SemiAnalysis

If you want to learn more about RNG, please check our networking model. (7/7) semianalysis.com/ai-networking-…

Overall, with the reduction in layers, Amazon is claiming up to 45% cost saving compared to fat trees on a similar oversubscription ratio. As of now, RNG is the first network based on random graphs concept deployed at scale. Amazon has currently deployed only 2 version of RNG: “server mesh” that connects servers to ToRs, and “edge mesh” that connects server mesh to the peering datacenters. (6/7)

In May, Amazon unveiled its latest network topology called RNG for Resilient Network Graphs. Unlike traditional fat trees, RNG relies on random graphs concept. (1/7)🧵

@dylan522p Plus the storage story nobody's watching closely enough: PCIe Gen 6 headroom, E1.S/E3.S density, KV cache offload, and software-defined storage partnerships that keep customers out of traditional lock-in. Watch the full interview: (3/3) youtu.be/YzFWTBfA2vs?si…

Some of what's inside: launch partner for AMD Helios with public customer traction, why customers aren't pushing back on non-standard double-wide racks. Traditional data centers can't fit next-gen anyway. Plus a global supply chain that delivers in the US, Asia, Europe, and Latam. (2/3)

AMD Helios. NVIDIA Vera Rubin. Double-wide racks. Liquid cooling. And why storage just became the main event. @dylan522p sat down with @Supermicro_SMCI Chief Business Officer Vik Malyala for a wide-ranging conversation. Full interview below. (1/3)🧵

The rough framework we use, and the one the article works through end-to-end, is cost-based floor vs value-based ceiling. On VR NVL72 those numbers are roughly $4.92/hr/GPU and $12.25/hr/GPU. Theres a lot of headroom in that range, and where the market settles is the actual question for the next two years. (4/4)

A few implications we work through in the piece: 🟠 The labs can cut sticker prices on tokens, hold or expand gross margin, and still take share. That has been the actual playbook for the Opus pricing reset, not a margin sacrifice. 🟠 Memory vendors (SanDisk, Western Digital, Seagate, Micron) ran 200%+ in 2025 because they captured the throughput-scaling upside without the labs being in a position to push back. 🟠 Nvidia has more pricing power than its using, especially on SOCAMM into Vera Rubin, and we lay out where on the curve they sit and why they haven't pulled the lever yet. (3/4)

Our recent AI Value Capture piece tries to answer the question we get asked more than any other in calls with allocators: who is actually keeping the margin in this cycle, now that the easy infrastructure trade has been priced in? (1/4)🧵

State of AI Models

When will Muse Spark be open-sourced? Is Zuck trying to pull a fast one and stop open-sourcing models altogether?

It's hilarious that you think it's unfair to say vLLM maintainers have dropped AMD support in their RL framework because AMD shifted engineers away from vLLM towards ATOM. Good one, buddy. Community vLLM maintainers want to help AMD, and we want to help AMD too, but as the saying goes, you can lead a horse to water, but you can't make it drink.

yes, that's my point.. it was not from day 1 ATOM seems to be new effort from AMD team, once project matures maybe it can become community project too or maybe not and stay independant like trt-llm ATOM is also seems to be using vllm plugin architecture which is purposely built for projects like these ROCm criticism is fine, but ATOM criticism in your tweet seems unfair to me x.com/SemiAnalysis_/…

ALERT: On vLLM Kimi, which has the same model architecture as xAI’s Cursor Composer 2.5, NVIDIA mogs AMD. Also great to see B300 faster than B200.

@venkat_systems @roaner @vllm_project What are u talking about? tpu-inference is literally in the vllm-project org & endorsed by community vLLM maintainers?

@SemiAnalysis_ @roaner @vllm_project if vllm-tpu can exist, why not vllm-amd (yes not officially in vllm-project yet)? what am i missing here ? have you seen this ? rocm.blogs.amd.com/software-tools…

This is why we built ChipBook: our internal OSINT machine for the semiconductor supply chain. Trade flows, inventory, production — tracked continuously, from WFE to PCBs, so the signal shows up before the change in narrative does. (10/10) semianalysis.com/chipbook/

OSINT never gives you the full picture — it gives you the puzzle pieces. It won't tell you the date of the strike or the exact quarter the cycle turns. But it generates signals — early, verifiable, and yes, at times a little rough — that give you an intelligent edge over everyone waiting for the press release or puzzle assembly manual. (9/10)

On Feb 28th 2026, the United States launched Operation “Epic Fury” against Iran. Markets and media were caught off guard, but careful and diligent observers were hardly surprised. The signals sat in plain sight for weeks, you just had to know where to look. A thread on OSINT — and why we run some of our research the same way. (1/10)🧵