Fujitsu's Monaka AI Chip to Ship to Super Micro Next Year, Manufactured by TSMC on 2nm Process - finance.biggo.com

Skyler Monroe (0:10)

Hey everyone, welcome back to the AI Hardware Show. I'm your host, Skyler Monroe. And if you're into the chips, the silicon, the data centers, and all the hardware that's actually making AI happen, you're in exactly the right place. Big thanks to our sponsors keeping the lights on here. Ada, helping businesses actually integrate AI into their real workflows. Ago Consulting, that's Ah-Go, your go-to for silicon development from AI accelerators to full-sock design. And Zen Semiconductor, the umbrella company building the processors and AI fabric behind modern data centers, with ventures like their Sierra RISC-V CPUs and Loom AI fabric, plus Ada for deploying AI at scale in any business. Today, we've got a packed show. Fujitsu is getting serious with a two-anime AI chip headed to Super Micro. There's a wild GPU supply story out of Dubai. Samsung is throwing $230 million behind an Nvidia rival. We're digging into some dense but fascinating chip research, and Meta is making a major move with in-house AI silicon. Let's get into it. All right, kicking things off, Fujitsu's Monaka chip. So if you haven't been tracking this one, now is the time to start paying attention, because this is a really interesting development in the global AI chip landscape.

Fujitsu has announced that their Monaka AI chip is going to be shipping to Super Micro next year.

And here's the headline spec that makes this worth talking about. It's being manufactured by TSMC on a 2nm process node, 2 nanometers.

That is cutting edge. That is as leading edge as it gets right now in semiconductor manufacturing.

Let's put that in context for a second. TSMC's 2nm process, sometimes referred to as N2, is the same node that Apple is expected to use for its next generation of chips, and it represents a significant leap in transistor density and power efficiency over the current 3nm generation. When you're running AI workloads, especially large model inference or training, power efficiency is not just a nice to have, it is a fundamental constraint. Data centers have physical power limits, so if you can do more compute per watt, you can fit more intelligence into the same facility without melting everything.

Now, Fujitsu isn't exactly a household name in the Western AI chip conversation, but they have serious credibility. These are the folks behind the A64FX processor that powered the Fugaku supercomputer, which was, for a time, the fastest supercomputer in the world, so they know how to design high-performance silicon. Monaka is their pivot toward the AI accelerator market, and it's clear they're not playing around if they're going straight to TSMC's most advanced node. The Super Micro angle is also interesting. Super Micro is one of the biggest server system integrators out there. They build the actual rack-mounted server systems that go into data centers, and they've historically been very Nvidia-friendly. The fact that they're now lined up to receive Monaka chips suggests they're hedging, or at least keeping their options open. That's smart business, and honestly, it's good for the ecosystem too. More competition at the silicon level means more options for hyperscalers and enterprises trying to build out their AI infrastructure. What I'll be watching is benchmark data. The process node tells you about manufacturing quality, but the architecture matters just as much.

How does Monaka's memory bandwidth look? What's the interconnect story? Those are the things that determine whether a chip actually performs well for transformer-based workloads. We don't have full specs yet, but the TSMC2Animiefoundation gives Fujitsu a really strong starting point. This is one to watch heading into 25 if Monaka delivers on its promise. It could be a genuine contender in the AI accelerator space. That's good news for anyone who doesn't want Nvidia to be the only game in town. Okay, story number 2, and this one is a little different in flavor. There's a report out about an Nvidia H100 and H200 wholesale supplier based in Dubai announcing bulk GPU availability specifically targeting buyers in Hong Kong. And I want to talk about this because it's a window into something that's been reshaping the global AI hardware market in ways that don't always make the front page. So first, let's establish why this is even a thing. The United States has export controls on advanced AI chips specifically. Chips like the H100 and H200 are restricted from being sold to certain countries and regions due to national security concerns. These controls have been tightened and adjusted over the past couple of years, and they've created this fragmented global market where the same GPU that you can buy relatively straightforwardly in the US or Europe becomes a much more complicated procurement challenge elsewhere in the world. Dubai has emerged as a kind of hub in this dynamic. The UAE is not on the restricted list, so AI chips can legally flow there. And then you get these wholesale operations that are at least on their face, offering to bridge supply gaps for buyers in places like Hong Kong, where the regulatory picture is more complicated because of Hong Kong's relationship with mainland China. Now, I want to be clear, I'm not here to adjudicate the legality of any specific transaction. These things exist in genuinely complex regulatory gray, areas that depend heavily on end-use certifications and the specifics of each deal. But the broader point is this, the export control regime is creating arbitrage opportunities and alternative supply chains that are real and active. And this matters for the AI hardware market because it means GPU supply is not just a manufacturing and logistics question anymore, it's also a geopolitical and compliance question. If you're a data center operator in the Asia Pacific region trying to build out AI infrastructure, you're navigating a genuinely complicated procurement landscape. From a hardware perspective, the H100 and H200 remain among the most sought after accelerators on the planet. The H200 in particular added HBM3E memory, which gives you significantly higher memory bandwidth compared to the H100's HBM2E, and that bandwidth is crucial for running large language models efficiently. The demand is real, it's intense, and it's global. So these kinds of gray market or parallel market operations are really a symptom of demand colliding with artificial supply constraints. It's a story about how hardware geopolitics work in the real world, and it's going to keep being a story for as long as AI compute remains strategically important, which by all indications is going to be a very long time. Next up, Samsung is making a bold move. The company has backed an Nvidia rival with a $230 million funding round, and this is part of a broader trend that I think is genuinely exciting for the AI hardware ecosystem.