Subprime AI Crisis: Rethinking Crypto x AI
TechFlow Selected TechFlow Selected
Subprime AI Crisis: Rethinking Crypto x AI
AI is closely tied to large corporations, which means that its prolonged inability to generate profits could trigger a chain reaction.
Navigating Web3 tides with focused insightsAuthor: Edward Zitron
Translation: Block Unicorn
If you're paying attention to AI in crypto or traditional tech, you need to seriously consider the future of this industry. This article is long—if you lack patience, feel free to leave now.
What I'm writing here isn't meant to spread skepticism or "attack" anyone, but rather a sober assessment of where we are and where our current trajectory might lead. I believe the AI boom—more precisely, the generative AI boom—is unsustainable and will eventually collapse. I also fear this collapse could devastate big tech companies, severely damage the startup ecosystem, and further erode public support for the tech industry.
I’m writing this today because things feel like they’re shifting rapidly. Multiple signs of an AI apocalypse have already emerged: OpenAI’s (rushed) release of its “o1” model (“Strawberry”) being called “a large, dumb magic trick”; rumors of future price hikes for OpenAI and others; layoffs at Scale AI; leadership departures from OpenAI. These are all indicators that things are starting to fall apart.
So I think it's necessary to explain the crisis unfolding and why we've entered a phase of disillusionment. I want to express my concern about the fragility of this movement, about how obsession and directionlessness brought us here, and I hope some people can do better.
Additionally—and perhaps something I haven’t emphasized enough before—I want to highlight the human cost if the AI bubble bursts. Whether it’s Microsoft and Google (and other major generative AI backers) slowing investment, or diverting internal resources to sustain OpenAI and Anthropic (and their own generative AI projects), I believe the outcome will be the same: thousands will lose jobs, and much of the tech industry will realize the only thing that grows forever is cancer.
This article won’t offer much comfort. I’ll paint a bleak picture—not just for big AI players but for the entire tech sector and its workforce—and explain why I think this chaotic, destructive end will come sooner than most expect.
Keep reading. Enter thinking mode.
How Does Generative AI Survive?
Currently, OpenAI—the ostensibly nonprofit organization likely transitioning to for-profit status—is raising a new round at a valuation of at least $150 billion, aiming to raise between $6.5 billion and $7 billion. The round is led by Josh Kushner’s Thrive Capital, with rumored participation from NVIDIA and Apple. As I’ve analyzed before, OpenAI must keep raising unprecedented amounts of capital just to survive.
Worse, Bloomberg reports OpenAI is also trying to secure $5 billion in debt via revolving credit lines—typically higher-interest borrowing.
The Information reported that OpenAI is negotiating with MGX—a UAE-backed fund with $100 billion in assets—for investments in AI and semiconductor firms, possibly also drawing funds from Abu Dhabi Investment Authority (ADIA). This is a massive red flag, because no one voluntarily turns to the UAE or Saudi Arabia for funding unless they desperately need money and can’t get it elsewhere.
Note: As CNBC pointed out, Mubadala—one of MGX’s founding partners—holds around $500 million in equity stakes in Anthropic, acquired from FTX’s bankruptcy estate. One can only imagine how thrilled Amazon and Google must feel about that conflict of interest!
As I discussed in late July, OpenAI needs at least $3 billion, more likely $10 billion, just to stay alive. It’s projected to lose $5 billion in 2024, a number expected to rise as increasingly complex models demand more compute and training data. Anthropic CEO Dario Amodei predicts future models could cost up to $100 billion to train.
Incidentally, that “$150 billion valuation” refers to how OpenAI prices shares for investors—even though “shares” here is fuzzy. In a typical company, investing $1.5 billion at a $150 billion valuation would get you 1% ownership. But with OpenAI, it’s far more complicated.
Earlier this year, OpenAI tried to raise funds at a $100 billion valuation, but some investors balked at the high price, partly due to growing concerns—per The Information’s Kate Clark and Natasha Mascarenhas—about inflated valuations for generative AI startups.
To close this round, OpenAI may convert to a for-profit entity. But the most confusing part is what investors actually receive. According to Kate Clark of The Information, investors were told they “won’t get traditional equity… Instead, they’ll receive profit-participation units—entitled to a share of profits once the company starts making them.”
It’s unclear whether switching to for-profit resolves this, given OpenAI’s odd “nonprofit + for-profit subsidiary” structure. Under its 2023 deal with Microsoft, the latter has rights to 75% of OpenAI’s profits—so even if the new structure includes equity, early investors still get Profit Participation Units (PPUs), not stock. As Jack Raines wrote in Sherwood: “If you own OpenAI PPUs but the company never profits, and you can’t sell them to someone who believes it eventually will, your PPU is worthless.”
Last weekend, Reuters reported any $150 billion valuation hinges on OpenAI restructuring and lifting investor profit caps currently capped at 100x initial investment. That cap was set in 2019, when OpenAI said excess profits would “go back to the nonprofit for humanity.” Recently, the rule changed to allow annual increases of 20% starting in 2025.
Given OpenAI’s existing profit-sharing with Microsoft—and its deep losses—any return is theoretical at best. Risking flippancy: zero multiplied by any amount is still zero.
Reuters added that converting to for-profit (raising valuation above recent $80 billion estimates) would force renegotiation with existing investors due to dilution.
Also reported: Financial Times noted investors must “sign an operating agreement stating ‘any investment in [OpenAI’s for-profit arm] should be viewed in the spirit of a donation,’ and that OpenAI “may never become profitable.” These terms are insane. Anyone losing money investing under such conditions deserves it—it’s an absurd investment.
In reality, investors gain no equity or control over OpenAI—just a claim on future profits from a company losing over $5 billion annually, likely to lose even more in 2025 (if it survives that long).
OpenAI’s models and products—whose utility we’ll discuss later—are operationally unprofitable. The Information reported OpenAI will pay Microsoft ~$4 billion in 2024 to run ChatGPT and underlying models, already discounted at $1.30/hour/GPU vs. standard rates of $3.40–$4. Without Microsoft’s partnership, OpenAI’s server costs alone could hit $6 billion annually—excluding employee expenses (~$1.5B/year). Training costs are ~$3B/year and almost certainly rising.
While The Information estimated OpenAI’s 2024 revenue at $3.5–4.5 billion in July, The New York Times last week reported annual revenue “now exceeds $2 billion,” suggesting year-end figures may land near the low end.
In short: OpenAI is burning cash—and will burn more. To keep doing so, it must raise funds from investors who signed agreements saying “we may never be profitable.”
As I’ve written, another issue with OpenAI is that generative AI (including GPT models and ChatGPT) fails to solve complex problems justifying its enormous cost. These models are probabilistic, leading to fundamental, unsolvable issues—they don’t “know” anything; they generate responses (text, images, translations, summaries) based on training data, which developers are rapidly exhausting.
The “hallucination” problem—where models confidently produce false information (or distorted visuals/videos)—cannot be fully resolved with current mathematical tools. While hallucinations can be reduced, their persistence makes generative AI unreliable for critical business applications.
Even if generative AI solves technical challenges, its business value remains unclear. Last week, The Information reported Microsoft 365 customers—including Word, Excel, PowerPoint, Outlook, and enterprise bundles tied to consulting services—show minimal adoption of its AI-powered “Copilot.” Only 0.1% to 1% of 4.4 million users (paying $30–$50/month) use these features. One company testing AI said, “Most people don’t find it very valuable right now.” Others noted “many enterprises haven’t seen breakthrough productivity gains”—and aren’t sure when they will.
How much does Microsoft charge for these marginal features? Astoundingly, $30/user/month, or up to $50/user/month for “Sales Assistant.” That’s effectively doubling customer costs—on top of existing annual contracts—for products that seem barely useful.
A caveat: Microsoft’s situation is so complex it may warrant its own dedicated coverage soon.
This is generative AI today: the leader in productivity and business software can’t find customers willing to pay for its product—partly because results are mediocre, partly because costs are too high to justify. Either Satya Nadella wants Microsoft to hit $500 billion in revenue by 2030 (as revealed in Activision Blizzard acquisition hearings), or costs are simply too high to lower prices—or both.
Yet nearly everyone insists AI’s future will blow our minds—the next generation of large language models is just around the corner and will be astonishing.
Last week, we got our first real glimpse of that supposed “future.” And the result was deeply disappointing.
A Dumb Magic Trick
On Thursday evening, OpenAI launched o1—codenamed “Strawberry”—with excitement rivaling a dentist appointment. Sam Altman described o1 across tweets as OpenAI’s “most capable and aligned model.” While admitting o1 “still has flaws, limitations, and performance degrades after extended use,” he promised better accuracy on tasks with clear correct answers—like programming, math, or science.
This admission alone is telling—but let’s unpack how it works. I’ll introduce new concepts but avoid excessive technical depth. For full details, see OpenAI’s blog post: *Learning to Reason with LLMs*.
Faced with a question, o1 breaks it down into steps—hoping each leads toward the correct answer. This is known as “Chain of Thought” (CoT). Think of o1 as two parts:
- One part applies reinforcement learning.
- The output-generating part rewards or penalizes itself based on correctness of reasoning steps, adjusting strategy upon penalty.
This differs from standard LLMs, which generate outputs directly without reflecting on intermediate steps. Here, the model evaluates and reinforces good reasoning paths before producing a final answer.
While this sounds revolutionary—another step toward Artificial General Intelligence (AGI)—the fact that OpenAI released o1 as a standalone product, not a GPT update, says otherwise. OpenAI’s demos focus on problems with pre-known answers (math/science), where correctness allows guiding the CoT process.
Noticeably absent: examples of o1 solving open-ended, unknown problems—mathematical or otherwise. OpenAI admits o1 suffers more hallucinations than GPT-4o and is less likely to admit ignorance.Why? Because even the “checking” component hallucinates—sometimes fabricating plausible-sounding answers.
According to OpenAI, o1 feels more persuasive to humans due to detailed outputs, increasing trust—even when answers are wrong.
If my critique seems harsh, consider how OpenAI markets o1. It frames reinforcement learning as “thinking” and “reasoning,” but really it’s guessing whether guesses are right, with outcomes often predetermined.
This insults actual thinkers—humans. Human cognition draws on personal experience, lifelong knowledge, neurochemistry. We may “guess” during problem-solving, but our guesses are grounded in facts, not clumsy math like o1.
And oh, the cost.
o1-preview charges $15 per million input tokens, $60 per million output tokens—triple the input cost and quadruple the output cost of GPT-4o. But there’s a hidden cost: data scientist Max Woolf points out that OpenAI’s “reasoning tokens”—intermediate outputs used to reach conclusions—are invisible in the API. So not only is o1 pricier, but its design forces users to pay repeatedly. All content generated while “contemplating” answers (note: the model isn’t actually “thinking”) gets billed—making complex tasks like coding extremely expensive.
Now, accuracy. On Hacker News—a Reddit-like site run by Y Combinator, founded by Sam Altman—users complained o1 invented non-existent libraries/functions during coding tasks and failed on questions not easily searchable online.
On Twitter, startup founder and ex-game developer Henrik Kniberg asked o1 to write a Python program multiplying two numbers and predict output. Though code was technically correct (though unnecessarily verbose), the actual output was completely wrong. Karthik Kannan, an AI founder, tested coding tasks: o1 hallucinated a nonexistent API command.
Sasha Yanshin tried playing chess with o1—the model conjured an extra piece on the board, then lost.
I cheekily asked o1 to list U.S. states with an “A” in their name. After 18 seconds, it listed 37—including Mississippi. Correct answer: 36.
When I asked for states with “W,” it included North Carolina and North Dakota after 11 seconds.
I asked how many times “R” appears in its codename “Strawberry.” It answered two.
OpenAI claims o1 performs at PhD-level on benchmarks in physics, chemistry, biology. Yet clearly, it struggles with geography, basic English, math, and programming.
Precisely what I predicted earlier: a “large, dumb magic trick.” OpenAI launched “Strawberry” to convince investors and the public that the AI revolution continues—delivering instead a clunky, dull, expensive model.
Worse, it’s hard to explain why anyone should care. Despite Altman boasting “reasoning capabilities,” wealthy backers see only 10–20 second delays, factual inaccuracies, and no exciting new features.
Nobody cares about “better” answers anymore—they want something truly new. I don’t think OpenAI knows how to deliver that. Altman anthropomorphizing o1 as “thinking” and “reasoning” clearly signals it as a step toward AGI, yet even staunch AI advocates struggle to feel excited.
In fact, I believe o1 reveals OpenAI’s desperation and lack of creativity.
Prices aren’t falling, software isn’t becoming more useful, and the much-hyped “next-gen” model since November has turned into a dud. Models also desperately need training data—so much so that nearly every large language model ingests copyrighted material. This urgency drove Runway—one of the largest video AI firms—to launch a “company-wide effort” collecting thousands of YouTube videos and pirated content. A federal lawsuit in August accused NVIDIA of similar practices using creators’ work to train its “Cosmos” AI.
The legal strategy relies on sheer willpower—hoping lawsuits fail to set precedents defining model training as copyright infringement. That’s exactly what a recent interdisciplinary study by the Copyright Initiative concluded.
Lawsuits are advancing: in August, a judge allowed plaintiffs to proceed with additional copyright claims against Stability AI and DeviantArt, and trademark/copyright claims against Midjourney. If any succeed, the impact on OpenAI and Anthropic would be catastrophic—and even worse for Google and Meta, which trained models on datasets containing millions of artists’ works. Since AI models cannot “forget” training data, they’d need retraining from scratch—costing billions and drastically reducing performance on tasks they already handle poorly.
I worry the industry rests on sandcastles. LLMs like ChatGPT, Claude, Gemini, and Llama are unsustainable, seemingly incapable of profitability. Their compute-intensive nature demands hundreds of millions or billions to train, requiring vast datasets—often stolen from millions of artists and writers—with hopes of evading legal consequences.
Even setting aside legal risks, generative AI hasn’t delivered revolutionary breakthroughs. The hype cycle doesn’t match the meaning of “artificial intelligence.” At best, generative AI occasionally generates accurate text, summarizes documents, or speeds up research slightly. Microsoft’s Copilot for 365 boasts “thousands of skills” and “infinite possibilities,” yet demoed functions include email drafting, prompt-based presentations, and querying Excel tables—useful, but far from revolutionary.
We’re not in the “early days.” Since November 2022, Big Tech has spent over $150 billion on infrastructure and AI startups, plus internal model development. OpenAI raised $13 billion, able to hire anyone it wants—same for Anthropic.
Yet this industry-wide “Marshall Plan” yielded four or five nearly identical LLMs, the world’s least profitable startups, and thousands of overpriced, underperforming integrated apps.
Generative AI is sold on multiple lies:
1. It’s artificial intelligence. 2. It will get better. 3. It will become real AI. 4. It’s unstoppable.
Forget vague terms like “performance”—usually referring to output “accuracy” or “speed,” not skill. Large language models have plateaued.“More powerful” rarely means “can do more”—it usually means “more expensive,” creating costlier tools with no added functionality.
If combined forces of every VC and tech giant still haven’t found a meaningful, widely paid-for use case, none will emerge. LLMs—yes, where those billions go—won’t suddenly become capable just because Big Tech pours another $150 billion in. No one is making them more efficient, or at least succeeding. If they did, they’d shout it from rooftops.
We face a collective delusion—a dead-end technology built on copyright theft (common in every era), requiring endless capital to sustain marginally useful services disguised as automation, costing billions and continuing indefinitely. Generative AI runs not on money (or cloud credits), but on confidence. The problem? Confidence—like investment capital—is finite.
I fear we’re heading toward an AI version of the subprime crisis—thousands of companies integrating generative AI into operations, priced unstably and far from profitability.
Almost every “AI-driven” startup relies on some combination of GPT or Claude. These models come from two deeply unprofitable companies (Anthropic expects $2.7B loss this year), pricing aimed at attracting customers, not profit. As noted, OpenAI depends on Microsoft’s support—cloud credits and discounted pricing—its entire pricing model reliant on Microsoft’s dual role as investor and service provider. Similar dynamics apply to Anthropic’s deals with Amazon and Google.
Given their losses, I estimate if OpenAI or Anthropic charged actual costs, API prices would increase 10x–100x (exact figures unavailable). Consider The Information’s report: OpenAI spends $4B/year on Microsoft servers—already half-price compared to other clients—plus loses over $5B annually.
OpenAI likely charges only a fraction of operational costs, surviving only by raising ever-larger VC rounds and maintaining Microsoft’s discounts—which Microsoft recently signaled may change, calling OpenAI a competitor. While uncertain, it’s reasonable to assume Anthropic receives similar favorable pricing from AWS and Google Cloud.
Suppose Microsoft gives OpenAI $10B in cloud credits, OpenAI spends $4B on servers, plus $2B (conservative estimate) on training—costs certain to rise with o1 and “Orion” models. By 2025, OpenAI may need more credits or start paying cash.
While Microsoft, Amazon, and Google may continue discounts, the question is whether these deals benefit them. As seen in Microsoft’s latest earnings call, investors increasingly worry about CapEx for generative AI infrastructure, doubting its profitability.
We don’t know generative AI’s true profitability for these giants, as they bury costs within broader revenues. If these ventures were profitable, they’d loudly advertise the income—but they don’t.
Markets are extremely skeptical of generative AI’s boom. NVIDIA CEO Jensen Huang offered no real ROI on AI investments, causing NVIDIA’s market cap to drop $279 billion in one day—the largest single-day decline in U.S. history, equivalent to nearly five Lehman Brothers at peak value. (The analogy ends there—NVIDIA faces no existential risk, and systemic fallout wouldn’t be comparable.) Still, it’s staggering—and shows AI’s distorting power on markets.
In early August, Microsoft, Amazon, and Google suffered market backlash due to AI-related CapEx. If next quarter they can’t show significant revenue growth from $150+ billion invested in new data centers and NVIDIA GPUs, pressure will intensify.
Remember: besides AI, Big Tech lacks other growth markets. When companies like Microsoft and Amazon show slowing growth, they rush to prove competitiveness. Google—a multi-faceted monopoly dependent on search and ads—needs flashy new things to attract investors. Yet these products deliver little utility, with much revenue coming from companies that “try” AI and find it not worth it.
Two scenarios now:
1. Big Tech realizes it’s trapped, fearing Wall Street backlash, and cuts AI-related CapEx.
2. Big Tech, desperate for growth, slashes costs to fund generative AI’s “death race,” laying off staff and diverting funds from other businesses.
Unclear which path unfolds. If Big Tech accepts generative AI isn’t the future, they have nothing else to show Wall Street—but may adopt Meta’s “Year of Efficiency” playbook: reduce spending (and lay off), promising “lower investment.” This is most likely for Amazon and Google, who—while eager to please Wall Street—still have profitable core monopolies.
But coming quarters must show real, substantial revenue growth from AI—not vague claims about “mature markets” or “annualized growth rates.” If CapEx rises, contribution must rise significantly.
I don’t believe this growth will materialize. Whether Q3/Q4 2024 or Q1 2025, Wall Street will punish Big Tech for AI greed. The punishment will be harsher than for NVIDIA—even though Huang’s empty slogans make NVIDIA the only company demonstrating AI-driven revenue.
I fear scenario two is more likely: these companies are convinced “AI is the future,” culturally detached from solving real-world software problems, potentially burning down the whole enterprise. I worry mass layoffs will fund this crusade, and past years suggest they won’t make the right choice—walking away from AI.
Big Tech is thoroughly poisoned by management consultants—Amazon, Microsoft, Google all run by MBAs—and monsters like Google’s Prabhakar Raghavan, who ousted engineers who built Google Search to seize control himself.
These people don’t face human problems. They foster cultures obsessed with fictional issues software supposedly fixes. For those whose lives consist of meetings and emails, generative AI seems magical. Satya Nadella’s success mindset is largely “let technologists solve problems.” Sundar Pichai could’ve ended the generative AI frenzy by mocking Microsoft’s OpenAI bet—but didn’t, because these leaders lack real ideas, and companies aren’t run by people who’ve faced real problems, let alone solved them.
They’re also desperate. Nothing has been this dire since Meta burned billions on the metaverse.But this is worse and uglier—they’ve poured immense resources into AI, embedding it deeply. Pulling back would be embarrassing, hurt stock prices, and implicitly admit it was all wasted.
If media held them accountable, this could’ve stopped earlier. The narrative sells via the same hype-cycle scams—media assumes these companies will “solve problems,” despite obvious failure. Think I’m pessimistic? Then tell me: what’s generative AI’s next move? What comes after? If your answer is “they’ll solve problems” or “they have amazing stuff behind the scenes,”you’re unwittingly participating in a marketing operation (ponder that).
Author Aside: We must stop being fooled. When Mark Zuckerberg claimed we were entering the metaverse, major outlets—The New York Times, The Verge, CBS News, CNN—promoted an obviously flawed concept, visually ugly and built on outright lies about the future. It was clearly just bad VR, yet The Wall Street Journal six months after the hype faded still called it “a vision for the future of the internet.” Same happened with crypto, Web3, NFTs! The Verge, NYT, CNN, CBS—again promoting clearly useless tech. Special mention to The Verge—really Casey Newton—who, despite solid reputation, in July still claimed “having the strongest LLM might give a company a foundation for various profitable products,” while the tech loses money and offers no truly useful, lasting products.
I believe Microsoft will start cutting costs elsewhere to sustain the AI boom. Earlier this year, a source shared an email where Microsoft’s senior leadership requested (though plan was shelved) reducing power usage across divisions to free up GPU capacity—including shifting computing for other services overseas to prioritize AI workloads.
On Blind—an anonymous network requiring corporate email verification—a Microsoft employee in mid-December 2023 complained “AI is taking their money,” saying “AI costs are too high, eating into raises, and it won’t get better.” Another in mid-July expressed anxiety, sensing Microsoft’s “marginal addiction to funding NVIDIA’s stock price with operating cash flow,” harming company culture.
Another added they believe “Copilot will destroy Microsoft in FY2025,” with “Copilot focus dramatically decreasing in FY2025,” revealing “major Copilot deals in their country, after nearly a year of PoC, layoffs, and adjustments, saw less than 20% usage,” warning “the company took too much risk,” and that Microsoft’s “massive AI investment won’t pay off.”
Blind is anonymous, but hard to ignore widespread posts describing cultural decay at Microsoft Redmond—senior leaders disconnected from real work, funding only projects labeled “AI.” Many posts express disappointment with CEO Satya Nadella’s “nonsensical jargon,” complaining about lack of bonuses and promotions in an org chasing a possibly nonexistent AI trend.
At minimum, deep cultural sadness exists internally—posts echoing “I hate working here,” confusion over why so much is spent on AI, yet resignation because Satya Nadella clearly doesn’t care.
The Information highlighted a worrying sign: Microsoft hides poor adoption of Office Copilot. It reserved server capacity in data centers sufficient for millions of daily users. Actual usage remains unclear.
Estimates suggest current Office Copilot users range from 400K to 4M—meaning Microsoft may have built vast idle infrastructure.
Some argue Microsoft is betting on future growth, but consider this: what if that growth never comes? What if—however crazy it sounds—Microsoft, Google, and Amazon built massive data centers anticipating demand that never arrives? Back in March,I argued I couldn’t find any company achieving significant revenue growth from generative AI. Nearly six months later, the problem persists. Companies are mostly attaching AI features to existing products, hoping to boost sales—but nowhere does this strategy show success. Like Microsoft, their “AI upgrades” bring little real business value.
Hence the bigger question: Are these AI investments sustainable? Did tech giants overestimate demand for AI tools?
While some companies’ “AI integration” may drive partial spending on Azure, AWS, or Google Cloud, I suspect much demand stems from investor sentiment. “Investing in AI” serves to placate markets, not stem from cost-benefit analysis or actual utility.
Yet these companies have spent immense time and money embedding generative AI into products. Now they face possible outcomes:
1. They launch AI features, but customers refuse to pay—as Microsoft experienced with 365 Copilot. If they can’t monetize during the AI hype, when executives demand employees “keep up with AI,” the situation worsens once the trend fades.
2. They launch AI features but can’t charge extra, forcing them to bundle AI into existing products without boosting margins. Eventually, AI becomes a “parasite,” eroding revenue.
Goldman Sachs’ Jim Covello noted in a generative AI report: if AI’s benefit is merely efficiency (e.g., faster document analysis), competitors can replicate it. Almost all generative AI integrations are similar: collaboration assistants answering internal/external queries (Salesforce, Microsoft, Box); content creation (Box, IBM); code generation (Cognizant, GitHub Copilot); and upcoming “intelligent agents”—basically customizable chatbots connecting to other website functions.
This exposes generative AI’s biggest challenge: while “powerful” in generating content from existing data, it lacks true “intelligence.” Hence why many AI landing pages overflow with buzzwords—their main selling point being “uh… figure it out yourself!”
I fear a domino effect. Many enterprises are “trying” AI now. Once trials end (Gartner predicts 30% of generative AI projects abandoned post-PoC by end of 2025), they’ll likely stop paying for extras or integrating generative AI.
If so, hyperscalers providing cloud infrastructure for generative AI apps—and LLM suppliers like OpenAI and Anthropic—will see already weak revenues shrink further. This pressures pricing, worsening already negative margins. Then OpenAI and Anthropic will almost certainly raise prices, if they haven’t already.
Though Big Tech can fund the boom—after all, they created it—this won’t help small startups accustomed to discounts, leaving them unable to survive. While cheaper alternatives exist—like independent providers running Meta’s LLaMA models—it’s hard to believe they won’t face the same profitability issues as hyperscalers.
Also note: hyperscalers deeply fear angering Wall Street. They could theoretically (as I fear) improve margins via layoffs and cost-cutting, but these are short-term fixes—only viable if they can squeeze some money from this barren generative AI tree.
Either way, it’s time to accept a truth: the money isn’t here. We must pause and recognize we’re in tech’s third age of illusion.Unlike crypto and metaverse, this time everyone joined the money-burning party chasing an unsustainable, unreliable, unprofitable, environmentally harmful project—packaged as “AI,” sold as “automating everything,” yet never having a real path to deliver.
Why does this keep happening? Why crypto, metaverse, now generative AI—technologies seemingly not designed for regular people?
It’s the natural evolution of an industry now solely focused on extracting more value per customer, not delivering more value to customers. Or worse—they don’t even understand who their customers are or what they need.
Products sold today almost certainly try to lock you into an ecosystem—consumers dominated by Microsoft, Apple, Amazon, Google. Exiting becomes increasingly costly.Even crypto—ostensibly “decentralized”—quickly abandoned libertarian ideals, concentrating users through a few major platforms (Coinbase, OpenSea, Blur, Uniswap), backed by familiar VCs (e.g., Andreessen Horowitz). Crypto failed to become a new, independent online economy, instead expanding through networks and capital from previous internet waves.
As for the metaverse, a scam—but Mark Zuckerberg’s attempt to dominate the next internet, positioning “Horizon” as the main platform. We’ll revisit generative AI shortly.
All this ties to further monetization—increasing average value per customer, whether through more platform usage (more ads), pushing “semi-useful” new features, or creating new monopolies/oligopolies accessible only to well-funded tech giants, offering little real utility or value to customers.
Generative AI excites (some) because Big Tech sees it as the next major moneymaker—adding revenue streams to every product from consumer tech to enterprise services. Most generative compute flows through OpenAI or Anthropic, cycling back to Microsoft, Amazon, or Google as cloud revenue, sustaining growth narratives. The biggest innovation isn’t what generative AI does—but creating a dependency loop controlled by a few hyperscalers.
Generative AI may not be very useful, but it’s easy to integrate, letting companies charge for “new features.” Whether consumer apps or enterprise software, these products earn millions or billions by upselling to as many customers as possible.
Sam Altman is smart—he realized the tech industry needed a “new thing”—a technology everyone could join and sell. While he may not grasp the tech, he understands the economy’s hunger for growth, packaging transformer-based generative AI as an easy-to-integrate “magic tool” adding novelty to most products.
But the rush to embed generative AI everywhere reveals a massive disconnect between these companies and real consumer needs or sound business operations. For 20 years, “doing new things” worked—launch features, push sales teams to sell them, sustain growth. This trapped tech leaders in a harmful, unprofitable business model.
Executives running these firms—mostly MBAs and consultants who never built products or tech companies from scratch—either don’t understand or don’t care that generative AI lacks a path to profit. Maybe they think it’ll naturally become profitable like AWS (which took nine years), though these are entirely different beasts. Past things “worked out somehow,” so why not now?
Except interest rate hikes drastically changed venture capital, reducing VC reserves and shrinking fund sizes. Plus, attitudes toward tech have never been more negative. Add numerous other factors—why 2024 differs from 2014—too many to cover in this 8,000-word piece.
Truly alarming: beyond AI, these companies seem to have no other new products. What else do they have? What drives future growth? Any alternatives?
No. That’s the problem.Because once AI fails, the fallout will inevitably ripple across the entire tech sector.
Every major tech player—consumer or enterprise—sells some AI product, integrates LLMs or proprietary models, usually running on Big Tech’s cloud systems. In essence, they all depend on Big Tech subsidizing the entire industry.
I suspect a subprime-style AI crisis looms, where nearly the entire tech sector participates in a technology sold at absurdly low prices, highly centralized and subsidized by Big Tech. Eventually, generative AI’s shocking, wasteful burn rate catches up—forcing price hikes or new products with extreme fees (like Salesforce’s “Agentforce” charging $2/conversation), making even well-funded enterprise clients unable to justify the cost.
What happens when the entire tech industry relies on software that only loses money and offers little real value? What happens when pressure mounts, AI products become untenable, and these companies have nothing else to sell?
I don’t know. But the tech industry is marching toward a terrible reckoning—a lack of creativity bred by an economic environment rewarding growth over innovation, monopoly over loyalty, management over actual creation.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@edzitron
