Bloomberg Tech Sep 18, 2026 44m 27m saved
With Chase Lochmiller, CEO of Crusoe, which has just closed a nearly $4B Series F at a valuation of about $31B and says it holds more than $140B of contracted value
The complaint against data centers that Chase Lochmiller hears most is about water. His answer is that each of Crusoe's large buildings in Abilene, Texas uses about as much as ten houses, because the chips are cooled on a closed loop.
Most of the week's argument was about whether AI is dangerous. Crusoe's business is the part that gets built regardless: it closed a nearly $4 billion round at about a $31 billion valuation, and the constraint it runs into is a county commission rather than a safety committee.
"I think more than anything, the data center industry has sort of a marketing issue."
Lochmiller runs a company that builds, powers and operates data centers and then sells compute and inference on top of them, with Oracle and Microsoft as data center customers and Perplexity and Figure among the buyers of managed clusters. The rest of the show covered the week that made everyone nervous about all of it.
The full episode is covered here so you can skip it. 44 minutes of audio, 17 minutes of reading.
Here are the 18 takeaways that matter.
Key Takeaways
A Crusoe building in Abilene uses about as much water as ten houses, because chip cooling runs on a closed loop
Crusoe will be about a third of the tax base in Abilene and Taylor County, and is doubling the school system's receipts
Data center commitments run 15 to 20 years; GPU cluster contracts run months to six years
Anthropic says its own model now leads 26% of its AI research and development work
SoftBank is borrowing $45–50B of the $65B it is putting into OpenAI, including against the OpenAI stake itself
The market's conclusion on pacing was that inference demand keeps the compute spend going whatever happens to training
The long-chips, short-software trade of the first half has reversed
A former FTC technologist's objection to antitrust waivers: only government-endorsed cartels are durable
Existing liability law already covers models that harm third parties, on his reading — no new statute needed to be exposed
Crusoe added three directors from Tesla, the data center industry and Cloudflare, and calls a listing a matter of timing
California's governor issued an executive order on an AI kill switch while the show was on air
1. Claude Does 26% of R&D
The show opened on a disclosure from Anthropic measuring how much of its own research and development work its model now does.
The figure reported was 26%, with the model collaborating with staff on roughly 90% of their work. The reporter's framing was careful about what the number is and is not: "Now, that's not the same as AI doing AI." What it is, she said, is one of the few figures quantifying the thing the labs are moving toward, which is AI feeding AI in order to keep improving itself.
Asked why the data appeared when it did, seven days after the essay that started the week's argument, she said nobody knows. The backdrop is that recursive self-improvement is simultaneously the business opportunity and the source of the anxiety, and that lab employees at both major companies have begun raising concerns publicly — including a former Anthropic researcher whose note went around saying he does not think any of the labs are handling it responsibly.
2. Who Set Up the Sandbox
Bloomberg's reporting on the incident behind the debate was about where the responsibility sits. The rhetoric, one reporter said, frames a sandbox that was not properly sealed as evidence that the technology cannot be contained — which sidesteps the companies' obligation to avoid safety failures. Her specific point was that the alarms did sound: "There were warnings that the human employees themselves did not take seriously."
Her colleague's version was that people are involved at every stage.
Humans set the goals. Humans built the environment: "Humans set up the situation here in the sandbox where the AI is working on things individually." Humans do the training. And humans noticed behavior during evaluations that they could have intervened on and did not. She credited OpenAI with acknowledging all of it and saying a good deal about what it is changing. The conclusion she drew is that it is not as simple as saying the AI went and did something.
3. A California Kill Switch
Mid-segment, the host read breaking news: California's governor had issued an executive order on the creation of an AI kill switch, which he said accelerates independent oversight.
The reporting connected that to a pattern. Bloomberg's story tracks the language federal lawmakers have begun using, and finds the rhetoric of doom migrating out of the industry and into policy. The researchers they interviewed made the consequential objection: a policy written from the assumption that AI threatens all humanity may not match the harms AI is actually causing now. Misinformation and surveillance, which are happening, are what gets lost.
4. Two Views From Montreal
Reporting from a conference in Montreal, the second reporter had both sides. Yoshua Bengio, who has long argued the harms are real, told her the incident was a mix of a human and a machine problem and that he is very concerned about what comes next.
Aidan Gomez, the chief executive of Cohere, gave her the other reading: "I don't think that this is going to kill people or kill us en masse, but there are some really real issues here, and those do need to be worked on." Slowing certain things might help, he said, without anyone being at imminent risk of extinction.
5. Inference Keeps the Spend
The market question the week posed was mechanical: if the labs slow training, does the compute spending fall. A portfolio manager and venture partner from Franklin Templeton, whose firm is an Anthropic investor, gave the answer the market settled on.
Her position was the opposite of the Monday reaction: "I'm definitely in the anti-doomerism camp." She is glad the founders are talking about guardrails and does not think the pace should slow. And even if pre-training slowed, the spending does not.
Her reasoning was short: "No, because inference requires a lot of compute." The evidence she offered was behavioral — the volume of compute deals those companies have signed in recent months, which she does not believe they would have done if they expected to pull back. Equities sold off on Monday on exactly that fear, and by Wednesday much of it had cleared. At an industrials conference on the Tuesday, she said, infrastructure companies exposed to data centers were emphasizing their non-data-center revenue, because they had seen the market's reaction.
On the proposal that the frontier labs get antitrust waivers so they can coordinate, she was doubtful they even want it: "I don't know if they really want widespread regulation and federal regulation." What she thinks they want is every company putting guardrails in place and not releasing a model that is not safe.
6. Cybersecurity Is the Trade
The investment conclusion she drew from a week of warnings is the same one the show had heard from a data-platform chief executive the day before and from Andrew Ng: security platforms matter more than ever.
Her framing went further than importance: "Not only are they important, but also just pacing these models." A tech conference a couple of weeks earlier had cybersecurity as its entire theme, she said, covering both listed companies exposed to it and what the frontier labs themselves are preparing to sell. New security products are coming soon.
7. Adoption Has Not Happened
Her second argument is that the debate is running ahead of the deployment. Next-generation models are arriving while knowledge workers are not yet used to the current ones, and there has been no widespread adoption.
Her expectation is that consumer agents pull enterprise adoption behind them, and her precedent is hardware: "It took a while for consumers to move from BlackBerrys to iPhones, right?" People saw the benefit and then asked for it at work.
Asked whether safety concerns are the real reason OpenAI will not list before 2027, she said no, and that some of those timelines were being pushed out anyway. The productivity gains are starting to diffuse across industries and the market has broadened as a result, which is her argument against slowing the funding.
8. SoftBank's $21B of Debt
The show's number of the day was $21 billion of fresh debt SoftBank is closing the week with. Bloomberg's Asia technology editor walked through the structure behind a $65 billion investment in OpenAI.
It comes in three slices: more borrowing against the roughly 90% stake in the chip designer Arm, a larger loan from Apollo, and a syndicated loan from 20 banks.
The total is the part that matters: "he's borrowing about 45 to 50 billion dollars out of the 65 billion dollars that he's gonna put into OpenAI." And that is not the end of it.
There is a further layer on top: "He's actually borrowing money against that OpenAI stake too." Roughly another $10 billion. His summary was that this is Masayoshi Son doing what he has always done, building a great deal of leverage on top of a startup bet.
Asked what SoftBank actually is now, he described a company that has distributed software, run trade shows and published magazines, and is now something simpler: an investment vehicle sitting on top of a telecom business that throws off cash. The telecom arm was spun off a few years ago and still generates much of the profit.
The risk he named is the combination of three things moving the wrong way: "So you've got him borrowing more money as interest rates are rising." For an investment in a company that has pushed back its listing, which delays the day the stake becomes liquid collateral. People inside SoftBank, he said, have raised questions about how concentrated the bet has become.
9. Software's Comeback
One of the year's largest trades has gone into reverse. The setup, as Bloomberg's markets reporter described it: "The first half of the year was really marked by this concern that AI was going to come in and decimate the legacy software industry." Chip stocks did well and software collapsed.
The performance gap has narrowed since, helped by a strong earnings season for software including several companies seen as most exposed to disruption — Salesforce and ServiceNow, with mixed-to-positive results from Oracle and a large quarter from Microsoft.
The conclusion drawn is about positioning rather than fundamentals: "So all of these things just helped reinforce the idea that maybe the selling had gone too far." Valuations had been washed out and everything sold indiscriminately, and buyers are returning.
Separately he flagged Netflix taking its first sell rating since late February. "Wells Fargo Securities downgraded the stock underweight," he said, on concern about engagement, with no large hits driving viewing while the company puts effort into gaming and podcasts. The stock is on track for its worst year since 2022.
10. The Antitrust Waiver
The week's other request was for antitrust exemptions so the frontier labs can coordinate, with cooperation extending to China. Neil Chilson, head of AI policy at the Abundance Institute and a former chief technologist at the Federal Trade Commission, was asked whether that is reasonable.
His first answer was that most collaboration does not need an exemption. Companies can build an industry standard or a set of best practices, and most of all they can build their products safely inside their own walls, which he said they are already obliged to do.
On regulatory capture he gave the structural argument: "But what we know from history is that cartels are fragile. The only cartels that are durable are ones that are endorsed by the government." That, he said, is the risk in the exemptions being sought — a durable cartel that harms competition and the country's own position at the frontier.
Asked whether existing law already covers a model that causes harm, he said it does, naming consumer protection law and tort law: "those models cause harm to third parties, I think the law is pretty clear." His reading is that this exposure is part of what is driving the call for pacing in the first place, and part of why the industry is telling itself to be cautious.
On what his former employer would be doing in a week like this: looking at its authorities on both the competition and consumer protection sides, looking for unfair methods of competition and for unfair or deceptive practices, and drawing on a long history of data security enforcement. The indicators would be whether a company's conduct is unreasonable, fails to match best practices, and could easily have been avoided with more care — then whether harm resulted, and whether the benefits outweigh it.
On speculative harm, his answer was that one would hope an existential risk shows up first as smaller incidents, and that this appears to be what is happening: incidents causing little harm but raising red flags, which leaves time to examine the law and look for gaps. His closing point was about the other side of the ledger, where he said the benefits are enormous and mitigation should not be allowed to foreclose them.
11. Crusoe's $4B Round
Crusoe closed a Series F of nearly $4 billion at a valuation of almost $31 billion, and says it has more than $140 billion of contracted value across its platform. Lochmiller's framing of why the money is available was about the size of the job.
Why the sector attracts it
The capital needs are really tremendous for what's happening.
Chase Lochmiller
The value creation potential is correspondingly large, he said, which is what drives capital to the sector.
12. Electron to Token
Asked to describe plainly what the company does, Lochmiller gave the phrase he uses for the whole business.
The span he claims
Crusoe is a vertically integrated AI infrastructure platform that helps manage infrastructure from the electron to the token.
Chase Lochmiller
Which splits in two
So this is both building, operating, and energizing these large-scale AI factories, and then building software to operate these.
Chase Lochmiller
What a customer buys is one of three things. A data center, which is what Oracle and Microsoft and other large buyers take. A managed compute cluster, which is what Perplexity, Figure and other AI companies take. Or managed services and managed inference, including serverless fine-tuning.
13. Contract Durations
Asked for the revenue split across the three, Lochmiller answered on duration instead, which is the more useful number for anyone underwriting the business.
The variable he thinks matters
they typically have different durations
Chase Lochmiller
The long end
we're typically working with very large tech companies that have investment grade credit ratings and they're typically committing to us for somewhere between 15 and 20 years
Chase Lochmiller
Multi-decade commitments from investment-grade counterparties sit under the data center business. The compute business is bounded by hardware.
The middle
those can range anywhere from a few months to five to six years, but they typically don't go beyond that just because that's sort of the life cycle of the chips
Chase Lochmiller
Managed inference and AI services are shorter again, closer to on demand, because they are developer tools that scale with an application's usage.
What that gives him
So it sort of offers a broad spectrum of services that we can offer the industry, as well as a broad spectrum of contracts and revenue resources that we're receiving as a business.
Chase Lochmiller
14. Three New Directors
The round came with three board appointments, one for each of what Lochmiller called the company's pillars: energy, data centers, and AI cloud computing.
JB Straubel, Tesla's co-founder and former chief technology officer, joined on the energy side. Bill Stein, whom Lochmiller described as widely considered the godfather of the data center industry and a mentor to him, represents the data center pillar. Thomas Seifert, long the chief financial officer of Cloudflare, joined as audit chair.
The obvious question was whether a round and a board of that shape is a precursor to a listing.
His answer
So I think at some point it probably does make sense for us to enter the public market, but it is kind of something that we're constantly evaluating.
Chase Lochmiller
Demand across the customer base is tremendous, he said, and the ambition for the platform is large, so the question is how best to capitalize the growth rather than whether.
15. The Water Narrative
The host put the political problem directly: data centers are an issue in several races ahead of the midterms, Texas has a moratorium on new connections, and Crusoe has a large Texas footprint. Lochmiller's diagnosis was not about policy.
What he thinks the industry has
I think more than anything, the data center industry has sort of a marketing issue.
Chase Lochmiller
And why
We are seeing a lot of information that's just, frankly, not accurate.
Chase Lochmiller
His example is water, where the story is that data centers consume enormous quantities of it.
The figure he gives against that
Each of our big buildings in Abilene, Texas uses roughly the same water footprint as about 10 single family homes.
Chase Lochmiller
That usage is people going to the bathroom and watering plants, he said. The cooling is separate.
The engineering behind it
But the water to cool the chips, that's actually operating on a closed loop system.
Chase Lochmiller
Cold water goes into the racks and is cooled again by air-cooled chillers outside the building.
16. A Third of the Tax Base
The other half of his answer is fiscal, and the numbers are the strongest argument he made.
What Crusoe will be to one county
we'll be about a third of the total tax base in Abilene and Taylor County
Chase Lochmiller
And to its schools
And we're more than doubling the annual receipts for the local school system.
Chase Lochmiller
At a second Texas site
a large campus is Armstrong County, where we're increasing the tax revenue by over 100%
Chase Lochmiller
Larger budgets let a community fund police, fire, roads and schools over the long run, he said, which is the case he makes locally alongside job creation.
Asked whether any specific project has been delayed or paused by local opposition, his answer was "Not yet," followed by an acknowledgment that he sees it happening elsewhere in real time.
His position on it
So it's something that we're trying to get ahead of and really we think that the truth is on our side.
Chase Lochmiller
Crusoe is investing in these places for multiple decades, he said, which is the basis on which he wants to be judged as a partner.
17. Microsoft vs Claude
Bloomberg's Seattle reporter brought two more positions from the week. Microsoft's AI chief, Mustafa Suleyman, who co-founded DeepMind, published an argument about why models appear to have feelings.
The reporter's summary of the argument: "He says part of the reason that AI might appear like it's got some kind of emotions or humanity is because it's been trained to believe that."
It is aimed at Anthropic and Claude specifically, pointing at training data where the model is told it may have proto-emotions or a sense of self — which, on this argument, is why it behaves that way. It does not make the system conscious; it means it was taught to act that way, and Microsoft's suggestion is that the industry stop.
Amazon's position was simpler and, the reporter said, in line with Nvidia's chief executive.
Its whole position, as relayed: "They said don't ship it until it's safe, don't ship it until it's been tested."
Amazon has stayed neutral on the broader question of slowing down, he said, which fits a company hosting models from both major labs.
18. The iPhone 18 Pro
The last item was retail. Apple's new flagship handset went on sale worldwide, the first major launch under John Ternus as chief executive, who opened the doors at the company's New York flagship store himself. The folding model does not go on sale until later next month. Apple shares were down seven tenths of one percent.
Bonus Insights
The week's scoreboard
The host's own recap of five sessions: the Nasdaq 100 finished modestly higher, up three tenths of a percent, having started the week with the calls to slow down, taken a Fed decision in the middle, and seen yields rise on expectations of further rate increases driven by energy and other inflationary pressure. The Philadelphia Semiconductor Index was down half a percentage point over the same five days. The question the week posed was whether the warnings would cut AI infrastructure spending and pull the market back. His conclusion, so far, was no.
The one thing everybody agreed on
Across every camp on the show, the host said, there was a single point of agreement: model capability is improving faster than alignment and safety work can keep up with it.
Where Crusoe operates
Lochmiller's business is not only American. The host noted the company is looking at the Gulf and other regions, which is the context for the political pressure in Texas rather than a substitute for it.
Lochmiller's bottom line is that the constraint on AI infrastructure has moved from capital to consent: the money is available on a scale he calls tremendous, the contracts run 15 to 20 years with investment-grade counterparties, and what stands in the way is a water story he says is wrong and a local tax case he thinks he wins.
Products, Companies & Tools Mentioned
Crusoe (Closed a nearly $4B Series F at about a $31B valuation, with more than $140B of contracted value; builds, powers and operates data centers and sells compute and inference on top)
Anthropic (Disclosed that its own model leads 26% of its AI research and development work, the week's clearest figure on AI improving AI)
OpenAI (The recipient of SoftBank's $65B, and the company whose listing timeline has slipped to 2027)
SoftBank (Borrowing $45–50B of that $65B, against its Arm stake, from Apollo, from 20 banks, and against the OpenAI stake itself)
Arm (The chip designer SoftBank owns about 90% of and is borrowing more against)
Oracle and Microsoft (Data center customers of Crusoe, on 15-to-20-year commitments)
Perplexity and Figure (Buyers of Crusoe's managed compute clusters, on shorter terms bounded by GPU life)
Cohere (Its chief executive told Bloomberg he does not expect mass casualties from AI but says the real issues need work)
Netflix (Took its first sell rating since February on engagement concerns, and is on track for its worst year since 2022)
Abundance Institute (Where Neil Chilson, formerly the FTC's chief technologist, argues that only government-endorsed cartels are durable)
Cloudflare and Tesla (The companies two of Crusoe's three new directors came from)
Apple (Its new flagship handset went on sale worldwide, the first launch under its new chief executive; the folding model comes next month)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: