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.

And crucially, chip imports are building the US capital stock — presumably increasing future GDP as AI capabilities expand. (6/6)

But chip imports from Taiwan are fundamentally different from tube socks from China. Chips have an enormous amount of US IP value embedded in them that gets added back in when GDP is calculated. (5/6)

We imported more from Taiwan than China as tariffs and AI trends collide. (1/6)🧵

QUESTION MARK: One of Elon’s companies has limited employee token usage to just $200 per week, while another Elon company is saying everyone will use so much AI that we’ll need space data centers because Earth will run out of power, plus TeraFabs to produce a terawatt of chips per year. How does this even make sense?

Hey Siri, sue OpenAI

Correct call by NVIDIA tbh. Why “waste” wafers and GDDR on fun when Rubin pays triple? So gamers get the only GeForce cards the margin math allows: the paper ones. Enjoy the mealworm buffet, fun is a datacenter-exclusive feature this gen, and the rest of us are getting GeForced onto the bug diet.

The bottom line is this: even though we expect co-packaged optics (CPO) to eventually be the go-to bandwidth scaling solution for optics, scale-up CPO will have to work through the integration, yield, and ecosystem maturity curve before it is ready for volume manufacturing and adoption ahead of the 2029 volume adoption timeline that we are anticipating. In the meantime, coherent-lite technology, which has better signal-to-noise ratio (SNR) tolerance and scales bandwidth using a higher-order modulation scheme (DP-16QAM) than traditional IM-DD PAM4 transceivers, is one of the many pragmatic solutions for optical bandwidth scaling, particularly in the high-bandwidth scale-up domain. That said, coherent-lite carries several structural advantages beyond the Google ICI scale-up use case — 2-20km reach versus IM-DD's sub-2km ceiling, fewer lasers per bit of bandwidth, and reduced fiber count — that make it a credible deployment option across other intra-datacenter interconnect tiers as well. (2/2)

The coherent-lite transceiver technology was presented by Google in 2021 for an IEEE meeting as an alternative to the shorter-reach, traditional CWDM4 IM-DD transceiver, and we are finally expecting to see volume adoption at Google. As flagged in May to our institutional clients, Innolight is set to ship 2.4T coherent lite transceiver modules for TPUv9's scale-up network, with Marvell working on the custom DSP. (1/2)🧵

@alexandr_wang @creatine_cycle

@creatine_cycle ok… and?

he's not gonna fuck you bro

The Future of Meta Superintelligence: A 1 Year Progress Update A top tier RL environment startup spawns out of thin air, the most aggressive compute ramp we've ever seen, 2000km+ scale-across, and some advice for Google DeepMind semianalysis.substack.com/p/the-future-o…

We do a full breakdown of the BoM and power budget in the VR NVL72 Component BoM Model. It provides a system-level bill of materials and power budget analysis for the VR NVL72 system. 7/7🧵

The compute tray topology of VR NVL72 is roughly similar to that of GB200 and GB300, though there are some differences. 6/7🧵

Vera Rubin NVL72 is a beast, thanks to its extreme co-designed system architecture, and First At-Scale Production Token (FASPT) is fast approaching. Below, we break down the architecture and some of the latest manufacturing innovations.👇️ 1/7🧵