Mark @GPUArbiter

neoclouds that figure out how to be power-first (not just gpu-first) will do much better than the ones that stay nvidia resellers this will help them get customers and give them freedom (both in not being locked into just nvidia, and finally not depending on a landlord for their data center) they're following the playbook of the ones who figured this out first, @CrusoeAI and @IREN_Ltd now we see other neoclouds trying to do the same: - @CoreWeave tried to buy @Core_Scientific for $9B, partly to get more power and sites (core scientific’s shareholders voted no) - @nebiusai is building its own data centers in missouri, pennsylvania, and finland instead of renting space it's the same reason inference companies want to own their gpus instead of renting them. for neoclouds, owning power matters just as much - it will get them customers faster. AI companies need more and more compute and will do deals with whoever has compute. a neocloud that owns power won't need to wait - it gives them the option to swap in new gpus, especially if they start using more than one kind of chip (which is already happening and will happen even more) - no landlord! if you own the site, you’re not paying part of what you make to a landlord

We got a delivery of our first @nvidia Vera CPU, soon to help power Modal Sandboxes.

ALERT🚨: NVIDIA SGLANG INFERENCE RUBIN IS 4.23x BETTER 🚀 PERF PER DOLLAR THAN EVEN GB300 NVL72. This is on the widely used production LLM engine called @sgl_project. (1/2)🧵

SpaceX as a neocloud

keyboard_arrow_left Previous keyboard_arrow_right Next

zerohedge @zerohedge

3 days ago

Spot Is Four Times The Hurdle: What $50 Billion-A-Gigawatt Neocloud Rents Say About The AI Revenue Gap zerohedge.com/markets/spot-f…

19 24 156 295K 103

There aren't enough performance engineers in the world to hand-tune every model, on every chip, for every workload. So we gave that job to agents. You tell us your workload and performance goals. Our agents search for the best configuration and build the entire inference stack just for your use case. Watch @gpuemi walk @furrier through how it works on @theCUBE.

there’s a reckoning in AI infrastructure inference is still printing money, but a lot of companies raised cheap capital on hype with extremely long-horizon execution risk (see companies with only 25k GPUs online filing to go public at $35B) these are obvious and important bets, but their execution needs to be flawless for these companies to grow into their valuations at higher interest rates, that’s a high bar - investors aren’t going to underwrite speculative data center bets that require everything to go right, because it won’t there are already bottlenecks across the entire supply chain, from permitting to getting PDUs delivered on-time IMO we’re going to see a correction that vindicates the belief that AI is a bubble, but customers are going to continue buying AI products, and smart capital will continue funding smart infrastructure plays however, a lot of the infrastructure tourists are going to be sent home it’s a healthy reckoning, but many will overreact and try convince you it’s something bigger it’s not

zerohedge @zerohedge

16 hours ago

Data Center debt getting smoked today. Coreweave unsecureds tag 1000 g-sprd bps for the first time ever. Oracle unsecureds trading like deep junk. Even Meta's Project Beignet about to hit record wide

78 289 2K 280K 457

BREAKING: Abdulrahman Alfozan, a 35-year-old Saudi national employed by OpenAI, has been arrested on suspicion of secretly installing an audio recording device at the company’s San Francisco headquarters, officials say.

keyboard_arrow_left Previous keyboard_arrow_right Next

“Vertical integration” is too broad for neoclouds. the two different axes: the physical stack you control, the compute/cloud/platform stack you own. everyone is integrating, but in different directions.

128 x B200 336 x RTX Pro 6000 10.8 PB Flash That’s a lot of NVIDIA compute, well done Georgia Tech!

Georgia Tech @GeorgiaTech

a day ago

Georgia Tech officially opens the Nexus AI Supercomputer, a $20 million, @NSF-funded supercomputer, creating new possibilities for researchers across the country. With 90% of its computing capacity available to universities across the United States, Nexus offers a flexible

13 58 387 21K 54

Georgia Tech officially opens the Nexus AI Supercomputer, a $20 million, @NSF-funded supercomputer, creating new possibilities for researchers across the country. With 90% of its computing capacity available to universities across the United States, Nexus offers a flexible suite of AI hardware and software to support research across a wide range of science and engineering fields. 🐝 #WeCanDoThat

以前のポストで考えてたnvme構成、DeepSeek v4.1 Flash をRTX 5090 + RAM128GB + gen5 nvme RAID0でデコード40tok/s出せました！ プレフィルはまだ調整してない推論基盤なのでプレフィル激おそですがRAM64GBに絞ってもnvmeでそれなりに速度確保できるので割と速度出ますね👀 #地元のLLM #LocalLLM

PentaCoxian @pentacoxian_dev

6 days ago

そういえばpcie Gen 5のm.2が8TB分刺さってるカード部屋に放置してたの思い出したんだけど、これでraid組めばnvmeストリーミングでも意外と速度でたりするかな

0 1 31 12K 7

The data center bottleneck may no longer be demand, but permission. While planned MWs represent data center capacity in the queue, permissioned MWs represent capacity that could realistically be built. Here’s why that distinction matters for investors. ⬇️ io-fund.com/ai-stocks/proj… $ORCL $APLD $BE

How Musk and Benioff pledged to rename SpaceXAI and AIForce, while Nadella and Nvidia adopted "super intelligence", as Silicon Valley falls in line with Trump (@TerrenceOBrien / The Verge) (Visit Techmeme dot com for the link and full context!)

I am not joking when I say this. Traditional game development is over. I asked Opus-5.5 to make Attack on Titan in Minecraft. Took ~1hr and cost less than $25 start to finish. The end result is extremely fun. And multiplayer! Guide + link to play below.

data centers are targets and inference needs to be as sovereign and decentralized as possible

Agent swarms send most new token volume to cheaper models, so agent token counts will outrun frontier-model revenue. In a swarm, a parent agent plans while child agents handle narrow jobs like searching files or running tests. Narrow jobs are usually easy to check, so developers can route them to the cheapest model that passes. And this massive token-explosion is also the major reason, flat-rate and per-seat pricing for AI software will give way to metering, because it relied on human prompting speed to cap usage. A person sends only so many prompts a day, so vendors could charge per user and still predict inference bills. But that equation will not hold with agent swarms working 24x7.

a16z @a16z

5 days ago

Your next users are machines Agents burn nearly 5x the tokens people do, up 14x in six months

78 138 1K 227K 507

🚨 CHINA JUST OPEN SOURCED AN OCR THAT EATS 100 PAGE PDFS IN ONE PASS It’s called Unlimited OCR. Only 3B params. Runs locally. Every other OCR tool chops your doc into pages and loses the thread. This one reads the whole thing in one continuous pass. → One shot full context parsing (32K context window) → Multilingual out of the box → 93% on the standard parsing benchmark (6 points over baseline) → Under 0.11 error rate past 40 pages → Runs 100% locally on your own hardware → Works with Transformers, vLLM, SGLang, Docker, Ollama, llama.cpp Traditional cloud OCR (Textract, Google Vision, Azure Doc Intelligence) costs $1.50 to $15 per 1,000 pages. This runs on your machine. For free. Forever. Baidu built it explicitly to push DeepSeek OCR one step further. Already at 1.9M downloads on Hugging Face and most people have no idea it exists yet. 100% open source.

Makes sense if you: > have enough $$$ to put down on a long term gpu rental or outright buy it + > your team has systems engineering expertise + > understand your workload really well + > can keep up with the architectural changes of models dropping every two weeks + > can keep up with the hardware cycles every two years + > do all the above well enough to extract enough tokens per GPU to justify buying/renting the GPU vs just buying tokens

Reverse engineer everything. Watch this 👇

Matthew Berman @MatthewBerman

6 days ago

Reverse Engineer Anything 😂 Software is done.

663 3K 39K 3.7M 45K

The software/API/language (CUDA) moat has collapsed thanks to modern LLM's - so trying to segment the market like this (into "gaming" vs. "pro AI") is bound to flounder long term. I can rapidly get 32GB AMD cards for peanuts and they run great, especially in pairs.

Chris Mizo @MizoChris

4 days ago

🚨 I have TWO theories on why NVIDIA may be killing the RTX 5090 and moving GB202 entirely to RTX PRO. 👀 • Theory #1: NVIDIA may be trying to tighten the GB202 supply chain and reduce the chances of RTX 5090s being diverted into China • One of the reports behind the RTX 5090

8 9 71 36K 30