Part IThe Story
The first problem was the heat.
Sometime in 2016, in an office in Manhattan rented by a natural gas hedge fund that was not, by most accounts, setting the world on fire, a small group of former commodities traders began plugging graphics cards into every available wall outlet. The cards were Nvidia GPUs, the chips that make video games look like video games. A GPU does an enormous amount of simple arithmetic in parallel, and in the mid-2010s it turned out that this was also the arithmetic of cryptocurrency mining. The traders ran the math the way traders do, decided there was a spread worth capturing, and bought hardware. Years later Michael Intrator, then 47 and the fund's chief executive, put it plainly to Forbes: "We founded the company on a credit card. It was like, 'Okay, I'll buy a couple GPUs. What the hell?'"
A couple became more than a couple. The machines ran hot. An office designed for Bloomberg terminals and conference calls has a fixed and fairly modest tolerance for thermal load, and that tolerance was approached and then, by Intrator's own later description, nearly exceeded. He calls it "nearly cooking a Manhattan office full of GPUs." In August 2017 the servers were moved to a garage in New Jersey. On September 21, 2017, the operation was incorporated as Atlantic Crypto.
Fix that image in your mind, the overloaded power strips and the warm air pooling against the ceiling tiles, because nearly everything that follows is a scaled-up version of it. The hot room became a garage, then warehouses, then a 450,000-square-foot supercomputer facility in Plano, Texas, then 43 data centers drawing more than 850 megawatts of active power. The credit card became a $2.3 billion loan collateralized by the chips themselves, then a $7.5 billion facility billed as the largest private debt financing ever, then total debt of roughly $35.1 billion as of June 30, 2026. The arithmetic stayed the same. The constraint stayed the same too: power, heat, and the money to buy more of both before anyone else could.
On the morning of March 28, 2025, Intrator stood on the podium at Nasdaq headquarters in Times Square next to Brian Venturo, the friend who had been beside him at two previous firms. Venturo wore a sweatshirt. They rang the opening bell and slapped five, and a Getty photographer caught the moment. The company was by then called CoreWeave, ticker CRWV. It had just completed the largest U.S. venture-backed tech IPO since 2021, and it had done so badly. The offering had been cut from $2.7 billion to $1.5 billion. The marketed range of $47 to $55 a share had collapsed to $40, and the deal had cleared at all only because Nvidia, CoreWeave's supplier, investor and customer, stepped in to buy at that price. The stock opened at $40 and closed at $40. The tape did nothing all day.
Three months later the shares had nearly quadrupled, and Intrator's net worth had gone from $5 billion to $10 billion in twelve days. A year after that, the stock had lost more than half its value from the peak, the company was carrying a debt load comparable to some sovereign nations, and Intrator was telling a Wall Street Journal columnist, "I'm pretty comfortable that I'm on the right side of this."
Each of these numbers is real. What they amount to depends on whether you believe the man who has spent his whole career selling things you can't see.
By the Numbers
The CoreWeave Machine
$5.13B2025 revenue, up 168% year over year
$66.8BContracted revenue backlog at year-end 2025, up from $15B
~$104BRevenue backlog reported in Q2 2026
$35.1BTotal debt as of June 30, 2026
850+ MWActive power across 43 data centers
$40IPO price, against a marketed range of $47–$55
9 of 10Leading AI model providers running on CoreWeave Cloud
Selling Smoke
Before compute there was carbon, and before carbon there was a political science degree.
Intrator was born in March 1969 and studied political science at Binghamton, the State University of New York campus in the Southern Tier. He then went to Columbia's School of International and Public Affairs and finished a master of public administration in 1995. SIPA trains diplomats, development economists and municipal finance officers. It does not usually produce cloud computing executives. A classmate, Julie Min Chayet, wrote decades later that she still remembered "those SIPA midterms and group projects" with him. Nothing in that record suggests GPUs.
In September 1998 he joined Natsource Asset Management in New York. Natsource worked in the emissions business, which was then new, odd and mostly theoretical. Its products were allowances, offsets and credits: legal rights to put a ton of sulfur dioxide or carbon into the atmosphere, created by regulation and given value by scarcity. The thing being traded was, physically, nothing. It was a permission. Intrator moved up through the firm to principal portfolio manager and managing director, investing in what his CoreWeave biography calls "global environmental markets and energy products." When TIME covered this emerging trade in March 2003, a piece cited in the record of his career, the headline was "Selling Smoke."
That was the job for sixteen years. He left Natsource in July 2014, already in his mid-forties, having spent his entire adult working life in one marketplace for invisible goods.
Brian Venturo was an energy trader at the same firm. The pair were good enough together that a colleague reached for the obvious sports comparison.
— Alex Baldassano, former Natsource portfolio manager, to CNBCThey were the Michael Jordan and Scottie Pippen of the carbon markets at the time.
The analogy says more than Baldassano probably meant. Jordan and Pippen won six titles by doing one thing obsessively, inside a structure that rewarded doing it obsessively. The carbon market was a structure of exactly that kind. It was built by policymakers, priced by traders, and turned on the gap between what the rules required and what the physical world could supply. A political scientist who understood how regulation created scarcity and a trader who understood how to price it made a natural pair. Bruce Usher, a former carbon-markets colleague who later became a professor at Columbia, welcomed Intrator back to campus in the fall of 2025. Intrator described the visit on LinkedIn as a chance "to share how I put my degree to good use." He meant it lightly. It was also accurate. The public-administration graduate's career rests on the idea that a government, a utility or a regulator can make an intangible thing scarce, and that whoever owns the scarce thing gets paid.
Neil Cohn, another Natsource alumnus, congratulated him in 2024 after one of CoreWeave's record debt raises: "Fun to see friends blazing a trail again." The word again deserves notice. To the people who knew him first, CoreWeave was not a reinvention. It was the second time he had found a market before it had a name.
The Downtime
In January 2013, Intrator and Venturo co-founded Hudson Ridge Asset Management, an algorithmic hedge fund focused on natural gas. Intrator ran it as CEO until January 2018. The official biographies say little more than that. TechCrunch, less politely, later described the fund as having "failed."
Intrator's own account is more interesting because it's about boredom.
"When you build an algorithmic hedge fund," he told Jason Calacanis on the All-In podcast in March 2026, from the floor of Nvidia's GTC conference, "once the algorithms are built, you're really just monitoring it and testing different thesis and doing all right. But there's also a lot of downtime and we got super interested in crypto. We're pretty nerdy. We dig under the hood."
There are two founding myths of CoreWeave, and they don't agree, which is part of their charm.
Venturo's version, told to Wired in 2024, is a joke that got out of hand. In late 2016 he and some hedge-fund friends had been staking their betting pool and fantasy football league in bitcoin instead of cash, "for the fun of playing with something volatile that could be either trash or treasure." Bitcoin went up. Someone said, "Maybe we should take this seriously." They rush-ordered GPUs from Amazon for same-day delivery while the enthusiasm was still hot. Beer money.
Intrator's version is a thesis. He says they looked at bitcoin mining and specifically rejected it. Bitcoin is mined most efficiently on ASICs, application-specific chips built to do one thing, and "we just thought that there's some brilliant engineer that built the ASIC and they're probably going to be better at running it than we are." So they chose GPUs, which could mine Ethereum but could also do "all these other things." In his telling, "right from the start, we looked at the compute as an option to be able to deploy our computing power to different use cases."
Option is a trader's word with a precise meaning: the right, but not the obligation, to do something later. An ASIC is a forward contract on one market. A GPU is a call option on every market that might someday need parallel arithmetic. Whether a 47-year-old hedge fund manager really thought about it this precisely in 2016, or whether a 57-year-old billionaire tidied the memory once he knew how things turned out, can't be settled from the outside. Both stories are probably true in the way founding stories are true. The fantasy football money got them in the door, and the options logic explains why they stayed.
One detail is the same in both versions: the hardware was bought before anyone knew what it was for. That habit is the whole company in miniature.
Brannin McBee, also a commodities trader and later CoreWeave's chief development officer, joined the venture and named it Atlantic Crypto with them in September 2017. Peter Salanki, who became chief technology officer, is the fourth name on the founders' list. None of the three traders had worked in technology. Fortune, reporting on their sudden fortunes in 2025, ran the headline "despite having zero tech background," and it read half as admiration and half as alarm.
Pennies on the Dollar
The garage in New Jersey was the right place to be when crypto collapsed.
In 2018 the price of Ethereum fell off a cliff, and with it the economics of every mining operation that had bought hardware near the top. Miners who had financed GPUs at high prices were suddenly holding depreciating equipment that could no longer cover its electricity bill. They needed to sell, and the buyer was a garage in New Jersey run by people who had spent two decades on trading desks.
Atlantic Crypto bought thousands of GPUs from distressed miners "for pennies on the dollar," as Forbes later reported, and used the hardware to mine Ethereum, bringing in an estimated $80 million in revenue. In a crypto winter, in what people who had been in crypto mining for 18 months were calling a catastrophe, three commodities traders did what commodities traders do when a market is liquidated: they took the other side.
Intrator credits this to the earlier careers. "Because we had come from a hedge fund, we have real chops in risk management and how we think about capital and risk exposure and allocation," he said on All-In. "So we weathered crypto winter really well and began to scale the company and immediately started to look for other use cases that you could use this compute for because crypto was pretty volatile."
That last clause is a little funny coming from someone who had bought in specifically because crypto was volatile. The volatility was the attraction until it became the risk, and the reverse was also true. A trader holds both views at once without strain.
The company's first contract with Core Scientific, a large crypto-mining host, was signed in 2018. Seven years later CoreWeave would try to buy that company for $9 billion. In 2019 Atlantic Crypto renamed itself CoreWeave and began leasing out the compute it had, first to visual effects and CGI rendering ("folks that were trying to animate and render images," in Intrator's phrase), then to batch computing and medical research. By 2020, according to CNBC, it had a real business renting Nvidia GPUs in the cloud. The customers were small and specific: studios, researchers, a few AI startups nobody had heard of.
What it lacked in scale it made up for in attention. "We would resolve things in 15 minutes," a manager from that period told Wired. "The dedication to customers was phenomenal." Venturo's version of the same quality is less flattering and more revealing: "We're lucky enough to be in this position because of this crazy origin story where we had so much expertise building at this rapid, frenetic pace." When coin prices were climbing, time was literally money, and the technicians had learned to install GPUs quickly and run them at the edge of failure. A crypto mine turned out to be an excellent training ground for a cloud company, mostly because nobody would have designed one that way on purpose.
In 2021 CoreWeave built its first real AI cluster, 96 Nvidia A100 chips for EleutherAI, an open-source collective of researchers training large language models outside any big lab. Ninety-six chips is a rounding error by later standards. The company's eventual fleet would number in the hundreds of thousands. But it was the first time CoreWeave's machines did the work they would eventually be valued for.
"I kinda feel like buying those initial GPUs," Intrator said in 2026, "was the tuition we paid to learn how to run this business."
That is the trader's reading of the crypto years. The mining income was real, but the real return was operational knowledge, acquired cheaply because the market thought the hardware was worthless. Nobody else had a reason to learn how to run tens of thousands of GPUs at the edge of their thermal limits in 2019. CoreWeave did, because it was already sitting on the chips.
An Asset You Can Unplug
In the summer of 2022, before ChatGPT, CoreWeave spent about $100 million on Nvidia's newest and fastest chip, the H100. Late that year Magnetar Capital invested $100 million in the company, and that money also went into H100s. The timing still looks almost suspicious. OpenAI released ChatGPT on November 30, 2022. Within weeks, the whole technology industry discovered that it needed exactly the hardware CoreWeave had spent the summer stockpiling.
This is the bet at the center of the story, and the reason it has been read both as genius and as luck. In hindsight it looks like prophecy. In context it looks like the same move as 2018: buy the scarce input before the market reprices it, because if the input has many possible uses, you'll find buyers. The crypto collapse had made GPUs cheap. The arrival of generative AI made them nearly impossible to get. CoreWeave held them through both.
The GPU shortage of 2023 was in large part a shortage of Nvidia's best chips, and CoreWeave, as the trade press repeatedly noted, had "unique access" to them through a relationship with Nvidia going back years. In January 2023 it signed a $300 million deal with Inflection AI to build a 22,000-GPU cluster. In April, Nvidia invested $100 million directly. Microsoft, OpenAI's main sponsor, became a customer. So did Stability AI. Venturo's biggest regret from this period sounds like something a trader would say: "If I had unlimited amounts, I could have it all sold in two weeks." The constraint was never demand. It was supply, and the supply cost money.
That is where the hot room stops being a metaphor.
Intrator went to Silicon Valley for venture capital and largely failed to raise it. The record is spare on why, but it isn't hard to guess. A company founded by three New Jersey commodities traders that had been a crypto miner eighteen months earlier, was buying hardware rather than writing software, and had no credible path to software-like gross margins does not match the venture template. Venture capital funds teams and software. CoreWeave was a warehouse of depreciating machines.
So he went back east to the debt markets, where he had spent his career.
The Financial Times later described the change by saying Intrator "started to treat CoreWeave's growth like a structured credit play." He began treating the company's assets as securities that could be packaged, rated in effect, and sold. In the summer of 2023 he met with Blackstone. He also persuaded Magnetar, already an equity investor, to join the lending group. On August 3, 2023, CoreWeave announced a $2.3 billion debt facility led by Magnetar and Blackstone, collateralized by Nvidia H100 GPUs. No one had pledged H100s as collateral before. The FT called it one of the largest private financings in U.S. corporate history and Blackstone's biggest single loan commitment. The Verge's headline was nearly a poem: "For this startup, Nvidia GPUs are currency."
It is worth spelling out what Intrator had done, because the rest of the company comes from it. A chip under heavy demand has a resale value. It also produces cash, because a creditworthy customer has contracted to rent it. Put those together and the chip stops being equipment and becomes a receivable with a physical backstop, closer to an aircraft lease or a mortgaged house than to a server. A project-finance professional named Samuel Wytenburg made the point more vividly than any analyst, in a comment under one of Intrator's LinkedIn posts in 2026: "Fully amortising debt backed by hyperscaler offtakes, secured against an asset you can unplug and resell. In renewables project finance we can only dream of collateral that cooperative. Try repossessing a wind farm!"
An asset you can unplug. The garage miners in 2018 had unplugged theirs and sold them to Intrator for pennies. He had seen exactly how liquid a GPU could be in a panic. Now he was asking lenders to believe that liquidity in the other direction: a chip in a rack in Texas, contracted to Microsoft, was worth something no matter what happened to the startup that owned it.
Lenders believed it. In September 2023, CoreWeave and Nvidia announced that the $1.6 billion Plano facility, around 450,000 square feet holding more than 3,500 H100s, was the fastest AI supercomputer in the world. That year CoreWeave's technicians installed 6,000 miles of fiber-optic cable, roughly the distance from New York to Berlin and partway back. In December, Fidelity and Jane Street invested at a $7 billion valuation, up from about $2 billion in the spring. Business Insider named Intrator one of the top ten people in AI hardware. Intrator's LinkedIn post on the Fidelity round was unusually plain: "It has been a truly extraordinary year for us."
He didn't say anything about the heat.
A Dash of Jersey Flair
On August 25, 2023, the Wall Street Journal ran a story by Ben Cohen under the headline "A Startup in the New Jersey Suburbs Is Battling the Giants of Silicon Valley." Intrator shared it on LinkedIn with a seven-word caption, "Great article in the Wall Street Journal today," and the comments did the rest. "It reads like the plot of a Silicon Valley TV show," wrote one well-wisher, "with a dash of Jersey flair!"
The geography mattered more than a dateline usually does. CoreWeave has always been, almost aggressively, a New Jersey company. It began in a garage there, later took offices in Roseland, and by 2026 was headquartered at 290 West Mount Pleasant Avenue, Suite 4100, in Livingston, an Essex County suburb of strip malls and good schools whose best-known export is not cloud computing. Its S-1, filed March 3, 2025, lists that Livingston suite as the address for service of process on Michael Intrator, Chief Executive Officer, phone number (973) 270-9737. Area code 973 is North Jersey, not Palo Alto.
In October 2024 the company signed a lease for 280,000 square feet at the Northeast Science and Technology Center in Kenilworth, the old Merck research campus, and committed $1.2 billion to building a data center there. New Jersey's political class treated this like a franchise relocation. Governor Phil Murphy had talked for years about making the state the East Coast's AI hub, and ROI-NJ, ranking Intrator No. 17 on its 2025 Power List, wrote that his decision to "double down on the state" was a huge step in that direction. CoreWeave joined Microsoft, Princeton and the New Jersey Economic Development Authority as a founding partner of the New Jersey AI Hub and seeded a $20 million fund alongside it. By February 2026, NJBIZ ranked him No. 5 among the most powerful people in New Jersey business.
You can make too much of this. You can also make too little. Silicon Valley's mythology is built on equity: seed round, Series A, Series B, IPO, dilution as a sacrament. New York and New Jersey are where the debt lives. The banks are there, the insurance companies are there, the private credit funds are there. What made CoreWeave possible was a financing structure that Sand Hill Road found strange and Park Avenue found familiar. A commodities trader from Natsource didn't need to be converted to the idea that a physical asset under contract can carry leverage. He had been pricing intangible permissions for sixteen years. A GPU you could actually touch was almost too easy.
There's also the matter of temperament. When Forbes put Intrator on its cover in September 2025 under the headline "GPU Gambler: Michael Intrator, the Great Compute Crapshoot," he sounded pleased rather than wounded. "People who know me won't be shocked to see a little passion and profanity slip into the story," he wrote. "In fact, one of my 5th grade classmates commented on the story, 'To be fair — Mike gesticulates defiantly when he orders a sandwich.'"
He posted that himself, which tells you something. Hard to imagine a Stanford-trained founder in a Patagonia vest volunteering it. (Intrator does own a vest, for the record. Under his May 2024 announcement of the $7.5 billion facility, a commenter wrote, "Love the news, love the vest." The record doesn't say what kind.)
The defiance is real and it's useful. CoreWeave's first shareholder letter, published in 2026, describes the company's founding choice as "audacious and defiant: a full-stack AI cloud, conceived when others doubted the opportunity and, frankly, us." Then comes the line he repeated in the LinkedIn post introducing the letter, as if he couldn't help himself: "Doubt, it turns out, is a powerful motivator."
It is not an elegant sentence. It is a sentence from someone who has been underestimated in rooms where everyone else went to Stanford, and who remembers.
Twelve Billion in Twelve Months
The year 2024 was when the hot room got very big and very expensive.
On May 1, CoreWeave announced a $1.1 billion Series C led by Coatue at a $19 billion valuation, nearly triple what Fidelity and Jane Street had paid five months earlier. On May 17 came the debt: a $7.5 billion facility led by Blackstone, with Magnetar and Coatue participating. In the announcement, Intrator called it "the largest private debt financing of all time" and added that it "brings our total capital raised to $12 billion in 12 months." In October, Goldman Sachs, JPMorgan and Morgan Stanley led a $650 million revolving credit line. That month Bloomberg reported Cisco was close to investing at $23 billion. In November a $650 million secondary sale brought in Jane Street, Magnetar, Fidelity and Macquarie, with Cisco now confirmed. CoreWeave opened a European headquarters in London and committed £1 billion (reported as $978.6 million) to British data centers. It was among the first clouds to receive Nvidia's new Blackwell hardware. In December, Cohere announced that CoreWeave would build and operate a multibillion-dollar data center for it in Canada.
The pace showed inside the company. In May 2024, Wired reported that CoreWeave had about 550 employees and that some of them, six former workers who spoke on condition of anonymity, described "a culture of grueling hours and high-pressure management that feels unsustainable." The company's official values list, from its "About Us" page, includes Act like an owner with this gloss: "Wait for others to solve problems? Not us." The sentence could be read as aspiration. It could also be read as a description of what happens when a company triples in value every few months and hires faster than it can train.
The headcount rose to roughly 2,189 by the end of 2025, about four times the 2024 figure. The new hires weren't only engineers. Intrator began recruiting the incumbents' operators. Nitin Agrawal came from Google Cloud, where he had been vice president of finance, after earlier stints at Amazon Web Services and Microsoft, and became CFO. Sachin Jain left Oracle's AI organization in August 2024 to become chief operating officer. Chen Goldberg came from Google as senior vice president of engineering. Michelle O'Rourke was promoted to chief people officer in October. In February 2025, Glenn Hutchins became lead independent director. Intrator wrote a blog post about the hiring, titled "Building Our Team — And How We Know Our Strategy is Working," which says, without quite saying it, that the people building the cloud businesses at Google, Amazon, Microsoft and Oracle were now willing to work for three traders in Livingston.
The founders recruited people who looked like the companies they meant to beat. Agrawal's résumé reads like a list of CoreWeave's competitors and customers. Intrator told Fortune in March 2024, "At the end of the day, it's the people that matter, so make sure you take people along with you in your journey." The other half of that advice, given in the same interview, carries more of his particular voice: "always keep a healthy disregard for the impossible while being grounded in reality."
A healthy disregard for the impossible, grounded in reality. Those two clauses pull against each other, and he holds both.
Not all of 2024 went his way. In June, CoreWeave offered about $1 billion for Core Scientific, the crypto-mining host it had worked with since 2018, and was turned down. And when the S-1 became public in March 2025, it showed what had been an open secret: Microsoft had accounted for 62 percent of CoreWeave's 2024 revenue. Fortune's headline on the filing was "The first big tech IPO of 2025 is here — and it's got some serious red flags." For all the demand, the company that had made its name on GPU scarcity depended to a large degree on one buyer.
The filing also showed that the three trader-founders had already taken some chips off the table. Each had sold more than $150 million in shares before the IPO.
Forty Dollars
The IPO window CoreWeave chose in late March 2025 was poor. Tariff anxiety was building ahead of what the White House was calling "Liberation Day." DeepSeek's January release had made some investors doubt how much compute frontier AI really needed. And the offering asked public investors to buy a company with roughly $8 billion in debt, servicing costs of at least $1 billion a year, a single customer accounting for most of its revenue, and a business built entirely on one supplier's chips.
In the days before pricing, CoreWeave announced two things meant to answer those worries. On March 5 it agreed to buy Weights & Biases, the machine-learning developer platform, for about $1.7 billion, a step up the software stack. On March 10 it signed a five-year contract with OpenAI worth about $11.9 billion, under which OpenAI would also take a $350 million equity stake through a private placement at the IPO. The OpenAI deal was the counterargument to concentration. If a 62-percent customer is a red flag, add a second giant.
It wasn't enough. On March 27 the offering was cut. The price came in at $40, $7 below the bottom of the range, and Nvidia anchored the book. CNBC's account opened bluntly: "If not for Nvidia, there would be no CoreWeave IPO." On Friday, March 28, the stock opened at $40 and closed at $40.
Venturo was on the podium in his sweatshirt. That afternoon he gave CNBC what is still the best description of the company's character, partly because of the pun he didn't seem to notice.
— Brian Venturo, co-founder and chief strategy officer, to CNBC, March 28, 2025I feel like this company has been built on a long list of events where the chips are down and the people at the top being tough as nails, but gets it through.
The chips are down. In a GPU business, the chips are literally the collateral. When they were down in 2018, CoreWeave bought them. When they were down at the 2025 IPO, Nvidia bought shares. Each time the company leaned on someone with a stake in the hardware being worth something.
Intrator's public line was steadier and more practiced. He told Axios the IPO was "a victory." He told Fortune's Diane Brady it was "a means to an end." "It puts us on the path towards what we need to accomplish as a business. A little bigger, a little smaller, a little higher, a little lower. That's not going to matter. What's going to matter is: how do we execute on our business?" He told Jim Cramer that night, on Mad Money, that "debt is the fuel for this company," which is roughly what Henry Ford might have said about gasoline. And when Brady asked whether AI data centers were a bubble, he said something that would become his standard answer for the next eighteen months:
"There's a divergence between what the capital markets and what the media is thinking, and what I am feeling down in the trenches. What I am feeling is relentless demand."
He offered CNBC an image, too: "We are built for speed. That is all we do, right? And when you build something to do one thing, you end up with a Lamborghini, not a minivan." Lamborghinis are fast and are also notoriously expensive to own. He didn't take the metaphor that far.
The founders' combined paper wealth that Friday was about $5.3 billion. Intrator owned roughly 15 percent of the company.
Twelve Days in June
Then the market changed its mind.
Through April and May, the contracts kept arriving and the AI trade revived. By late June, CoreWeave's stock had roughly tripled from its listing price, up nearly 300 percent. Bloomberg calculated that Intrator's net worth went from $5 billion to $10 billion in twelve days. Fortune noted that the average time to make that jump was three years and four months, and Intrator had done it about 101 times faster. TechCrunch's headline read, "In just 3 months, CoreWeave CEO, once a crypto-mining bro, becomes a deca-billionaire." Venturo was worth $6.4 billion. McBee was worth $4.7 billion. Forbes put Intrator at No. 219 on its 2025 list of the richest Americans.
Intrator's comment on all this, given to Fortune before the surge, read like a trader who'd learned not to look at the screen. "I don't really care where it is today or tomorrow or the day after, but I believe fundamentally, the business model that we have, the software solutions that we have, the capacity to build and deliver this and the demand we see in front of us will lead to enormous value to our clients over time."
The public role came with the market cap. On May 8, 2025, in the Hart Senate Office Building, he testified before the Senate Commerce Committee at a hearing titled "Winning the AI Race: Strengthening U.S. Capabilities in Computing and Innovation." He sat between OpenAI's Sam Altman and Microsoft's Brad Smith, with AMD's Lisa Su on the same panel. The SIPA graduate who had spent sixteen years trading permits created by regulation was now telling legislators what regulation he needed. In a follow-up op-ed he called for "clear regulation, energy policy reform, global market access, and bold workforce development," and ended with the kind of line that public-administration programs train people to write: "This isn't just a tech challenge, it's an American imperative."
Then he went back to buying.
On July 7, 2025, CoreWeave agreed to acquire Core Scientific in an all-stock deal worth about $9 billion. This was the company that had turned down $1 billion a year earlier and whose relationship with CoreWeave, as Intrator pointedly reminded Bloomberg's Caroline Hyde the next day, went back to "our first contract... in 2018." The logic was vertical integration. Weights & Biases had moved CoreWeave up the stack toward its customers. Core Scientific would move it down, into "the physical data centers, the power requirements, all of those things." When Hyde cited Stifel's downgrade and the worry that owning data centers would hurt CoreWeave's valuation multiple, Intrator replied with a phrase that captures him well: "Yeah. So I would take issue with that assessment."
"Let's do that," Hyde said.
The market didn't buy it. CoreWeave's stock fell after the announcement. A shareholder commented under Intrator's LinkedIn post about the Mad Money interview: "Hopefully the market reaction turns around... The deal is reasonable, but man has it hurt to see." Another, a more technical admirer, wrote that the deal "flipped a large liability (lease) into an asset," which was exactly the argument and exactly the kind of balance-sheet reasoning public markets tend to discount.
On July 15 he announced up to $6 billion for a new campus in Lancaster, Pennsylvania: roughly 78 acres at 216 Greenfield Road, reusing two vacant printing-press facilities, starting at 100 megawatts and scalable to 300, with a $4 billion joint venture from Blue Owl, Chirisa Technology Parks and Machine Investment Group behind the financing and about $200 million set aside for grid upgrades. "This data center will accelerate innovation and drive economic growth across the region," he said. Some residents of the region pushed back on the scale and location.
By autumn Core Scientific's shareholders had turned against the deal. Both major proxy advisers, ISS and Glass Lewis, recommended voting against it because CoreWeave's own share price had fallen and the all-stock offer was now worth less. Intrator insisted the deal was "by far the best alternative for the long term interests of Core Scientific stockholders." Then, on CNBC on October 21, he gave up the position he'd just defended: Core Scientific was "not a need to have." Nine days later, on October 30, its shareholders rejected the offer.
That reversal is the most revealing moment of 2025. For sixteen weeks Intrator argued for the deal with complete conviction. Then, within one sentence on live television, he reclassified it from strategic necessity to discretionary trade. That isn't hypocrisy. It's how a portfolio manager talks about a position that has stopped working. You don't fall in love with your book.
The fall's other news continued without a pause. CoreWeave bought OpenPipe, a startup building reinforcement-learning tools for AI agents, in September, and Monolith AI, which applies machine learning to physics and engineering, in October. It launched a venture fund. At Goldman Sachs' Communacopia conference that September, Intrator told CNBC that CoreWeave had raised "in excess of $25 billion" in eighteen months. Then he admitted something unusual for a CEO on television, a small crack in the composure: what the market was "really internalizing," he said, was "something that we have been trying to communicate. I guess sometimes I'm good at it, sometimes I'm bad at it."
At Bloomberg Tech in London that October, he offered six words that work as his credo: "Speed's easy. Vision's rare."
Before We Were Cloud Builders
CoreWeave's first annual letter to shareholders, published in spring 2026, is a strange document. It's half sales copy (the company now calls itself "The Essential Cloud for AI™," trademark symbol included) and half confession. The confession is in a heading that reads, simply, Discipline to build and scale, followed by a sentence that may be the truest thing Intrator has written:
— Michael Intrator, CoreWeave 2025 Letter to ShareholdersBefore we were cloud builders, we were risk managers. That instinct to size opportunities carefully, build against what we can see, and never outrun our contracts shaped how we financed CoreWeave from the beginning.
The numbers that follow are meant to show it. Revenue in 2025 was $5.13 billion, up 168 percent, which made CoreWeave, the letter claims, "the fastest cloud platform in history to reach $5 billion in annual revenue." Contracted backlog rose from $15 billion to $66.8 billion, and the weighted average contract length went from four years to five. The concentration problem that had worried IPO investors was, by the company's account, solved: at the start of 2025 one customer represented 85 percent of backlog, and by the end no single customer exceeded 35 percent. The company raised about $18 billion in debt and equity during the year and cut its weighted average cost of debt by more than 300 basis points, worth roughly $700 million a year in interest savings. Nine of the ten leading AI model providers ran on its cloud.
The letter ends its finance section with a phrase that is both a boast and a thesis: CoreWeave, it says, had shown "how this historic buildout can be financed, tying capital deployment directly to long-term contracted demand, and creating a new asset class in the process."
A new asset class. The man who began his career turning smoke into a tradable instrument now said he had done the same with computation.
The 10-K covering the same year showed the rest of the picture. Operating income was negative $46 million. The net loss was $1.2 billion. Total assets were $49.3 billion against total equity of $3.34 billion. A business that rents out machines bought with borrowed money, at the scale CoreWeave had reached, spends a large share of each revenue dollar on interest.
The stock had already turned. In a Wall Street Journal column that year, Steve Rosenbush reported that CoreWeave's shares had fallen about 58 percent from their June 2025 high, pulled down by disappointing guidance and a sector-wide worry about capital spending and leverage. YouTube's archive of the period has headlines like "CoreWeave Lowers Outlook on Capacity Crunch" and "CoreWeave Shares Drop After Forecast Sparks Growth Fears." Alex Kantrowitz's Big Technology Podcast had Intrator and Venturo in the studio on January 7, 2026, for an episode titled "Coreweave: AI Bubble Poster Child Or The Next Tech Giant?" The Verge's Decoder was blunter: "The company at the heart of the AI bubble." Under the shareholder-letter post on LinkedIn, one reader wrote: "Now let's see if the insider trading filings match the 'long-term confidence' sentiment. Real conviction doesn't need a secondary sale exit this early." Another, writing in Chinese, put it more sharply: Interest this high and not the least bit panicked — GPUs are like Chinese real estate fifteen years ago.
The capital, meanwhile, kept coming. In January 2026, Nvidia invested another $2 billion at $87.20 a share, more than double its IPO-day price of $40, to support CoreWeave's plan for more than five gigawatts of AI "factories" by 2030. In February the company sought an $8.5 billion loan backed by its contracts with Meta. In early April, within 48 hours, it announced a $21 billion expansion of the Meta relationship, a multi-year deal to provide GPU capacity for Anthropic's Claude inference workloads, and an upsized $3.5 billion convertible note offering, which Intrator told Bloomberg was part of "the largest dual offering ever." He described the Meta-backed debt, by his account, as "A-minus paper," credit quality borrowed from the creditworthiness of the customer. Revenue for the first quarter of 2026 was $2.08 billion, more than double the year before. In the second quarter it was $2.58 billion, up 112 percent, with backlog around $104 billion and a $626 million net loss driven mainly by interest expense. As of June 30, total debt was about $35.1 billion: $31.4 billion recourse and $3.7 billion non-recourse, structured against hyperscaler contract commitments.
The bear case can be stated simply. The collateral depreciates. The customers are a handful of companies with their own capital-spending cycles. The supplier is also an investor and could become a competitor. And the binding constraint has moved from chips to power, as Texas showed in 2026 by halting new grid connections for large loads. A company whose first real crisis was an overheating office now faces a whole state telling it there isn't enough electricity.
Intrator's answer, on All-In and elsewhere, is that the depreciation argument comes largely from traders with short positions, and that the market's evidence is the contracts: customers keep signing five-year deals for the compute, so the compute is worth something for five years. It's a trader's argument. It ignores the abstraction and points to the bid. "I always think of inference," he said, "as the monetization of the investment in artificial intelligence." Training builds the models. Inference, the endless daily querying of them, is the recurring payment. If the payment arrives, the debt gets serviced. If it doesn't, the collateral gets unplugged.
To Rosenbush he made the longer historical argument, the one that comes closest to explaining what he thinks he has built. Debt, he said, built homes, highways and railroads, financed schools, mobilized nations for war. "All of these things use a combination of both sides of the balance sheet. This particular episode of technology build-out required mobilization of a scale of capital that the world rarely sees. And with that, you have to bring to bear every tool to drive down the cost." The technology industry, raised on equity, "needs to be much more cognizant of capital efficiency. It wasn't required to do that when the capital needs of the company were not so gargantuan, right?"
And then the sentence he seems to have been building toward since the Manhattan office:
"AI is a generational, maybe multigeneration, change. We know what we're building. We know why we're building it. We know how we're building. I'm pretty comfortable that I'm on the right side of this."
Somewhere on Amazon there's already a paperback biography, Michael Intrator: Betting Big on AI: The CEO Who Turned CoreWeave into a $29 Billion Powerhouse. The valuation in its subtitle was out of date almost as soon as it was printed, first too low and then, for a while, too high. That's a hazard of writing about a stock.
The more durable printing is in Pennsylvania. On 78 acres off Greenfield Road in Lancaster stand two empty buildings that once housed printing presses. More than two million square feet of floor were built to hold heavy machines that ran all day and turned paper into something people would pay for. The presses are gone. Utility crews are working on roughly $200 million of grid upgrades to bring in power the old plant never needed. When the first phase is finished, scheduled for summer 2027, someone will walk into one of those halls, find a socket on the wall, and plug something in.
Part IIThe Playbook
What follows are the operating principles that can be drawn from Intrator's path from carbon permits to compute. Some are about financing, some about timing, and a few about temperament. Taken together they describe a way of building that doesn't come from Silicon Valley: a trader's playbook applied to infrastructure, in which the main question is never "what is the product?" but "what is scarce, who needs it, and how can it be financed before anyone else notices?"
The principles
- 1.Read every new technology as a commodity market first.
- 2.Buy the option, not the forward.
- 3.Be the bid in the liquidation.
- 4.Pay your tuition before the market opens.
- 5.When equity says no, find the lender who understands the collateral.
- 6.Never outrun your contracts.
- 7.Borrow your customer's credit rating.
- 8.Treat concentration as a phase, not a fate.
- 9.Hire the incumbents' operators.
- 10.Build the Lamborghini, then price it like infrastructure.
- 11.Treat every transaction as a means to an end.
- 12.Defend the position hard; exit it without sentiment.
Principle 1
Read every new technology as a commodity market first
Intrator came to computing with sixteen years at Natsource and five at a natural gas fund behind him, and he seems to have seen GPUs the way he had seen emission allowances and gas futures: as a scarce input whose price would be set by the gap between demand and deliverable supply. Analytics Insight's summary of his strategy fits: "CoreWeave began treating computing capacity the way a trader treats a scarce commodity. GPUs were acquired ahead of demand and contracted against future workloads, rather than simply purchased as equipment."
This framing changes the questions you ask. A software founder asks what users want. A commodities trader asks where the bottleneck is, who controls supply, how long it takes to bring new supply online, and what the forward curve looks like. In 2022 the bottleneck was Nvidia's H100. In 2026 it is power. Intrator has organized CoreWeave around the bottleneck every time, not around the product.
The lesson for operators in any boom is that the most valuable position is often not the application everyone is excited about but the constrained input everyone needs to build it.
Tactic: For any market you're entering, write down the single physical or regulatory input with the longest lead time to new supply, and build your strategy around owning access to it.
Principle 2
Buy the option, not the forward
The decision to mine with GPUs instead of bitcoin ASICs looks minor and turned out to matter enormously. Intrator's reasoning, that an ASIC engineer would always run ASICs better and that GPUs could "do all these other things," amounts to choosing optionality over efficiency in a single market. An ASIC is a forward contract on bitcoin. A GPU is a call option on any future workload that needs parallel computation.
When Ethereum collapsed, the option paid off: the same hardware moved to CGI rendering, then batch computing and medical research, then AI. A competitor holding ASICs had nothing to move to. Optionality costs something when you buy it. GPUs were the less efficient mining tool. But it pays out exactly when your first thesis fails.
Tactic: When choosing between a specialized asset optimized for your current use case and a general asset that is slightly worse at it, ask what each is worth if your current use case disappears tomorrow.
Principle 3
Be the bid in the liquidation
The 2018 crypto crash was the most important purchasing window in CoreWeave's history, and most people in the industry treated it as a disaster. Atlantic Crypto bought thousands of GPUs from distressed miners "for pennies on the dollar" and earned an estimated $80 million mining with them. Intrator credits the firm's "real chops in risk management" for its ability to buy instead of sell.
Being the buyer in a liquidation requires two things most startups lack: a balance sheet that isn't overextended when the crash comes, and a thesis about what the asset will be worth after the panic. The hedge fund background supplied both. The same instinct shows in the 2022 H100 purchase, made before ChatGPT, when demand for top-end GPUs was still a matter of conviction rather than consensus.
Tactic: Keep enough unencumbered capital and a written post-crash thesis that you can buy your industry's core assets when forced sellers appear.
Principle 4
Pay your tuition before the market opens
"Buying those initial GPUs was the tuition we paid to learn how to run this business," Intrator said in 2026. Between 2017 and 2022, CoreWeave learned to install, cool, network and run GPUs at the edge of their limits under the time pressure of crypto mining, where an idle chip was lost money. Venturo describes that "rapid, frenetic pace" as the reason the company could scale when AI demand arrived.
When the market finally opens, incumbents with more capital still have to learn the operational craft. A company that learned it earlier, cheaply, on a different business has a lead that capital can't easily buy. CoreWeave's 2021 cluster of 96 A100s for EleutherAI was trivial in revenue and invaluable as practice.
Tactic: Find a cheaper, adjacent business that forces you to master the operations your target market will eventually require, and run it until the target market arrives.
Principle 5
When equity says no, find the lender who understands the collateral
Silicon Valley venture capital largely passed on CoreWeave. Intrator turned to East Coast debt markets and, in the Financial Times's words, "started to treat CoreWeave's growth like a structured credit play." The August 2023 facility of $2.3 billion, collateralized by H100s and led by Blackstone and Magnetar, was the first time that hardware had been pledged that way. It was followed by $7.5 billion in May 2024 and roughly $18 billion in debt and equity in 2025 alone.
The insight is that the right investor depends on what you own. Venture capital pays for software margins and network effects. Private credit pays for contracted cash flows against assets with resale value. CoreWeave had the second, so it went to people who price the second. Intrator's broader argument to the Wall Street Journal, that debt built railroads, highways and homes and that the tech industry "needs to be much more cognizant of capital efficiency," is an argument about matching the type of capital to the nature of the asset.
$
Two Ways to Fund a Tech Company
How CoreWeave departed from the venture template.
| Silicon Valley default | Intrator's approach | Phase |
|---|---|---|
| Raise equity, accept dilution as the cost of growth. | Raise debt against contracted assets; use equity to complement it. | Scale |
| Pitch software margins and network effects. | Pitch collateral: "an asset you can unplug and resell." | Scan |
| Build capacity ahead of demand, funded by the next round. | "Never outrun our contracts." Build against signed offtake. | Try |
| Valuation is the headline metric. | Cost of capital is the headline metric: down 300+ bps in 2025. | Scale |
| Leverage is a sign of weakness. | Leverage is a tool; interest expense is the visible price. | Risk |
Tactic: Before your next raise, describe what you actually own in the language of the investor best placed to value it, and pitch them instead of the investors who are merely fashionable.
Principle 6
Never outrun your contracts
The shareholder letter's central claim, that CoreWeave would "size opportunities carefully, build against what we can see, and never outrun our contracts," is the discipline that makes the leverage survivable. Capacity is financed against signed, multi-year commitments from creditworthy customers. By the end of 2025, backlog was $66.8 billion with a weighted average term of five years, and by mid-2026 it had passed $100 billion.
Whether CoreWeave always held to this standard can be argued. The $6 billion Lancaster campus and the five-gigawatt ambition suggest at least some building ahead of what can be seen. But the principle is right, and it separates infrastructure finance from speculation. A data center built against a five-year contract is a project. A data center built on hope is a bet.
Tactic: Set a hard internal rule that a stated percentage of any new capacity must be contracted before capital is committed, and publish the rule to your lenders.
Principle 7
Borrow your customer's credit rating
CoreWeave's most sophisticated financial move is to finance itself largely on its customers' creditworthiness rather than its own. The February 2026 effort to raise $8.5 billion against Meta contracts, which Intrator described as "A-minus paper," is the clearest example: a company with net losses borrowing on terms that reflect the strength of the counterparty it's selling to.
The same logic runs through the equity side. OpenAI took a $350 million stake at the IPO. Nvidia anchored the IPO at $40 and later invested $2 billion at $87.20. Every major stakeholder in CoreWeave has some interest in its hardware remaining valuable. That's a fragile structure in some ways, but it means the people best placed to judge the business have their own money in it.
Tactic: Structure your largest contracts so they can serve as collateral, then take them to lenders as evidence of your customer's credit, not yours.
Principle 8
Treat concentration as a phase, not a fate
Microsoft accounted for 62 percent of CoreWeave's 2024 revenue, and at the start of 2025 one customer represented 85 percent of backlog. These numbers fueled the "red flags" coverage at the IPO. Twelve months later, by the company's account, no customer exceeded 35 percent of backlog, the customer base had grown about 150 percent, and nine of the ten leading model providers were on the platform.
The sequence matters. CoreWeave didn't try to diversify early by chasing many small customers. It used one giant customer to finance scale, then used that scale to win the others: OpenAI in March 2025, Meta through 2025 and 2026, Anthropic in April 2026. Concentration was the bootstrap, not the business model.
◷
From One Customer to Nine of Ten
How CoreWeave's dependence on a single buyer was diluted.
2021
96-GPU A100 cluster for EleutherAI, the first AI customer.
2023
Inflection AI ($300M, 22,000 GPUs); Microsoft and Stability AI sign on.
2024
Microsoft accounts for 62% of revenue.
2025
OpenAI contract (~$11.9B, five years) signed weeks before the IPO; Meta contracts follow.
2025
Year-end: no customer above 35% of backlog, down from 85% in January.
2026
$21B Meta expansion and Anthropic deal announced within 48 hours in April.
Tactic: Accept a dominant early customer if it finances your scale, but set a dated target for reducing its share and treat every new logo as a step toward that target.
Principle 9
Hire the incumbents' operators
Three traders with "zero tech background," in Fortune's words, knew what they didn't know. Starting in 2024, Intrator hired executives straight from the companies CoreWeave was competing with: Nitin Agrawal, a Google Cloud finance veteran with AWS and Microsoft history, as CFO; Sachin Jain from Oracle's AI organization as COO; Chen Goldberg from Google as senior vice president of engineering. Intrator's own blog post on the hires was titled, tellingly, "How We Know Our Strategy is Working."
This does two things at once. It brings in institutional knowledge about running a hyperscale cloud, and it signals to customers and lenders that people with the most informed view of the industry chose to leave the incumbents. Each hire is both a capability and an endorsement.
Tactic: For each critical function you're scaling, identify the incumbent that runs it best and recruit the person who ran it there, then let the market notice that they left.
Principle 10
Build the Lamborghini, then price it like infrastructure
"When you build something to do one thing," Intrator said on IPO day, "you end up with a Lamborghini, not a minivan." CoreWeave's argument against the hyperscalers is specialization: a Kubernetes-native cloud built only for GPU-intensive AI work, first to deploy Nvidia's GB200 NVL72 and GB300 NVL72 systems, without the burden of serving every type of enterprise workload. By 2026 Intrator described the company as operating "above the hardware and below the models, where the real complexity shows up."
The tension is that Lamborghinis carry premium prices and CoreWeave's financing depends on utility-like predictability. Intrator's answer has been to sell the specialized product on long-term, utility-style contracts, delivering performance through a capital structure designed for durability. Specialization wins the customer, and the contract structure keeps the lender comfortable.
Tactic: Pick the single workload your product does better than any generalist, then sell it on terms long and predictable enough to finance as infrastructure.
Principle 11
Treat every transaction as a means to an end
The IPO was cut by almost half, priced $7 below range, and rescued by a supplier. Intrator's response: "A little bigger, a little smaller, a little higher, a little lower. That's not going to matter. What's going to matter is: how do we execute on our business?" When the stock tripled three months later, his public stance was the same: "I don't really care where it is today or tomorrow or the day after."
This is a trader's detachment applied to corporate milestones. An IPO is not a coronation. It's a financing event that cuts the cost of future capital. The real scoreboard was cost of capital, which fell more than 300 basis points in 2025, not the share price on the first day. Founders who treat their listing as a verdict often make defensive decisions afterward. Intrator treated his as one trade in a long series.
Tactic: Before any major financing, write down the specific operational outcome it is meant to enable, and judge the deal afterward only against that outcome.
Principle 12
Defend the position hard; exit it without sentiment
For sixteen weeks in 2025 Intrator argued for the $9 billion Core Scientific acquisition with full conviction. He told Bloomberg he "would take issue" with analyst skepticism and publicly disputed the proxy advisers' recommendations. Then, on October 21, with the vote turning against him, he called Core Scientific "not a need to have." The deal failed nine days later, and CoreWeave moved on.
This is the portfolio manager's discipline: argue for a position as hard as it deserves, while the thesis holds, and drop it the moment it stops holding, without loyalty to sunk effort. The same pattern shows across the company's history: bitcoin dropped for Ethereum, Ethereum dropped for rendering, rendering dropped for AI. What Intrator holds onto is never the position. It's the underlying asset and the optionality it carries.
!
The Open Questions
Where the playbook is still being tested.
| Risk | Intrator's position | Status |
|---|---|---|
| GPU depreciation erodes collateral value. | Five-year client contracts show the compute retains value. | Risk |
| $35.1B debt; Q2 2026 net loss of $626M, mostly interest. | Debt is non-dilutive "fuel"; cost of capital keeps falling. | Risk |
| Power scarcity (e.g., Texas halting new grid connections). | Vertical integration into physical sites; multi-state footprint. | Try |
| Customer concentration. | No customer above 35% of backlog at end of 2025. | Scale |
| Supplier dependence on Nvidia. | Nvidia is also an investor, repeatedly increasing its stake. | Scan |
Tactic: For every major strategic position, set in advance the condition under which you'll abandon it, and when that condition arrives, say so publicly and move on.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Michael Intrator, to Forbes, 2025We founded the company on a credit card. It was like, "Okay, I'll buy a couple GPUs. What the hell?"
— Michael Intrator, All-In Podcast, March 2026We looked at Bitcoin and the mining for Bitcoin and we didn't like it. We just thought that there's some brilliant engineer that built the ASIC and they're probably going to be better at running it than we are.
— Michael Intrator, to Fortune, March 28, 2025There's a divergence between what the capital markets and what the media is thinking, and what I am feeling down in the trenches. What I am feeling is relentless demand.
— Michael Intrator, to the Wall Street Journal's Steve Rosenbush, 2026This particular episode of technology build-out required mobilization of a scale of capital that the world rarely sees. And with that, you have to bring to bear every tool to drive down the cost.
— Michael Intrator, to Fortune, March 2024Always keep a healthy disregard for the impossible while being grounded in reality.
Maxims
- Find the bottleneck. In every boom, the durable position belongs to whoever controls the scarce input, not the popular application.
- Prefer the option. A general-purpose asset that is slightly worse today is worth more than a specialized one when your first thesis dies.
- Buy from forced sellers. A crash is a transfer of assets from the overextended to the prepared; arrange to be prepared.
- Learn on someone else's market. The cheapest place to master an operation is a business that needs it before your real market does.
- Match capital to collateral. Fund machines with debt and ideas with equity, and pitch each to the investor who knows how to price it.
- Never outrun your contracts. Leverage is survivable only when every dollar of capacity is backed by a signed obligation to pay.
- Borrow strength. A creditworthy customer's contract can lower your cost of capital more than your own balance sheet can.
- Concentration can be a ladder. One giant customer can finance the scale that wins the next nine.
- Milestones are means. An IPO, a valuation, a share price: judge each by what it enables, not by how it looks on the day.
- Hold the asset, not the position. Argue hard while the thesis holds, then walk away without ceremony when it doesn't.
