David Huang was trying to remember when he’d stopped believing.

He was thirty-seven, had a PhD in computer science from Berkeley, had spent five years at Nvidia working on GPU architecture, and was currently—well, “currently” was complicated.

Six months ago, he’d been the VP of Engineering at Prometheus AI, a startup that was going to revolutionize natural language processing. They had $180 million in funding, 120 employees, and a model that could supposedly understand context better than GPT-4. David had believed in it. He’d turned down offers from Google and Anthropic to join Prometheus because he thought they were building something real.

Then reality had intervened.

The problem—David had realized too late—was that their model wasn’t actually better than GPT-4. It was different. Which sounded like a distinction without a difference, but in Silicon Valley, “different” meant “no compelling reason to switch.” And if customers didn’t switch, you didn’t have a business.

They’d spent $120 million training their model. Another $40 million on sales and marketing. Another $20 million on the beautiful office in San Mateo with the artisanal coffee bar and the meditation room and the massage chairs. They’d gotten their revenue up to $800,000 a quarter, which their CFO kept saying was “hockey stick growth” in percentage terms while carefully not mentioning that it meant they were losing money on every single customer.

Two weeks ago, Prometheus had missed payroll. Not by much—just forty-eight hours—but that was forty-eight hours of everyone checking their bank accounts and realizing that something was very wrong.

David had been in a meeting with the CEO, the CFO, and their lead investor when they’d gotten the email from Sequoia Harbor Capital. It was professionally worded but the message was clear: you owe us $85 million in forty-one days, and we want our money.

“What happens if we don’t have it?” the CEO had asked.

The CFO had pulled up the loan agreement. “They seize the collateral.”

“Which is?”

“All of our GPUs. The entire training cluster.”

The CEO had looked confused. “But we need those to run the service.”

“I know.”

“So if they take them, we can’t operate.”

“I know.”

“But if we can’t operate, we can’t generate revenue.”

“I know.”

“But if we can’t generate revenue, we can’t pay them back anyway.”

“I know.”

That was when David had understood that the incentives were all wrong. Sequoia Harbor didn’t care about getting paid back anymore. They cared about getting something. Anything. Even if it meant killing the company to get it.

Now David was sitting in his apartment in San Francisco, updating his LinkedIn profile and trying to figure out how to explain to recruiters that his last company had imploded. His phone rang. Unknown number. He almost didn’t answer, but something made him pick up.

“David Huang?” A woman’s voice.

“Yes?”

“My name is Sarah Chen. I’m a portfolio manager at CalPERS. I’m trying to understand the private credit exposure in AI companies, and I was told you might be able to help me.”

David had never heard of Sarah Chen. He had no idea how she’d gotten his number. But he recognized the tone in her voice. It was the same tone he’d had when he’d first realized Prometheus was doomed. The tone of someone who’d just figured out something terrible and needed to know if they were crazy.

“What do you want to know?” he asked.

“I want to understand how equipment collateralization works. Specifically, I want to know if it’s common for the same assets to be pledged to multiple lenders.”

David laughed. It wasn’t a happy laugh.

“Common?” he said. “It’s standard practice. Every AI company I know has the same setup. Equipment lender, venture debt, bank line. All of them secured by the GPUs. Nobody cared as long as the companies were growing. But now…”

“Now they’re failing,” Sarah said.

“Now they’re failing,” David agreed. “And in about six months, every private credit fund in Silicon Valley is going to realize they’re holding collateral that’s been pledged three times over and is depreciating by 30% a year.”

There was a long silence on the line.

“David,” Sarah said finally, “how many companies are in this situation?”

“You mean how many AI companies have cross-collateralized equipment loans?”

“Yes.”

“All of them,” David said. “Every single one.”

Senator Patricia Mills did not have time for this bullshit.

She was in her Senate office in Washington DC, supposed to be reviewing the infrastructure bill markup, but instead she was reading testimony for tomorrow’s hearing on public pension sustainability. The hearing was her idea—she chaired the Banking Subcommittee on Financial Institutions—and she’d spent six months putting it together.

Nobody had paid attention. The Washington Post had run a three-paragraph story on page A19. CNN hadn’t covered it at all. Even her own staff seemed to think it was boring.

The problem with pensions, Patricia had learned, was that they were simultaneously the most important thing in millions of people’s lives and the most boring topic in American politics. Nobody cared about pensions until their pension disappeared. And by then it was too late.

She was sixty-one, had been in the Senate for twelve years after three terms in the House, and she’d developed a reputation as someone who actually read the bills and asked uncomfortable questions. This had not made her popular with leadership, but it had made her popular with her constituents in Michigan, who appreciated that she gave a shit.

The pension hearing had grown out of a conversation she’d had eighteen months ago with the Detroit fire chief. He’d mentioned, casually, that the firefighters’ pension fund was only 68% funded. Patricia had asked what that meant. He’d explained: for every dollar the fund owed to retirees, it only had 68 cents. Where was the other 32 cents coming from? Good question.

That had led her down a rabbit hole. Detroit’s firefighters weren’t unique. Across the country, public pensions were underfunded by $1.4 trillion. And most of them had made the same bet: they’d moved money out of boring bonds and into exciting alternatives. Private equity. Private credit. Real estate. Anything that might earn the 7% returns they needed to meet their obligations.

The problem, Patricia had discovered, was that nobody really knew what those alternative investments were worth. The funds said they were worth X. The pension managers believed them. The actuaries put X in their models. And everyone felt good.

But what if X was wrong?

Her chief of staff, Michael Chen—no relation to Sarah Chen, though Patricia would meet Sarah in a few weeks and they’d discover they were both asking the same questions—knocked on her open door.

“Senator, your witness list for tomorrow.”

She glanced at it. Four pension fund managers. Two academics. One whistleblower from a failed PE firm. All of them prepared to say, in careful bureaucratic language, that everything was fine and the system was working as intended.

“Michael, who’s the most interesting witness?”

“Define interesting.”

“Who’s going to say something that pisses people off?”

Michael smiled. “Professor Zimmerman from Princeton. He’s going to testify that public pensions are carrying at least $300 billion in phantom gains from overstated private market valuations.”

“Will he say it that directly?”

“I talked to him yesterday. He’ll be diplomatic but firm.”

“Get me on the phone with him. Tonight. I want to know exactly what he’s going to say and what backup he has.”

Michael made a note. “Also, you got a request for a meeting from someone at CalPERS. Sarah Chen. She’s a portfolio manager, wants to talk about alternative investment exposure.”

Patricia looked up. “CalPERS is sending a mid-level portfolio manager to Washington to talk to a senator? That’s unusual.”

“She didn’t go through official channels. She called your personal cell phone.”

“How did she get my personal cell?”

“No idea. But she said it was urgent. Something about systemic risk in private credit markets.”

Patricia felt the familiar tingle she got when someone said something important. The same tingle she’d felt when the fire chief mentioned 68% funded. The same tingle she’d felt when she’d first read about Enron’s special purpose entities.

“Set up the meeting. Tomorrow, before the hearing. I want to know what she knows.”

Rebecca Torres had always been the person who asked uncomfortable questions. At parties, she was the one who pointed out that the retirement plan numbers didn’t add up. At Thanksgiving, she was the one who fact-checked her uncle’s political rants in real-time. It had not made her popular.

But it had made her an excellent risk officer.

She was forty-two, had spent eighteen years at Consolidated American Insurance—CAI was one of the fifty largest insurance companies in the country—and for the past two years, she’d been in charge of alternative investment risk. Which meant that when CAI invested its $84 billion in assets, Rebecca was supposed to make sure they didn’t do anything stupid.

She was currently failing at her job.

The problem had started three months ago, when one of their private credit funds—Sequoia Harbor Capital—had sent their quarterly report. Everything looked fine on the surface. Portfolio performing. Net asset value up 8%. Returns on track.

But there was a footnote. A tiny footnote that most people would have skipped right over: “As of Q4 2025, certain portfolio companies in the technology sector have experienced delays in achieving revenue targets. Management remains confident in the long-term prospects of these investments. Some portfolio companies have requested extensions on payment terms, which have been granted at revised rates.”

Rebecca had read that footnote seventeen times. She’d then spent six weeks quietly investigating what “revised rates” meant.

What it meant was that the loans were going bad, and instead of admitting it, the fund was “extending” them—lending the companies more money at higher rates so they could make their interest payments. It was the financial equivalent of paying off one credit card with another credit card. It could work for a while, as long as someone kept extending credit.

The moment the music stopped, everyone would realize they were all broke.

She was sitting in her office in Hartford, Connecticut—CAI had never bothered to move to a fashionable city—staring at a spreadsheet that showed their exposure to private credit. They had $8.4 billion invested across thirty-seven funds. She’d been trying to map the underlying portfolio companies, just like Sarah Chen at CalPERS had been doing on the other side of the country, though neither of them knew about the other yet.

What Rebecca had found was worse than she’d imagined.

CAI wasn’t just invested in Sequoia Harbor. They were in eight other private credit funds, all of which had made similar loans to similar companies. AI startups. Enterprise software companies that used AI. Data infrastructure companies. Cloud providers.

She started adding up the numbers. Not the fund-level numbers—those looked fine. The company-level numbers. The real exposure.

When she got to the total, she had to check it three times because she didn’t believe it.

$3.2 billion.

CAI had $3.2 billion in exposure to AI-related companies across their private credit portfolio. That was 38% of their total alternative investment allocation. And according to industry regulations, CAI needed to maintain a Risk-Based Capital ratio of at least 200%. They were currently at 287%, which seemed healthy.

Until you realized what would happen if that $3.2 billion got marked down.

Rebecca opened a new spreadsheet and started running scenarios.

Scenario 1: 20% markdown

– Loss: $640 million
– New RBC ratio: 264%
– Status: Uncomfortable but manageable

Scenario 2: 40% markdown

– Loss: $1.28 billion
– New RBC ratio: 233%
– Status: Regulatory scrutiny, possible restrictions

Scenario 3: 60% markdown

– Loss: $1.92 billion
– New RBC ratio: 198%
– Status: Below regulatory minimum, forced to raise capital or sell assets

Scenario 4: 80% markdown

– Loss: $2.56 billion
– New RBC ratio: 161%
– Status: State regulator takeover, forced liquidation

She sat back in her chair and looked out the window at Hartford’s skyline. It was not an impressive skyline. Hartford had peaked economically in about 1965 and had been on a slow decline ever since. Half the insurance companies that had once called Hartford home were gone now, merged or acquired or simply failed.

CAI had survived because they’d been careful. Conservative. They didn’t make stupid bets.

Except apparently, they had.

Her phone rang. It was Gerald Stevenson, the Chief Investment Officer. Rebecca’s boss.

“Rebecca, have you seen the news?”

“What news?”

“Inflection AI just filed for bankruptcy protection. Chapter 11. Couldn’t make their debt payments.”

Rebecca’s blood went cold. “Which lenders?”

“Sequoia Harbor was one of them. And two others we’re also invested in.”

“How much was outstanding?”

“$150 million across the three lenders. All of it secured by equipment. Nvidia GPUs.”

Rebecca pulled up her spreadsheet and searched for Inflection. There it was. CAI was an LP in all three funds that had lent to Inflection. Their proportional exposure was $8.2 million.

Not enormous. Not catastrophic. But…

“Gerald,” she said slowly. “What happens to those GPUs now?”

“What do you mean?”

“I mean, who gets them? All three lenders have security interests. Someone has to get paid first.”

There was a pause. “I… I assume that’s defined in the intercreditor agreements.”

“And if it’s not?”

“Then it goes to bankruptcy court.”

“Which takes how long?”

“Six months? A year?”

“And while they’re fighting, the GPUs sit idle?”

“I suppose so. Why?”

Rebecca looked at her screen. At the numbers. At the scenario analyses. At the little model that showed their regulatory capital evaporating.

“Gerald, I think we need to call an emergency board meeting,” she said. “Because if Inflection is the first domino, and if those GPUs are stuck in bankruptcy court, and if the other AI companies are in similar situations…” She trailed off.

“Rebecca, finish the thought.”

“Then we’re fucked, Gerald. We’re completely fucked.”

Dr. James Morrison—everyone called him Jamie—was thirty-four years old and running on four hours of sleep. This was normal. What wasn’t normal was the email from his CFO with the subject line: “We need to talk. NOW.”

Jamie was sitting in Conference Room 3B of his company’s Palo Alto headquarters, which was really just a renovated warehouse with exposed brick and aspirational whiteboard walls covered in equations. Through the glass walls he could see forty-seven of his employees—engineers mostly, a few PMs, one designer who claimed to understand “AI UX” whatever that meant—all of them working on laptops, drinking cold brew, believing in the vision.

The vision was elegant: an AI model that could understand medical imaging better than any radiologist. They’d trained it on 400 million X-rays, CT scans, and MRIs. It could detect lung cancer eighteen months earlier than human doctors. It could spot brain aneurysms that showed up on scans as barely-visible shadows. It was, by any objective measure, a medical miracle.

It was also burning $12 million a month.

His CFO, Priya Chandran, appeared in the doorway. She was thirty-one, had worked at Goldman before business school, and had joined the company because she believed in saving lives through AI. Now she looked like she was trying not to cry.

“Conference room,” she said quietly.

They moved to the small interior room with no windows. The room where you delivered bad news. Jamie had done three layoffs in this room. He hated it.

Priya opened her laptop, pulled up a spreadsheet, and rotated it to face him.

“We have eleven million in the bank,” she said. “At current burn, that’s forty-two days.”

Jamie felt the room tilt slightly. “What about the Series C?”

“Khosla passed. NEA passed. A16Z said they’d ‘circle back in Q3.’ That’s VC for ‘no.’”

“The revenue pipeline—”

“Jamie.” Her voice was gentle but firm. “We did $800,000 in revenue last quarter. We burned $36 million. The unit economics don’t work. Every hospital we sign costs us $200,000 to onboard and they pay us $30,000 a year. We lose money on every single customer.”

“But we’re saving lives—”

“I know. I know we are. But we’re also running out of money.” She pulled up another tab. “And there’s something else. The equipment loans.”

Jamie felt his stomach drop further. The equipment loans. Of course.

Eighteen months ago, when they’d needed to scale up their training infrastructure, venture debt had seemed like a good idea. Borrow $50 million against the GPUs, pay it back when the Series C closed, everyone wins. The interest rate was high—14%—but that was fine. They were going to be profitable by Q4 2025.

Except they weren’t profitable. They weren’t even close.

“How much do we owe?” he asked, though he knew the number.

“$52 million. Principal plus interest. Due in forty-seven days.”

“Can we extend?”

“I’ve been trying. Sequoia Harbor says no. They want payment in full or they’ll seize the collateral.”

Jamie laughed. It came out harsh and bitter. “They’ll seize the GPUs? What are they going to do with them? We need those GPUs to run the service. Without them, we have no product. Without a product, we have no revenue. Without revenue, we can’t pay them back anyway.”

Priya looked at him with something like pity. “Jamie, I don’t think they care about getting paid back at this point. I think they want to seize the assets before someone else does.”

“Someone else?”

She pulled up an email. “I got this this morning from our equipment financing company. Apparently, when we bought the GPUs, we financed them through Nvidia’s preferred vendor. They’ve got a lien on the equipment too. And there’s a third lien from our bank line of credit, which is secured by ‘all business assets including equipment.’”

Jamie stared at the screen. “The same GPUs are collateral for three different loans?”

“That’s what our lawyer says. She’s looking into the intercreditor agreements but…” Priya trailed off.

“But what?”

“But she thinks if we default, all three lenders will try to seize the equipment simultaneously. And it’ll go to bankruptcy court. Which takes six to eighteen months. During which time the GPUs will sit idle while lawyers fight over who gets them.”

“But… but the GPUs will be worthless by then. Nvidia’s already announced the B100. The H100s will be obsolete in a year.”

“Exactly.”

Jamie put his head in his hands. Outside the glass walls, his employees kept working, kept believing. One of them—Sarah Kim, their lead ML engineer—was laughing at something on her screen. She’d turned down offers from OpenAI and Anthropic to work here. She believed they were going to save lives.

“What do we do?” he asked.

Priya closed her laptop. “I don’t know, Jamie. But I think we need to tell the team.”

Marcus Webb had the best office in the Embarcadero Center, which was saying something. Forty-second floor, floor-to-ceiling windows, view of the Bay Bridge that made real estate agents weep. He was paying $85,000 a month for it, which his partners thought was insane until they realized how useful it was for closing deals. Nothing said “we’re making money hand over fist” quite like casual opulence.

Of course, they weren’t actually making money hand over fist. Not anymore. But the office helped maintain the illusion, and in private equity, illusions were half the business.

Marcus was fifty-three, looked forty-five (personal trainer, good genes, expensive dermatologist), and had spent twenty-seven years in finance without ever quite becoming rich enough to stop working. He’d made it to partner at Sequoia Harbor Capital, a $12 billion private credit fund that had spent the last three years lending money to AI companies at rates that would have made loan sharks blush.

Fourteen percent interest. Warrants for 5% equity. Balloon payments in eighteen months. And—this was the clever part—equipment as collateral. All those beautiful, expensive Nvidia GPUs sitting in data centers, training models that would definitely, absolutely, certainly generate enough revenue to pay back the loans.

Marcus had structured forty-seven of these deals personally. He was good at it. He knew all the AI founders—the Stanford kids, the ex-Google engineers, the DeepMind refugees. He spoke their language. He understood transformer architecture and attention mechanisms and could nod intelligently when they explained why their approach to reinforcement learning was revolutionary.

What he also understood, and what none of them seemed to grasp, was that at 14% interest, you needed to grow revenue 40% year-over-year just to keep up with your debt service. And that assumed your costs didn’t grow. Which in AI, they always did.

His phone buzzed. Sarah Chen. He almost didn’t answer—he’d been avoiding her calls for three weeks—but something made him pick up.

“Marcus,” she said, without preamble. “I need you to explain something to me.”

“Good morning to you too, Sarah.”

“I’m looking at our exposure. Our real exposure. Not the fund-level summaries. I went through every portfolio company in every PE and PC fund we’re in.”

Marcus felt his stomach drop. “That must have taken a while.”

“Three weeks. Marcus, we’re in twenty-three different funds. Those funds have positions in 180 different companies. And you want to know what I found?”

He didn’t want to know. He absolutely did not want to know. But he said, “Tell me.”

“Forty-two of those companies are in AI or AI-adjacent businesses. Same companies, different funds, different structures. We’ve got direct exposure, indirect exposure, exposure to their suppliers, exposure to their customers. It’s like looking at a house of cards where every card is the same picture.”

“That’s… that’s diversification,” Marcus said, knowing it sounded weak even as he said it.

“Diversification is supposed to reduce risk. This is concentration masquerading as diversification. And Marcus? I called you because I know Sequoia Harbor is in at least eight of these deals. So I’m going to ask you a question, and I need you to be honest with me.”

Marcus looked out his expensive windows at the Bay Bridge, glinting in the morning sun. Somewhere out there, in a data center in Santa Clara or Palo Alto or South San Francisco, were millions of dollars worth of Nvidia H100 GPUs, all of them pledged as collateral on loans that would never be repaid.

“I’m always honest with you, Sarah.”

“Bullshit. But try this time. Those AI companies you’ve lent to—the ones with equipment collateral—do any of them actually own the equipment free and clear?”

The silence stretched out for five seconds. Ten. Fifteen.

“Marcus?”

“Define ‘free and clear.’”

“Jesus Christ.”

“Sarah, it’s more complicated than—”

“How many lenders per asset?”

“That’s not exactly—”

“How many, Marcus?”

He closed his eyes. “It varies.”

“Give me an average.”

“Two point seven.”

“Two point seven lenders per asset. You mean the same GPU is collateralizing multiple loans.”

“Not exactly the same—”

“Don’t parsing words with me. You’ve got cross-collateralization, don’t you? Multiple lenders, same assets, different seniority claims that aren’t clearly defined.”

“The inter-creditor agreements are very specific about—”

“Marcus, what happens when one of these companies fails?”

And there it was. The question he’d been asking himself at 3 AM for the past six months. The question that made him pour three fingers of Macallan instead of two. The question that his partners had told him not to worry about because they had “sophisticated risk models.”

“Sarah, these are good companies. Strong teams. Proven technology—”

“What happens when one fails?”

He took a breath. Outside his window, a container ship was moving slowly under the Bay Bridge, heading for Oakland. Thirty years ago, he’d wanted to be a writer. He wondered sometimes what that life would have been like. Probably involved fewer 3 AM whiskey sessions.

“We’ll find out soon enough,” he said finally. “Inflection AI just missed their debt payment.”

March 15, 2026

Sarah looked up from her wall of monitors to see if anybody else was in the office yet The thing about financial catastrophes, she’d later reflect, is that they never announce themselves with a bang. They whisper. And if you’re not listening—really listening—you miss the moment when you could have done something about it.

She was listening that Monday morning, sitting in her grey-walled office on the third floor of CalPERS headquarters in Sacramento, staring at a spreadsheet that made no sense. The numbers were right there, validated by three different systems, signed off by two portfolio managers and a compliance officer. Everything was in order.

That was the problem.

Sarah had spent fifteen years managing pension money, and she’d developed what her ex-husband called her “spider sense”—a visceral reaction to things that looked right but felt wrong. The spreadsheet in front of her showed that CalPERS had $48 billion in alternative investments, split between private equity and private credit. Standard stuff. Well-diversified across sectors. Clean.

But then she’d done something unusual. She’d started mapping the underlying portfolio companies across all their different PE and PC holdings. Not the fund level—everyone looked at the fund level. She’d gone deeper, company by company, asset by asset. And that’s when she saw it.

The same names. Over and over. Not exactly the same—the fund managers were too clever for that. But variations. AI infrastructure companies. Data center operators. GPU financing vehicles. Cloud service providers. Enterprise AI software. It was like looking at a hundred different jigsaw puzzles and slowly realizing they were all pieces of the same picture.

She picked up her phone and called Marcus Webb.