was fumbling around trying to build a crypto mining rig. What will NVIDIA pull out this time? Will Jensen Huang finally swap out the leather jacket? Probably not. Either way, the entire industry is holding its breath as NVIDIA prepares to unveil its next-generation GPU roadmap.

But honestly, I'm less interested in the announcements themselves and more curious about what's happening underneath. That's usually where the real investment insight lives.

Why did GB200 have such a rough landing?

The GB200 took far longer to gain traction than most people expected. You'll hear all sorts of explanations — supply chain issues, software optimization challenges — but my gut says a big part of the problem was heat. If you've ever stepped inside a crypto mining facility, you know exactly what I mean. The air temperature can push close to 60 degrees Celsius in there. Anyone who's been inside knows. You can almost feel the machines aging in real time.

In a data center, heat isn't just uncomfortable — it's an uptime problem. These facilities run 24/7, and when a chip crosses a certain temperature threshold, it either throttles its clock speed or shuts down entirely. I vaguely remember reading in Sedra's textbook thirty years ago that channels start saturating around 90 degrees — that memory surfaced briefly and then disappeared. The point is: the moment a chip misbehaves, SLAs get broken, and customers start calling. No data center operator takes that lightly. It's also why they're so reluctant to adopt new technology — the stakes are real.

That context makes it easy to understand why the industry is chasing solutions like glass substrates and new packaging materials. JEDEC's recent decision to loosen the thickness specifications for next-generation HBM caught my attention too. Thicker HBM means more flexibility in thermal design. I read that as a deliberate move to reduce heat-related errors. And as a side effect, it seems to give Samsung — a relative latecomer in this space — a bit more room to breathe.

The structural shift inside the GPU — SRAM is getting bigger

There's another, more fundamental design direction quietly taking shape. GPU makers are increasing on-chip SRAM capacity to reduce how often data has to travel to and from external memory. I've also heard that the speed requirements for next-gen HBM are being relaxed somewhat — again, likely with thermals in mind. Every time data leaves the chip and comes back, it burns energy. If you can keep more of the work inside the chip itself, efficiency goes up meaningfully. As AI workloads continue to scale, this kind of architectural evolution feels inevitable.

Which brings me to an interesting question.

Could Samsung Foundry actually catch a break here?

Dense, high-quality SRAM integration is a core foundry capability — one that Samsung is widely regarded as strong in. Meanwhile, NVIDIA has to be at least a little uncomfortable with how concentrated its supply chain has become around TSMC. Apple, NVIDIA, AMD — they're all competing for the same leading-edge capacity. For NVIDIA, diversifying that supply isn't just a nice-to-have; it's basic risk management. I have a lot of friends at Samsung Foundry, so I'm rooting for them. You know who you are — fighting!

If Samsung Foundry can leverage its SRAM integration strength to win even a slice of NVIDIA's GPU production, that's a meaningful catalyst for the stock.

So... should you just go buy Samsung stock? Honestly, I'd slow down a little. Foundry yield still needs to get there, and the real momentum will come when a concrete order gets announced. For now, this is a thesis worth tracking — not necessarily acting on yet.

And then there's Alsemy

There's someone I really wanted to bring to GTC this year: Hyunbo Cho, CEO of Alsemy, one of our portfolio companies. March is peak season for shareholder meetings and fund reporting — one of the busiest stretches of the year — so making the trip just wasn't possible. That part stings. But what matters more than being at the conference is what kind of news comes out of it.

Here's the thing: when chip architecture changes, the design blueprint changes with it — and that means a whole cascade of process parameters has to be recalibrated. That's exactly where design automation tools earn their keep. In a period like this, when GPU architectures are evolving as fast as they are, the ability to rapidly translate new designs into process-ready specifications becomes incredibly valuable. That's what Alsemy does, powered by AI. I'm convinced they'll become an indispensable company in semiconductor process design. I just wish I could have experienced those conversations happening in real time.

I keep saying I'm coming up on twenty years in venture investing — it's practically a reflex at this point. Over that time, the moments I'm most proud of are, of course, the exits. But the moments that feel most real are when I hear that a company I backed played a meaningful role in changing how the world works. That's what it's really about. I'm hoping to hear a lot more of those stories from Alsemy. Many, many more.

This was translated by Sonnet 4.6.