Patel explains agentic training only needs to sync tokens every few minutes versus weights every seconds in pre-training.
“When you're doing these rollouts and especially as things get more and more agentic and training, you might not only need to send not the entire weights but just the tokens that are relevant.”
Patel notes semiconductor industry doubles transistors yearly but energy industry wasn't prepared for similar growth.
“Whereas the energy industry in America wasn't. And and so, like, initially, people were, like, not creative. They're like, let's do let's do these kinds of gas plants.”
Patel quantifies InferenceMax uses tens of millions of dollars in GPU hardware with daily software updates.
“we have tens of millions of dollars of hardware of the current and last generation GPUs. As I mentioned before, the software updates every single day.”
Patel contrasts traditional data centers with minimal retrofit costs against modern GPU deployments requiring major upgrades each generation.
“it used to be you'd build a data center and there wouldn't really be much retrofits cost, right? You would slide servers in and out, you would rent racks to people.”
Patel describes Meta's temporary tent-like data centers designed for single GPU generation lifecycles instead of 30-year buildings.
“Meta has got a data center design that's only built to last for a single generation. They put them in these tent buildings.”
Patel reports new Trump tax bill allows year-one GPU depreciation, worth $10 billion annually to Meta alone.
“the new the new Trump, you know, tax bill institutes something really incredible, which is that you can depreciate all of the GPU cluster cost in year one, which we put out, like, a note about how, like, the tax implications to, like, meta are, like, $10,000,000,000 a year. And across each of the major hyperscalers, it's, like, massive.”
Patel reveals Meta is building temporary tent structures to house GPUs because permanent buildings take too long to construct.
“Meta is literally building these, like, temporary, like, tent structures to put GPUs in because building the building takes too long, and it takes too much labor, as you mentioned labor.”
Patel describes former Yandex workers paid bonuses for speed who use drugs to complete data center buildouts faster.
“Like and they get paid bonuses for being faster, and therefore, do, like, certain drugs to be able to finish the build outs faster.”
Patel notes CoreWeave has fully liquid cooled data centers while hyperscalers are still working on designs.
“They're doing full liquid cooling as well and they're working on those designs, but, you guys have fully liquid cooled data centers.”
Patel argues that efficiency gains alone without capability increases cannot justify massive AI infrastructure investments.
“That would not pay for all of these build outs. Right? AI is useful today, but it's not capable of doing a lot of things.”
Patel says certain states will see over half their power consumed by AI within a few years.
“When we look at power consumption by state, something pretty simple, there are certain states where more than half their power is going be consumed by AI just in the next few years.”
Patel identifies a 40 gigawatt power shortfall for US data centers even with all planned capacity additions.
“There is an over 40 gigawatt hole in power for The US. Even if we don't turn off any more coal, in fact we restart some, we take all the natural gas dual combined cycle reactor deliveries that we're going to have, and you add in renewables, you have a 40 gigawatt hole of power production in The US.”
Patel identifies a 40 gigawatt power deficit in US even with all planned gas and restarted coal.
“Even if we don't turn off any more coal, in fact we restart some, we take all the natural gas dual combined cycle reactor deliveries that we're going to have,”
Patel says in some Midwest regions, grid transmission costs exceed power generation costs, constraining data center buildouts.
“It's like, what the flip? Right? Like, it's like so, like, the grid needs huge investments.”
Patel forecasts data center power will grow from 150 megawatts today to gigawatt-scale clusters by 2026.
“We're seeing, like, you know, 400, 500 megawatts cluster sort of being built out now. Right? Like you know? And and then, you know, in 2026, I'd imagine it'll be, like, gigawatt scale clusters.”
Patel reports Microsoft plans 1.5 gigawatts in one site within two and a half years across four 300 megawatt buildings.
“And so the scale that they wanna reach to, 1.5 gigawatts in two and a half years in one site.”
Patel claims China could build a gigawatt data center in six months near the Three Gorges Dam but hasn't.
“China could just build it in six months, I think, around the 3 Gorges Dam or many other places. Right?”
Patel says Microsoft and OpenAI have 500,000 GB200 GPUs totaling one gigawatt coming online next year.
“Microsoft slash OpenAI slash their partners for them. And then and then potentially even more, 500 k GB two hundreds, right, is a gigawatt. Right? And that's, like, online next year.”
Patel predicts OpenAI will raise 50 to 100 billion dollars by early next year to fund planned cluster builds.
“there's no fucking way you can pay for the scale of clusters that are being planned to be built next year for OpenAI until unless they raise, like, 50 to $100,000,000,000, which I think they will raise that, like, end of this year, early next year.”
post Patel asks Douglas if he can use the model without privacy protections
https://x.com/dylan522p/status/2092262188409119109