Something bad for the (existing) class of mathematicians as they lose their privilege, but not necessarily bad for mathematics & our body of knowledge?
The Pentagon is about to lend $5 billion to a company owned by a hedge fund that just blew up, a research analyst who rates it Gold, and three guys who met at Oxford.
WSJ broke it Thursday, Reuters confirmed Friday. The Pentagon's Office of Strategic Capital is in talks to lend roughly $5B to Fluidstack. It would be the largest loan in the office's history by a wide margin.
The loan:
- Roughly $5B from the Office of Strategic Capital, which lends to companies deemed critical to national security
- Previous recipients were rare earth firms: Vulcan Elements, Phoenix Tailings, Energy Fuels
- The money is for domestic manufacturing capacity for data center components, power and cooling gear, not for building a facility
- The Pentagon asked Congress for more than $20B for the OSC loan programme in 2027
- Context: Trump signed an executive order last month declaring a national emergency and banning some foreign equipment in the US electricity grid
- JPMorgan forecasts lenders will supply $4.1 trillion of AI infrastructure financing by 2030
Who is advising the application: Erebor, the national bank Palmer Luckey chartered in February 2026 for AI, defense, crypto and manufacturing companies. Six months in it holds more than $4.6B in deposits. A bank built for firms traditional lenders avoid is now advising on the Pentagon's largest ever AI loan.
The company:
- Founded 2017 as an Oxford spinout, originally an Airbnb for GPUs matching idle gaming rigs with researchers who could not get compute
- Gary Wu, 31, CEO, studied economics. Jamie Cox, 29, dropped out of classics and became a Thiel Fellow. Cesar Maklary, 29, worked on Formula One aerodynamics before joining in 2020
- Revenue: $1.8M in 2022, $66.2M in 2024, roughly $660M projected for 2026
- Moved its headquarters from London to New York in December 2025
- Building across New York, Indiana, Louisiana and Texas
- All three founders are now billionaires. Had they stayed in Britain they would have been among Europe's youngest self-made billionaires
The cap table, which is the story:
- January 2026: $830M Series A at $7.5B, led by Situational Awareness, Leopold Aschenbrenner's fund. Backers included the Collison brothers, Nat Friedman and Daniel Gross
- Recently: a previously unreported $1.5B round led by Jane Street at more than $18B, per Forbes on September 3
- Plus more than $15B of debt with partners including Coatue's data center arm, per SEC filings
- Valuation went from $7.5B to $18B+ in roughly eight months
Dylan Patel's position, and why it matters:
- Patel, 29, runs SemiAnalysis, the most-read chip research shop in the industry. Revenue went from about $20M last year to a projected $100M+ this year
- He raised a $50M SPV inside Fluidstack's round. Forbes estimates that SPV, named after Patel and SemiAnalysis president Doug O'Laughlin, could now hold about $140M of Fluidstack stock
- That is roughly 2.8x in under a year
- He reportedly shared a San Francisco office with Aschenbrenner
- He is now raising SemiAnalysis Capital Fund I, targeting $400M per a securities filing, with stakes already in about 20 startups including Thinking Machines Lab and Enfabrica
- Background: rural Georgia, parents ran a motel, rejected by MIT and Stanford, University of Georgia
The conflict nobody has resolved:
- Patel's ClusterMax ranking rates cloud providers. Fluidstack, which he is invested in, received Gold
- He points out that Prime Intellect, also his investment, received only Bronze, as evidence he does not favour affiliated companies
- A former employee, Wei Zhou, terminated in January, alleges Patel pushed him to include sensitive Fluidstack TPU data in SemiAnalysis reports and that he was fired for refusing
- SemiAnalysis sued Zhou first for breach of contract and trade secrets. Zhou countersued. Patel says the dismissal was for rudeness and drinking
- Clients have called SemiAnalysis asking how it handles their information. The firm is commissioning a third-party compliance report
The structure underneath all of it:
- Google backstops Fluidstack's data center deals so it can borrow cheaply. If Fluidstack fails to perform, Google assumes the lease or pays the termination fee
- Anthropic picked Fluidstack to build custom data centers in Texas and New York as part of its $50B US infrastructure commitment, leasing capacity including Google TPUs
- Other customers: Mistral, https://t.co/wGcs0jpzkB, Poolside, Black Forest Labs
- Nvidia has backed CoreWeave, Nebius, Crusoe and most other neoclouds. It has not backed Fluidstack. Fluidstack is the one helping Google sell TPUs against Nvidia
Two footnotes worth keeping. Situational Awareness sold its entire public portfolio in a single block trade weeks ago after margin calls. Fluidstack is private, so that position survived the unwind and is now marked at 2.4x. And the $18B figure comes from Forbes sourcing; the only valuation Fluidstack has publicly confirmed is $7.5B.
The frame: in May the Pentagon signed agreements to deploy AI from Nvidia, Microsoft, AWS, Google, OpenAI, SpaceX and Oracle on classified systems. That made it a customer. A $5B loan makes it a lender. A customer can walk away when a contract ends. A lender is exposed to whether the borrower survives.
The US government is now underwriting the balance sheet of a nine-year-old company whose largest shareholders are a blown-up hedge fund, a quant trading firm, and the analyst whose ratings the industry reads to decide where to put its compute.
I bought a Fable dataset from one of the top Chinese LLM routers yesterday.
With just 6TB data, I can take over 7 Chinese/CIS gov entities & 19 top Chinese firms like Xiaomi, Huawei, NIO, Minimax using SSH keys, VPN configs, Aliyun keys, GitLab tokens sent to the router.
I like to point out that in the modern world the method of influencing someone to do something is money.
Looked at that way the economy has clearly shifted in favor of things an AI superintelligence would like for eg datacenters and away from things that humans need like housing and office space.
@ikirigin that was the beauty of the original constitutional convention, drafting and then the ratification by the states. the process manufactured consensus. New York didn't really want to ratify and was forced to ratify after Virginia ratified.
If men learn this, it will implant forgetfulness in their souls; they'll cease to exercise memory because they rely on that which is written, calling things to remembrance no longer from within themselves, but by means of external marks.
- Plato
- About learning writing
"the big labs don't quite have an answer"
Baseten co-founder and CTO Amir Haghighat says sandboxing can isolate code, but providers cannot guarantee what an AI agent does inside.
"The security that is hard to guarantee is the agent that is running in this sandbox, what is it doing?"
"it seems like the big labs don't quite have an answer to either."
"by definition, like, we're zero data retention."
@amiruci@baseten
"the big labs don't quite have an answer"
Baseten co-founder and CTO Amir Haghighat says sandboxing can isolate code, but providers cannot guarantee what an AI agent does inside.
"The security that is hard to guarantee is the agent that is running in this sandbox, what is it doing?"
"it seems like the big labs don't quite have an answer to either."
"by definition, like, we're zero data retention."
@amiruci@baseten
"I don't see them volunteering to have some type of, like, let's say, arms control agreement between AIs"
Independent researcher Collin Hogue-Spears predicts US-China AI talks will produce, at most, an incident channel unless Washington offers major trade or GPU concessions.
"incident channel that might come up about based on some shared agreement of certain incidents"
"Our military people try to call their counterparts on the Chinese side and nobody answers."
@BlackDuck_SW
"I don't see them volunteering to have some type of, like, let's say, arms control agreement between AIs"
Independent researcher Collin Hogue-Spears predicts US-China AI talks will produce, at most, an incident channel unless Washington offers major trade or GPU concessions.
"incident channel that might come up about based on some shared agreement of certain incidents"
"Our military people try to call their counterparts on the Chinese side and nobody answers."
@BlackDuck_SW
"There's no power greater than the party"
Independent researcher Collin Hogue-Spears predicts repeated AI failures in China would trigger immediate CEO summons and enforced remediation, not a GPU shutdown.
"There's no power greater than the party within the People's Republic Of China."
"It will be you're being called to the capital now, and you need to explain how this is not gonna happen again and how, we can monitor this situation better."
"The first response would be collaboration consultation and probably some hand wringing."
@BlackDuck_SW
"There's no power greater than the party"
Independent researcher Collin Hogue-Spears predicts repeated AI failures in China would trigger immediate CEO summons and enforced remediation, not a GPU shutdown.
"There's no power greater than the party within the People's Republic Of China."
"It will be you're being called to the capital now, and you need to explain how this is not gonna happen again and how, we can monitor this situation better."
"The first response would be collaboration consultation and probably some hand wringing."
@BlackDuck_SW
I asked my ~25-student Probability and Decision Theory class today how many of them had read AI news this last week. Only 2 of them said they had, though a few more recalled seeing something about Navier-Stokes when I mentioned it. Only 1 has an X account. The news segments are good but the top stories at the NYT are all about the Iran War or Republican Midterm Convention. I think this issue still only has, say, 1% of the salience that the George Floyd story had a week after his death.