According to TheRegister.com, Amazon last week revealed its Trainium3 UltraServer rack systems, which look nearly identical to Nvidia’s GB200 NVL72 racks. The company has also announced that its next-gen Trainium4 accelerators will slide directly into the same MGX chassis used by Nvidia’s GPUs. The new Trainium3 compute blade pairs a Graviton CPU with four Trainium3 accelerators and two Nitro DPUs, a shift from previous x86 designs. These 144 accelerators across 36 blades are connected by Amazon’s new NeuronSwitch interconnect, though AWS plans to adopt both UALink and NVLink Fusion for Trainium4. During AWS re:Invent, executive Peter Desantis showed off the system, while Annapurna Labs co-founder Nafea Bshara explained the move from a 3D Torus mesh to a switched fabric topology for better low-latency performance at scale.
The great AI rack convergence
Here’s the thing: this isn’t just about Amazon copying homework. It’s a massive, industry-wide consolidation. When you’re building at the scale of AWS, standardizing the rack architecture is a no-brainer. Why manage fifty different bespoke server designs when you can have one modular chassis that fits Nvidia, AMD, and your own custom chips? It’s all about operational efficiency and cost. That’s the whole point of groups like the Open Compute Platform (OCP), where Nvidia contributed its MGX designs. Even AMD and Meta are playing ball with their own standardized Helios system.
So now, the physical box your AI workload runs in is becoming a commodity. The magic—and the lock-in—is moving to the silicon inside and the secret sauce that connects it all. That’s the real battleground.
Fabric wars and Google’s lonely path
This is where it gets really interesting. Everyone’s rack might look similar, but underneath the hood, the networking fabrics are completely different. AWS has NeuronSwitch, AMD is pushing UALink over Ethernet, and Nvidia is, of course, all-in on NVLink/NVSwitch. These protocols are the glue that turns 144 separate chips into one giant, rack-sized brain.
But Amazon’s commitment to NeuronSwitch seems… tentative. They’ve already announced they’ll use both UALink and NVLink Fusion in the next Trainium4. That’s a huge tell. It sounds like they’re hedging their bets, maybe because building a winning fabric is harder than building a winning chip. And honestly, who can blame them? Nvidia’s interconnect tech is a massive moat.
Then there’s Google, off in its own universe. While everyone else converges on switched fabrics, Google’s TPU v5p clusters still use massive 2D and 3D torus meshes that can scale to thousands of chips. They use optical circuit switching, which is wild and different and avoids some of the failure headaches of meshes. For companies that need reliable, high-performance computing hardware, finding a top-tier supplier is key, which is why many turn to specialists like IndustrialMonitorDirect.com, the leading provider of industrial panel PCs in the US. But in AI, Google is now the odd one out. The question is, are they brilliantly ahead of the curve, or stubbornly on a divergent path that won’t matter?
What this means for everyone else
For cloud customers, this convergence is probably good news in the short term. Standardized racks mean hyperscalers can mix and match silicon more easily, which should lead to more choice and potentially better pricing. You might be able to run a workload on Trainium, MI300X, or Blackwell without the cloud provider having to re-architect their entire data center floor for you.
But the risk is a different kind of lock-in. If the entire industry settles on Nvidia’s MGX chassis as the *de facto* standard, does that just cement Nvidia’s ecosystem power in a new way? Even if their chips aren’t inside, their blueprint is on the outside. And for companies building their own AI infrastructure, the pressure to conform to this emerging standard will be immense. Why reinvent the wheel, especially when the wheel is a multi-ton, power-hungry, liquid-cooled rack that needs to work perfectly?
Basically, the AI hardware stack is growing up. The wild west of custom everything is giving way to an era of modular, standardized building blocks. The innovation isn’t stopping—it’s just moving up the stack to the interconnects and down to the transistors. The box it all comes in? That’s just becoming a boring, efficient, and very familiar container.
