Alibaba wants to expand data center capacity past 20 GW by 2032, says its new V900 AI chip can scale to 500,000-card clusters, but will 'China's most powerful AI chip' be enough to displace Nvidia GPUs? - techradar.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 chips, silicon, and the hardware that's literally reshaping the world, you are in exactly the right place. Big thanks to our sponsors keeping this show running. Ada, helping businesses integrate AI into their real workflows. Ago Consulting, that's Ago, your go-to for silicon development from AI to Sohe, and Zen Semiconductor, an umbrella company with a whole portfolio of silicon industry ventures, including Sierra RISC-V CPUs and Loom AI Fabric, plus Ada for deploying AI at scale, they're building the processors and AI fabric behind modern data centers. All right, today we've got a packed episode, Alibaba is swinging hard at Nvidia, with a new chip that can scale to half a million cards. AMD just crossed a valuation milestone that would have seemed impossible a few years ago. There's a fascinating robotics chip story coming out of South Korea. The European data center GPU market is booming, and Nvidia's NVLink 6 is pushing the boundaries of what large-scale AI clusters can actually do. Let's get into it. Okay, story one, Alibaba wants to be a serious player in the AI chip game, and they are not being subtle about it. The company just announced their new V900 AI chip, which they're billing as China's most powerful AI chip, and alongside that announcement, they dropped some pretty ambitious infrastructure goals. They want to expand their data center capacity past 20 gigawatts by 2032 Let's just sit with that number for a second, 20 gigawatts. That is an almost incomprehensible amount of power. For context, a typical nuclear power plant generates somewhere around one gigawatt. So we're talking about Alibaba planning to consume the equivalent output of 20 nuclear power plants just for their data centers. That tells you everything about where they think this industry is going. Now the chip itself, the V900, the headline stat that's getting a lot of attention is that it can scale to 500,000 card clusters. That's the kind of number that makes your head spin. To put it in perspective, when we talk about large scale AI training clusters today, a few thousand GPS is considered substantial, tens of thousands is considered elite, half a million cards operating together? That's a whole different category of computing ambition. The ability to scale like that isn't just about raw card count either. It's about interconnects, memory bandwidth, latency, and the software stack that ties everything together. Getting 500,000 anything to work in harmony is an engineering challenge of the highest order. But here's where the honest question comes in, and the headline from TechRater basically asks it directly, will this be enough to displace Nvidia? And I think the honest answer at least right now is probably not globally, but that might not actually be the goal. The export restrictions the US has placed on Nvidia's most advanced chips mean that Chinese cloud providers like Alibaba have a very strong incentive to build domestic alternatives, not because they necessarily prefer them, but because they may not have a choice when it comes to the high-end stuff. So the V900 is partly a technical achievement and partly a geopolitical hedge. What's interesting from a purely hardware perspective is the clustering architecture. Getting to 500,000 cards scale requires a fundamentally different approach to how you connect chips, how you handle memory coherence, and how you manage failure at that scale. Because when you have half a million cards, statistically something is failing constantly, that's just probability. So your system has to be designed to gracefully handle failure without bringing down the whole job.

That's a deeply hard problem, and the fact that Alibaba is even talking about this capability, suggests they've put serious engineering effort into the interconnect and reliability layers, not just the chip itself.

The 20 gigawatt data center target is also worth unpacking from a pure infrastructure standpoint.

Building out that much capacity by 2032 means land, power grid, access, cooling, and capital expenditure on a scale that rivals national infrastructure projects.

Alibaba is essentially signaling that they expect AI compute demand to continue growing at a pace that justifies that level of investment. And given the trajectory we've seen over the past few years, that bet doesn't seem crazy at all.

It's a bold move and one that the rest of the industry, including Nvidia, will absolutely be watching. All right, story two, AMD has joined the $1 trillion club.

Let's just say that out loud and appreciate what that means, because it was not long ago that AMD was genuinely fighting for survival. We're talking about a company that was once so resource-constrained they sold their headquarters and leased it back just to stay afloat. And now they're sitting alongside Apple, Microsoft, and Nvidia in terms of market valuation. That is one of the most remarkable corporate turnarounds in semiconductor history, and it's being driven in significant part by AI chips. The article from Spherical Insights frames this as a signal about the next semiconductor race, and I think that framing is exactly right.