July 8, 2026

Rebecca Torres was in a meeting about flood insurance when her phone started buzzing. Not just one alert – a cascade of them. Bloomberg. Reuters. Financial Times. All with the same headline.

She glanced down: “China Announces ‘Enhanced Safety Inspections’ for Taiwan Strait Shipping”

The meeting continued around her. Someone was talking about actuarial tables. Rebecca opened the Bloomberg story.

BEIJING – The People’s Republic of China announced today that all commercial shipping passing through the Taiwan Strait will be subject to “enhanced safety and customs inspections” beginning July 15th. The Ministry of Transport cited “increased maritime safety concerns” and the need to “ensure compliance with international shipping standards.”

The announcement includes:
– Mandatory 72-hour advance notification for all vessels
– Physical inspection of cargo manifests at designated checkpoints
– “Expedited processing” for ships carrying “non-sensitive materials”
– Indefinite delays possible for vessels carrying “dual-use technologies”

Taiwan’s government condemned the move as an “illegal blockade.” The U.S. State Department called it “concerning and destabilizing.”

Rebecca’s hands went cold. She knew exactly what “dual-use technologies” meant. Semiconductors. Advanced chips. GPUs.

She stood up. “I need to step out.”

In the hallway, she called Sarah Chen.

“Sarah, are you seeing this?”

“The Taiwan thing? Yeah. I’m on a call with our Asia portfolio managers right now. They’re panicking.”

“Sarah, 90% of advanced GPUs are manufactured in Taiwan. If shipping gets delayed or disrupted-”

“-the supply chain freezes. I know. We’re trying to figure out what this means for our portfolio companies.”

Rebecca pulled up her risk model on her phone. She’d built in a scenario for “supply chain disruption” but she’d assumed it would be a natural disaster or a factory fire. Not a quasi-blockade.

“How long until companies start running out of chips?” she asked.

“Depends on inventory. Most companies keep 60 to 90 days of buffer. If ships get delayed by weeks instead of days, we’re looking at shortages by September, October latest.”

“And if there are shortages?”

“GPU prices spike. Companies that are barely surviving now go under. Companies that were planning to expand can’t get equipment. The whole sector freezes.”

Rebecca thought about Pacific Life’s stress test. The one that showed a 15% probability of systemic crisis. That model had assumed steady equipment supply.

“Sarah, this changes everything. If GPU supply gets constrained, even the healthy AI companies are at risk.”

“I know. I’m running new scenarios now. Rebecca, this could be the thing that tips it over.”

After they hung up, Rebecca went back to her office and pulled up the portfolio. Pacific Life had $4.2 billion in AI-related exposure. She’d been modelling that as three separate risk buckets: company-specific risk, sector risk, and systemic risk.

She’d been treating them as independent variables. But they weren’t independent. They were all connected by the same supply chain. The same chips. The same Taiwan Strait that Chinese naval vessels were now “monitoring.”

She opened a new spreadsheet and started modeling.

Scenario: Taiwan Strait Delays

Assumptions:
– Average shipping delay: 2-3 weeks
– Duration: 3-6 months
– GPU price increase: 200-400%
– Companies with <60 days inventory: immediate crisis
– Companies with 60-90 days inventory: crisis by Q4

Impact on portfolio:
– 15 companies go critical immediately
– 20 more companies by end of Q3
– Total exposure at risk: $3.8 billion

She ran the numbers three different ways. They all came out in the same range.

Pacific Life’s surplus would drop from $7.25 billion to $4.1 billion. Below the regulatory minimum. Not approaching insolvency. Actually insolvent.

Her boss, Robert Chen, appeared in her doorway. “You saw the Taiwan news.”

“I’m modelling it now.”

“And?”

“If this lasts more than sixty days, we’re insolvent.”

Robert was quiet for a moment. Then he stepped into her office and closed the door.

“Rebecca, I need you to do something. Don’t send that model to anyone yet. Don’t email it. Don’t save it to the shared drive. Just keep it on your laptop.”

“Robert, if we’re heading toward insolvency, we have a duty to disclose-”

“We have a duty to not cause a panic. Right now, this is a hypothetical. Ships are still moving. Companies still have inventory. If we announce that we might be insolvent based on a China trade policy that might not even be enforced, we trigger the very crisis we’re trying to avoid.”

Rebecca looked at him. She’d worked with Robert for eight years. He’d always been the guy who followed the rules. Who erred on the side of caution. Who valued transparency.

“You’re asking me to hide a risk assessment.”

“I’m asking you to wait forty-eight hours. Let’s see if ships actually get delayed. Let’s see if this is real or just saber-rattling. Then we’ll decide what to disclose.”

After he left, Rebecca sat alone in her office, staring at her spreadsheet. She thought about Sarah Chen, testifying before the Senate. About how she’d been told her analysis was “too pessimistic.”

She thought about Marcus Webb, who’d seen the cross-collateral problem in December and been told he was being “too conservative.”

She thought about all the risk managers and analysts who’d seen pieces of this crisis coming and been told to wait. To see if it got better. To not cause a panic.

And now here she was, being told the same thing.

She looked at her model. At the numbers that showed Pacific Life heading toward insolvency. At the 2.8 million policyholders who deserved to know the truth.

Then she saved the file. Encrypted it. And sent a copy to her personal email.

Because when this was over – when Pacific Life failed and everyone asked why nobody warned them- Rebecca wanted there to be a record that someone had tried to tell the truth.

Even if nobody had wanted to listen.

—–

That evening, David Huang was at dinner with his girlfriend when she asked him about Taiwan.

“Should we be worried?” Lisa asked. “My dad has TSMC stock. He’s freaking out.”

David had been following the news all day. He’d been tracking shipping data, reading analyst reports, talking to his former colleagues at other AI companies.

“In the short term? Maybe a dip. Long term, I don’t think China actually blockades Taiwan. Too much risk of U.S. response.”

“But the shipping delays?”

“Those are real. China’s playing games. They’ll inspect some ships, delay some cargo, make everyone nervous. But they won’t actually stop trade. They can’t afford to.”

Lisa looked at him. “You don’t sound convinced.”

David took a sip of wine. The truth was, he wasn’t convinced. Because he’d seen how fragile the AI supply chain was. How everything depended on TSMC manufacturing in Taiwan and ships making it through the Strait and nothing going wrong.

And now something was going wrong.

“Lisa, you know how I told you about Artemis failing?”

“The GPU freeze thing?”

“Yeah. The reason that killed us so fast is that AI companies have no slack in the system. We’re all running at the edge of sustainability. We’re all dependent on equipment we don’t own, capital we’re about to run out of, and a supply chain that assumes nothing goes wrong.”

“And now something’s going wrong.”

“And now something’s going wrong.”

His phone buzzed. A text from Jamie, his old CEO: “You seeing this Taiwan situation? Companies are going to start hoarding chips. Prices are going to spike.”

David texted back: “How fast?”

“60 days. Maybe 90. Anyone who needs new equipment after September is screwed.”

David put his phone away. He’d been unemployed for two months. He’d had three interviews at Meta, all of them going well. He was expecting an offer next week.

But now he was thinking about something else. He was thinking about all the companies that were barely surviving. The ones with 90 days of runway and 60 days of GPU inventory. The ones that had been hoping to raise another round or get acquired or somehow make it through Q3.

They weren’t going to make it.

“David?” Lisa was looking at him. “You okay?”

“Yeah. Sorry. Just thinking.”

“About what?”

“About how many more companies are going to fail in the next three months.”

—–

In Washington, Janet Rodriguez sat in her office at the FDIC, reading briefing memos about Taiwan. She was a Regional Director, which meant she supervised banks and insurance companies in six states. She’d been with the FDIC for nineteen years. She’d worked through the 2008 crisis. She knew what contagion looked like.

And this looked like contagion.

Her phone rang. Her boss, the Deputy Director for Risk Management.

“Janet, I’m sending you a list of twelve insurance companies. I need risk assessments by Friday.”

“All twelve?”

“All twelve. They all have significant exposure to private credit funds with AI portfolios. With the Taiwan situation, we need to know who’s vulnerable.”

“That’s a forty-eight hour turnaround for twelve complex assessments.”

“I know. Get your team on it. This is priority one.”

After she hung up, Janet pulled up the list. She recognized three of the names immediately. Pacific Life. MetLife. Prudential. The others were smaller regional carriers.

She called her senior analyst. “Mike, I need you to drop everything. We’re doing deep-dive risk assessments on these twelve insurance companies. I need to know their AI exposure, their capital ratios, their liquidity positions, and their stress test results.”

“What’s the concern?”

“Taiwan. If GPU supply gets disrupted, AI companies fail. If AI companies fail, private credit funds take losses. If funds take losses, insurance companies mark down. If enough insurance companies mark down at the same time, we’ve got a solvency crisis.”

Mike was quiet for a moment. “How bad are we talking?”

“Bad enough that the Deputy Director wants assessments in forty-eight hours.”

After the call, Janet sat at her desk, looking at the list of twelve companies. She’d spent nineteen years at the FDIC. She’d seen banks fail. She’d seen insurance companies get into trouble. But she’d never seen anything quite like this.

Because this time, everyone had made the same bet. Everyone had piled into the same sector. Everyone was exposed to the same risk.

And that risk just got a lot bigger.

—–

By the end of the week, the first shipping delays were reported. A container ship bound for Los Angeles had been “requested” to stop for inspection in the Taiwan Strait. The inspection had taken four days. The ship was still there.

Two more ships diverted their routes entirely, adding ten days to their journey.

GPU futures prices jumped 40% in two days.

And in insurance companies, pension funds, and private equity firms across the country, risk managers started updating their models.

Because the thing they’d all been worried about – the thing that could turn a manageable crisis into a systemic one—had just happened.

The supply chain was breaking.

June 3, 2026

Marcus sat in Tom Hendricks’s office, looking at the spreadsheet that would end their fund. The numbers were worse than he’d modeled. Much worse.

“Walk me through it again,” Tom said quietly.

Marcus pulled up the portfolio view. “We have forty-seven portfolio companies. Twelve have failed or are in bankruptcy. That’s $680 million in exposure.”

“Recovery?”

“TechLease marked their collateral at 25 cents on the dollar last week. That’s our best guide. If we use their numbers, we recover $170 million. We lose $510 million.”

“On a $2.2 billion fund.”

“Yes. But that’s not the worst part.”

Tom looked at him. “There’s a worse part?”

Marcus clicked to the next tab. “Of the remaining thirty-five companies, eighteen are in severe distress. They’re negotiating forbearance agreements. They’re burning through cash. They’re laying off employees. If even half of those fail—and I think more than half will—that’s another $420 million in losses.”

“Total?”

“$930 million. On a $2.2 billion fund. That’s a 42% loss.”

Tom stood up and walked to the window. It was a beautiful June day. The Bay Bridge sparkled in the sun. Somewhere out there, thousands of pension funds and insurance companies were looking at the same numbers and having the same conversation.

“We have to tell the LPs,” Tom said.

“I know.”

“They’re going to panic. They’re going to try to redeem. We’ll have to gate the fund.”

“I know.”

“Marcus, when you first told me about the problem, I didn’t listen. I said you were being too conservative. I was wrong.”

Marcus had been waiting for this conversation for months. He’d run the numbers in December. He’d warned Tom in January. He’d built the contagion model in February. And every time, Tom had said: “Let’s wait and see.”

Now they were seeing.

“What do we tell them?” Tom asked.

Marcus had already drafted the letter. He’d been working on it for three weeks. “We tell them the truth. We tell them that the AI sector has experienced unprecedented stress. We tell them that our portfolio has been materially impacted. We tell them that we’re marking down our positions to reflect current market realities. And we tell them that we’re suspending redemptions to protect remaining LPs from a fire sale.”

“They’re going to sue us.”

“Probably. But if we don’t gate, they’ll all redeem at once and we’ll have to liquidate at the worst possible time. At least this way we can manage the losses.”

Tom sat down heavily. “How many other funds are in the same position?”

“All of them. Everyone was in AI. Sequoia Harbor, Khosla, Andreessen, Founders Fund—they all have massive exposure. Some of them have already gated. More will follow.”

“So it’s systemic.”

“It’s systemic.”

Tom looked at Marcus for a long moment. Then he said: “You tried to warn me. I should have listened.”

“Would it have mattered?”

“Probably not. By the time you saw it clearly, the losses were already baked in. We just hadn’t admitted them yet.”

That afternoon, they sent the letter to LPs. Marcus had written it carefully, trying to balance honesty with reassurance. But there was no good way to say: “We’ve lost half a billion dollars of your money.”

 

TO: Limited Partners
FROM: Sequoia Harbor Capital Management
RE: Portfolio Update and Redemption Suspension

Dear Partners,

We are writing to inform you of significant developments in our portfolio and important changes to fund operations…

Sector Developments: The AI sector has experienced severe stress in Q2 2026. Multiple portfolio companies have filed for bankruptcy protection. Equipment financing disputes have frozen assets across the sector. This has resulted in significant mark-downs across our portfolio…

Portfolio Impact: We have marked down twelve positions to zero, representing $680 million in capital. An additional eighteen positions have been marked down by an average of 55%. Total estimated losses: $930 million…

Redemption Suspension: Effective immediately, we are suspending redemption rights for 180 days. This suspension is necessary to protect remaining LPs from forced liquidation of assets at distressed prices…

 

Within an hour, the phone started ringing. Marcus let it go to voicemail. There was nothing he could say that the letter didn’t already explain.

By 6 PM, three other major funds had announced similar suspensions. By midnight, it was eleven funds. Total capital frozen: $14 billion.

Marcus sat in his expensive office, looking at the view he wouldn’t be enjoying much longer. His firm was going to survive—barely. But his career as a partner was effectively over. Nobody would trust him with capital again.

His phone buzzed. Sarah Chen: “I saw the letter. I’m sorry. How are you holding up?”

He typed back: “I’ve been better. CalPERS marking down yet?”

“June 30th. We’re taking $6.2 billion in losses. Funded ratio drops to 66%. There are going to be benefit cuts.”

Marcus stared at that message. $6.2 billion in losses at one pension fund. And CalPERS wasn’t even the biggest. New York State was bigger. Texas was bigger. All of them would have to mark down.

He thought about Rebecca Torres’s model. About the cascade. About how everyone had seen this coming and nobody had done anything until it was too late.

“This is how it happens,” he typed back. “Not all at once. One admission at a time. Until everyone’s admitted it and we realize we’ve lost hundreds of billions.”

“I know,” Sarah replied. “I testified about this in February. Nobody listened.”

“Would it have mattered if they had?”

“Probably not. By the time I saw it clearly, the losses were already baked in. We just hadn’t admitted them yet.”

Marcus put his phone down and looked out at San Francisco. The city glittered in the twilight. People were coming home from work, going to dinner, living their lives. None of them knew yet what was coming.

But they would soon. Because this was just the beginning. The AI sector collapse was going to spread to private credit. Private credit would spread to insurance companies. Insurance companies would spread to pension funds. Pension funds would spread to public markets.

It was a cascade. And they were only at the first level.

The real crisis was still ahead.

May 8, 2026

David’s phone rang at 6:47 AM. Jamie’s name on the screen. This was going to be bad news. Good news didn’t call before 7 AM.

“Hey,” David said.

“It’s over. We’re shutting down today. I need everyone in the office by 9 AM.”

David sat up in bed. “Today? I thought we had until June—”

“Microsoft pulled out of acquisition talks yesterday. Said the GPU situation was too complicated. Without them, we’re out of options. We’ve got five weeks of cash and no path forward. The board voted last night to wind down.”

“What about the other companies that were interested?”

“They were never serious. They were waiting to see if we’d go bankrupt so they could buy the IP for nothing. Which is exactly what’s going to happen.”

After Jamie hung up, David lay in bed for a few more minutes, staring at the ceiling. Four years. Four years of his life. Four years of 60-hour weeks and believing in the mission and telling himself that the equity would make it all worthwhile.

$270,000 in equity. Four years of foregone salary—he’d been making $180,000 when he could have been making $350,000 at Meta. That was $680,000 in opportunity cost. Plus the equity. Plus the stress.

Zero return.

He got dressed and drove to the office for the last time. When he arrived at 8:30, the parking lot was already full. Nobody wanted to be late to their own funeral.

Jamie stood in front of the whiteboard one last time. The engineering team sat in the same places they always sat. But the energy was different. No laptops open. No side conversations. Just people waiting to hear how their dream had died.

“Okay,” Jamie started. “I’m not going to sugarcoat this. As of 5 PM today, Artemis Medical AI is ceasing operations. We’re filing for Chapter 7 bankruptcy protection. That’s liquidation, not reorganization. The secured creditors will take possession of the assets. The company will be dissolved.”

He paused, looking around the room. “Here’s what this means for everyone. Your last day of employment is today. You’ll be paid through May 15th. Your health insurance continues through the end of the month. After that, you’ll need to find COBRA coverage or new employment.”

Someone asked: “What about our equity?”

“Common stock receives zero. The secured creditors are owed $90 million. The assets are worth maybe $20 million at auction. Even the preferred stockholders are getting wiped out. I’m sorry. I know that’s not what anyone wanted to hear.”

The room was silent.

“I want you to know,” Jamie continued, “that what we built here was real. We created an AI system that can detect lung cancer eighteen months earlier than human doctors. We saved lives. That matters. It doesn’t matter to the balance sheet, but it matters.”

David felt a sudden surge of anger. It mattered? It mattered that they’d saved lives but couldn’t figure out how to not lose $200,000 on every hospital they signed? It mattered that they’d built something revolutionary but worthless?

He stood up. “Jamie, can I say something?”

Jamie looked surprised but nodded. “Of course.”

David faced the room. “I want everyone to know that this wasn’t your fault. You all did incredible work. You built something that actually worked. We failed because the current economics of AI are broken. Not because our technology was bad. Not because we didn’t work hard enough. Because the entire sector is built on a false assumption—that you can scale compute-intensive AI fast enough to make up for losing money on every transaction.”

He turned to Jamie. “And you, you did everything right. You raised capital. You hired great people. You built an amazing product. You just couldn’t change the fundamental math that makes AI companies unprofitable.”

Jamie nodded slowly. “Thanks, David. I think—” his voice caught slightly, “—I think we all needed to hear that.”

After the meeting, people started packing up their desks. David watched his coworkers taking down photos, packing up monitors, saying goodbye to each other. Some were crying. Most just looked shell-shocked.

His phone rang. Sarah Chen.

“David, I heard about Artemis. I’m sorry.”

“Thanks. How did you hear?”

“It’s going to be in TechCrunch in about two hours. They called me for comment.”

“What did you say?”

“I said that Artemis was a well-run company that fell victim to structural problems in AI financing. And that there would be more failures like this.”

David walked out to the parking lot. The sun was shining. Birds were chirping. Everything looked normal except that his career had just ended.

“Sarah, how many more companies are going to fail?”

“I’ve been tracking forty-two AI companies in CalPERS’s portfolio. Artemis makes three that have shut down. I think we’ll lose at least ten more in Q2. Maybe fifteen.”

“And that triggers the cascade you warned about?”

“That is the cascade. David, can I ask you something? Would you be willing to talk to regulators about what happened at Artemis? About the cross-collateral situation, the leasing freeze, all of it?”

David thought about TechLease. About how they’d frozen Artemis’s GPUs for a technicality they’d ignored for eighteen months. About how that freeze had killed the company faster than running out of money would have.

“Yes. I’ll talk. What do you need to know?”

“Everything. I want to understand exactly how the equipment financing worked. How the cross-collateral was structured. How the freeze happened. Because Artemis won’t be the last company that TechLease kills. There are going to be more. And if I can document the pattern, maybe we can do something about it.”

After they hung up, David went back inside. The office was almost empty now. Just Jamie and Priya, sitting in the conference room, going through final paperwork.

“Hey,” David said. “I’m heading out.”

Jamie looked up. “Thanks for everything, David. You were one of the best engineers I’ve ever worked with. I’m sorry it ended like this.”

“Me too. What are you going to do now?”

“I don’t know. Take some time off. Maybe write a blog post about what we learned. Maybe try to warn the next cohort of founders that AI startups are a trap.”

“Nobody will listen.”

“Probably not. But maybe someone will.”

David walked to his car for the last time. He sat in the driver’s seat, looking at the building. Four years ago, he’d walked through those doors for the first time, excited to be part of something revolutionary. Today he was walking out with nothing.

His phone buzzed. An email from Meta’s recruiting team: “We’d love to schedule a final round interview for next week.”

He stared at the message. Three years ago, he’d turned down Meta because Artemis had more upside. Now Meta was offering him a second chance.

He hit reply: “I’m available Tuesday.”

Because that’s what you did when your startup died. You took the job at the big company. You gave up on the dream of equity riches. You became another engineer with a salary and benefits and zero chance of ever making life-changing money.

But at least you paid your rent.

April 15, 2026

Sarah sat in the CalPERS board room, presenting her Q1 portfolio review. Eighteen board members sat around the table—elected representatives, gubernatorial appointees, labor leaders. All of them knew this meeting mattered.

She clicked to slide twelve: “Private Market Valuations – AI Exposure.”

“As of March 31st, our AI-related exposure stands at $18.3 billion across 42 portfolio companies. Of those companies, three have filed for bankruptcy protection in Q1: Inflection AI, CloudScale Technologies, and Artemis Medical AI.”

Board member Patricia Wong, representing the California Teachers Association, raised her hand. “What’s our exposure to those three specifically?”

“$380 million marked value. Our fund managers are estimating recovery of 30 to 40 cents on the dollar. Expected losses: $240 million.”

“Expected. Not actual?”

“The bankruptcies are ongoing. Final recovery won’t be known for twelve to eighteen months. But based on intercreditor disputes and collateral values, we’re modeling significant losses.”

Board member James Rodriguez, a gubernatorial appointee, leaned forward. “Sarah, you testified before the Senate six weeks ago. You said we had $18 billion in AI exposure and that the sector was facing systemic stress. Since then, three companies have failed. What about the other thirty-nine companies?”

Sarah pulled up her tracking spreadsheet. She’d been updating it daily.

“Of the remaining 39 companies: sixteen are in active refinancing discussions. Eight have missed debt payments and are operating under forbearance agreements. Seven have conducted layoffs of 30% or more. Three have shut down new customer acquisition. The remaining five appear stable.”

“So forty-three out of forty-eight companies are in distress.”

“Yes.”

The room was quiet.

Martin Zhao, CalPERS’s Chief Investment Officer, spoke up. “The question before the board is whether we mark down our positions now or wait for the Q2 close in June. Our fund managers are arguing strongly that the current distress is temporary. They believe that as the market stabilizes—”

“The market isn’t stabilizing,” Sarah interrupted. She immediately regretted the tone, but it was too late. “I’m sorry, Martin. But we’ve been tracking daily developments. TechLease marked down their entire AI portfolio by 65% last week. Sequoia Harbor Capital suspended redemptions. Coatue pulled out of every AI deal in their pipeline. This isn’t temporary distress. This is a sector collapse.”

Martin’s expression was carefully neutral. “Sarah, we have a fiduciary duty to not panic. Marking down now, in the middle of Q2, sends a signal to the market that we’ve lost confidence. That could trigger—”

“It could trigger honesty,” Sarah said. “Every large institutional investor is having this exact conversation. Everyone’s waiting for someone else to mark down first. But someone has to go first. Why not us?”

Patricia Wong spoke up. “Sarah has a point. We’ve been in this situation before—2008, the subprime crisis. Everyone waited. Everyone hoped values would recover. And when they finally marked down, the losses were twice what they would have been if we’d acted early.”

“But this isn’t 2008,” another board member said. “AI is real. The technology works. These companies might recover—”

“Some might,” Sarah agreed. “But most won’t. And while we’re waiting to see which ones survive, we’re misleading our stakeholders about our funded status. We’re telling teachers and firefighters that we’re 72% funded when the real number is probably 66% or 67%.”

The debate went on for another hour. Finally, Martin suggested a compromise.

“Here’s what I propose. We maintain current valuations through Q2. We ask our fund managers to provide updated assessments by June 15th. If those assessments show significant deterioration, we mark down on June 30th. That gives companies another quarter to stabilize and gives us better information.”

“That’s just delaying the inevitable,” Sarah said.

“Perhaps. But it’s prudent delay. Sarah, you’ve done excellent work identifying these risks. But we can’t mark down based on projections. We need actual data points. June 30th gives us that.”

The board voted. Fourteen to four in favor of waiting until June.

After the meeting, Sarah found Martin in his office.

“You’re making a mistake,” she said.

“Maybe. But Sarah, you have to understand—marking down is an irreversible decision. If we’re wrong, if these companies recover, we’ll have unnecessarily panicked our members and hurt our credibility.”

“And if we’re right? If we wait and the losses get worse?”

“Then we’ll deal with it in June. Three months, Sarah. That’s all I’m asking for. Three months to see if the situation stabilizes.”

Sarah thought about Rebecca Torres’s model. About the contagion timeline. About how three months could turn moderate stress into severe stress.

“Martin, in three months, the situation won’t stabilize. It’ll be worse. A lot worse.”

“Then we’ll mark down in June and everyone will see we were responsible and careful.”

After she left his office, Sarah sat in her car in the parking lot, staring at her phone. She had Rebecca’s number. She had David Huang’s number. She had Marcus Webb’s number. All of them were watching the same cascade unfold.

She called Rebecca.

“They voted to wait,” Sarah said.

“Until when?”

“June 30th.”

Rebecca was quiet for a moment. “Sarah, a lot can happen in three months.”

“I know. What’s Pacific Life doing?”

“We’re presenting to our board next week. I’m recommending we mark down immediately. But I’ll probably get the same answer you did—wait and see.”

“Wait and see while the losses compound.”

“Exactly.” Rebecca sighed. “You know what the crazy thing is? Every risk manager I talk to sees the same thing we do. We all know what’s coming. But we’re all waiting for permission to acknowledge it. It’s insane.”

Sarah thought about her Senate testimony. About standing in front of the cameras and saying out loud what everyone already knew. It had felt terrifying and necessary.

This felt like the opposite. Like being forced to lie by staying silent.

“Rebecca, what happens if we’re both right and nobody marks down for six months?”

“Then the losses double. And when everyone finally admits it at the same time, the market crashes. It’s 2008 all over again, except this time everyone saw it coming.”

After they hung up, Sarah sat in her car for another twenty minutes. Then she opened her laptop and started drafting a memo. Not to the board—they’d already made their decision. This memo was for the record.

MEMORANDUM
TO: File
FROM: Sarah Chen, Risk Analysis
DATE: April 15, 2026
RE: AI Exposure Risk Assessment – Dissenting View

Summary: The Board’s decision to delay mark-downs until Q2 close represents, in my professional judgment, an imprudent deferral of necessary action. Current evidence suggests that postponement will result in significantly larger losses and reduced transparency to stakeholders…

She wrote for an hour. She documented every warning sign. Every company failure. Every indicator that the sector was collapsing faster than anyone had modeled.

Then she saved it, encrypted it, and sent a copy to her personal email.

Because when this was over—when CalPERS finally admitted the losses and everyone asked why they’d waited—Sarah wanted there to be a record that someone had tried to tell the truth.

Even if nobody had wanted to listen.

March 20, 2026

Rebecca Torres had been working on the stress test for three weeks. Her boss had finally agreed to let her run it after Pacific Life took a $300 million mark-down on their private credit portfolio. The mark-down had been quiet, disclosed in a footnote of their quarterly filing, but Rebecca knew what it meant.

It meant they’d finally admitted the losses were real.

Now she sat in a conference room with four other risk managers, staring at a model that showed Pacific Life’s path to insolvency.

“Walk me through the assumptions again,” said Robert Chen, the Chief Risk Officer. He’d been at Pacific Life for twenty years. He’d survived the 2008 crisis. He did not look happy.

Rebecca pulled up the first slide. “Scenario One: Moderate Stress. We assume 30% of AI companies fail over the next eighteen months. Private credit funds mark down by 25%. Our exposure is $4.2 billion. We take losses of $1.05 billion.”

“Impact?”

“Statutory surplus drops from $8.3 billion to $7.25 billion. That’s still above the regulatory minimum, but it triggers the RBC Action Level, which means we have to file a plan with the California Department of Insurance explaining how we’ll restore capital.”

“Scenario Two?”

“Severe Stress. 50% of AI companies fail. Funds mark down by 45%. We take $1.9 billion in losses. Surplus drops to $6.4 billion. We’re now in the Company Action Level territory. We have to restrict new business, stop paying dividends, and potentially raise new capital.”

Robert was making notes. “And Scenario Three?”

Rebecca had been dreading this one. “Systemic Crisis. 70% of AI companies fail. The equipment leasing companies—TechLease, Velocity Capital—become insolvent. The private credit market freezes. Funds are forced to sell assets at distressed prices. We take $3.2 billion in losses. Surplus drops to $5.1 billion.”

“Below the regulatory minimum,” Robert said quietly.

“Below the minimum. At that point, California DOI takes over. They appoint a conservator. We stop writing new policies. Existing policyholders face benefit reductions. It’s effectively insolvency.”

The room was silent.

Finally, someone asked: “What’s the probability of each scenario?”

Rebecca had been working with their actuarial team on this. The numbers had kept her up at night.

“Based on current default rates, comparable historical episodes, and correlation analysis: Moderate Stress is 35% probability. Severe Stress is 25%. Systemic Crisis is 15%.”

“Those don’t sum to 100%,” someone noted.

“Because there’s a 25% chance we’re fine. That everything stabilizes, companies raise new capital, the marks recover, and we take manageable losses.”

Robert stood up and walked to the window. Outside, it was a clear spring day in Newport Beach. Pacific Life’s headquarters overlooked the harbor. Boats bobbed in the water. Everything looked normal.

“Twenty-five percent chance we’re fine,” he said. “Seventy-five percent chance we’re in some level of crisis. Fifteen percent chance we’re insolvent.”

“Yes.”

“And every other major insurance company has similar exposure?”

“I’ve been talking to risk managers at MetLife, Prudential, and New York Life. They’re all running similar numbers. Nobody wants to be first to announce problems, but everyone’s preparing for the same scenarios.”

Robert turned back from the window. “Rebecca, I want you to present this to the board next week. Full presentation. All three scenarios. Probability-weighted expected losses. Action plans for each outcome.”

“That’s going to leak,” someone said. “The board has seventeen members. If we tell them we might be insolvent—”

“Then maybe we should be insolvent honestly rather than solvent dishonestly,” Robert said sharply. “We have a duty to policyholders. That duty includes telling the truth about our financial condition.”

After the meeting, Rebecca sat in her office, updating the model one more time. She’d added a new variable: contagion speed. How fast would the cascade move once it started?

Her model suggested six to nine months from the first major failure to systemic crisis. Inflection had failed two weeks ago. By her math, they were already two weeks into the six-to-nine-month window.

Her phone buzzed. Sarah Chen, the CalPERS analyst who’d testified before the Senate.

“Sarah, hi. Thanks for reaching out.”

“I saw Pacific Life’s quarterly filing. You took a $300 million mark-down.”

“We did.”

“Was that you? Did you push for it?”

“I ran the numbers. Leadership made the call.”

“Thank you,” Sarah said quietly. “Thank you for telling the truth. Do you know how rare that is right now?”

Rebecca thought about all the conversations she’d had with risk managers at other insurance companies. How many of them had admitted privately that their AI exposure was underwater but were waiting to mark down until “everyone else moved first.”

“Sarah, can I ask you something? After your testimony, what happened at CalPERS?”

“They didn’t fire me. That was surprising. Instead, they put me in charge of ‘AI exposure remediation.’ Which is a fancy way of saying I get to be the person who tells everyone how much money we’ve lost.”

“Have you marked down yet?”

“Not officially. We’re still ‘evaluating valuations.’ But between you and me? We’re going to take $4 to $6 billion in losses when we finally admit it.”

Rebecca pulled up her model again. She added CalPERS’s numbers to her contagion scenario. If CalPERS marked down $6 billion, other pensions would have to follow. If other pensions followed, the denominator effect would force them to sell public equities to rebalance. If they sold public equities, the stock market would drop. If the stock market dropped, Pacific Life’s equity portfolio would take losses too.

She ran the calculation. In the Systemic Crisis scenario, Pacific Life’s total losses weren’t $3.2 billion. They were $4.8 billion.

They weren’t just insolvent. They were deeply insolvent.

“Sarah,” Rebecca said carefully, “when are you going to mark down?”

“June. After Q2 closes. We’re trying to coordinate with New York State and Texas. Nobody wants to be first, but somebody has to be.”

“Make it sooner.”

“What?”

“Make it sooner. The longer we wait, the worse it gets. Every month we delay is another month of losses piling up. Just mark down. Tell the truth. Let everyone else follow.”

Sarah was quiet for a long moment. Then: “You know what’ll happen if we do that?”

“Yes. But it’s going to happen anyway. At least this way we’re honest about it.”

After they hung up, Rebecca stared at her model. She’d been a risk manager for twelve years. She’d modeled a thousand different scenarios. She’d seen portfolios blow up, seen companies fail, seen markets crash.

But she’d never seen anything quite like this. Because this time, everyone knew it was coming. Everyone had seen the same warning signs. Everyone had run the same numbers.

And everyone had decided to wait and see if someone else would blink first.

This was how every financial crisis started. Not with surprise, but with collective pretending. With everyone hoping that if they just waited a little longer, maybe the problem would solve itself.

It never did.

March 15, 2026

James Morrison was in a meeting with Microsoft’s corporate development team when Priya texted him: “Emergency. Need you back at office immediately.”

He excused himself and called her from the hallway. “What’s wrong?”

“TechLease just froze our GPUs.”

Jamie felt his stomach drop. “What do you mean, froze?”

“I mean they’ve declared an Event of Default under our lease agreement. They’re claiming we violated Section 8.4 by failing to maintain adequate insurance on the equipment. They’ve locked our access to the data center. Our training runs just stopped.”

“That’s insane. We have insurance. We’ve always had insurance.”

“I know. But the lease agreement requires $5 million in coverage and our policy is for $3 million. It’s been $3 million for eighteen months. They never said anything before.”

“Because they’re looking for an excuse,” Jamie said. He was already walking toward the elevator. “CloudScale failed and TechLease is panicking. They’re trying to repossess assets before we fail.”

“Jamie, if we can’t access the GPUs, we can’t fulfill our customer contracts. We can’t train new models. We’re dead in the water.”

“I’m on my way.”

The drive from Microsoft’s campus in Mountain View back to Artemis took twenty-five minutes. Jamie spent the entire time on the phone with their lawyer, who explained that yes, technically they were in violation of the lease agreement, and no, there wasn’t much they could do about it immediately.

“Can we get an injunction?” Jamie asked.

“We can try. But TechLease has a secured interest in the equipment. If they can demonstrate that you’re in default—and you are—a judge is unlikely to force them to maintain your access.”

“How long would litigation take?”

“Months. Maybe years if it goes to appeal.”

“We don’t have months. We have eight weeks of cash.”

The lawyer was quiet for a moment. “Jamie, you need to start thinking about bankruptcy options. Chapter 11 might let you—”

“I’m not giving up yet.”

But even as he said it, he knew. This was the end. You couldn’t run an AI company without compute. And TechLease had just taken theirs away.

When he got back to the office, the engineering team was standing around looking shell-shocked. David was at his desk, staring at his monitor, which showed error messages from their training pipeline.

“How bad?” Jamie asked.

“Everything mid-training is frozen. We can’t access any of the models. All our customer inference is running on cached versions but those will be stale in forty-eight hours. After that, we start missing SLAs.”

Priya appeared from her office. “I just got off the phone with TechLease’s counsel. They’re willing to restore access if we pay $8 million—the full remaining lease value—in cash, up front, plus penalties.”

“We don’t have $8 million.”

“I know.”

Jamie looked around the office. Forty-seven employees, most of them under thirty, most of them with equity that was about to become worthless. He thought about the medical imaging software they’d built. The lung cancers they’d detected early. The lives they’d probably saved.

None of it mattered anymore.

“Call an all-hands,” he said. “I need to tell them.”

Twenty minutes later, the entire company was gathered in the main room. Jamie stood in front of the whiteboard where they’d mapped out their product roadmap for 2026. It seemed absurd now.

“Okay,” he said. “I’m going to be direct. TechLease has frozen our access to our GPUs. They’re claiming we’re in default on our lease agreement. Without access to compute, we can’t fulfill customer contracts. We can’t train new models. We can’t operate.”

The room was silent.

“I’ve been on the phone with Microsoft, Google, and Amazon about acquisition possibilities. All three have expressed interest in the technology. But the situation with TechLease complicates things. Any acquirer would have to either pay off the lease or negotiate a new arrangement, and TechLease is playing hardball.”

He took a breath. “Here’s what’s going to happen. We have eight weeks of cash. We’re going to use that time to try to find a buyer who will acquire the company and hire the team. If we can’t find a buyer, we’ll wind down operations and file for bankruptcy protection. In a bankruptcy scenario, the secured creditors—TechLease and the venture debt lenders—get paid first. Preferred stockholders get what’s left. Common stockholders—that’s everyone with equity—get nothing.”

Someone in the back said: “So we’re dead.”

“We’re not dead yet. But we need to be realistic. Everyone should start interviewing. If you get an offer, take it. Don’t wait for an acquisition that might not happen.”

After the meeting, people drifted back to their desks in a daze. Some started updating their LinkedIn profiles. Others just sat there, staring at nothing.

David found Jamie in his office.

“That went well,” David said.

Jamie almost laughed. “I’ve had better days.”

“What’s the actual probability of an acquisition?”

“With frozen GPUs? Maybe 10%. Nobody wants to buy a company that can’t operate. They’d rather wait for bankruptcy and pick up the IP for pennies.”

“So we’re done.”

“Yeah. We’re done.”

David sat down. “You know what’s funny? Two years ago, we were the hot company. Everyone wanted to invest. Every engineer wanted to work here. We were going to change healthcare.”

“We did change healthcare. We built something real.”

“Just couldn’t figure out how to make money doing it.”

Jamie looked at the framed poster on his wall—the first X-ray their AI had analyzed, showing a tiny tumor that three radiologists had missed. They’d sent the scan back to the hospital. The patient had gotten surgery. She was alive now.

“You know what nobody tells you about startups?” Jamie said. “It’s not enough to build something that works. It’s not enough to save lives. You have to build something that works and saves lives and makes money. If you can’t do all three, you fail. And we couldn’t do all three.”

David was quiet for a moment. Then: “What are you going to do?”

“After this? I don’t know. Maybe take some time off. Maybe join a bigger company where someone else worries about unit economics.” He smiled tiredly. “Maybe get a job at TechLease. They seem to be the only ones making money in this ecosystem.”

After David left, Jamie sat alone in his office, looking at the poster. Then he took it down and put it in a box. Along with his Stanford diploma, his Y Combinator certificate, and a coffee mug that said “Move Fast and Break Things.”

He’d broken something, all right. He’d broken his company. He’d broken the equity of forty-seven employees who’d believed in him. He’d broken the trust of customers who’d bet on their technology.

All because he’d assumed that if you built something good enough, the money would follow.

But the money never followed. The money just burned.

March 2, 2026

Marcus was in a partner meeting when his phone started buzzing. Not just one call—a cascade of them, one after another. He glanced down and saw twelve missed calls in three minutes. All from portfolio companies.

Tom Hendricks, managing partner, noticed. “Marcus, if you need to—”

“Sorry. Give me two minutes.” Marcus stepped out into the hallway and called back the most recent number.

It was Kevin Zhang, CEO of TensorFlow Capital, one of their portfolio companies. His voice was tight. “Marcus, have you seen the filing?”

“What filing?”

“Inflection AI. They filed for Chapter 11 twenty minutes ago. The bankruptcy petition lists eleven creditors. We’re number three.”

Marcus felt the floor tilt slightly. “How much?”

“$62 million. Senior secured debt. But Marcus—there are four other lenders ahead of us claiming senior secured status. Same collateral. The intercreditor agreement is ambiguous about priority.”

“How ambiguous?”

“The phrase ‘first priority lien’ appears in six different documents referring to six different lenders. Our lawyers are on it but—” Kevin’s voice cracked slightly, “—this is going to be a bloodbath.”

After he hung up, Marcus stood in the hallway for a moment, staring at nothing. Then he went back to the conference room.

“We need to stop the meeting,” he said.

Tom looked at him. “What’s wrong?”

“Inflection AI just filed for bankruptcy. We have exposure through TensorFlow Capital. $62 million senior secured. But the collateral—the GPUs—are pledged to six different lenders who all think they’re first in line.”

The room went silent.

Karen Reyes, the fund’s general counsel, pulled out her laptop. “How did we not know there were six lien holders?”

“Because the intercreditor agreements don’t require disclosure of all parties,” Marcus said. “Each lender only has to be notified of new liens at their level or above. But everyone’s claiming the same level.”

“That’s fraud.”

“That’s private credit in 2025. Everyone knew it was happening. Everyone assumed it wouldn’t matter because the companies would grow into their valuations.”

Tom pulled up the fund’s portfolio on the screen. “Okay. What’s our total exposure to Inflection?”

Marcus had already run the numbers in his head during the thirty-second walk back from the hallway. “Direct: $62 million through TensorFlow. Indirect: another $30 million through two other portfolio companies that have their GPUs financed by the same equipment lessors that Inflection used. If the leasing companies fail, those companies lose their compute.”

“Total: $92 million.”

“On a $2.2 billion fund. It’s manageable.”

“Unless more dominoes fall,” Tom said quietly.

Marcus pulled up his tablet and opened the file he’d been building for three months. The one he’d titled “Contagion Model” and hadn’t shown anyone yet because the numbers were too scary.

“Here’s what I’ve been mapping,” he said. “There are seventeen AI companies that went through hypergrowth in 2024-2025. All of them financed equipment the same way—debt secured by GPUs, multiple lenders, cross-collateralized. The three largest equipment lessors are CloudScale, TechLease, and Velocity Capital. Between them, they financed about $8 billion in GPU purchases.”

He clicked to the next slide. “CloudScale failed in January. That froze $2.4 billion in GPU assets across eleven companies. Those companies are now operating under forbearance agreements while they try to refinance. But nobody’s lending to companies with frozen assets.”

“How many of those eleven companies are going to fail?” Karen asked.

“At least four in the next quarter. Maybe seven.”

“And when they fail?”

“Their assets go into bankruptcy. Which triggers cross-default provisions in their other credit agreements. Which freezes more assets. The GPU leasing companies take losses. They’re all leveraged 6-to-1. If they lose 20% on their loan book, they’re insolvent.”

Tom was staring at the screen. “You’re describing a cascade.”

“I’m describing what’s going to happen. Inflection was $1.2 billion in annual revenue. They had 200 enterprise customers. They were one of the ‘safe’ ones. If they can’t make it, who can?”

The room was quiet except for the hum of the HVAC system.

Finally, Tom said: “What’s the total exposure if this cascade happens?”

Marcus clicked to his final slide. He’d been working on this model for six weeks. He’d stress-tested it twelve different ways. The numbers always came out in the same range.

“If three more companies fail—companies at Inflection’s scale or larger—the total losses across the private credit ecosystem are $8 to $12 billion. That’s assuming 40% recovery on collateral, which might be optimistic given the equipment disputes.”

“That’s across all private credit?”

“That’s just AI-related exposure. But those losses flow through to the funds. Which triggers mark-downs. Which triggers denominator effects at the pension funds and insurance companies that are LPs. Which makes them pull back from new commitments. Which freezes the entire private credit market.”

Karen was making notes. “What’s our fund’s exposure to that scenario?”

“If it’s three companies? We take $200 to $250 million in losses. If it’s seven companies? $400 to $500 million.”

“On a $2.2 billion fund.”

“Yes.”

“That’s catastrophic.”

“Yes.”

Tom got up and walked to the window. It was a beautiful March day in San Francisco. Clear skies, sailboats on the bay, people jogging along the Embarcadero. None of them knew that a $300 billion ecosystem was starting to collapse.

“What do we tell the LPs?” he asked.

Marcus had been thinking about this for weeks. “We tell them that we’re experiencing market volatility in the AI sector. We tell them we’re actively managing our portfolio. We tell them that we expect some near-term mark-downs but remain confident in the long-term value of our investments.”

“That’s a lie.”

“That’s what every fund is going to say.”

“And when the losses materialize?”

“We mark down in Q3. We say it was unforeseen. We say the market moved faster than anyone expected. We stay calm and professional and we absolutely do not panic because if we panic, the LPs panic, and if the LPs panic, they trigger redemption rights, and if they trigger redemption rights, we have to sell assets in a falling market, which makes everything worse.”

Tom turned back from the window. “Marcus, how long have you known this was coming?”

Marcus thought about lying. Then he thought about Sarah Chen, testifying in front of the Senate. About Rebecca Torres, running stress tests at Pacific Life. About all the people who’d seen this coming and been ignored.

“Since December,” he said. “I tried to tell you in January. You said I was being too conservative.”

“You should have tried harder.”

“Would it have mattered?”

Tom sat down heavily. He looked older suddenly. “No. No, it wouldn’t have. Because even if we’d pulled out of every AI deal in January, the losses were already baked in. We just hadn’t admitted them yet.”

Karen closed her laptop. “What do we do now?”

Marcus looked at his partners. Twenty-seven years in finance had taught him that in a crisis, you had three options: get ahead of it, ride it out, or pretend it’s not happening. Most people chose option three.

“We run the numbers,” Marcus said. “We figure out exactly how bad this can get. We build a plan to communicate with LPs before they hear it from someone else. And we hope to God that Inflection is the only one that fails this quarter.”

But he knew it wouldn’t be. He’d run the numbers too many times. The cascade had already started.

They just hadn’t admitted it yet.

February 25, 2026

David sat in his kitchen at 6 AM, reading TechCrunch on his laptop while his coffee got cold. The headline made his stomach drop:

“Private Credit Funds Suspend Redemptions Amid AI Valuation Concerns”

Below it, a smaller story: “CalPERS Analyst’s Senate Testimony Triggers Market Jitters”

He’d known this was coming. He’d known it since the day he’d run the numbers on Artemis’s burn rate and realized they had six months, not twelve. He’d known it when CloudScale missed their debt payment. He’d known it when Jamie had pulled him aside and said, very carefully, “If you’re thinking about leaving, you should start interviewing.”

But knowing something and seeing it happen were different things.

His equity was worth $270,000 on paper. Four years of 60-hour weeks, four years of believing in the mission, four years of telling himself that the next funding round would make it all worthwhile. He’d turned down a job at Meta three years ago because Artemis had “so much upside.”

The upside was vanishing.

His phone rang. Jamie.

“You’re up early,” David said.

“I didn’t sleep. Did you see TechCrunch?”

“Yeah.”

“We need to talk. Can you come in?”

“It’s 6 AM.”

“I’m already here. So is Priya.”

Twenty minutes later, David walked into Artemis’s headquarters. The usual startup buzz was absent. Half the desks were empty—they’d laid off 30% of the team in January. The remaining engineers looked hunched and tired, like plants that hadn’t been watered in weeks.

Jamie and Priya were in the conference room with the door closed. Through the glass, David could see spreadsheets projected on the screen.

He knocked. Jamie waved him in.

“How bad?” David asked.

Priya turned her laptop so he could see. “We have $8 million in the bank. At current burn, that’s nine weeks. Maybe ten if we’re very careful.”

“What about the Series C?”

Jamie rubbed his eyes. “We had a term sheet from Coatue. Had. They pulled out yesterday after the Senate hearing. Said they’re ‘pausing new AI investments pending market clarity.'”

“So we’re dead.”

“We’re not dead yet,” Jamie said, but his voice said otherwise. “We’re talking to strategic acquirers. Google, Microsoft, maybe Amazon. We might be able to sell the technology—”

“For how much?”

Priya and Jamie exchanged a look. “For enough to pay back the senior debt,” Priya said carefully. “Maybe.”

David did the math in his head. The senior debt was the venture debt—the $50 million loan secured by the GPUs. The intercreditor agreement put them first in line. After that came the preferred stockholders. After that came common stockholders like David.

“So equity gets zero.”

“David—” Jamie started.

“Just tell me. Equity gets zero.”

“We don’t know that yet,” Priya said. “If we can sell for more than the debt—”

“We won’t. The GPUs are pledged to three different lenders. There’s going to be a fight over who owns what. By the time that’s sorted out, the technology will be obsolete and the talent will be gone. This is a zero.”

Jamie stood up and walked to the window. Outside, the sun was coming up over Palo Alto. Teslas were pulling into the parking lot of the startup next door. A different company, a different dream, probably the same ending.

“I’m sorry,” Jamie said quietly. “I really thought we had something.”

“We did have something,” David said. “We built a medical imaging AI that can detect lung cancer eighteen months earlier than human doctors. We built something that saves lives. We just couldn’t figure out how to not lose $200,000 on every hospital we signed.”

“The unit economics—”

“—don’t work. I know. Sarah Chen said it in her testimony. Everyone’s known it for months. We all just kept pretending because stopping meant admitting we were wrong.”

Priya closed her laptop. “David, I need to ask you something. Are you going to stay?”

The question hung in the air. David thought about his equity. His four years. His belief that they were building something that mattered.

Then he thought about his rent. His student loans. His girlfriend who wanted to get married but kept saying “after you get acquisition money.” His parents who’d immigrated from Taiwan and worked sixty-hour weeks so he could go to Stanford.

“I’ll stay until the end,” he said. “But I’m going to start interviewing.”

Jamie nodded. “Smart. Everyone should. I’m going to be honest with you—we’ve got maybe 30% odds of making it to June. If you get an offer, take it.”

“What about you?”

“I’m the captain. I go down with the ship.” He smiled, but it didn’t reach his eyes. “Besides, who’s going to hire a founder whose company burned $120 million and failed?”

After the meeting, David went back to his desk. He opened LinkedIn and updated his profile for the first time in three years. Within an hour, three recruiters had messaged him. One was from Meta.

He stared at the message. Four years ago, he’d turned down Meta because Artemis had upside. Because equity in a startup was worth more than salary at a big company. Because he believed.

He hit reply.

That evening, he met Rebecca Torres for coffee. She’d reached out on LinkedIn after Sarah’s testimony, asking if he wanted to talk. He’d said yes, figuring she might know something about the AI sector that he didn’t.

She looked tired. Her Pacific Life badge was still clipped to her belt.

“Thanks for meeting,” she said. “I wanted to ask you about Artemis. Off the record.”

“What do you want to know?”

“Sarah mentioned cross-collateralization in her testimony. Your company has equipment debt, right?”

“$50 million. We pledged our GPUs.”

“To how many lenders?”

David was quiet for a moment. Then: “Three.”

Rebecca closed her eyes. “Jesus.”

“It’s not fraud,” David said defensively. “There are intercreditor agreements. Everyone knows—”

“Everyone knows and everyone assumes they’re first in line. David, when your company fails—I’m sorry, if your company fails—do you know what happens to those GPUs?”

“They get sold and the lenders get paid back.”

“No. They get frozen while three lenders fight over who has priority. The intercreditor agreement probably has language about ‘good faith negotiations’ and ‘commercially reasonable efforts’ to sell the assets. Know how long that takes?”

“How long?”

“In the bankruptcy I’m working on? Eleven months so far. The GPUs are sitting in a data center, depreciating, while lawyers bill $1,500 an hour to argue about what ‘pari passu’ means. By the time they get sold, they’ll be worth half what they were when the company failed.”

David felt something cold in his chest. “Which means all three lenders take losses.”

“Which means Pacific Life marks down our investment. Which means every insurance company with exposure marks down. Which means pension funds have to mark down. Which means—”

“—everyone panics and stops lending.”

“Exactly. And then every AI company that needs to refinance their debt can’t. And then they fail. And then more GPUs get frozen. It’s a doom loop.”

David thought about Artemis. About CloudScale, which had been their GPU supplier until they’d gone bankrupt in January. About the other AI startups in his network, all of them burning cash, all of them assuming they could raise another round.

“How many companies fail before it becomes systemic?” he asked.

Rebecca pulled out her tablet and showed him a spreadsheet. “I’ve been modeling this. If three more companies fail in the next quarter—companies at the CloudScale level or larger—the total losses across private credit are $8 to $12 billion. That’s enough to trigger covenant breaches in multiple funds. Which triggers redemption requests. Which forces asset sales. Which drives down marks further.”

“A cascade.”

“A cascade.” She put her tablet away. “David, I need to ask you something. If Artemis fails, will you talk to regulators? Will you explain what happened? Because right now, everyone’s protecting their position. Nobody wants to be the first to tell the truth.”

David thought about Jamie, who’d hired him and believed in him. He thought about his equity, already worthless. He thought about Sarah Chen, who’d testified in front of the Senate and probably ended her career.

“If we fail,” he said, “I’ll talk. But Rebecca—we’re not going to fail alone. We’re all going down together.”

She nodded slowly. “I know.”

They sat in silence, drinking bad coffee, watching people walk past on University Avenue. Students, tourists, other startup employees who still believed their equity would make them rich.

None of them knew yet. But they would soon.

February 20, 2026

The Senate Banking Committee hearing room was nothing like what Sarah had imagined. She’d expected wood paneling and gravitas. What she got was C-SPAN cameras, harsh fluorescent lighting, and senators who kept checking their phones.

Senator Mills sat in the center of the dais, flanked by twelve other senators, half of whom hadn’t shown up yet. It was 2:15 PM. The hearing had started at 2:00.

Sarah sat at the witness table with a microphone in front of her and a glass of water she was afraid to drink because her hands were shaking. David sat behind her in the gallery. Marcus had sent a text that morning: “Good luck. Also, my lawyers said I can’t be there. Sorry.”

To her right sat Dr. Kenneth Rashid, Chief Investment Officer of the New Jersey State Pension Fund. To her left, Thomas Werner, representing the American Investment Council—the private equity lobby. Both men had already testified. Rashid had assured the committee that public pensions were “well-diversified and prudently managed.” Werner had explained that private markets provided “essential capital allocation for economic growth.”

Neither had mentioned AI. Neither had mentioned cross-collateralization. Neither had mentioned that the entire alternative investment complex was built on the same thirty companies being valued twenty different ways by twenty different fund managers who all had incentives to keep the marks high.

Senator Mills was looking at Sarah now.

“Ms. Chen, you submitted written testimony that differs substantially from the previous witnesses. You claim that CalPERS has 38% of its alternative investments exposed to AI-related companies. Can you explain how you arrived at that figure?”

Sarah took a breath. “Senator, I spent three weeks mapping our underlying portfolio companies across all of our private equity and private credit holdings. Not at the fund level—everyone reports at the fund level. I went company by company. I found that 42 out of 180 portfolio companies were either AI businesses, AI infrastructure providers, or substantially dependent on AI customers.”

“And you believe this is a problem?”

“I believe it’s concentration risk masquerading as diversification.”

Werner, the private equity lobbyist, leaned into his microphone. “If I may, Senator—Ms. Chen’s methodology seems questionable. Just because companies operate in the same sector doesn’t mean they’re correlated. That’s like saying every software company is the same investment.”

Sarah turned to face him. “Mr. Werner, do you know what cross-collateralization is?”

“I’m familiar with the concept.”

“Then you know that when the same physical asset secures multiple loans across multiple lenders, the actual recovery value in bankruptcy is far below the marked value. You know that intercreditor agreements are often ambiguous about seniority. And you know that when one company fails, the collateral dispute freezes the assets for other borrowers who might have otherwise survived.”

Werner’s expression didn’t change. “That’s an extremely pessimistic view.”

“It’s the accurate view. Would you like to see the loan documents?”

Senator Mills intervened. “Ms. Chen, you brought documents?”

Sarah pulled out the folder Marcus had given her. Her hands were steadier now. Anger helped.

“These are redacted copies of loan agreements for three different AI companies. Company A borrowed $50 million from a private credit fund, secured by GPUs. Company B borrowed $30 million from a different lender, secured by the same GPUs. Company C—” she paused, “—is actually the same company as Company A, which took out a second loan six months later under a subsidiary structure, pledging the same equipment.”

The room got very quiet.

“You’re describing fraud,” Senator Mills said.

“I’m describing standard practice. The loan documents all have intercreditor agreements. Everyone technically knows about everyone else. But the intercreditor agreements use phrases like ‘pari passu treatment’ and ‘pro rata distribution’ without specifying what happens if the asset value is less than the combined loan value. Which it always is.”

Senator Hartwick, a Republican from Texas who’d arrived ten minutes late, leaned forward. “Ms. Chen, are you suggesting that private credit is systematically mismarking assets?”

“I’m saying that everyone’s marks assume they’re first in line. When Company A fails, the lenders will spend two years fighting over who has priority while the GPUs sit in a warehouse losing value. The market for used AI chips is thin. Everyone will take losses. But right now, everyone’s balance sheet assumes 100% recovery.”

Rashid, the New Jersey CIO, couldn’t stay quiet. “Senator, Ms. Chen is presenting an extreme scenario. New Jersey has conducted extensive due diligence—”

“Did your due diligence involve calling the other lenders to confirm collateral arrangements?” Sarah asked.

“We rely on fund managers to—”

“So no.”

Senator Mills held up her hand. “Gentlemen, let’s let Ms. Chen finish. You mentioned unit economics in your written testimony. Explain that.”

Sarah pulled up her tablet, glad to have something to look at besides the cameras.

“The largest AI company, OpenAI, burned $8 billion last year. They need to reach $125 billion in annual revenue just to break even at current costs. That’s more than Netflix, Adobe, and Oracle combined. There’s no clear path to get there. Competition from open-source models is driving down prices. Enterprise adoption is slower than projected. Every paying customer increases compute costs faster than revenue.”

“But surely,” Senator Hartwick said, “if the technology is valuable—”

“The technology is valuable. The companies are unsustainable. Senator, artificial intelligence is real. The business models are not. And we’ve built a $300 billion private market ecosystem around companies that cannot generate positive cash flow at any realistic scale.”

Werner tried again. “Ms. Chen’s analysis ignores the possibility of—”

“Mr. Werner, have you read OpenAI’s debt agreements?”

“I’m not familiar with the specifics of—”

“I have them here. Would you like me to read Section 12.4, which requires them to raise $10 billion in new capital by Q4 2026 or face mandatory conversion terms that would effectively dilute existing equity to zero?”

Senator Mills leaned back. “Ms. Chen, you’re painting a very dire picture. What happens if you’re right?”

Sarah thought about her prepared answer. The careful one. The one that wouldn’t get her fired or destroy confidence in public pensions or trigger a panic.

Then she thought about the teachers.

“If I’m right, Senator, CalPERS drops from 72% funded to somewhere between 61% and 52%, depending on how bad the losses are. That triggers benefit restrictions under California law. Retirees get 80% of promised benefits. Current workers face higher contribution rates. The state has to increase its annual payment by $4 to $6 billion. And that’s just CalPERS.”

“How many other pensions have similar exposure?”

“All of them. Every major public pension, every insurance company, every endowment—we all made the same bet at the same time because we all needed the same returns. The target return rate for most public pensions is 7%. We can’t get 7% from bonds anymore. So we went to private markets. And private markets went to AI.”

The hearing room was silent except for the scratch of reporters’ pens.

Senator Mills closed her folder. “Ms. Chen, thank you for your candor. This committee will take your testimony seriously.”

After the hearing, Sarah stood in the hallway, waiting for David. Her phone had forty-three missed calls. Twelve from CalPERS executives. Eight from fund managers. Three from journalists. One from her mother, who’d watched on C-SPAN.

David appeared, holding two coffees. “Well,” he said, “you definitely made an impression.”

“I’m fired, aren’t I?”

“Probably. But you were right to do it.”

Her phone rang again. This time she answered. It was Martin Zhao, CalPERS’s Chief Investment Officer and her boss’s boss.

“Sarah, I need you on a plane tonight. Don’t talk to any reporters. Don’t answer any calls except from me. We need to discuss your future with the organization.”

“Martin, everything I said was—”

“Accurate. I know. That’s why we need to talk.”

After he hung up, Sarah looked at David. “What did he mean, ‘that’s why we need to talk’?”

David took a long sip of coffee. “Either they’re firing you for making them look bad, or they’re going to make you the person in charge of fixing it. There’s no middle option.”

Sarah’s phone buzzed with a text from Marcus: “Holy shit. You just moved the market. Three funds suspended redemptions in the last hour.”

She stared at the message. Three funds. Already.

This was how it started.

Sarah Chen had never been to Washington DC. She’d spent her entire career in California—CalTech for undergrad, Stanford for business school, straight to CalPERS after graduation. DC was foreign territory. All those marble buildings and men in suits who thought they were important.

But here she was, sitting in a conference room in the Russell Senate Office Building, waiting for Senator Patricia Mills.

She’d brought David Huang with her. After their phone call, she’d realized that she needed someone who could explain the technical details of how the AI infrastructure worked. David had been reluctant—he was in the middle of job interviews, and a trip to DC meant missing meetings—but Sarah had been persuasive. She’d also offered to pay him as a consultant, which helped.

Marcus Webb was supposed to be here too, but he’d backed out at the last minute. “I can’t testify against my own fund,” he’d said. Sarah had understood. Marcus had a mortgage and kids in private school. You didn’t bite the hand that fed you.

But he’d given her something better: documents.

Loan agreements. Intercreditor agreements. Email chains between fund partners discussing how to structure the deals. Presentations to LPs showing returns that Marcus had admitted were “aspirational.” It was enough to hang someone, if anyone cared to look.

Senator Mills walked in at exactly 8:00 AM. She was tall, grey-haired, wearing a dark blue suit that probably cost more than Sarah’s monthly rent. But her handshake was firm and her eyes were sharp.

“Ms. Chen. Thank you for coming.” She nodded at David. “And you are?”

“David Huang. I’m an engineer. I worked at several AI companies. Ms. Chen asked me to explain the technical side.”

“Perfect.” The senator sat down, opened a legal pad, and looked at Sarah. “You told my chief of staff this was about systemic risk. Explain.”

Sarah had prepared a presentation. Forty-seven slides explaining the interconnections between private credit, private equity, pension funds, insurance companies, and AI startups. But something about Senator Mills’ directness made her put the laptop away.

“Senator, CalPERS has $48 billion in alternative investments. When I mapped the underlying exposure, I found that $18 billion—38% of our alternatives—is tied to AI-related companies. Either directly or through their suppliers or customers or creditors.”

The senator made a note. “38% seems high.”

“It is high. But it gets worse. Those investments are cross-held across multiple funds. Same companies, different structures. And many of those companies have cross-collateralized equipment loans. The same GPUs are pledged to three or four different lenders.”

“Explain cross-collateralized.”

Sarah looked at David, who leaned forward.

“Senator, imagine you buy a car. You finance it with a bank loan, using the car as collateral. That’s normal. Now imagine you also take out a second loan from a different lender, using the same car as collateral. And a third loan from another lender, also secured by the car. All three lenders think they have first claim on the car. That’s cross-collateralization.”

“That’s fraud,” Senator Mills said flatly.

“It’s not technically fraud if everyone knows about it and there are intercreditor agreements defining who gets paid first. But…” David trailed off.

“But?” the senator prompted.

“But in practice, the intercreditor agreements are ambiguous. They use phrases like ‘pari passu’ and ‘pro rata’ and ‘waterfall provisions’ that sound precise but aren’t. And when a company fails, all the lenders hire lawyers and fight for months or years while the collateral sits idle and loses value.”

Senator Mills wrote this down. “How common is this?”

“In AI? Universal. Every company does it. The equipment is expensive—$30,000 to $40,000 per GPU—and companies need hundreds or thousands of them. They can’t pay cash. So they finance it. Multiple ways. Multiple lenders.”

“And what happens when they fail?”

Sarah pulled out her laptop after all. She opened a spreadsheet and rotated it toward the senator.

“I modeled different scenarios. This shows what happens to CalPERS if the AI sector experiences stress.”

The senator put on reading glasses and studied the screen.

Base Case (10% of AI investments fail):

– Loss: $1.8 billion
– Funded ratio impact: 72% → 70%
– Status: Uncomfortable but manageable

Moderate Stress (30% fail):

– Loss: $5.4 billion
– Funded ratio: 72% → 66%
– Status: Below 70% triggers benefit restrictions

Severe Stress (50% fail):

– Loss: $9 billion
– Funded ratio: 72% → 61%
– Status: Critical status, possible benefit cuts

Systemic Crisis (70% fail + public market declines):

– Loss: $15 billion+
– Funded ratio: 72% → 52%
– Status: Potential insolvency

Senator Mills looked up. “These numbers assume CalPERS is unique. Are they?”

“No,” Sarah said quietly. “Every large public pension has similar exposure. Every insurance company. Every endowment. We all made the same bet at the same time because we all needed the same returns.”

The senator was quiet for a moment. Then she asked the question Sarah had been dreading.

“What’s your probability estimate for each scenario?”

Sarah had been working on this for three weeks. She’d run Monte Carlo simulations. She’d built models based on historical tech busts—2000, 2008, the crypto crash. She’d talked to economists and statisticians and risk managers.

The numbers terrified her.

“Base case: 20% probability,” she said. “Moderate stress: 35%. Severe stress: 30%. Systemic crisis…” She paused. “15%.”

“Those probabilities sum to 100%,” Senator Mills said. “You’re saying there’s zero chance everything works out fine.”

“Yes.”

“Why?”

Sarah looked at David, who shrugged. “Go ahead,” he said.

“Because the unit economics don’t work,” Sarah said. “OpenAI is burning $8 billion a year. Anthropic is burning $3 billion. Every AI startup is losing money on every customer. The only way this ends well is if they scale fast enough to make it up in volume. But they can’t scale because the chips are too expensive and the compute costs increase with scale. It’s a trap. And everyone’s in it.”

Senator Mills closed her legal pad. “Ms. Chen, I’m holding a hearing this afternoon on pension sustainability. I want you to testify.”

“I… I’m not authorized to speak for CalPERS.”

“Then speak for yourself. As an expert witness. Tell them what you just told me.”

Sarah felt her heart racing. This would end her career at CalPERS. You didn’t go to Washington and tell senators that your own fund was overleveraged. But she thought about the firefighters and teachers and state workers whose retirements were in her hands. The people who’d worked thirty years and deserved to retire with dignity.

“I’ll do it,” she said.