Thread by @wmhuo168

Nvidia is an embarrassment.

Nvidia-backed startup invents Ethernet memory pool to help power AI — claims it can add up to 18TB of DDR5 capacity for large-scale inference workloads and reduce per-token generation costs by up to 50% | Tom's Hardware share.google/F5sTi1m7ormiYI…

Nvidia’s new Ethernet memory pool is being paraded around as a game-changing AI breakthrough. In reality, it’s a kludge. China solved this better, years ago. Let’s break the illusion. (1/13)

Western media is drooling over a startup backed by Nvidia that attaches 18TB of DDR5 over Ethernet to servers to "reduce token cost." Sounds impressive. Until you realize what it actually is. (2/13)

It’s a memory patch for a system bottleneck Nvidia created. You burn through HBM fast, can’t scale past GPU limits, so now you offload to a glorified RAM shelf over RDMA. (3/13)

Latency is still measured in microseconds. Bandwidth is a fraction of what HBM or even a proper internal fabric delivers. With some fancy software to "tier" the memory, it sort of works. (4/13)

Meanwhile, Huawei solved the same challenge at scale. Not with Ethernet. With optical interconnects between clusters. No retrofit. No workaround. Just proper engineering. (5/13)

Huawei’s Atlas 900 superclusters use optical links to stitch AI accelerators into a single-system image. Your model doesn’t even know it’s distributed. No tiering tricks required. (6/13)

And they’re already in production. These systems are powering models like Pangu. They run long-context inference with less latency, more bandwidth, and better energy efficiency. (7/13)

While Nvidia is still piloting DDR5 over Ethernet, Huawei is scaling light-speed clusters for national labs and sovereign AI stacks. The contrast is brutal. (8/13)

Not a single tech outlet in the West will mention this. Not Tom’s Hardware. Not Ars Technica. Not SemiAnalysis. They write like China doesn’t exist. (9/13)

Because the myth must be protected. Nvidia must remain the Oracle. The U.S. must remain the innovator. China must stay invisible. Otherwise someone might start asking hard questions. (10/13)

When Nvidia builds an Ethernet RAM box, it’s genius. When Huawei builds an optical AI cluster, it’s authoritarian overreach. That isn’t analysis. That’s Cold War propaganda. (11/13)

The real story isn’t Nvidia’s workaround. It’s that the West has no answer to China’s vertically integrated, optically connected AI stack. And won’t admit it. (12/13)

Exhibit Q. Innovation didn’t die. It moved to Shenzhen. And Western media still pretends not to see it. (13/13)