The New York Times just published an opinion piece arguing that Meta is entering its zombie era, the slow decline phase where a once dominant internet company becomes irrelevant like AOL or Yahoo. User numbers drop for the first time ever, the stock is slumping, Zuckerberg is burning cash on failed AI bets while the core business withers. But I think the New York Times piece, despite getting the trajectory right, misses what's actually happening here. And what's happening at Meta isn't just founder makes bad bets and company declines, it's something more specific and more interesting, strategic incoherence driving financial desperation driving illegal shortcuts.

Let's start with the metaverse as promised because it sets the pattern for everything that came after. On October 28th, 2021, Mark Zuckerberg appeared on a virtual stage and declared that Facebook, the social network with over 3 billion users, would henceforth be called Meta Platforms. The company was betting its entire identity on the metaverse, a virtual world where people would don VR headsets and exist as avatars, attending meetings, concerts, and social gatherings in immersive 3D spaces. He described it as the next frontier, predicting the metaverse would eventually reach a billion people and generate hundreds of billions in digital commerce. The flagship product was Horizon Worlds, a social VR platform where you could build virtual spaces, attend events, and hang out with friends as a cartoon avatar. But here's what actually happened. Horizon Worlds launched in late 2021. By February 2022, it had approximately 300,000 monthly active users across Horizon Worlds and Horizon Venues combined. For context, Facebook had 3 billion users, Instagram had over a billion, WhatsApp had 2 billion, the metaverse product got 300,000 people, and most of them left within weeks. And here's the detail that became a meme within days. The avatars launched without legs, just floating torsos. Screenshots of these legless blocky figures circulated online immediately and became the visual shorthand for the entire project's absurdity.

Meta kept pouring money into it anyway. From 2020 to 2026, Reality Labs, the division responsible for VR and metaverse development, accumulated over $80 billion in operating losses. In 2025 alone, the division lost $19.2 billion on just $955 million in revenue. Let me just put that into perspective for a second. $80 billion is more than the entire GDP of Luxembourg. It's roughly what Apple spent to develop every iPhone model ever made combined. And Meta spent it on a product that never cracked 300,000 monthly users and became best known for its legless avatars. On March 18th, 2026, Meta announced it was shutting down Horizon Worlds. VR access dies permanently on June 15th. The metaverse vision that consumed over $80 billion, employed thousands of engineers, and inspired an entire corporate rebrand is, for all practical purposes, dead.

Now, here's what's interesting. The technology itself wasn't the problem. Quest headsets sold reasonably well for gaming. People who bought VR headsets wanted to play Beat Saber. They wanted immersive games. What they didn't want was to attend virtual meetings with legless avatars in empty conference rooms. Meta built a solution to a problem nobody had. Think about the products Meta actually succeeded with. Facebook solved I want to keep in touch with people I know. Instagram solved I want to share photos easily. WhatsApp solved I need free messaging that works internationally. These were market pull innovations. People already wanted these things and Meta just built better versions. But the metaverse was technologically pushed. We build this VR thing, now we need to convince everyone they want to live in it. There was no existing demand for virtual concert attendance as a legless avatar. They tried to manufacture desire for a product nobody really asked for. And the wild part is the Meta stock price hit near all-time highs when they announced the Horizon Worlds shutdown. The market basically rewarded them for killing the project because it signaled leadership was finally allocating capital towards something with proven demand. Which brings us to AI.

After burning 80 billion dollars on legless avatars, Zuckerberg pivoted to AI and here's where the pattern repeats just with much bigger numbers. In 2022 ChatGPT launched and suddenly everyone in tech was racing to build large language models. Zuckerberg jumped in with an open checkbook announcing Meta would build Llama, an open-source AI model that anyone could run on their own machine. The pitch was about democratizing AI, making powerful models available to everyone not locked behind corporate APIs. And Meta spent approximately 100 billion dollars building Llama as an open-source offering. Here's what happened. The model turned out to be kind of slow, inaccurate, and too unwieldy for most people to operate on their own. Running a large language model locally requires significant compute resources, GPUs, memory, technical know-how. The democratization pitch ran into the reality that most people, developers excluded, just want to type a question into a web interface and get an answer, not set up local model infrastructure.

So, Zuckerberg abandoned that effort and committed another $115 billion minimum for 2026 on a new AI push. This time building models that compete directly with ChatGPT, Claude, and other leading systems. And here's the thing, their current models perform worse than the competition. Now, stop and think about that for a second. Meta is a company with 3 billion users, the world's most sophisticated ad targeting infrastructure, decades of behavioral data about what people click, buy, and engage with, very deep understanding of content virality and engagement patterns. If they were strategically coherent, what would they be building, really? AI that makes ad targeting even more effective, their actual business. Content moderation AI, their biggest cost center and legal liability. Recommendation systems that keep people on platform longer, their core engagement problem. Creator tools that make Instagram and Facebook stickier than TikTok. And instead, they're building general-purpose language models to compete with OpenAI.

Why? Who at Meta actually benefits from having an open-source LLM that runs on consumer hardware? What problem does this solve for Meta's business model, which is selling ads based on user attention? This is kind of like if Coca-Cola decided to spend $100 billion building smartphones because tech is the future, while letting their beverage business rot. The core competency doesn't transfer, the strategic logic doesn't connect. And here's where it gets even weirder, Meta is an advertising company. They have proprietary user data that no one else can access. The social graph, behavioral patterns, what content drives engagement. If they wanted to build AI that leverage their actual moat, they would be building domain-specific models for the following: AI-powered advertising tools that small businesses and creators can use to automatically generate optimized campaigns, identify target audiences, and manage the entire conversion funnel, taking maybe 2 to 5% of revenue generated. That is a multi-hundred-billion-dollar business built on infrastructure they already own. Then, content moderation AI that actually works, reducing their biggest cost center and legal exposure while making platforms usable again. Maybe creator monetization tools that keep influencers on Instagram instead of jumping to TikTok, built around AI-assisted content creation and audience analysis. These would be multi-hundred-billion-dollar businesses built on infrastructure they have, but instead, let's spend $150 billion per year building ChatGPT except worse and give it away for free. And look, I understand Meta can build multiple things simultaneously. They have the resources, but here's what I don't understand. Why isn't there a public-facing advertising product built on that data set? The strategic incoherence is just staggering.

Which brings us to the lawsuit that dropped 2 weeks ago, because I think it explains why Meta keeps making these kinds of choices. On May 5th, 2026, five major publishers, Hachette, Macmillan, McGraw Hill, Elsevier, and Cengage, plus author Scott Turow, filed a class action lawsuit against Meta and Zuckerberg personally, alleging they pirated millions of copyrighted books and articles to train Llama. The lawsuit includes internal memos that are generally extraordinary. Here's what happened. According to the filing, after releasing Llama 1, Meta briefly considered licensing training data from publishers. Internal discussions from January to April 2023 show Meta considered increasing their data set licensing budget to as much as $200 million. Then, in early April 2023, the question was escalated to Zuckerberg. Should they license content or just take it? After Zuckerberg weighed in, Meta's business development team received verbal instructions to stop licensing efforts. The internal rationale, according to one Meta employee quoted in the lawsuit, was the following: "If we license one single book, we won't be able to lean into the fear use strategy."

They decided not to license content specifically because licensing even one book would undermine their legal argument that they don't need to license anything. It's not that they couldn't afford $200 million. Meta made $200 billion in ad revenue just in 2025. They're spending $115 billion on AI infrastructure. $200 million for licensing is kind of like a rounding error. They chose piracy because they're betting the entire AI strategy on fair use, the legal doctrine that allows limited use of copyrighted material without permission for purposes like criticism, education, or transformative works.

And the lawsuit alleges they went hard on the piracy. According to the complaint, on December 13th, 2023, Meta employees internally circulated a memo about LibGen, a notorious repository of pirated academic papers and books. The memo described it as "a data set we know to be pirated" and noted "we should not disclose use of LibGen data sets used to train." Despite those concerns, Zuckerberg and other Meta executives authorized torrenting over 267 terabytes of pirated material. The lawsuit says that's equivalent to hundreds of millions of publications and many times the size of the entire print collection of the Library of Congress. Just for context, 267 terabytes, the entire text of Wikipedia is about 20 GB. Meta allegedly torrented the equivalent of 13,000 Wikipedias worth of pirated books and journal articles.

Now, previous lawsuits against AI companies for copyrighted infringement have failed. In June 2025, a federal judge ruled that Meta engaged in fair use when training Llama on a data set of nearly 200,000 books, but this lawsuit is different because of the internal memos. The previous cases were about whether AI training constitutes fair use in principle. This case is about whether you can deliberately circumvent copying protections and actively choose piracy over licensing as a cost-saving measure, then claim fair use as a shield. The lawsuit argues that Meta stripped copyright management information from the copyrighted works it stole to conceal training sources and facilitate unauthorized use. That's not transformative use or incidental copying during research. That is deliberate infringement at industrial scale with internal documentation showing it was a conscious strategic choice. And here's what I personally find fascinating. Meta knows how to license content. According to the lawsuit, they signed licensing agreements with Fox News, CNN, and USA Today. They signed four licenses in 2022 with African language book publishers. They understand the market for licensing AI training materials. They just decided it was cheaper to pirate and fight in court later. And I think that decision reveals something very important. Meta can't afford to build competitive AI the legal way.

If you're building domain-specific AI for advertising using your own proprietary user data, you don't need to scrape the Library of Congress. But if you're trying to build general-purpose models that compete with OpenAI and Anthropic, you need all of human knowledge, which means either paying for it or stealing it. The piracy strategy is a symptom of the strategic incoherence. They're building the wrong thing, which requires data they don't have, which means they either spend hundreds of millions on licensing or they gamble on a fair use defense while torrenting pirated material. They chose the gamble, and while Meta is pirating external data, they're also extracting training data from an even more accessible source, their own employees.

In April 2026, Reuters reported that Meta quietly deployed monitoring software on US employees work computers through a program called the Model Capability Initiative. The software captures every single keystroke, mouse movement, click, and takes periodic screenshots as employees work across hundreds of websites and apps, including Google, LinkedIn, GitHub, Slack, and even personal Gmail accessed on work devices. According to a Meta spokesperson quoted by Reuters, "If we're building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them." That sounds reasonable until you think about what it actually means. Meta is training AI agents to replicate human computer work. To do that, they need to watch humans do computer work, so they're recording every drop-down menu selection, every copy-paste between tools, every Jira workflow that Meta's roughly 72,000 employees perform. CTO Andrew Bosworth confirmed there is no opt-out. If you work for Meta in the US, you're being monitored. European employees are exempt because GDPR won't allow it, but American workers just get a notice that their work laptop isn't really theirs.

And here's the darkly hilarious part. Meta announced this the same week they laid off 8,000 employees. So, the workflow is basically monitor employees doing their jobs, use that data to train AI agents that can do those jobs, lay off the employees, sell the AI agents to other companies. The workers generating the training data don't get a cut. They get a monitoring notice and potentially a severance package if they're lucky. Now look, I understand the business logic of this. Every company right now is trying to figure out how to automate knowledge work. I get it. Collecting data on how employees actually use software is a reasonable R&D strategy. If Meta wants to build workplace AI, they need examples of workplace tasks, sure, but the execution is just kind of tactically insane. You've now got 72,000 people who know they're training their own replacements. Morale is of course destroyed, trust is just out the window, and for what strategic advantage? Because here's the thing, if Meta were building domain specific AI for their actual business, app targeting, content moderation, creator tools, they wouldn't need to monitor employee keystrokes. They would train on their proprietary user data, which they already have copious amounts of. But they're trying to build general workplace AI agents to compete in the enterprise software market, which means they need training data about how people use Google Docs, Slack, LinkedIn, email, none of which is Meta's core business. So once again, wrong product, wrong lane, desperate measures to get the data they need, and the cost isn't just employee morale, it's legal exposure. Think about what flows through those tools on a normal workday, by the way, code commits with proprietary logic, Slack messages about medical leave, personal emails accessed on work devices. You're capturing potentially very sensitive information at massive scale, feeding it into AI training pipelines, and hoping nothing goes wrong. That's not just ethically questionable, it is a compliance nightmare waiting to happen. But they're doing it anyway, because they need the data, and they cannot really afford to acquire it through other means.

Which brings us to the financial engineering holding this whole operation together. Meta is spending staggering amounts on AI. Capital expenditures for 2026 are projected at 115 to 135 billion dollars. directed primarily at data centers AI chips and model development. To put that in context, that's roughly 67% of Meta's projected annual revenue. Most tech companies spend about 15 to 25% of revenue on CapEx. Meta is spending more than half. Where is that money coming from? Increasingly, debt. Meta's long-term debt closed out 2025 at 59 billion dollars, double the prior year's total. And that doesn't count the creative accounting by the way.

In Louisiana, Meta is building a data center called Hyperion, a 27 billion dollar facility that at full build-out could scale to five gigawatts and cover an area Zuckerberg says would stretch across a significant part of Manhattan. Here's how they're financing it. Meta partnered with Blue Owl Capital, a private credit firm. They created a joint venture where Meta owns 20% and Blue Owl owns 80%. Blue Owl formed a legal entity called, and I'm really not making this up, Beignet Investors LLC. They named their 27 billion dollars debt vehicle after a donut. Beignet Investors sold 27 billion dollars in bonds to Wall Street investors. Meta leases the data center from Blue Owl. The rent money flows to Blue Owl which uses it to pay back bond holders. And here's the clever part. Meta gets to renew its lease every four years. If they change their mind on AI, they can just walk away. Because Meta only owns 20% and doesn't control the entity, they keep the 27 billion dollars in debt off their balance sheet.

The Wall Street Journal called this Frankenstein financing. Ernst & Young, Meta's auditor, approved the arrangement while designating it as a critical audit matter, accounting speak for this is complicated, involves significant judgment, and warrants investor attention. Some analysts are comparing this structure to the special purpose entities that led to the Enron collapse, and Meta just announced they're doing it again, raising another $13 billion through a similar structure for a data center in Texas.

Meanwhile, in Louisiana, the Hyperion data center is causing real-world consequences. The tiny town of Holy Ridge, population 6,500, is now home to construction for the largest data center in the world. Residents report a 600% increase in vehicle crashes over the last year from the nonstop parade of construction trucks. Three truck crashes occurred just outside Holy Ridge Elementary School. Fourth graders told reporters the trucks are so loud and frequent that classroom walls shake. And if Meta decides to walk away after 4 years, the three new natural gas plants built to power the data center have a 30-year lifespan. If the data center closes early, the cost of those plants will appear on household utility bills for Louisiana ratepayers. So, Meta gets the infrastructure, keeps the debt off their books, retains the option to walk away, and if things go bad, ordinary people in Louisiana pay the bill. This is what financial desperation looks like at scale.

So, let's just connect the pieces because we've talked about a lot of things here. Meta spent $80 billion on the metaverse trying to create a market that didn't exist. People wanted VR for gaming, not for hanging out as legless avatars in virtual conference rooms. The product failed because it solved a problem literally nobody had. Then they pivoted to AI, but instead of building domain-specific models that leverage their actual competitive advantages, user data, ad targeting, content moderation, they're building general-purpose LLMs to compete with OpenAI. To train those models, they need massive amounts of data they don't actually have, so they pirated 267 terabytes from sites they internally describe as data sets we know to be pirated and bet on a fair use defense. Internal memos show Zuckerberg personally killed licensing discussions because paying for content would undermine that legal strategy. And simultaneously, they're surveilling their own employees' keystrokes to generate training data for AI agents that will replace those workers. American employees cannot opt out. European employees are exempt because GDPR protects them. To fund all this, they're taking on massive debt, $59 billion and growing, and using creative accounting structures to keep additional billions off the books. They're building data centers financed through entities named after pastries with lease terms that let them walk away while local communities bear the infrastructure risk.

And the core business that's supposed to fund all of this is declining. Daily active users across Meta's properties dropped for the first time ever last quarter from 3.58 billion to 3.56 billion. To compensate for this, Meta is cramming more ads onto platforms and charging advertisers more, moves that increase revenue per user by 27% in Q1 2026, but are actually likely to further alienate the users who are already leaving. This is what strategic incoherence looks like. Meta has technical execution capability. They can build very sophisticated systems, deploy infrastructure at scale, hire very talented engineers, but they've completely lost clarity on what problem they're solving and for whom. Facebook, Instagram, WhatsApp succeeded because they served existing needs. People wanted to stay connected, share photos, message internationally. These were market pull innovations. The metaverse and general-purpose AI are technology push. We build this, now we need to convince everyone they want it. But there is no clear user need being served.

And here's what I personally find remarkable. Meta is sitting on the most valuable advertising data set in human history. 3 billion users, decades of behavioral data, complete social graphs, deep understanding of what makes people click, buy, engage. It is the holy grail of advertising. If they were structurally coherent, they would be building AI-powered advertising tools that small businesses and creators can use to automatically generate optimized campaigns, identify target audience, manage the entire conversion funnel. But instead, they're just building ChatGPT, but kind of like worse, give it away for free, pirate all the training data, and just burn 115 billion dollars per year doing it. The desperation isn't just financial, it is strategic. They don't know what they're building or why. And the piracy, the employee surveillance, the debt, the accounting tricks, these are all downstream consequences of that fundamental confusion.

So, where does all of this go next? The New York Times opinion piece argues Meta is entering its zombie era, the slow decline into irrelevance like AOL or Yahoo. I think that's probably right about the trajectory, but it undersells how much damage a dying internet giant can do on the way down. Yahoo in its decline failed to invest in cybersecurity and suffered the largest data breach in history. 500 million accounts compromised with Russian hackers targeting dissidents and journalists. Meta's platforms are already riddled with fraud and scams. The company has been slashing staff in areas focusing on AI safety and identifying dangerous content. As the financial pressure increases, that's only going to get worse. We are likely to see more AI-generated deepfakes, more child exploitation material slipping through understaffed moderation, more toxic content optimized for engagement because the AI systems prioritize watch time over safety.

And Meta is still Meta, even after losing a bellwether lawsuit alleging their addictive design choices triggered anxiety and depression in teenagers with over 100,000 similar cases waiting in the wings seeking tens of billions in damages, the CFO recently bragged to Wall Street that the company's using AI to increase the amount of time users spend watching videos. The incentives haven't changed. Extract engagement, sell ads, maximize revenue. The fact that it's destroying the platform just means they will extract harder while they still can. On the copyright front, this lawsuit is going to be fascinating to watch, I think. Previous cases ruled that AI training constitutes fair use, but those were about the principle, basically whether training AI on copyrighted material is inherently transformative. This case is about deliberate circumvention, internal memos showing conscious choices to pirate rather than license, stripping copyright management information to conceal sources, torrenting from sites the company internally described as pirated. If Meta loses this, it doesn't just cost them money, it establishes precedent that choosing to infringe at scale is not protected by fair use, which has implications for the entire AI industry. And if Meta wins, then we've essentially ruled that AI companies can pirate anything as long as they call it training data and claim transformative use, which is on its own kind of alarming.

Either way, the lawsuit is a symptom of a deeper problem. AI development at the scale Meta is attempting is either exceptionally expensive or fundamentally incompatible with doing it legally if you're building general-purpose models. If you're building domain-specific AI using your own proprietary data, you're fine. Do that. Tesla can train on their fleet data, Meta could train on user behavioral data for ad targeting, but if you're trying to build models that know everything competing with ChatGPT and Claude, you need all of human knowledge, which means either paying for it or stealing it. And paying for it at the scale required might actually be economically impossible. So, Meta chose theft and is gambling on the legal system protecting them. Meanwhile, the debt keeps growing, the accounting gets even more creative, the core business keeps declining, and the strategic bets keep failing. Zuckerberg cannot be fired due to dual class shares, so there is no forcing function to correct course.

We're just going to watch this play out in real time. And here's the broader lesson, and this connects to what I talk about often on this channel. Everyone right now is throwing money at AI as a category, treating it like a universal solution to every single problem. But AI is not one thing. It's an incredibly broad field with thousands of specific applications and use cases. The companies that will succeed are the ones building domain-specific AI that solves actual problems for actual users. Tesla's AI works because it's trained on real-world driving data from millions of vehicles, solving the specific problem of autonomous navigation. The companies building AI for medical imaging, drug discovery, material science, climate modeling, these are focused applications where you can measure success very cleanly. But build next-gen AI to compete with OpenAI is not a strategy.

What does next-gen even mean? What problem does it solve? Who needs this? Meta keeps chasing the shiniest object, VR, then open-source LLMs, then general AI agents, without asking whether these things connect to their business or serve their users. And the result is strategic incoherence driving financial desperation driving illegal shortcuts. This is not a story about one company making bad bets. It's a story about what happens when you lose sight of the fundamental question every business should be able to answer, "What problem are we solving and for whom?" Meta used to know the answer. Facebook solved exactly a specific use case. So did Instagram, so did WhatsApp, but somewhere along the way they started chasing vision instead of solving problems. The metaverse was a vision, general purpose AI is a vision now. Visions are great for fundraising and headlines, but they don't become sustainable businesses unless they're grounded in real user needs.

And when you spend 80 billion dollars on legless avatars, 100 billion dollars on AI models nobody really can run, and another 115 billion dollars on systems that perform worse than competitors, all while pirating training data, surveilling employees, and using Enron-style accounting to hide debt, you are not really building the future. You're just burning money in increasingly desperate ways while the actual business withers.