Head of Research @Silicon_Data. @UMICH/@ETH_en/@UVA alum. Perpetually curious, but also "incredibly unsophisticated" (according to Chamath Palihapitiya)
New York, USA
- Tweets 92,996
- Following 5,655
- Followers 61,448
- Likes 169,637
I've often been asked why I joined @silicondata. Was it bc I was extremely convinced of the compute market? While I was bullish, that's not why. I've only seen in my life maybe a handful of people that work as hard as @carmenli. But she is the only one, who does so effortlessly.
Our CEO @carmenli was featured in cover story in the @WSJ Markets & Finance section today. Thanks to @rudegeair and @candicechoi for covering Carmen's life journey and the AI compute market! [Link below]
2
12
1,872
Steve Hou retweeted
Bears out discussing AI's vulnerability to a VC funding retrenchment
IF VC funding tracks the last hike cycle, ~25% of lab ARR is at risk
However, the likelihood that the current tightening tracks 2022 is v low + ex-labs VC funding is below last cycle's peak (excerpt 9/17 3FR)
21
23
217
25,161
Dynamic pricing is cool 😎. It’s good for market efficiency, promotes price discovery, increases liquidity, and lowers costs for all. Kudos to @nebiusai for starting this and hope more will join soon! 😊
🚨New: Our Head of Research @stevehou breaks down the implications of @nebiusai’s new dynamic pricing for spot/preemptible GPU compute.
silicondata.com/blog/gpu-spo…
1
2
18
2,460
Steve Hou retweeted
🚨New: Our Head of Research @stevehou breaks down the implications of @nebiusai’s new dynamic pricing for spot/preemptible GPU compute.
silicondata.com/blog/gpu-spo…
1
5
17
3,779
Steve Hou retweeted
🐐
Only index provider at NVIDIA’s GTC analyst Q&A, so of course I asked @JensenHuang about GPU residual values. 😆
His take: newer is better, but older GPUs matter when everything is sold out.
He knows exactly what we do. Then he asked: “How did you get your data, and how can we help?”
Surreal. 💚 Also love running into Alison!
1
19
3,051
Steve Hou retweeted
We trained the #1 decision models in the world, mammoth-1 and elephant-1!
This is our first Base Labs model release and we're excited to open-source all weights and training details. We really focused on scaling high-quality data in these training runs – not only synthetically generating new decision data but stacking how we validated our data. This follows something called our 'swiss cheese' approach to looping over and improving training data.
There are also a bunch more low hanging fruit when it comes to training these models, and we expect that we'll improve quickly over the coming weeks.
Today we release mammoth-1 and elephant-1!
We post-trained mammoth-1, the #1 ranked best decision model in the world, and elephant-1, which is the best model under 15B (and beats Jev).
We're open-sourcing all model weights and provide a technical report detailing all our training and data methods.
2
7
45
2,885
👀
1/ Are we overbuilding compute for AI?
Introspection on eve of release of DeepInfer - the most economically valuable software release of history (on par with Claude Code).
DeepInfer has even a bigger money printer, if your are to believe latest numbers compared to what @SemiAnalysis_ reported.
They can make 149B revenue from 1GW of GB200 without slack, and 100B with very big slack at DeepSeek's current prices, while serving fast tokens at 230 tokens/sec per user at 38$ per GPU hr of GB200. The performance is reproducible. (I will explain the calculations in following tweets)
These numbers are much higher than numbers reported by SemiAnalysis below at 15B revenue per GW with B200 GPUs (not GB200) with the same model DS Flash 4.1. The difference shows the power of NVL72 architecture are large scale up domain.
Large domains afforted by NVL72 allow to have large scale up domain for decode of 64 GPUs (WideEP). This allows for much less use of HBM on each GPU for weights (spread out), and more use for long sequences and batches. This allows them to serve 200+ tokens per second per user and still make 38$ revenue per GPU.
If those numbers check out, implications are massive!
Impact on demand and supply:
---------------
Now imagine with 50GW compute by 2027 that SemiAnalysis claims will come online (I trust them), even with 50%/25GW going to training and post training, how much would have to be token consumption in USD to use up all 25GW at DeepSeek rates? 2.5T USD off course (25 GW * 100B / GW)
Leading US labs are more expensive in terms of token rates (say 4 times), so revenue would have to be 10T USD ARR to consume 25GW inference compute by 2027 end.
That is 8% of Global GDP of 125T USD. This is at the time revenue may be slowing down, if we are to believe all the rumours.
Industry Implications
---------------
But, this newfound DeepSeek efficiency will make Meta Muse much cheaper to serve, for example.
Overall, assuming US labs already have such efficiencies or they incorporate them, this is superbullish for US labs, as well as, existing hyperscalar who may be able to charge much more for existing GPUs by slowing down build outs beyond 2027 (with LTAs it is hard though).
Coming back to question: How does 1 GW of GB200 compute turn into ~$148B/year (100B with slack) in revenue?
The math behind the token economics vs. standard cloud rental 🧵👇
CC: @firstadopter @bubbleboi @GavinSBaker @zephyr_z9 @teortaxesTex
kingy.ai/blog/deepinfer-deep…
4
2
17
6,303
Steve Hou retweeted
Something I haven't explored in my posts on RSI and would love ideas on: Using AI to create synthetic training data and RL environments.
We hear that most AI recent AI gains are coming from RL and synthetic data. Could AI be used to accelerate the production of either of those?
If you have opinions or data, I'd love to hear them!
🧵New post in my AI series: Can a million parallel agents create superintelligence? TLDR: More agents buy speed, but intelligence gains diminish fast, and much faster than RSI models. Thanks @noahpinion for the encouragement! 1/25
25
10
56
9,663
Nonsense is exactly right.
Maybe in some future time, we could have a scenario where supply outstrips demand (hard to imagine based on currently trajectories but anything is in theory possible), but "at the moment" we absolutely do not have more compute capacity than we have demand.
Clearly a lot of compute capacity has come online and even more will come in coming years. Meanwhile, compute costs are also surging and have been surging all year with on-demand GPU rental rates hitting ATHs. Econ 101: when price and quantity are both increasing, it means demand increases (demand curve has shifted outward such that there's more demand at every price)!
“At the moment, we have more compute capacity than we have demand [...]”
Ok @Gartner_inc, this is nonsense. We are *massively* compute constrained. Our portfolio companies are struggling to find capacity.
Full article here: cnbc.com/2026/10/19/gartner-…
12
11
84
9,908
Steve Hou retweeted
I find it very comical that people are debating whether AI is a bubble. Ofc it’s a bloody bubble! The right question is how proverbially old it is and when it’ll drop dead of old age and ill health. Personally I think we may be still just going through some volatile adolescence.
I’d be very disappointed in y’all if the AI bubble isn’t bigger than the crypto bubble. Let’s make AI the biggest bubble of all time, and party like it’s 1999.
5
1
64
19,945
AI buildout is happening at a breakneck speed. Some may question why we need to be in such a rush or urgency. The fact that this is a gold rush is certainly one reason. But perhaps the following is a good reason. Literally lives may be on the line if AI could expedite life-saving drugs or therapeutics. I speak as someone who knows absolutely nothing about medicine and whether AI can actually do that.
“Our virtual biologist may have compressed a year's worth of work into a couple days, but biology works at the speed of life…Over two days, we watched Claude work through problems at a speed that was difficult to comprehend. For the next two months, we spent much of the time becoming familiar with laboratory turnaround times (This mostly involved sending memes in Slack).”
Where do you go next after reporting from the Strait of Hormuz? For the past three months, Citrini Research, an independent research firm, has been working with our team and an internal research model to see if they could use AI to create a new drug. What follows is an entertaining, candid, and humbling account of a group of motivated nonexperts (or “filthy casuals”, as Citrini would say) trying to tackle a real problem in human health with Claude.
We landed on scorpion neurotoxins, which kill over 3,000 people a year. Of the 211 designs we produced and tested, 17 met our criteria for a binder. None of the designs bound more than one toxin and none have been tested for neutralizing a toxin.
One of my favorite quotes from the piece is: “Our virtual biologist may have compressed a year's worth of work into a couple days, but biology works at the speed of life…Over two days, we watched Claude work through problems at a speed that was difficult to comprehend. For the next two months, we spent much of the time becoming familiar with laboratory turnaround times (This mostly involved sending memes in Slack).”
6
2
34
6,386
Damn @citrini got cited in an Anthropic science blog! They are famous now!
New on our Science Blog: designing protein binders against scorpion neurotoxins.
Scorpions sting about 1.2m people every year, and antivenoms are often costly or in short supply. Claude has made an early step toward finding new treatments.
anthropic.com/research/claud…
7
2
73
9,671
That’s why they are called neoclouds! There are always new ones coming up.
There's a bunch of neoclouds out there now that nobody's ever heard of. bloomberg.com/news/articles/…
3
4
31
9,419