AI Infra @awscloud · Silicon, Scaling, and Economics of Compute · Notes from depth of 100k node GPU clusters, RL optimization & inference scaling · Views my own
@TheStalwart As they say, "Your margin is my opportunity". And NVIDIA' margins (well deserved) create this opportunity.
Remember that Apple did the same in around 2008, in what was the most pivotal decision for the company.
@bpodgursky Google is buying these chips at OD prices in a deal that can be terminated any month. Either it is related to the IPO or maybe, they urgently need those GPUs for some product launch or acquisition they are going to announce.
@drewhahn He forced us to give our APIs for free to our largest customer. Zero revenue is apparently better than a million dollar ARR. He was right of course.
OPC - One Person Company - is the new big thing in China right now, with plenty of promotional credits, compute, and cowork facilities for solo founders. They will be drive by AI tools and agents.
Maybe it is a way to deal with the rising unemployment in the country!!
AI appears expensive because it is suboptimal at every layer - from NVIDIA's margins, the MFU, size of the models, the number tokens it requires to get reasonable output from the model, API margins, the model architectures - will all be at a more reasonable range soon.
@realdealpatil Wow! Thats a great analogy; the design automation company for synthetic biology. EDA is 2% of semiconductor industry. Pharma is more R&D intensive, so the tam would be much much bigger!
The AI companies going public are at the layers where capital burn is the highest. The app layer is just getting started and will probably stay private for much longer.
Chips: Cerebras, AsteraLabs, Biren, MetaX
Infra: Coreweave, Nebius
Model: Zhipu, MiniMax, Anthropic, OAI
For maximizing consumer surplus, US needs to allow its citizens to buy Chinese car brands. And China needs to allow Google in their country. Tesla feels like a clanker and Baidu is yahoo-search-lite!
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://t.co/ut6GGWQTVO
China in 2026 is one of the hardest place to travel. Very few foreign tourists, younger generation doesn’t speak English anymore, international apps don’t work and Domestic apps are suboptimal for non Chinese. Didi l/wechat are great though and people and food are amazing.