Yesterday evening, a teammate opened a huge spreadsheet that had been created for a client.

It looked substantial. It looked organized. It looked finished.

The client had already found something important that it failed to address.

My teammate’s private message to me was blunt:

“AI slop. Zero human review.”

Harsh. But true.

The criticism was aimed at another teammate. I was the one who turned it back toward me.

“I might be the worst offender,” I admitted.

They did not exactly disagree.

They said they do review some of my work before presenting it to clients, but that this was “a known Mark factor.”

Fair. 😅

The joke contained an uncomfortable truth. I had recently responded directly to a client using AI output I had not properly reviewed. Like the spreadsheet, it looked finished. It sounded confident. I assumed it was good enough.

It was not.

The client was pissed off. Rightfully.

That is the deceptive thing about AI-generated work. The failure is not always obvious. The output can be polished, thorough, and professional while still missing the central point.

We are accustomed to unfinished thinking looking unfinished. AI breaks that relationship.

The obvious lesson is that AI output needs human review. But that explanation feels incomplete. We all know these systems make mistakes. We have repeated “human in the loop” so often that it has become almost meaningless.

The more difficult question is why intelligent people who understand these systems keep sending their output into the world without fully understanding it.

I think part of the answer is that AI has changed the economics of producing plausible work without changing the economics of judgment.

Production became cheap

Before generative AI, producing a detailed analysis, client response, or nine-tab spreadsheet imposed natural limits.

The work required enough time that the person creating it usually developed some understanding of it along the way. Writing was not separate from thinking. The friction of production forced a certain amount of contact with the material.

AI weakens that connection.

We can now produce the visible evidence of thought much faster than we can perform the thought itself. A coherent document can appear in minutes, complete with headings, recommendations, caveats, and a confident conclusion.

It looks like the end of a process, even when it is closer to the beginning.

That surface quality matters because we often use polish as a proxy for completeness. AI output does not arrive looking like a rough draft. It can look better than many human final drafts.

But polish is not judgment.

A document can be clear, professional, and completely misunderstand the assignment. A spreadsheet can contain nine beautifully organized tabs and still fail to answer the client’s actual question. A response can sound empathetic while missing the one fact that matters.

The problem is not merely that AI can be wrong. People are wrong all the time.

The problem is that AI can be wrong in a form that looks unusually complete.

The pressure is coming from both directions

None of this is happening in isolation.

Our clients know these tools exist. Many are using them themselves. They reasonably expect us to move faster, operate more efficiently, and accomplish things that would have required more time and larger budgets a few years ago.

The pressure is also coming from inside the company.

I am actively pushing our team to use more AI. I want us automating repetitive work, researching more broadly, testing more thoroughly, and finding ways to deliver better results without simply adding more hours.

I believe that is the right direction. Refusing to use these tools would not preserve some higher standard of craftsmanship. It would make us slower, more expensive, and eventually less useful to our clients.

But I also have to acknowledge the tension I am creating.

I am asking the team to increase its output while expecting the same people to preserve the judgment, context, and attention that made the work valuable in the first place. Clients are raising the bar from the outside, and I am pushing for greater efficiency from the inside.

The team is caught between those expectations.

When something goes wrong, it is easy to point at the individual and say, “Zero human review.” Sometimes that criticism is deserved. People remain responsible for the work they send.

But if leadership rewards speed, expands the amount of work people are expected to handle, and immediately fills every hour AI saves, inadequate review is not only an individual failure. It may also be the predictable result of the system we created.

AI can reduce the cost of production. It does not eliminate the cost of verification.

If I want the team to use more AI, I also need to make room for the work that follows: checking assumptions, tracing conclusions, questioning whether the output answers the actual question, and deciding whether we are willing to put our name on it.

We cannot demand acceleration and treat review as free.

Clients increasingly expect both speed and quality. Internally, I expect us to use the best tools available. But if every efficiency gain becomes an excuse to take on more work, then the time saved by AI never becomes time available for judgment.

The bottleneck has not disappeared.

It has moved from producing the work to understanding it.

Every output creates review debt

I have started thinking about this as review debt.

Every AI-generated artifact creates an obligation to verify it. The faster we produce artifacts, the faster those obligations accumulate.

Like technical debt, review debt can remain invisible for a while. The documents exist. The tasks are marked complete. The client receives something on time. From the outside, the system appears more productive.

But the uncertainty has not disappeared.

Someone still needs to confirm that the analysis used the right assumptions, that the spreadsheet answers the actual question, and that the client response reflects what really happened.

If that work is not performed before delivery, the debt remains embedded in the output. Eventually, someone pays it through rework, confusion, damaged trust, or an uncomfortable client call.

Sometimes the work has not even been eliminated. It has simply been displaced.

When my team quietly reviews and repairs my AI-assisted work, I may experience a productivity gain that does not exist at the company level. My time was saved by consuming someone else’s attention, often later in the process and under greater pressure.

The work moved from the visible author to an invisible reviewer.

Calling that efficiency would be misleading.

Hierarchy makes the problem worse

There is another uncomfortable dimension to my experience: I am the CEO.

People are less likely to tell me directly that my work looks like AI slop. They may revise it, compensate for it, or develop an affectionate term like “the Mark factor.”

Hierarchy makes direct feedback harder. The people with the most authority may receive the least correction.

That is especially dangerous with AI because confidence scales more easily than competence.

AI allows me to involve myself in more subjects, generate more opinions, and produce more material. Each output carries the implied authority of the person sending it, even when that person has spent very little time developing the underlying judgment.

The result can look like expanded leadership capacity while functioning as expanded organizational noise.

That possibility bothers me more than the embarrassing client response.

A single bad message can be corrected. A company quietly learning to compensate for its leader is a structural problem.

I tried to solve judgment with more engineering

I also blamed the models.

As I moved between newer models, I noticed the output becoming more verbose and changing in ways I had not fully anticipated.

I eventually built a separate skill using GPT-5.6 to polish the output from another model.

There is something revealing about that response.

I had one system generating the work and another system improving its presentation. The final result read better, but the additional polish made it no more likely that I understood or agreed with the substance.

I was improving the signal that told my brain the work was finished.

The actual problem was not verbosity. It was that I had substituted a chain of increasingly polished outputs for my own judgment.

Better prompting can improve an answer. A second model can identify errors in the first. Automation can create valuable checks.

But none of those things resolves the question of who understands the work well enough to take responsibility for it.

“Human in the loop” is not a process

It is easy to conclude that a human should remain in the loop.

The phrase sounds reassuring but leaves the important questions unanswered.

Which human? At what point? With how much time?

Are they reviewing the presentation, checking the facts, or reconstructing the reasoning?

Do they have the authority to send the work back when the deadline is approaching?

A quick glance from an overwhelmed person does not become quality control merely because a human technically participated.

Real review has a cost. If we want it, we have to account for it when setting timelines, assigning ownership, and deciding how much work AI allows us to take on.

The standard cannot simply be that a person looked at the output.

Someone must understand it well enough to explain the reasoning, defend the conclusion, and accept responsibility when it is wrong.

The goal is not maximum output

AI is not making human judgment obsolete. It is making judgment the scarce resource.

That changes how I think about productivity.

The goal should not be to maximize the volume of work we can generate. It should be to maximize the volume of work we can responsibly stand behind.

Those are not the same thing.

The companies that use AI well may not be the ones producing the most. They may be the ones that learn where speed is valuable, where friction is protective, and where human attention cannot be removed without changing the nature of the work.

We went through a similar process with remote work. The technology made distributed work possible before companies understood the norms required to make it sustainable. Over time, we learned that trust, accountability, asynchronous communication, and real-time collaboration each had their place.

We are still developing those norms for AI.

For now, I know I need to change my own behavior.

AI can help me begin the work. It can challenge it, organize it, and test it. But if my name is on the output, understanding it is still my responsibility.

AI did not eliminate the final round.

It made the final round more important.

I flew Toronto to San Francisco this week. Five hours, gate to gate, on a United mainline aircraft with a Starlink antenna on the roof. I worked the entire way, on everything, with no allowance made for the fact that I was in the air, and I landed having had an ordinary working day that happened to take place at 38,000 feet.

It was the best flight I have ever taken.

I want to be honest about how low a bar that clears. The seat was a normal seat. Lunch was pretzels and water. The only thing different was the connection, and I want to be specific rather than saying it worked, because every airline on earth has claimed their Wi-Fi works for fifteen years. I had better than 50 Mbps down and better than 50 Mbps up, for five straight hours. Symmetrical, which is the part nobody mentions, because uploading is what you do when you are contributing rather than consuming. It was not a good version of airplane internet. It was indistinguishable from sitting at my desk in my own office.

I have flown that route, and the forty other versions of it I take in a year, on standard airline Wi-Fi: you buy the pass, get a connection that technically exists, and spend five hours discovering the list of things it cannot do. This time there was no list.

Which brings me to a bookmark that other people find odd and I find the most rational tab I keep. There is a website whose entire purpose is to count how many United aircraft have had a Starlink antenna bolted to the roof. It is called the United Starlink Tracker, and as of this week it reports 481 of 1,807 aircraft equipped, roughly 27% of the fleet, with more than 40 installs a month.

Here is the part that should embarrass the industry. That is the best information available, and it is not certainty. The aircraft is not firmly assigned until a few days before departure, so at the moment you buy the ticket the tracker gives you a probability and nothing more. United’s mainline fleet was about 12% equipped when I booked, so what it honestly told me was that I would probably lose. I drew a good tail, which is the only reason I have a story here and not a complaint. Five hours of my working life, on a route I fly constantly, and the best available instrument was a third-party website quoting me odds.

Let me state my willingness to pay plainly, because it matters later. I would pay $50 an hour for guaranteed high-speed connectivity, priced at checkout. On that Toronto flight that is $250 on top of the fare, for the same seat under a different antenna, and I would not think about it. Nobody is offering to sell me that, which is why nobody is collecting it. A free third-party database, maintained tail by tail because enough travelers want to know before they book, is a market pricing an amenity the sellers have not gotten around to pricing themselves.

We did, of course, once build a tracker for legroom. SeatGuru ran for almost 25 years as the place you went to find out whether 14C was a bulkhead or a lav-adjacent mistake, and it shut down on November 1, 2025, the same month United’s first Starlink-equipped mainline aircraft entered service. But notice the difference in kind. A seat map is reference data: you look it up once and the answer holds for years. A Starlink tracker is a progress bar, and people refresh it, because the underlying fact moves fast enough that a stale answer is actively misleading. The market builds infrastructure around whatever is both variable and decisive. Connectivity is no longer a feature of the cabin. It is the reason to pick the flight.

That would be a mildly interesting consumer observation if it were not also one of the more instructive capital allocation stories in aviation right now. The technology was available to everybody at roughly the same time, cost a rounding error against a fleet, and installs in a single overnight. What separated the winners from the losers was not money. It was the calendar.

The Airline That Went First Had 30-Seat Jets

The first airline in the world to sign a Starlink deal was not United, or Emirates, or Qatar. It was JSX, the public charter operator that flies out of private terminals and asks you to show up 20 minutes before departure.

JSX signed with SpaceX in April 2022, when Starlink’s aviation business essentially did not exist. Installation took two to three days per aircraft on 30-seat Embraer E135s, and the service was free from day one, no portal and no login. CEO Alex Wilcox’s framing of the whole business is the line worth stealing. The ultimate luxury is time.

So for roughly three years, a carrier with fewer than 60 small regional jets and coach-adjacent fares delivered a better working environment at altitude than nearly every flagship business class cabin on earth. A $12,000 lie-flat suite on a transatlantic 777 had worse internet than a $299 hop out of Burbank. The technology was equally available to both. That is not a technology story. That is a decision-speed story.

The Constraint Was Never Capital

United’s Starlink installations take about eight hours per aircraft. Eight hours. That fits inside a normal overnight maintenance window, which is why United can run 40 to 50 tails a month without meaningfully touching utilization.

United announced its deal on September 13, 2024. Testing started in early 2025, the FAA certified the first mainline aircraft that September, and passenger flights followed in October. United now expects around 1,000 aircraft equipped by the end of 2026 and the entire fleet, wide-bodies included, done before the end of 2027.

The fleet split is the tell on how a big carrier sequences this. United Express is better than half done, 342 of 669 aircraft. Mainline sits at 139 of 1,138. The regional fleet went first because regional jets are simpler, cycle through maintenance faster, and are where the product gap was most embarrassing.

None of this required an invention. It required somebody with authority to say yes eighteen months earlier than their competitors did.

Delta Optimized the Deal Instead of the Calendar

Delta, which flies about 900 mainline aircraft on Viasat, did not take the Starlink deal. It went with Amazon Leo, bundled into a broader arrangement including seat-back content and AWS. Installations begin in 2028, starting with 500 aircraft, on a constellation that is not yet delivering commercial in-flight service.

I want to be fair, because it is the kind of decision that looks perfectly defensible in the room. Amazon almost certainly offered better economics, bundling connectivity with cloud spend creates real synergies on a spreadsheet, and avoiding single-vendor dependence on SpaceX, a supplier with a habit of becoming a platform and then repricing, is legitimate strategy.

But run the timeline. United finishes in 2027. Delta starts in 2028. That is a multi-year window in which the two largest premium-focused US carriers offer categorically different products on the single amenity that customers demonstrably rank highest after price.

And they do rank it highest. Viasat’s own survey of 11,053 travelers across ten countries found free quality Wi-Fi to be the most influential booking factor other than price, at 22%, ahead of food and legroom. Twenty-nine percent said they would pay more for it. Those numbers come from Starlink’s competitor, which is what makes them useful.

Delta negotiated the better contract. United bought the better years. In a business where customers re-choose you every single booking, years are the scarcer asset.

The Gap Is Latency, and Latency Decides What Job You Can Do

It is worth being precise about what separates a good connected flight from a bad one. The industry keeps describing the gap in megabits, and megabits are the least interesting part of it.

Starlink runs at about 550 km. Geostationary satellites, which is what most legacy in-flight Wi-Fi uses, sit at 35,786 km. That is a round trip of roughly 143,000 km before your keystroke reaches a server, and no amount of engineering fixes it, because the constraint is the speed of light. LEO delivers 20 to 40 ms. GEO delivers 600 ms or more. Bandwidth splits the same way: 100 to 350 Mbps per aircraft against 5 to 25 Mbps, shared across everyone on board. And that shared number is the download. Upload on a legacy system is a fraction of it, which is why on old airline Wi-Fi you can watch a video and cannot be in one.

That is a difference in kind, not degree. At 600 ms a video call does not degrade, it fails. So does anything else that assumes a conversation: SSH, remote desktop, a VPN into your own network, a shared cursor in a doc, a database query. What survives is the asynchronous half of knowledge work. Email. Reading. Writing into a local file. All of which, and this is the part nobody says out loud, you could have done with the Wi-Fi turned off.

So the old product did not sell you productivity. It sold you the ability to do the portion of your job that never required connectivity in the first place, and charged you $19 for it. The new product sells you the other half: you take the call, you unblock the person waiting, you push the fix, you sit in the room.

That is exactly what Toronto to San Francisco was. Not a flight on which I got some work done, a sentence that has always meant I cleared an inbox. A flight on which nothing was deferred. This is not five hours of extra output added to a normal week. It is the removal of a five-hour hole from the middle of a working day, plus the pile that hole generates, plus the evening spent digging out of the pile. One flight erased maybe eight hours of downstream cost.

Now put a number on it. A partner billing at $500 to $1,500 an hour is looking at $2,500 to $7,500 of capacity per transcon that either exists or does not. If you are an operator, the right input is not your rate but the throughput of everyone waiting on you. Eight people blocked behind a decision only you can make do not lose your five hours, they lose forty. And the fare difference is usually zero. Same airline, same route, same fare bucket, different tail number.

I Would Rather Fly Commercial With Starlink Than Private Without

Private aviation is the most expensive way yet devised to move a human body, and it currently offers, on average, the worst connectivity per dollar in the sky.

Look at the installed base rather than the brochures. Gogo ended 2025 with 6,402 business aircraft online using air-to-ground service that tops out around 9.8 Mbps, priced at roughly $99 to $175 per hour of use. Its LEO product, Galileo, had about 120 aircraft flying as of March 2026, with a stated path to roughly 700 by year end. Even at the optimistic number, that is a tenth of the fleet.

Read that pricing again. Private aviation already sells connectivity by the hour, and already charges more per hour than the $50 I said I would pay, for a product roughly forty times slower. The hourly model is not exotic and the market plainly bears it. The only operators not using it are the ones whose product is finally worth paying for.

So the realistic comparison in 2026. Charter a midsize jet at $6,000 to $9,000 an hour and there is a solid chance you get 9.8 Mbps at 600 ms shared across the cabin, with the surcharge on top. Or board a United E175 in a 31-inch-pitch economy seat and get 250 Mbps at roughly 30 ms, free, gate to gate.

One of those is an office. The other is a very fast, very expensive waiting room with excellent snacks.

I am not making that trade hypothetically. Offer me a charter out of Toronto that morning in exchange for the connectivity I actually had and I decline on the way to the gate. The jet saves me two hours of airport. The antenna saved me the other five.

And this resolves the JSX story rather than contradicting it. Private aviation sells you time on the ground: no TSA line, no two-hour buffer, 20 minutes from curb to seat. Starlink sells you time in the air. Two halves of the same product, and almost nobody bought both. JSX did, in 2022, which is why a $200 seat on a 30-year-old Embraer out of a private terminal is, for a certain kind of working traveler, quietly the best value proposition in American aviation. Not because any part of it is luxurious. Because no part of it wastes the day.

Everyone Is Making the Same Pricing Mistake

The second allocation error in this story is one basically the whole industry is committing together: nobody is charging for it.

Remember the $50 an hour. On a five-hour flight that is $250 no airline is collecting, from a customer who volunteered the number, on inventory they already own. And it is worse than not charging, because they are also not selling. An airline mid-rollout has two kinds of seat worth materially different amounts to the same customer, prices them identically, and declines to say which one you are buying. Any carrier could fix this tomorrow by fencing off a subset of equipped tails, committing them to specific rotations, and selling the commitment. Nobody has. The most valuable attribute of the product is distributed by lottery, and a hobbyist website is doing the airline’s disclosure for it.

Every airline currently flying Starlink offers it free. Some gate it behind loyalty enrollment (United, Alaska and Air France all require a program signup) and the rest, JSX and Hawaiian and Qatar and Emirates among them, hand it to everyone. The gated version is defensible: you are converting a capital expenditure into program enrollments with measurable lifetime value, and the friction of a signup is trivial next to the friction of a paywall. Fine.

What nobody seems to be modeling is that this is a one-way door. Airlines spent twenty years training passengers that in-flight Wi-Fi is a paid product, then about eighteen months retraining them that it is free. You do not get to reverse that a third time. Checked bags went from free to paid and the industry absorbed the outrage exactly once. Going from free back to paid, on the amenity customers say matters most, on aircraft where a competitor is giving it away, is not a fight anyone will pick.

So the entire industry is permanently pricing its most valuable differentiator at zero, right at the moment it is most differentiated. The window to sell guaranteed connectivity at a premium is open right now and closing. My read is that it shuts around the time the third major US carrier finishes its rollout, after which connectivity stops being a product and becomes an expense line, permanently.

The Advantage Window Is About Three Years Wide

Count the rollouts. Hawaiian, JSX and Zipair are done. Qatar, Emirates, Air France, WestJet and British Airways are mid-installation, finishing between now and 2028. American and Singapore start in early 2027. Lufthansa Group is working through roughly 850 aircraft by 2029.

By 2029, satellite connectivity on a major carrier will be as noteworthy as a seat-back pocket. Which means the entire strategic value of this technology is concentrated in a window that opened in 2025 and closes around 2028, and your position in that window was determined by a procurement decision you made in 2022, 2023 or 2024.

Somebody Is Going to Sell the Working Seat

What I am buying on a six-hour flight is not the seat, it is the six hours. A lie-flat suite with 600 ms of latency is a very expensive nap. A middle seat where my entire toolchain works is a branch office at 39,000 feet.

Wilcox was right that the ultimate luxury is time. The carriers that installed early bought the air half of it by accident, and mostly have not noticed what they are sitting on.

So somebody is going to sell the working seat. The first thing it sells is not an antenna, it is a promise: this flight, this booking, guaranteed equipped, and if we swap the aircraft you are compensated. Then reserved bandwidth, power, and no middle seat. Price it by the hour, the way private aviation already does, and aim it at the travel manager who cannot justify a $4,000 business class fare but can very easily justify six recovered working hours at a fully loaded cost of several hundred dollars each. Whoever builds it first converts an antenna into yield instead of into goodwill.

Until then, we are all just refreshing a tracker to find out whether tomorrow’s plane is the good one.

For the better part of two decades, the advice was practically unanimous. Guidance counselors said it. Parents said it. Reddit said it. The message was simple: study computer science, get a six-figure job, live happily ever after. It was the closest thing America had to a guaranteed career path.

That guarantee has expired.

The Numbers Nobody Wants to Talk About

Here’s the stat that should make every parent rethink their assumptions: the unemployment rate for recent computer science graduates has hit 6.1%. Computer engineering graduates? 7.5%. To put that in perspective, fine arts graduates -- the very people CS majors were told they’d never want to become -- are now more employed than computer engineers.

These aren’t anomalies. They’re data points from the Federal Reserve Bank of New York, and they paint a picture of an industry undergoing something more fundamental than a cyclical downturn.

The raw numbers are staggering. Over 127,000 workers at U.S.-based tech companies were laid off in 2025. So far in 2026, another 55,000+ have been cut. According to Indeed’s 2025 Tech Talent Report, tech job postings have dropped 36% compared to pre-2020 levels. And the roles that are disappearing fastest? Entry-level ones -- precisely the jobs that new graduates need.

Meanwhile, the supply side has gone in the opposite direction. U.S. universities handed out roughly 110,000 CS bachelor’s degrees in 2022-2023, about double the number from a decade earlier. We spent years telling every ambitious 18-year-old to study CS, and they listened. Now they’re all showing up to a party that’s winding down.

The AI Elephant in the Room

The factor that makes this different from previous tech downturns is artificial intelligence, and not in the way most people think. AI isn’t just creating new jobs in tech while eliminating old ones. It’s fundamentally compressing the value of the skill that CS graduates were trained to sell: writing code.

A Stanford Digital Economy Study found that by July 2025, employment for software developers aged 22 to 25 declined nearly 20% from its peak in late 2022. But here’s the twist: employment for developers aged 35 to 49 actually increased by 9%. The industry isn’t dying. It’s hollowing out from the bottom.

The reason is straightforward. AI coding tools have gotten good enough that a senior engineer with Claude Code or GitHub Copilot can do work that previously required a team of three or four juniors. One engineer at a large San Francisco tech company told the SF Standard that all of his code is now written by AI: “I’m basically a proxy to Claude Code.” Companies don’t need fewer software engineers total -- they need fewer beginning software engineers. And those are exactly the people we’ve been minting by the tens of thousands.

Even graduates from elite programs aren’t immune. Data from SignalFire shows that the share of graduates from MIT, Stanford, Carnegie Mellon, and UC Berkeley employed as engineers at major tech companies dropped from 25% in 2022 to just 11-12% recently -- a decline of more than 50%. If a Stanford CS degree can’t reliably get you into Big Tech anymore, what chance does a degree from a mid-tier state school have?

The individual stories are brutal. Manasi Mishra, a CS major from Purdue, received exactly one interview offer after a year of searching -- from Chipotle. Zach Taylor, an Oregon State graduate, submitted over 5,700 applications with no success. These aren’t lazy people or weak students. They did everything they were told to do. The system just changed the rules while they were playing the game.

The Enrollment Cliff Is Already Here

The market is starting to send signals back to prospective students, and they’re listening. For the first time since the early 2010s, total enrollment in traditional CS undergraduate programs has dropped by 6%. The University of California system is reporting noticeable declines in CS majors even as overall university enrollment rose by 2%.

This is a meaningful shift. For years, CS departments couldn’t build lecture halls fast enough. Now the tide is turning, not because students suddenly lost interest in technology, but because the career calculus no longer adds up the way it used to. When 55% of hiring managers expect more layoffs and 44% say AI will be the top driver, students and their families are paying attention.

So What Do You Actually Tell a Teenager?

This is where the conversation gets uncomfortable, because the honest answer isn’t as clean as “just major in CS” used to be. But here’s my best attempt at advice that accounts for the world as it actually is, not as it was five years ago.

First, stop thinking in terms of “safe” majors. The entire concept of a safe major is a relic of a more stable economy. The same forces that disrupted CS -- AI automation, globalization, rapid industry shifts -- are coming for every field eventually. The goal isn’t to find the one discipline that AI will never touch. The goal is to become the kind of person who can adapt when your field inevitably changes.

Second, domain expertise is the new moat. The graduates who will thrive in an AI-saturated economy aren’t the ones who can write the best code. They’re the ones who deeply understand a specific problem domain -- healthcare, energy, logistics, finance, education -- and can use AI tools to solve problems within it. A nursing student who understands AI-assisted diagnostics is more valuable than a CS graduate who can sort algorithms on a whiteboard. An environmental scientist who can build data pipelines for climate modeling has a career that no chatbot is going to replace.

Third, the trades and physical-world careers deserve a serious look. Electricians, plumbers, HVAC technicians, construction managers -- these are careers with strong demand, rising wages, no student debt, and near-zero risk of AI displacement. The largest areas of total job creation right now are in care, construction, logistics, and education. These roles combine people skills with hands-on work that simply cannot be automated away. The cultural bias against trade careers has always been irrational, and the current moment makes it look even more absurd.

Fourth, if a student genuinely loves CS, they should still study it -- but differently. The degree itself still has long-term value. NACE data showed CS topped the starting salary list at $88,907 for the Class of 2024. Software engineering roles overall are projected to grow by 17% through 2033. The Bureau of Labor Statistics projects 317,700 annual openings in computer and IT occupations through 2034. The field isn’t dead; it’s restructuring. But students need to graduate with more than a transcript. They need a portfolio of real projects, experience with AI tools, and ideally a secondary area of expertise. The formula has changed from “degree equals job” to “degree plus portfolio plus real experience equals job.” Students who treat their CS education as a foundation rather than a finish line will be fine. Students who expect the degree alone to open doors are in for a painful surprise.

Fifth, consider the interdisciplinary path. Pair a major in something you find genuinely interesting -- biology, political science, economics, design -- with a minor or certificate in data science, cybersecurity, or applied AI. This combination produces exactly the kind of hybrid thinker that employers are scrambling to find. The fastest-growing roles include AI and machine learning specialists, sustainability specialists, and business intelligence analysts. Notice how each of those combines technical skills with deep knowledge of a specific domain.

The Bigger Picture

What’s happening in computer science is a preview of what’s coming for many white-collar professions. Anthropic CEO Dario Amodei has warned that AI could eliminate half of all entry-level white-collar jobs within one to five years. Geoffrey Hinton, the so-called godfather of AI, predicts that in a few years, AI will be able to handle software engineering tasks that currently take a month. Whether those predictions are precisely right or somewhat exaggerated, the direction is clear.

The old playbook -- pick the “right” major, get the degree, collect the job offer -- is breaking down. The new playbook is messier and less reassuring, but it’s also more honest. It says: develop real skills, not just credentials. Combine disciplines rather than siloing into one. Stay curious about how AI is reshaping your field instead of ignoring it. Build things that demonstrate what you can do, not just what you studied. And for the love of everything, don’t pick a major solely because someone told you it was “safe.”

The safest bet in an era of rapid change isn’t picking the right field. It’s becoming the kind of person who can learn a new one.

Dave Clark, the former CEO of Worldwide Consumer at Amazon, built a custom CRM in a night and a morning for his new startup.

Not a prototype. Not a demo. A real system that actually fits how his company sells.

He tried the off-the-shelf option first. Too many fields he did not need. Missing the ones he did. A pipeline that did not match reality.

His verdict: “I spent more time fighting the tool than using it.”

So he stopped fighting it and built exactly what he needed.

That should scare every SaaS company charging $50 to $500 per seat per month.

The SaaS bargain is breaking

The SaaS model rests on a simple trade. It is cheaper to pay us every month than to build this yourself.

That math worked when “build it yourself” meant hiring engineers, provisioning infrastructure, coordinating teams, and maintaining software for years.

That math is collapsing.

Over the past year, I have saved my company tens of thousands of dollars in SaaS fees by doing exactly what Dave did. I replaced internal tools that used to justify six-figure annual contracts with custom software built in days.

Not because I suddenly became a better developer. Because development itself fundamentally changed.

The real product was coordination

SaaS companies did not really sell software. They sold relief from coordination pain.

You did not buy Salesforce for a database. You bought it to avoid months of requirements gathering, vendor evaluations, implementation plans, and training sessions.

Dave put it perfectly: “Most friction is not technical. It is structural. Waiting on vendors. Scheduling demos. Debating requirements in meetings that create more meetings.”

When AI removes that friction, what remains?

A monthly bill for features you do not use and workflows that do not match how you actually operate.

The bloat problem

Every successful SaaS product follows the same arc.

They start focused and useful. Then they chase enterprise checklists. Features get added to satisfy procurement, not users. Five years later, you are paying for 200 features to use 12.

Custom tools have the opposite profile. They do exactly what you need and nothing more.

That used to be a weakness.

Now it is the advantage.

The open source middle path

You do not even have to start from scratch.

Open source alternatives exist for nearly every SaaS category. CRMs, project management, analytics, help desks. The code is free. The problem was always implementation and customization.

That problem just got solved.

Take an open source CRM. Fork it. Strip out what you do not need. Add the fields that match your actual sales process. Deploy it on your own infrastructure. A year ago, that was a multi-month project requiring specialized talent. Now it is a weekend with AI assistance.

You get the foundation someone else built, customized exactly to your workflow, with no per-seat fees and no vendor lock-in.

The SaaS pitch was always “why build when you can buy?” Open source offered a counter: “why rent when you can own?” But ownership came with complexity that made the rental worthwhile for most companies.

AI collapses that complexity gap. The open source option just became viable for companies that never would have considered it.

Who is most at risk

Not all SaaS categories are equally exposed.

The most vulnerable: horizontal tools with generic workflows like CRMs, project management, and basic analytics. High per-seat pricing with low switching friction. Products where the interface is the value, not the data or the network. Tools where customers can export their data and rebuild the experience on their own stack.

If your customer is paying $50,000 a year and can recreate 80 percent of your product in a long weekend, you are not selling software. You are renting time until replacement.

Who survives

Some products still have real moats.

Network-driven tools where value increases with adoption. Deeply embedded systems where replacement touches everything. Highly regulated workflows where homegrown software creates compliance risk. Products where the proprietary model is the product, not the UI wrapped around it.

These companies are not competing with weekends.

Everyone else is.

The venture math problem

Most SaaS companies are still priced and operated on growth-era assumptions.

High acquisition spend justified by long-term retention. Valuations built on 95 percent net revenue retention. Switching costs assumed to be structural.

Those assumptions are eroding fast.

If a customer can rebuild your core functionality in 72 hours, your churn model is already wrong. And if you raised at a double-digit revenue multiple assuming it would never happen, no amount of “AI features” will save you.

The only thing holding this together

Right now, the SaaS industry is not being sustained by product value. It is being sustained by friction.

Switching costs. Multi-year contract lock-ins. The pain of migrating data. The institutional inertia of “we already use this.”

These are not moats. They are delays.

Every contract that expires is a decision point. Every renewal is now a question that did not used to get asked: do we actually need to keep paying for this, or could we just build it?

SaaS companies know this. It is why sales teams push so hard for multi-year commitments. It is why auto-renewal clauses exist. It is why cancellation flows are designed to be as painful as possible.

The product is no longer the lock-in. The contract is.

That works until it does not. And when enough companies start doing the math at renewal time, the churn will not be gradual. It will cascade.

What this means for buyers

Before you sign your next annual contract, ask one question:

Could we build 80 percent of this in a weekend?

If the answer is yes, you probably should.

The SaaS industry spent 15 years convincing companies that building software was too hard to attempt.

That was true.

It is becoming less true every month.

I am not renewing three contracts this quarter. The replacements took less time to build than the sales calls took to schedule.

Michael Burry’s November 2025 accusation that Big Tech is perpetrating “one of the more common frauds of the modern era” through server depreciation manipulation has collided with an even more concerning development: the emergence of a $125+ billion debt-financed AI infrastructure boom that makes the telecom bubble’s financing arrangements look conservative by comparison.

This analysis examines the intersection of three critical trends: hyperscaler depreciation schedule extensions that have added $13+ billion in artificial annual earnings, a surge in GPU-collateralized debt financing that now exceeds $20 billion, and Oracle’s unprecedented $108 billion debt load as the company races to fulfill a $300 billion contract with an unprofitable customer (OpenAI). The implications extend beyond individual company risk to systemic concerns about asset-backed securities, private credit markets, and the fundamental economics of AI infrastructure.

Part I: The Depreciation Manipulation Case

What the Hyperscalers Actually Changed

The depreciation schedule extensions are documented fact. Between 2020 and 2025, all four major hyperscalers systematically extended server useful lives from industry-standard 3-4 years to 5-6 years, generating $13+ billion in cumulative annual earnings benefits.

Microsoft moved first in July 2020, extending server lives from 3 to 4 years, then jumped to 6 years in July 2022. The FY2023 impact: $3.7 billion in additional operating income. Microsoft’s depreciation rate as a percentage of net property and equipment fell from 30-34% in FY2014-2020 to approximately 15% by FY2024.

Google followed suit, extending server lives from 3 to 4 years in January 2021, then to 6 years in January 2023. The 2023 change alone reduced depreciation expense by $3.9 billion and boosted net income by $3.0 billion.

Amazon pioneered the trend, moving from 3 to 4 years in January 2020, then to 5 years for servers and 6 years for networking equipment by 2022. But Amazon’s February 2025 decision to reverse the extension for AI-specific servers-shortening useful life from 6 back to 5 years-represents the most significant validation of Burry’s concerns. The company cited “increased pace of technology development, particularly in the area of artificial intelligence.”

Meta extended non-AI server lives to 5.5 years in January 2025, projecting $2.9 billion in reduced depreciation expense. Notably, Meta explicitly excluded AI servers from the extension-an implicit acknowledgment that GPU-intensive infrastructure depreciates faster than traditional compute.

The Technical Case for Accelerated Obsolescence

Burry’s core argument rests on a fundamental mismatch: NVIDIA releases new GPU architectures annually (accelerated from 18-24 month cycles), yet hyperscalers depreciate these assets over 5-6 years. The technical evidence supports his concern.

A Google architect’s assessment found GPUs running at 60-70% utilization-standard for AI workloads-survive only 1-3 years due to thermal and electrical stress. Princeton’s CITP analysis of Meta’s Llama 3 training study documented a 9% annualized GPU failure rate, implying 27% failure over three years. H100 GPUs consuming 700 watts per chip create significant thermal degradation that legacy servers never faced.

Technological obsolescence compounds physical degradation. NVIDIA’s GB200 “Blackwell” chip delivers 4-5x faster inference than the H100. CEO Jensen Huang stated: “When Blackwell starts shipping in volume, you couldn’t give Hoppers away.” NVIDIA declared the A100 series end-of-life in February 2024, merely four years after its 2020 release.

Part II: The Debt-Financed Infrastructure Boom

The Scale of AI Infrastructure Debt

A UBS report from November 2025 revealed that AI data center and project financing deals surged to $125 billion in 2025, up from just $15 billion in the same period in 2024-an 8x increase. Morgan Stanley estimates private credit markets could supply over half the $1.5 trillion needed for data center buildout through 2028. JP Morgan now estimates AI-linked companies account for 14% of its investment grade index, surpassing U.S. banks as the dominant sector.

The financing has taken multiple forms: investment-grade corporate bonds (Meta’s $30 billion, Oracle’s $18 billion), private credit facilities (Meta’s $29 billion deal with PIMCO and Blue Owl), GPU-collateralized debt (CoreWeave’s $9.9 billion), and an emerging asset-backed securities market that BofA estimates could add $50-60 billion in supply in 2026.

Oracle: The $108 Billion Test Case

Oracle has emerged as the most extreme example of debt-financed AI infrastructure ambition. As of December 2025, the company carries approximately $108 billion in debt-up from $92.6 billion in May-making it the largest issuer of investment-grade debt among non-financial firms.

The debt is being deployed to fulfill a staggering $300 billion, five-year contract with OpenAI for cloud compute services, with payments expected to reach $60 billion annually starting in 2027. Oracle’s remaining performance obligations have exploded to $523 billion-up 438% year-over-year-as the company has signed deals with OpenAI, Meta, NVIDIA, xAI, and others.

The December 2025 Warning Signs: Oracle’s fiscal Q2 2026 earnings (reported December 10, 2025) revealed concerning dynamics. Revenue of $16.06 billion missed expectations of $16.21 billion. Free cash flow was negative $10 billion for the quarter-nearly double the consensus estimate of negative $5.2 billion. The company raised its full-year capex guidance to $50 billion, up from $35 billion just three months prior. The stock fell 11% after the report.

Citi analyst Tyler Radke estimates Oracle will need to raise $20-30 billion in debt annually for the next three years. Moody’s changed its outlook on Oracle to negative in July 2025, citing “the expectation of continuing elevated leverage and increasingly negative free cash flow.” Oracle’s 5-year credit default swaps have climbed to their highest level since 2009.

The Counterparty Problem: The most significant risk may be Oracle’s largest customer. OpenAI remains unprofitable and relies on continuous funding rounds. Sam Altman has stated OpenAI will reach $20 billion in annualized revenue in 2025 and projects “hundreds of billions” by 2030-but paying Oracle $60 billion annually starting 2027 requires extraordinary revenue growth. As Moody’s analysts noted: “Given the lack of financial information about the potential counter parties, this risk assessment is subjective at best.”

The $38 Billion Stargate Debt Package

In October 2025, banks led by JPMorgan Chase and Mitsubishi UFJ Financial Group began marketing a $38 billion debt offering-the largest AI infrastructure financing in history-to fund data centers tied to Oracle. The package is split across two facilities: $23.25 billion for a Texas campus and $14.75 billion for a Wisconsin project.

These facilities are part of the Stargate initiative, a $500 billion AI infrastructure project announced by President Trump in January 2025 involving OpenAI, Oracle, and SoftBank. The project aims to build 10 gigawatts of compute capacity across sites in Texas, New Mexico, Ohio, and Wisconsin. SoftBank has already borrowed $10 billion from Mizuho for its portion; Blue Owl raised $18 billion from Japanese banks for a New Mexico site.

The financing structure reveals the precarious nature of this buildout. Oracle has signed 17-year leases on sites to support the debt, with lenders stepping in to assume control if projects default. The dependency chain is stark: Oracle borrows to build infrastructure → OpenAI commits to pay Oracle → OpenAI must raise funds or generate revenue → investors must continue believing in the AI thesis.

CoreWeave: The GPU Collateral Experiment

CoreWeave has pioneered a novel and concerning financing model: using NVIDIA GPUs as collateral for massive debt facilities. The company has raised $25+ billion in total capital commitments, with approximately $9 billion in current and non-current debt secured primarily by its GPU inventory.

The Collateral Problem: Unlike real estate-which generally appreciates and can be amortized over decades-GPUs are rapidly depreciating assets. CoreWeave’s loan terms require quarterly payments based on cash flow and, critically, the depreciated value of the GPUs used as collateral. As new NVIDIA architectures launch and GPU rental prices fall, the collateral value shrinks while the principal remains fixed.

The H100 rental market has already seen dramatic price declines: from $8/hour at peak to $2.36-3.50/hour by late 2025-a 60-70% reduction. Analysis suggests that once H100 rental rates fall below $1.65/hour, revenues no longer recoup the investment. Prices need to remain above $2.85/hour to beat stock market returns.

The Depreciation Discrepancy: CoreWeave depreciates its GPUs over 6 years-the same aggressive schedule the hyperscalers use. Competitor Nebius uses a 4-year depreciation period. This longer schedule artificially suppresses CoreWeave’s operating expenses and inflates its operating income, while masking the true rate of asset value erosion.

Jim Chanos, the legendary short-seller who exposed Enron, has raised concerns about CoreWeave’s model. The company’s annualized interest expense of approximately $1.2 billion approaches its adjusted EBITDA of $3.4 billion, leaving minimal margin for GPU depreciation. CoreWeave’s stock has fallen approximately 57% from its June 2025 high, though it recovered after announcing a $14.2 billion contract with Meta.

Customer Concentration Risk: Just two companies drove 77% of CoreWeave’s 2024 revenues, with Microsoft accounting for 62%. These largest customers are also its biggest competitors-hyperscalers who could decide to build rather than rent at any moment.

The GPU-Backed Debt Contagion

CoreWeave’s model has spawned imitators. London-based Fluidstack secured over $10 billion in loans from Macquarie and other lenders using NVIDIA GPUs as collateral. Multiple AI cloud computing startups now use high-power chips as collateral, with total GPU-backed borrowing exceeding $20 billion.

Lenders are demanding interest rates in the double digits-exceeding most high-yield bond requirements-reflecting the depreciation risk. Some companies are exploring AMD chips as alternative collateral. TensorWave is actively seeking debt financing with AMD chips as collateral, which would be one of the first such deals.

The fundamental question remains: what happens when the next NVIDIA architecture launches and existing GPU collateral loses 40-50% of its value overnight? The loans require principal payments; the collateral doesn’t maintain principal value.

Part III: The Securitization Parallel

Data Center ABS and CMBS: 2008 Redux?

Asset-backed securities (ABS) and commercial mortgage-backed securities (CMBS) are emerging as significant funding sources for AI infrastructure. BofA notes that digital infrastructure-primarily data centers-accounts for $82 billion of the $1.6 trillion U.S. ABS market, having expanded more than 9x in less than five years. Data centers backed 63% of that digital infrastructure segment.

These securities bundle data center lease payments-rent paid by hyperscalers to facility operators-into tradeable bonds structured by risk tranche. The pitch is compelling: hyperscaler tenants have strong credit ratings, long-term leases, and mission-critical need for the facilities.

The 2008 Echo: ABS are viewed with caution since the 2008 financial crisis, when billions of dollars’ worth of products turned out to be backed by soured loans and highly illiquid assets. The data center version differs in that the underlying cash flows come from creditworthy tenants rather than subprime borrowers. However, the structures share a common vulnerability: they assume the underlying asset-whether a house or a GPU cluster-maintains value throughout the security’s life.

Al Cattermole, fixed income portfolio manager at Mirabaud Asset Management, told Reuters in November 2025 that his team had not invested in any AI-linked investment-grade or high-yield bonds. His reasoning: “Until we see data centres being delivered on time and on budget and providing the computing power that they are intended to-and there still being the demand for it-it is untested. And because it’s untested, that’s why I think you need to be compensated like an equity investor.”

Private Credit: The New Risk Reservoir

Private credit has become a crucial funding source for AI infrastructure. UBS estimates private credit AI-related loans nearly doubled in the 12 months through early 2025. Morgan Stanley projects private credit could supply over half the $1.5 trillion needed for data center buildout through 2028-approximately $750 billion.

The appeal for borrowers is clear: private credit offers fixed-rate structures, customized terms, and avoidance of public bond market scrutiny. Meta’s $29 billion deal with PIMCO and Blue Owl-structured as $26 billion in debt and $3 billion in equity-exemplifies the model. Microsoft struck a $30 billion partnership with BlackRock. xAI raised $5 billion in syndicated debt.

The Illiquidity Risk: Unlike traded bonds, private credit loans are harder to trade during market turmoil. The Bank of England has flagged that “pockets of risk are building in parts of the financial system populated by opaque, hard-to-trade illiquid assets.” If AI demand disappoints and borrowers struggle, private credit investors may find themselves holding assets with no market and rapidly deteriorating collateral.

Part IV: Historical Parallels and Differences

The Telecom Bubble Comparison

The telecom bubble offers the most relevant precedent. WorldCom’s $11 billion fraud included depreciation manipulation alongside capitalizing operating expenses. Waste Management stretched garbage truck depreciation periods to reduce annual expense, ultimately restating $1.7 billion in earnings.

More broadly, $500+ billion was invested in fiber optic infrastructure, 85-95% of which remained “dark” (unused) four years after the bubble burst. Global Crossing achieved a $47 billion market cap without ever turning a profit. Lucent Technologies offered $8.1 billion in vendor financing-about 24% of revenue-and collapsed when customers defaulted.

JP Morgan has explicitly made this comparison, estimating AI will need $650-800 billion in annual revenue by 2030 just to generate a 10% return on infrastructure capex. Bain estimates an $800 billion annual revenue gap between AI investment and revenues.

Critical Differences

Several factors distinguish the current situation from the telecom bubble. Hyperscalers possess massive cash reserves and sustainable core businesses unlike pure-play telecoms. Amazon’s AWS, Microsoft’s Azure, and Google Cloud generate hundreds of billions in annual revenue with strong operating margins. The AI infrastructure spending, while enormous, represents a fraction of their financial capacity. Meta issued its first dividend in 2024 despite the capex boom-impossible for debt-laden 1990s telecoms.

Additionally, infrastructure financing has matured as an asset class. Specialized lenders understand data center economics. Structures include ring-fenced cash flows, specific covenants, and security packages that didn’t exist in the 1990s.

However: Oracle is not a hyperscaler. It carries $108 billion in debt, has negative free cash flow, and is building capacity for a customer that has never been profitable. CoreWeave is not a hyperscaler. It has 77% customer concentration, GPU collateral that depreciates rapidly, and an interest coverage ratio of 0.17. The startups borrowing billions against GPU inventory are certainly not hyperscalers. The comparison to telecom overbuild is most apt not for Microsoft or Google, but for the second and third tier of infrastructure providers.

Part V: The NVIDIA Paradox

Why Obsolescence Might Benefit NVIDIA

Here’s where Burry’s thesis creates an unexpected implication for NVIDIA investors. If GPUs depreciate faster than accounting schedules suggest, this creates sustained demand for replacement chips rather than a one-time buildout.

CoreWeave CEO Michael Intrator provided direct evidence: “A batch of Nvidia H100 chips became available because a contract expired, and they were immediately booked at 95% of their original price.” He added that “all of our Nvidia A100 chips, which were announced in 2020, are all fully booked.”

The “value cascade” model explains how older GPUs retain economic utility despite obsolescence: Years 1-2 for frontier model training, Years 3-4 for high-value real-time inference, Years 5-6 for batch inference and analytics. Meta exemplifies this tiering-training on cutting-edge H100/H200 GPUs but running inference on AMD MI300X chips.

If Burry is correct that 6-year depreciation is aggressive, the implication is that hyperscalers must continuously purchase new NVIDIA chips at 3-4 year intervals rather than 6-year cycles. This doubles the replacement frequency and sustains demand indefinitely. NVIDIA’s $500 billion backlog through 2026 supports this thesis.

The Competitive Moat

NVIDIA’s defensive position remains formidable regardless of the depreciation debate. SemiAnalysis conducted a five-month benchmark of AMD’s MI300X in December 2024 and concluded: “For all models, the H100/H200 wins relative to MI300X. AMD’s software experience is riddled with bugs rendering out of the box training with AMD impossible... The CUDA moat has yet to be crossed by AMD.”

NVIDIA’s CUDA ecosystem encompasses 3.5 million developers with nearly two decades of optimization. Intel’s Gaudi holds less than 1% market share. Custom silicon from Google, Amazon, and Meta handles primarily inference and internal workloads-less than 20% of frontier model training runs on non-NVIDIA silicon.

Conclusion: Layers of Risk

The AI infrastructure financing landscape presents layered risks that compound upon each other:

Layer 1 - Depreciation Manipulation: Hyperscalers have documented $13+ billion in annual earnings benefits from extending depreciation schedules. Amazon’s reversal for AI servers validates that these schedules are aggressive.

Layer 2 - Debt Financing Surge: AI infrastructure financing surged to $125 billion in 2025 (8x 2024 levels), with projections of $750 billion in private credit through 2028. This creates massive leverage exposure if demand disappoints.

Layer 3 - Collateral Degradation: GPU-backed debt exceeding $20 billion depends on assets that have already declined 60-70% in rental value. The 6-year depreciation schedules used by borrowers like CoreWeave mask this reality.

Layer 4 - Counterparty Concentration: Oracle’s largest customer is unprofitable. CoreWeave’s two largest customers represent 77% of revenue. The dependency chains are fragile.

Layer 5 - Securitization Proliferation: Data center ABS/CMBS have grown 9x in five years. Private credit markets are increasingly exposed to AI infrastructure. The lack of transparency about actual utilization and demand echoes pre-2008 mortgage market opacity.

The key variable is whether AI generates sufficient return on investment to justify the infrastructure buildout. Combined hyperscaler capex guidance exceeds $300 billion for 2025 alone. Backlogs total $747 billion. Power constraints, not demand, appear to be the limiting factor.

For NVIDIA, the paradox of obsolescence remains intact: even if individual GPUs become worthless faster than accounting suggests, the aggregate effect is a treadmill that customers cannot exit. Oracle’s $50 billion annual capex, CoreWeave’s refinancing requirements, and the broader infrastructure boom all translate to continuous GPU orders.

The risk is not that AI demand disappears-it almost certainly won’t. The risk is that the second and third tier of infrastructure providers, loaded with debt secured by depreciating assets, cannot survive the gap between infrastructure investment and AI monetization. Amazon’s depreciation reversal may represent the market’s most honest signal: AI hardware depreciates faster than traditional servers, but the hyperscalers will keep buying. Whether Oracle, CoreWeave, and the GPU-backed lending ecosystem can say the same is the $176 billion question.

In June 2023, I publicly raised concerns about OpenStore. Not because I’m some oracle of business wisdom, but because the fundamentals were screaming red flags that anyone in the ecommerce space could see. The aggressive aggregation model, the lack of domain expertise, the hubris of believing algorithms could replace operational know-how.

I felt like Michael Burry in 2006, watching the mortgage market and wondering if I was the only one who could do basic math. The difference? Burry was betting against fraud hidden in complexity. I was just pointing out incompetence hiding in plain sight.

Now, in late 2025, OpenStore is essentially dead. They’ve shut down 40+ brands, keeping only Jack Archer, and reportedly raised at a $50 million valuation. That’s a staggering 95% drop from their previous nearly $1b valuation in September 2022. They raised over $150 million in investor capital buying brands they couldn’t profitably operate, all while their architect, Keith Rabois, promoted mediocre milestones on social media like they were revolutionary achievements.

But here’s the thing: OpenStore isn’t an anomaly. It’s a pattern.

When VCs Play PE in Industries They Know Nothing About

Here’s what really happened with OpenStore, and it’s a pattern repeating across multiple sectors: venture capitalists trying to execute private equity rollup strategies in industries where they have zero domain expertise.

Private equity firms have been doing rollups for decades. They work when you have operators who deeply understand the industry, know how to identify operational efficiencies, can integrate acquisitions effectively, and understand the unit economics down to the cent. PE firms hire industry veterans. They bring in people who’ve run similar businesses for 20+ years. They respect the operational complexity.

VCs like Rabois looked at this playbook and thought: “We can do that, but with ‘technology’ and ‘algorithms.’” Then they proceeded to do everything wrong.

OpenStore didn’t hire people with ecommerce experience. They didn’t bring in operators who had successfully scaled DTC brands, managed Amazon storefronts, or understood the nuances of customer acquisition costs, inventory management, and brand positioning. Instead, they hired from tech companies like Apple, Uber, and DoorDash.

Think about how insane that is. You’re buying 40+ ecommerce brands, and you staff your company with people from ride-sharing apps and tech hardware companies. People who’ve never had to worry about cost of goods sold, supplier relationships, seasonal inventory planning, or the difference between a sustainable CAC:LTV ratio and one that burns cash.

It’s the blind leading the blind, except the blind people are convinced they can see better than everyone else because they know nothing about the industry. They genuinely believed their ignorance was an asset. “We’re not constrained by industry thinking!” they’d say. “We can see opportunities others miss!”

No. You’re just ignorant. And your ignorance cost $150 million and destroyed 40+ businesses.

The arrogance is breathtaking. Imagine a world-class chef watching someone who’s never cooked before walk into a Michelin-star restaurant and declare: “I’m going to revolutionize fine dining because I’m not constrained by culinary training. I have an algorithm!” That’s essentially what happened here.

The people who actually understand ecommerce, the ones who’ve been in the trenches building and scaling brands, saw this coming from miles away. We weren’t smarter. We just knew what they didn’t know. And what they didn’t know could fill a warehouse.

When you combine VCs with no ecommerce experience, executives with no ecommerce experience, and a strategy that requires deep operational excellence in ecommerce, you get exactly what we got: spectacular failure.

This is what happens when venture capital, drunk on years of backing asset-light software businesses, decides to wade into asset-heavy operational businesses. Software can scale with minimal marginal cost. Real businesses with physical products, inventory, supply chains, and customer service teams? Those require actual operational expertise.

But admitting that would require humility. And humility doesn’t raise $150 million or get you on the Forbes Midas List.

The OpenDoor Parallel: A Masterclass in Value Destruction

Let’s talk about OpenDoor, another Keith Rabois creation, this time co-founded and currently serving on its board. In December 2020, Rabois partnered with “SPAC King” Chamath Palihapitiya to take OpenDoor public via Social Capital Hedosophia II at a $4.8 billion enterprise value. The market initially loved it. The valuation soared to nearly $18 billion on its first day of trading.

Today? The story is both more volatile and more revealing. OpenDoor’s stock hit a 52-week low of $0.51 per share in mid-2025. That’s a catastrophic 97% decline from its IPO peak. While the stock has since surged to trade around $6-7 (giving it a current market cap of approximately $6-7 billion as of November 2025), this represents classic meme stock behavior, not fundamental improvement.

OpenDoor has never turned a profit except for brief quarters in 2022-2023. In 2024, the company posted a net loss of $392 million. In Q3 2025 (their most recent quarter) they lost $0.12 per share and posted a net loss of $90 million. The company is targeting “adjusted net income breakeven by the end of 2026.” That’s six years after going public, and after cumulative losses in the billions.

The Math That VCs Don’t Want You To See

Here’s where it gets really ugly. Let’s talk about the actual numbers, because they tell a story of systematic value destruction that should make any investor sick.

Total capital raised by OpenDoor:

• ~$1.3 billion in pre-IPO venture capital funding

• $1 billion from the SPAC merger (2020)

• $850 million in a post-IPO equity round (August 2021)

• Total equity capital raised: approximately $2.1-2.2 billion

And what do investors have to show for it?

OpenDoor’s current book value (stockholders’ equity as of Q2 2025): $631 million.

Read that again. They raised over $2.1 billion in equity capital. The book value of the company (assets minus liabilities) is $631 million. That means they’ve destroyed approximately $1.4-1.5 billion in shareholder value. Gone. Vaporized.

That’s not just a negative return. That’s a negative IRR over the entire life of the company. Every single investor, from the earliest venture rounds through the SPAC and post-IPO financing, has collectively lost money on a book value basis. The only way anyone makes money now is if they can find a greater fool to buy the meme stock at an inflated price.

This isn’t disruption. This is destruction. And not the good kind of “creative destruction” that capitalism is supposed to deliver. This is pure capital incineration, dressed up in the language of innovation and sold to investors by people who confused their past successes in completely different industries with competence in real estate.

Think about what else could have been done with $2.1 billion. That’s real money. That’s funding for dozens of actually viable startups. That’s returns that could have gone to pension funds and endowments. Instead, it’s been systematically destroyed by a business model that economics 101 could have told you wouldn’t work.

Like the synthetic CDOs in The Big Short, everyone was so busy celebrating the innovation that nobody bothered to check if the underlying assets made any sense. The difference is that mortgage traders at least understood mortgages. These VCs didn’t even understand the basics of the industries they were “disrupting.”

The iBuyer “Lemons Problem”

The fundamental flaw in the iBuyer model (what economists call the “lemons problem”) was well-documented even before OpenDoor went public. When sellers know more about their homes than buyers (including algorithmic buyers), they have an incentive to offload their worst properties to instant-cash buyers. OpenDoor’s algorithms, no matter how sophisticated, consistently underperformed against this adverse selection problem.

Rabois, of course, dismissed these concerns. In late 2020, when he partnered with Chamath to take OpenDoor public via SPAC, he was all over social media promoting it as the future of real estate. The fact that he had no real estate experience? Irrelevant. The fact that the unit economics were questionable? Details. The fact that Zillow, with far more data and domain expertise, would shut down their iBuying operation just a year later? A validation that OpenDoor was “the winner.”

The hubris is breathtaking. While actual real estate professionals were raising red flags about adverse selection and capital intensity, Rabois was tweeting about “revolutionizing a $1.6 trillion market.” Now, five years later, OpenDoor has revolutionized nothing except new ways to destroy shareholder value.

Zillow shut down Zillow Offers in November 2021 after $881 million in losses. Redfin shut down RedfinNow in November 2022. But OpenDoor pressed on, burning through billions while Rabois championed it as a revolutionary real estate platform.

From “Tech-Enabled” to Plain Old Real Estate

Here’s what OpenDoor actually is: a poorly performing REIT masquerading as a technology company. They buy homes, hold inventory, and try to flip them. There’s no revolutionary technology here. Just a capital-intensive, low-margin business in one of the most cyclical sectors in the economy. When interest rates rose and the housing market cooled, OpenDoor was left holding billions in depreciating inventory and no path to profitability.

The recent stock surge? It’s not driven by improved fundamentals. The new CEO, Kaz Nejatian, announced a gimmick to squeeze short sellers by issuing tradable warrants to shareholders. It’s the kind of financial engineering that sends retail traders into a frenzy but does nothing to fix the underlying business model.

Speaking of Nejatian, here’s an interesting detail that fits the pattern: he came from Shopify where he served as COO, but his previous company Kash was reportedly “acquired by one of the largest fintech companies in the U.S.” with mysteriously no public record of this acquisition anywhere. No SEC filings, no press releases from the supposed acquirer, no deal terms, nothing. In an industry where founders inflate their exits like balloons at a children’s party, OpenDoor managed to hire a CEO whose biggest career achievement remains unverified. It’s perfect, really a company built on financial engineering and meme stock manipulation, run by someone whose own exit story appears equally engineered.

With approximately 25% of the float sold short, OpenDoor has achieved meme stock status, not business success.

The Aggregator Graveyard

OpenStore’s demise marks the effective end of the ecommerce aggregator boom. Thrasio (once the poster child with billions in funding) filed for bankruptcy in February 2024. Perch, Heyday, Unybrands: all either failed, merged in desperation, or cut their valuations dramatically. The entire sector raised over $16 billion collectively, and almost all of it has been vaporized.

But here’s the kicker: even as the entire aggregator sector was collapsing around him, Rabois was still promoting OpenStore’s “successes” on social media. In March 2023, while Thrasio was heading toward bankruptcy and other aggregators were desperately cutting valuations, Rabois was tweeting about OpenStore’s “operational excellence” and sharing articles about scaling brands from $1M to $10M as if this was revolutionary rather than table stakes for any competent ecommerce operator.

By July 2025, reality finally caught up. OpenStore’s valuation was cut by 95%, a new CEO was brought in, and Rabois quietly distanced himself from day-to-day operations. No mea culpa. No acknowledgment of the failure. Just a swift pivot back to Khosla Ventures where he could start the cycle anew with fresh LP money.

Why did all these aggregators fail? Because buying businesses isn’t the hard part. Operating them profitably is. And you can’t algorithm your way out of needing actual operational expertise.

The Rabois Pattern

What connects these stories is Keith Rabois. He’s a venture capitalist with an impressive resume (PayPal, Square, LinkedIn) who appears to believe his past successes make him infallible in any market he enters.

With OpenStore, Rabois had no ecommerce operational experience but confidently declared he could “acquire a business in a day” and eventually wanted to get to “one an hour.” He promoted the company relentlessly on social media, posting about “the best talent” and “the future of commerce online” even as the business was imploding behind the scenes.

When things went south? Silence. No accountability. No acknowledgement of the $150 million raised and 40+ brands destroyed. Just a pivot to calling it “10x focus on what is anomalously great.” As if concentrating on one surviving brand out of 40 was the plan all along.

With OpenDoor, the playbook is eerily similar: aggressive promotion, grandiose claims about revolutionizing a massive market, algorithmic overconfidence, and a consistent inability to turn a profit despite years and billions in capital. Rabois rejoined OpenDoor’s board in September 2025, just in time for the company to become a meme stock rather than a sustainable business.

The Art of Failing Up

There’s something particularly galling about Rabois’s trajectory. In any other industry, destroying $150 million at OpenStore while simultaneously presiding over $1.4 billion in value destruction at OpenDoor would end a career. In Silicon Valley? It gets you a promotion and access to $3.1 billion in fresh capital.

When Vinod Khosla announced Rabois’s return in January 2024, he gushed that Keith “knows how to advise entrepreneurs on hiring/firing, running teams, managing funding.” This is the same person whose own venture couldn’t manage any of these things successfully. OpenStore went from unicorn to essentially a single menswear brand. The open.store domain, once the flagship of their “revolutionary ecommerce platform”, now doesn’t even load.

The timing tells the real story. Rabois left Khosla for Founders Fund in February 2019, co-founded OpenStore in Miami in 2021 during the peak of cheap capital, watched it implode through 2023-2024, and then boomeranged back to Khosla in January 2024, just as the OpenStore disaster became impossible to ignore. Forbes called it a “surprise return.” The only surprise is how transparent the move was.

This is the Silicon Valley accountability problem in miniature: fail spectacularly, move laterally, raise more money, repeat. The same network that enables this behavior then wonders why “founders” with zero domain expertise keep burning billions on obviously flawed business models. These VCs spend their careers vying to be featured on the Forbes Midas List, but it seems their touch turns real, hard-working businesses to dirt instead of gold. King Midas at least had the excuse of a curse, these guys are just incompetent.

Not Gloating: Just Pattern Recognition

I’m not writing this to gloat. I’m writing it because there’s a lesson here about venture capital, expertise, and humility (or the lack thereof).

The venture capital model works when VCs back founders with domain expertise. It fails spectacularly when VCs believe they can parachute into complex operational businesses and use capital and “technology” to paper over their lack of expertise.

OpenStore wasn’t killed by market conditions. Plenty of ecommerce businesses thrived during the same period. It was killed by operational incompetence. Similarly, OpenDoor isn’t struggling because the real estate market is inherently unprofitable. Traditional brokerages and individual investors make money every day. It’s struggling because the iBuyer model, as architected, doesn’t work at scale.

The recent meme stock surge only obscures the fundamental reality: after 11 years in operation and five years as a public company, OpenDoor still loses hundreds of millions of dollars annually while having destroyed over $1.4 billion in shareholder value.

The Bigger Picture

These failures matter beyond just a few companies. They represent hundreds of millions in misallocated capital, hundreds of jobs lost, and the destruction of dozens of previously viable ecommerce brands that got rolled up and mismanaged into oblivion.

More importantly, they represent a broader phenomenon in tech and venture capital: the belief that “disruption” means you can ignore the fundamentals of the industries you’re trying to disrupt. That algorithms can replace judgment. That capital can substitute for competence. That viral marketing can replace product-market fit. That hiring from Big Tech is better than hiring people who actually understand the business you’re trying to run.

Sometimes the emperor really has no clothes. And sometimes the people pointing it out aren’t being cynical. They’re just paying attention to the fundamentals everyone else is ignoring in favor of a good story told by a charismatic founder with an impressive pedigree but zero relevant experience.

Postscript: Where Are They Now?

OpenStore is now essentially just Jack Archer, a menswear brand, run by a new CEO with no involvement from Rabois in day-to-day operations. The open.store domain now redirects to jackarcher.com. Not exactly the revolutionary “portfolio of serendipitous discovery” that was promised.

OpenDoor continues to lose money quarter after quarter, targeting profitability “by the end of 2026.” That’s a promise they’ve been making in various forms for years. The stock has become a meme stock playground with extreme volatility, recently trading in the $6-7 range on warrant gimmicks and short squeezes rather than actual business improvement. The company is now run by a new CEO who describes OpenDoor as “a software and AI company” rather than acknowledging the reality: it’s a capital-intensive real estate flipping operation that has yet to prove it can make money. With over $1.4 billion in shareholder value destroyed, it stands as a monument to what happens when VCs play in industries they don’t understand.

Keith Rabois? After OpenStore’s spectacular implosion became undeniable, he made a surprise return to Khosla Ventures in January 2024 as a managing director, conveniently leaving Founders Fund just as his ecommerce aggregator was collapsing. The official story? He didn’t like the commute from SF to Sand Hill Road back in 2019. The convenient timing? Founders Fund had cut back their fund size while Khosla had just closed on $3.1 billion in fresh capital.

Khosla even set up a new Miami office for him in Wynwood, a validation of his “commitment to the city”. The same city where he co-founded OpenStore in 2021 before watching it burn through $150 million. Now he’s back on OpenDoor’s board, presumably to oversee its continued money-losing operations as it promises profitability “by the end of 2026.” Same playbook, different fund, fresh billions to deploy.

The Real Cost

The pattern is clear: OpenStore, OpenDoor, and the broader failures they represent aren’t bugs in the Silicon Valley system, they’re features. They’re what happens when venture capitalists confuse confidence with competence, when viral marketing substitutes for viable business models, and when the ability to raise capital becomes more important than the ability to deploy it wisely.

The real tragedy isn’t just the billions destroyed. It’s that this capital could have funded hundreds of boring, profitable businesses run by operators with actual domain expertise. Instead, it went to feed the egos of VCs who thought they could algorithm their way through industries they didn’t understand.

Until there are real consequences for this kind of value destruction. Until LPs stop funding VCs who consistently burn billions, we’ll keep seeing the same pattern: charismatic founders with zero relevant experience raising massive rounds, burning through capital while posting victory laps on social media, then failing up to the next fund when reality catches up.