If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large (updated to version 3.0 a few weeks ago). My Hater's Guides To the SaaSpocalypse, Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater's Guide To Oracle.

This month, I published a two part series that took a deep-dive into the bubbles-within-a-bubble that make up the AI bubble — from the unsustainable and reckless growth of semiconductor companies, to the cults of personality surrounding Sam Altman and Dario Amodei. On Friday, I’ll publish my long-awaited Hater’s Guide to Softbank. You won’t want to miss it. 

Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week. 

Soundtrack — Queens of the Stone Age - Hideaway (Baloise Orchestral Arrangement)

On Sunday, the Bank of International Settlements (BIS) put out its annual report and said, well, a bunch of things that I’ve been saying:

In the near term, the ongoing AI investment boom raises questions about the sustainability of the current economic expansion. The five largest hyperscalers are set to spend over a trillion US dollars on AI-related capital expenditure from 2025 through 2026. These commitments are outpacing earnings and the free cash flow of these firms, leading some to issue debt to raise additional financing.

As edifying as it is to see the bank for central banks say exactly what I’ve been saying for the last few years, this part is the one that both rocks as far as being right goes and sucks for the world at large:

Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust, with potential knock-on effects on financial conditions…should hyperscalers slow or halt the aggressive pace of capex deployment, many borrowers across the supply chain could struggle to replace lost revenue and service their debt.

No shit. In April of last year, I wrote a piece called “AI is a systemic risk to the tech industry,” where I outlined how the failure of one model lab, OpenAI, would have seismic effects down its supply chain, delivering body blow after body blow to NVIDIA, Oracle, Microsoft, and the various Neoclouds that serve its compute, the most notable of which being CoreWeave. 

Since then, OpenAI’s slimy tendrils have sunk into even more facets of the tech industry, and it has signed deals with the likes of Google, Amazon, Cerebras, and Broadcom, while also taking on more investments, including mammoth commitments from Softbank, which is only able to meet them by selling off prized stock in companies like ARM and NVIDIA, and by raising debt. 

The idea of systemic risk has never quite left my work, and I’ve spent a lot of time thinking about it over the past year — and, as a result, my writing has examined the potential consequences of an AI spending pullback on those financing the sector, in particular private credit, as well as the semiconductor industry. 

The BIS’s concern wasn’t about revenues tanking — which would happen should, as it fears, hyperscalers decide to “slow or halt the aggressive pace of capex development” — but rather revenues tanking and the borrowers within the AI supply chain being unable to service their growing debt burdens. 

Again, this is something I’ve raised the alarm bells over a bunch of times. CoreWeave has been a favored popinjay of this newsletter, and in March of 2025, I published CoreWeave Is A Time Bomb, where I focused heavily on the company’s overwhelmingly toxic debt pile and its reliance on OpenAI as a customer. 

On a much grander scale, we have Oracle — which I exhaustively profiled in my Hater’s Guide to Oracle newsletter. 

Unlike neoclouds like CoreWeave, Oracle’s a much older company, having spent most of its existence selling database and ERP software to some of the world’s largest companies and public sector institutions. Oracle pivoted to serving AI compute at a time when its core business lines had started to stagnate, and thanks to its large scale, it was able to raise insane amounts of debt.

And Oracle, as I’ve noted previously, is a company that, even before the AI bubble, was massively indebted. It just so happens that, as a result of its tryst with OpenAI, Larry Ellison saw fit to twist the debt knob to eleven. 

Oracle’s spending has already pushed its free cash flow into negative territory — minus $23.7bn, as of the end of FY 2026 — and at the end of May, it had $129.5bn in outstanding debt. This doesn’t include its various lease commitments, which add up to nearly $38bn, nor the additional $260bn in lease commitments that have been signed, but haven’t actually started yet. 

All of this is to say that Oracle has massively leveraged itself for the benefit of one company, OpenAI, and if that company can’t pay its bills, it’s fucked. Oracle’s existence — and Larry Ellison’s personal wealth — hinges on whether OpenAI can make good on its promise to spend $300bn in compute. 

This is both the most-obvious and under-discussed part of the AI bubble — that the trillion-plus dollars of hyperscaler capex is feeding a massive semiconductor boom based on, at best, the very small likelihood that large language models will turn into something completely different. 

If Microsoft, Google, Amazon and Meta decide that it’s time to stop spending $30 billion or more a quarter on GPUs, RAM, storage, and data center construction, that’ll tear a hole in the side of what people assume is a permanent supercycle.

I need to state how fucking silly it is that anybody considered said semiconductor boom anything other than a brief chance to fill their boots before a global equity catastrophe so severe that the Futurum Group will be on suicide watch.

Hyperscalers — who will see their capex outpace their cashflows as of Q3 2026 — have had such poor returns on their investment in AI that none of them will actually disclose their revenues outside of vague “run rates,” which means that all of this investment is effectively based on the idea that something completely different will happen in the future. 

Said future will have to make them at least $2 trillion in brand new revenue by 2030, because if it doesn’t, effectively all of that capex will have been spent to prop up Anthropic, OpenAI, and whatever it is that Meta is doing with its chatbots. 

There is no cogent or rational argument in favor of continued capital expenditures, at least not one without a tacit acceptance that much of the current spend has been a waste outside of pumping equities and incubating two different large, unprofitable AI labs. Those millions of H100 and B200 and B300 GPUs are not going to usher in a digital God, they are not going to create recursive self-improvement, they are not going to be the fulcrum to adding $600 billion or more in brand new revenue to current services, and the only revenue they’re generating is compute spend from Anthropic and OpenAI, which I estimate makes up 20% or more of cloud revenues for Google, Amazon, and Microsoft. 

I must also be clear that the cost of these companies extends far beyond equity investment. While Microsoft invested $13 billion in funding OpenAI, Microsoft executive Michael Wetter revealed as part of the Musk vs Altman trial that the partnership has cost it more than $100 billion, suggesting infrastructure costs of at least $87 billion just for OpenAI. I imagine Amazon and Google have had to spend similar amounts to handle Anthropic’s similarly-rapacious compute demands, especially given the $11 billion-and-counting cost of Amazon’s Anthropic dedicated Project Rainier data center.

This is a criminally-underdiscussed part of the AI bubble. Anthropic and OpenAI have raised a little under $300 billion combined since 2019, but I estimate their true cost is at least $500 billion given hyperscaler capex investments that were necessary for them to exist, and that’s before you consider the $340 billion or more that Oracle is spending to build out the 7.1GW of “Stargate” data centers for OpenAI. These are not startups, but subsidiaries of big tech that only exist as separate arms as a means of pumping equity positions and hiding the truth: that AI capex has been a complete waste of money, even when you include two bulbous failsons that lose tens of billions of dollars a year.

As I reported two weeks ago, OpenAI spent $17.2 billion on Microsoft Azure in 2025, a year when it lost $20.9 billion on $13.04 billion in revenue. Even if that were profit (which it is not), that’s $4.2 billion less than the capital expenditures that Microsoft spent in the first quarter of 2025.

Outside of OpenAI, Microsoft may as well not have an AI business. While it boasted back in April about having a $37 billion AI revenue run rate (meaning a non-specific month multiplied by 12), that only works out to about $3.08 billion a month, or less than a tenth of the $31.9 billion that it spent on capital expenditures in the quarter. To make matters worse, Microsoft revealed that number was “up 12% year-over-year,” suggesting that its AI revenue run rate in Q3FY25 was $16.59 billion, or around $1.38 billion a month.  

Yet my own reporting on OpenAI’s inference spend from last November showed that it spent $2.947 billion in Q3FY25, representing about $11.7 billion on an annualized basis, meaning that, at least in that quarter, OpenAI likely represented around 70% of Microsoft’s AI revenue, and I’d be surprised if that dramatically changed in the year that followed, given that OpenAI’s inference spend was $3.648 billion in Q1FY26.

All of this is to say that the only real outcome from all of this capex spend appears to be propping up Anthropic and OpenAI, two deeply-unprofitable companies, and then receiving a small fraction of it back in the form of revenue that is only made possible through hundreds of billions of dollars of venture capital subsidies. 

Now OpenAI and Anthropic represent 50% or more of hyperscaler remaining performance obligations, or around $748 billion.

There is simply no logical or rational reason to invest any further capex in AI, outside of the mistaken belief that OpenAI or Anthropic could actually afford to pay without Google, Amazon, or Microsoft handing it to them. Hyperscalers do not have meaningful AI revenues of any kind outside of their own pseudo-startup investments, and it is equal parts ludicrous and irrational that A) they are continuing to invest and B) that the markets, analysts and journalists are acting as if everything is fine.

Sidenote: I haven’t discussed Meta, because Meta does not have an AI story. Mark Zuckerberg has wasted every ounce of its capex, outside of whatever it could get by reselling its capacity to somebody else — but don’t worry, he thinks (that’s a quote!) that Meta has a use for the compute! No, sorry, those GPUs are not driving meaningful increases in ad revenue, I already covered that in the past.

Record sales across NVIDIA, Micron, Sandisk, SK Hynix, and Samsung are a direct result of an entirely speculative asset bubble, driven by the reckless and directionless capital expenditures of some of the largest and richest companies in the world. 

Anyone investing in data centers is building speculative capacity for demand that does not exist outside of Anthropic and OpenAI. If said demand existed, AI data center neocloud company CoreWeave would have a healthy and diverse revenue stream, rather than 65% of its revenues coming from Microsoft (for OpenAI) and NVIDIA, and the rest coming from Google (for OpenAI), Anthropic, Meta, and, of course, OpenAI. There are simply no other massive consumers of AI compute, and the only reason we haven’t hit that harsh reality is that data centers take 18-34 months to finish. 

Even if there was, I can find little evidence of anyone but OpenAI, Anthropic and hyperscalers having the demand or funds necessary to substantiate the data center buildout. 

I really need to hammer this point home.

If we assume that NVIDIA CEO Jensen Huang’s prediction of $1 trillion in Blackwell and Vera Rubin sales comes true, that would be around 40GW of data center capacity with around 30GW of IT load, and if we assume that data centers get about $12 per-megawatt of revenue, that works out to about $435 billion in annual compute demand by, being generous, 2030.

Let’s be abundantly clear about something: the only companies that can afford to spend money on compute right now are either hyperscalers or the companies that hyperscalers subsidize. Even then, outside of OpenAI’s $50 billion in 2026 compute spend and what I estimate will be a similar amount from Anthropic, there doesn’t appear to be more than a few billion dollars of demand, and if there were, CoreWeave, IREN, Nebius, Cipher Mining, and other neoclouds would have hundreds of billions of dollars’ worth of remaining performance obligations rather than RPOs that expand only with hyperscaler backstops or the depths of Meta’s Zuckerbergian AI psychosis.

Let me put it even simpler: those hundreds of billions of data centers are being built for no-one, and the only companies that can “afford” to pay for even a fraction of the compute are unprofitable AI companies propped up by hyperscalers. 

While this might read as a radical position, I think it’s far more radical to look at the current state of affairs and say “fuck it, I think hyperscalers should spend a trillion dollars next year.”

There is no rational justification for doing so out of fantastical thoughts driven by a deranged market desperate to avoid thinking about how tech doesn’t have any hypergrowth ideas left. 

The current capital expenditures have, outside of the creation of OpenAI and Anthropic, been a near-complete waste. Microsoft 365 Copilot sucks. GitHub Copilot sucks. Google AI Overviews suck. Google Gemini is an also-ran LLM and thus, as a result, sucks. Meta’s LLMs are horrifyingly dangerous. Amazon Rufus sucks, and Amazon should be investigated by the SEC for suggesting it drove $10 billion in “annualized revenue” in Q3 2025, because it most assuredly did not. Alexa+ sucks. It all sucks, and it would suck just as badly if big tech had spent a quarter of the capex. 

These products are near-universally loathed, barely generate any revenue, and even in the case of the modestly-successful GitHub Copilot (around $1.08 billion in annualized revenue as of end of last year), it was only because users’ compute was heavily-subsidized, leading Microsoft to move users to token-based billing, outraging customers who were used to paying $39 a month to burn thousands of dollars of tokens.

Yes, All This Money Can Be Wrong

Sundar Pichai, Andy Jassy, Satya Nadella, and Mark Zuckerberg are losers. They may have billions of dollars, they may run giant tech companies, but they are losers selling a doomed technology based on unreliable, inefficient and overly-expensive technology ill-suited for the kinds of reliable, deterministic, “set it and forget it” tropes that people actually associate with AI.  

The Four Losers are the only reason that anyone has taken these Large Loser Models seriously, which is a sign that the tech industry and our economy are also piloted by losers. Every bit of “progress” that we’ve seen from LLMs has come from aggressively cramming a square peg into a round hole — billions of dollars of training costs, hundreds of billions of dollars of capex, endless harnesses and scripts and wrappers and layers to try and eek out anything approaching the supposed promise of autonomy. 

All the king’s horses and all the king’s men have sunk every dollar and ounce of brain matter into trying to make LLMs into something they’re not, and we, as a society, are expected to coddle these things and act like they’re exceptional, and give them credit for things that have yet to take place. I refuse to buy into the premise that LLMs’ ability to generate code or replicate open source software is proof that these things will become a powerful, autonomous tool in the future, and I think those that extrapolate to that point are either intellectually bankrupt, deeply cynical or so easily-fooled that they click every single email claiming their Paypal account has been compromised.

I assure you, all this money can be wrong! Hyperscalers can, in fact, spend a trillion dollars on something that doesn’t do what they say, because these companies are more than happy to mislead you, and, to quote Nik Suresh:

A huge amount of the economy is driven by people who are, simply put, highly suggestible. That is to say that it is very, very easy to get them excited and willing to spend money.

Why did everybody invest in data centers? Because the hyperscalers did so! Why are Micron and RAM companies selling so much RAM? Because A) GPUs use a ton of high-bandwidth RAM, B) said HBRAM consumes three times as much wafer space as normal DRAM, leaving less space for other kinds of cheaper, lower-margin RAM, and C) because the servers for said AI GPUs are, too, full of RAM! 

Those data centers aren’t being built because the creditors have any “insight” into the massive amounts of AI compute that generative AI tools need, and will need. They see the “success” of ChatGPT and Claude (two heavily-subsidized products) and think that because Anthropic and OpenAI need lots of compute, everybody will need lots of compute. And because banks and private credit crave ways to invest their money and everybody is so excited, it’s super easy to get them excited about the prospect of building something big, sexy and costly!

It doesn’t help that a lot of the information out there is deeply, deeply flawed.

Exponential View Should Be Ashamed Of Itself 

In Brief:

• Exponential View’s research rigs the dice using obfuscated, proprietary data.
• Anthropic and OpenAI represent at least 68% of the supposed $110 billion in AI revenue from the last 12 months. While the report claims to ‘deduplicate’ revenue across the AI stack, it does not provide any source data of any kind, making it impossible to verify.
• The report uses “annualized run rate” to try and make the AI industry’s revenue seem larger than it is.
• This report is industry marketing framed as research, but uses deliberately positive framing and questionable data sources.

Last week, research firm Exponential View put out a questionable report claiming that AI had $110 billion in trailing 12-month revenues (between what looks like June 2025 and mid-June 2026), and did so by smashing together all AI revenues, including both OpenAI and Anthropic’s customer spend and compute spend, While the report claimed to “deduplicate” the numbers somehow, Exponential View declined to explain how it had done so. It’s also deeply deceptive to include both revenues and compute spend to try and represent the material health of the AI industry.

This is because the AI industry is full of losers that cannot win without fiddling with the numbers, and because everybody is so excited, they’re ready to be fooled, and hesitant to dig an inch deeper. 

Not me! I don’t give a shit, and I hate the feeling of being lied to, so I dug in.

That’s because OpenAI and Anthropic represent as much as 75% of that revenue between their compute spend and revenues. Per The Information’s and my own reporting, OpenAI had around $8.77 billion in revenue and spent about $17.48 billion on compute in 2025, and per The Information had $5.7 billion in revenue and spent $17.8 billion on compute in the first quarter of 2026, for a total of around $44 billion (40% of Exponential View’s total), which doesn’t include any of OpenAI’s compute spend or revenue for the months of April, May or June, which likely inflates the total further.

While Anthropic is a little more-difficult to parse thanks to the Wall Street Journal’s unwillingness to make a readable chart, it had $4.8 billion in revenue in Q1 2026, and spent what I think is at least four billion on inference, and though its training costs are unreported, I think it’s reasonable to assume they’re at least $5 billion, for a total of $14.6 billion. If we, based on The Information’s reporting, take half (being generous, as most of this was weighed toward the end of the year) of Anthropic’s (all numbers are projections) $4.5 billion in 2025 revenue, $2.7 in inference costs and (I seriously question this number) $4.1 billion in training spend, we get $5.65 billion, for a total of $20.25 billion of contributions to Exponential View’s analysis, or around 18.4% of that $110 billion total.

So, yeah, not including anything from Q2 2026, Anthropic and OpenAI represent 68% of the $110 billion of AI revenue that Exponential View is trying to get people excited about. 

These are the actions of a loser propping up an industry of losers that cannot win by telling you the truth. This report exists entirely to fool the already-fooled and support an existing narrative, which is why Bloomberg covered it in the most obtuse, industry-servile way possible:

Revenue from artificial intelligence has reached a tipping point, showing that the hundreds of billions of dollars tech companies are spending on it may be economically sustainable, according to a report from research firm Exponential View.

Global AI sales, excluding China, reached $25 billion in the first quarter of 2026, exceeding the industry's estimated $21 billion in depreciation costs tied to investments in data centers and chips for the second consecutive quarter. While the milestone suggests that AI companies are beginning to cover the cost of their capital spending, the margins are thin. Depreciation charges still consume more than two thirds of revenue, leaving a small buffer to cover other costs such as power, labor and financing.

Here’s two reasons this is fucking silly!

1. You are comparing costs of the entire industry against depreciation costs of the few companies that actually buy AI GPUs.
2. In Q1 2026, Amazon had $18.94 billion, Microsoft $10.1 billion and Google $4.4 billion in depreciation. That’s $33.44 billion! That’s more than $25 billion! And I haven’t even included Meta, but don’t worry, as I’ll get to, neither does Exponential View!

Now, you may be wondering how they got that $25 billion number, and that’s because Exponential View gave it to them! 

The next question we wanted to track is whether AI revenues can cover the capital investment that’s required to build the infrastructure. Our model separates AI-oriented CapEx from ordinary CapEx across the major hyperscalers and neoclouds, the specialist AI cloud providers. This adjustment is important because hyperscalers were already spending around $120 billion annually on CapEx before ChatGPT.

We capture the additional investment in AI infrastructure, then depreciate compute assets over 6 years and other infrastructure over 14 years. Our modeling shows that revenues attributable to hyperscalers just about clear the depreciation expense.

Yeah, but now they’re spending $765 billion on capex. Anyway, as I mentioned above, Exponential View’s Magical Maths magically brings those capex charges down to $25 billion, and entirely removes Meta because "initiatives are focused on ad uplift, so not recognized as pure GenAI revenue, or currently have minimal direct monetization.” What a loser move! Meta has oriented its entire company around AI!

I refuse to waste too much more time on this piece, but I need you to see how deceptively it’s framed this supposed “good news” for the AI industry, comparing its own proprietary depreciation formula against its own proprietary AI revenue formula to get a chart that is built to make the AI industry look good. No need for sourcing! No need for data! Just put the hype in the bag and invest in AI stocks! 

I also find it despicable that Exponential View resorted to this weird, confusing “cumulative” AI revenues versus CapEx depreciation chart. The vast majority of this revenue is OpenAI and Anthropic’s compute spend, and I dunno, if you’re trying to do a report that gives the real state of the AI industry, maybe try and represent that anywhere in the report!

These are, as I’ve suggested, the acts of losers propping up other losers. In the event that this industry had a fundamentally-sound revenue story, it would be extremely easy to show profits versus losses, track revenue in a transparent way, and produce a report that showed AI’s remarkable ascent.

Instead, Exponential View says that AI is “real, big & fast” through a Pee Wee’s Playhouse of undefined models, datasets and alleged “quality grades” that helps feed a dangerous bubble further, and likely cons retail investors into further terrible decisions. 

The Large Loser Model

I know it sounds a little mean to call people losers, but what do I call an industry that sells itself on lies and deception? What do I call people that intentionally mislead people about the economics and outcomes of generative AI? If AI is so incredibly successful and impossibly brilliant, why does every explanation sound like it was written by The Riddler or somebody about to chug Jonestown Kool-Aid?

Because they’re losers that can’t win by actually winning. Their best (and only) hope is to overwhelm you with a 24/7 marketing campaign (powered by the media) that makes all of this seem inevitable, impossible-to-stop, and a rip-roaring success, even as every company loses money and every product rings with a soulless mediocrity.

That’s because LLMs are, while an interesting tool in a vacuum, currently being marketed by losers to losers using a mixture of Doom Trolling, insane extrapolations, and outright lies, manipulating people’s assumption that tech always gets better and that this much money can’t be wrong to create a marketing campaign fueled by deception. While using them doesn’t automatically make you a loser, you become one the very second you aggressively push somebody into doing so, as you have become the acolyte of the Loser Mafia.

I have never heard anyone that’s an AI booster advocate for a technology with any level of excitement in their life, because they’re excited about how these tools make them feel and what they represent far more than anything else. They’re also tools intentionally built to produce engagement, and to make you feel you’re productive, even if you’re not.

Just listen to this guy in this Bloomberg story about AI making people “productive, anxious and afraid to log off”:

Matt Van Horn, a serial entrepreneur and father of four, never turns his laptop off anymore. He has more than a half-dozen artificial intelligence agents running at all times in Anthropic's Claude Code.

Every 10 minutes or so, they ask him what to do next.

He keeps his laptop running at his kids' soccer practice, while dropping them off at school and in the hotel during vacations.

When he goes to sleep, one agent steps in to babysit the others.

Van Horn is one of many founders whose work has been transformed by Al. As he builds his latest company, he's used Al agents to help contribute to hundreds of projects on GitHub. But he and many other Al evangelists are also working longer hours than ever before as they grapple with anxieties about how Al might advance without them if they log off.

I’m sorry man, you have an addiction, and I worry it’s ruining your life. What is this producing? What are you actually doing with this time? Because if you’re allegedly 100 times more productive, wouldn’t that, y’know, produce something fairly incredible? I have no idea — and don’t want to put this man on blast — how significant his commitments on GitHub may or may not be, but the return on investment of “obsessively checking your laptop at all times in case you might not be productive” should be something on the order of curing a disease.

The story continues:

After 15 minutes of conversation with a Bloomberg reporter, he notes that most of his agents are probably waiting for his next prompt. "I don't have a therapist, but if I did, they'd be like, 'It's OK, Matt," he says with a laugh. "They said that agents were supposed to do our work for us, but I've never worked harder in my life. I just have 100 times the output that I had before."

This man is a victim of a con, an industry-wide psychosis where you’re judged for not constantly dedicating every single second of your existence to prompt a series of chatbots into making something, all under the mistaken belief that at one point it’ll be so smart you…won’t have to prompt them?  

Nevertheless, Van Horn is completely right — the sales pitch of AI is that agents were supposed to do the work for you, but billionaire losers are gaslighting you into believing that a digital busybox that requires constant vigilance to make sure it does what you ask or doesn’t spend too much money was somehow “autonomous.”

While it’s easy to make fun of Silicon Valley, what we’re witnessing is a widespread mental health epidemic caused by liars like Sam Altman, Dario Amodei, and their wealthy backers lying about the capabilities of AI, creating an abusive culture where humans become subordinate to unthinking, hallucination-prone agents either subsidized by OpenAI or their employer:

Engineers are working until 4 a.m. to demonstrate productivity on par with the agents they’re deploying. Startups are creating internal counseling programs so employees can vent about their AI-induced burnout or team up with a self-proclaimed AI ambassador who can help them learn how to better use the technology. In San Francisco, mental health walks are taken in the shadow of small planes flying banners to stop hiring humans, and Friday nights increasingly involve “touch grass” parties — intentional spaces to not talk about AI because everywhere else has been infected.

This is fucking horrible, and every loser who inflated this bubble should be ashamed of themselves. 

In fact, fuck it, I want to speak directly to the people working in Silicon Valley and the tech industry who have been ground down by this industry. 

A Plea To The The Tech Workers Victimized By The AI Bubble

I know not all of you are anti-innovation.

I know many of you feel suffocated.

I see you, I hear from you every day, and I find what is being done to you repulsive.

Your industry has abandoned you. 

Your investors are lying to you, and are getting rich while you can’t afford a studio apartment in the Tenderloin. AI does not do what you have been promised it does, and those who are excited about it are excited because they believe it will replace you. You are victims of a marketing campaign built to enrich a few people by sacrificing your time and energy to defend a doomed tool. 

You are using tools that are built to manipulate you into making you work longer hours in the name of automation. You are being abused. You are being tricked into fighting for the 1% in the name of democratizing software. Your agents are meant to set you free, but they chain your body and mind to a system built to exploit your labor, extract your value and leave you dead. The people who make these agents fantasize about replacing you with them, and want to use your data to do so. They are lying that it is possible, but they want you to be scared so you will use their products more. 

They have convinced you to fight on their side in a war where you will lose regardless of the victor. 

You are a victim. I am not your enemy. I love technology too, and I want the tech industry to make cool shit again. 

That will not happen under its current leadership.

This era is built to drain the life out of you, to suffocate you with endless tech chatter, to make technology every part of your life, to somehow sell you the promise of automation, but only a kind of automation that you have to monitor constantly, prompt constantly, built to be addictive and superficially productive, built to fuel a Bay Area culture steeped in a godless version of the Protestant Work Ethic. 

You must be a cracked engineer, you must work 15 hour days, you must have 8 subagents beating the absolute shit out of your codebase for one reason or another,  your Calendly must be open 8AM to 8PM, and you must be willing to work yourself to the bone for a chance to escape “The Permanent Underclass,” a misused term to refer to the world after an entirely-imaginary concept of Superintelligence, peddled by people who speak with a smugness that makes me want to spritz them like they jumped on the dinner table. 

The grotesque glee that some have at the idea of being the first to announce AI’s destruction of everything you hold dear are your enemy, as are those who are desperate to constantly lick the boots of the Altmans and Amodeis of the world. Do not trust those who say that being part of an in group requires you to use certain kinds of software or attack others in the name of Silicon Valley. 

The people encouraging you to work in this way do not care about you, or are being manipulated into believing this is how you all become rich by people exploiting their ignorance, fear or greed. 

The people at the top do not care about the future, or progress, or anything other than growth. They are acolytes of a egregore of capital that has no purpose other than to expand and maximum velocity at all times, everything is fine as long as something is always happening, because the moment you stop moving you remember that nothing you’re doing really matters, because you’re making software while working sweatshop hours. 

AI agents are built to make you interact with them. They are built to make you burn tokens. They are built to make you apologize for their mistakes and give them credit for your labor. Any “autonomous” tool that requires specific prompting, harnesses, scripts and tooling to make it sometimes work autonomously is conning you. 

I’m also sure that there are a few perfectly normal software people using this stuff locally or with an open source model who treat it as normal software, loathe the data centers and see no need for the capex or mass market version of LLMs. These people are drowned out by a worryingly large crowd that speaks like they’re in a cult that exists to prove that OpenAI and Anthropic are somehow something more than SaaS companies. To them, using AI is a way of virtue signaling that they’re a pure, productive spirit, a willing supplicant for a future where they assume they’ll ascend because they told enough people “we’re still early.”

The tech industry got taken in by a form of religious con, sold to them wrapped in atheistic “rationalism.” 

Some may or may not have AI psychosis — or at the very least a severe addiction — as a result of being forced to interact with these things day-in-day-out, and the easiest way to check is to try not to use them for a day, or to try and solve a problem without them. If this is you, please know that I am not attacking you, and see you as a victim of a con.

You are ingesting poison while being told it’s ambrosia. You are being made to work twice as much for roughly the same output, if not less. You are being humiliated or isolated for not using the right tools or saying the right things. Silicon Valley was built on the ideas of individualism and rationality, and the people at the top of your industry are telling you to fall in line and join an illogical consensus. You exist in a monoculture sold as anti-establishment as it mostly enriches Microsoft, Google and Amazon.

Your culture is being eroded by people who do not care about technology. You are unwitting pawns in a greater war against innovation, where billions are steered into the hands of those who only ever care about growth and “acceleration” that benefits only a small few. You are not alone if you feel scared, anxious, listless and drained, because you are being worked to the bone building layers on top of AI models owned by subsidiaries of the largest companies in the world. 

The fact that so many of you have to orient your products or fundraising around Twitter is a sign that your culture is decaying. A true meritocracy would reject the idea of “going viral on social media” like a virus, because it overwhelmingly benefits a monoculture that suppresses free thought and dissent. 

Tech workers are in a constant battle between imbeciles and monsters, or an Arnold Palmer of the two. Those who want to build useful software that customers like you are drowned out by a Greek chorus of unexceptional cretins that think they’re competent because they can bonk an LLM on the head to make an impression of competence. 

Generative AI is the Peter Principle on steroids, removing the friction points where a diplomatic moron might get caught out, making them far more mobile and extremely dangerous. Companies are run by men that don’t know what they’re doing, desperate to avoid anybody realizing that we’re at the end of software’s era of hypergrowth, increasingly aware of their own mortality and their lack of a culture that might actually build something a human being would want. 

For those of you still hanging in there, I see you and admire you, because if I worked at most tech companies right now I’d fucking quit. Seeing this entire industry bow at the feet of the great unprofitable mediocrity machine is sickening, and based on the many tech workers I talk to every week, the mood effectively everywhere is exhausting, demoralizing, manic, and horrible to watch. 

Everything must be done faster, with less people, with less organizational support, but more use of a tool best known for its hallucinations and ruinous cost, which you must use a lot, but also not too much. However much you use it, you must constantly celebrate it for fear a cult of personality and mediocrity will isolate or fire you for the crime of not wanting to “Do AI.”

Even if you are still trapped in this world for months or years to come, know that you’re not crazy for finding it revolting, exhausting and debilitating. You do not have to do things this way, but I understand if you’re made to by circumstance or social pressure.

The tech industry is in the throes of minor AI psychosis, or, put another way, it’s a way to scale the already-potent sense of make believe that has kept this industry afloat the last decade. 

The grander cargo cult of praying at the foot of whatever capital-lust the venture capitalists currently have has led everyone astray, to the point that companies worth billions — or even trillions — of dollars on things based on how they might play out on Twitter, a maligned representation of the tech industry that caters to Silicon Valley gossip and the derangement of the markets, intellectually stunting most who cater their business or marketing to it. 

Sidenote: You may just be a regular person in an unfortunate situation where your boss (or bosses) are demanding you adopt a tool that, at best, is kind of useful in specific situations. Your performance reviews or continued employment may be dependent on your use of AI tools, and if that’s the case, you must make it your mission to cost your company as much as humanly possible. I call this “rascal’s wager” — in a sufficiently AI-pilled organization you’ll be hailed as a hero, but burn tons of money, and likely get them to reduce their dependence on AI as a result. In a normal one, your CEO will see the astonishing cost of AI and, hopefully, some sense.  

The rest know exactly what they’re doing: appealing to an audience of venture capitalists convinced they’re “in the arena” by posting 12 hours a day writing 2000 word long posts using Claude. You must coddle these rich oafs, because it’s effectively impossible to raise money if you don’t. You must be able to recite the rituals — Hermes! Loops! Permanent underclass! — or you’re considered uncool by the least cool people alive. You, the great individualistic thinker of Silicon Valley, must convince wealthy oafs that you are an independent and rational person, but also that you will follow the greater consensus. 

It’s a really unfortunate time to have ideas, dreams or goals outside of some sort of Potemkin agentic startup or if you can do the hocus pocus to con a VC into thinking you — or anyone — will invent recursive self-improvement, or AI that teaches itself. 

You’re getting money right now if you can make noises that sound like you’ll be the next Baseten or whatever. It’s the era of inference I guess. Loops too. Keep cheering along! Never stop agreeing with what everyone else is doing, or if you do, only do so in a way that suggests that you all agree on the big stuff, which means you ultimately support either or both OpenAI and Anthropic, who companies that effectively operate as subsidiaries of the largest tech companies in the world. 

It will stay this way until something changes. 

The AI Industry Is Losing

As if I haven’t made it clear enough, the AI industry is losing. Their plans are not working, their products are not doing the things that they’ve promised, and though they intend to exhaust every available source of capital, they aren’t going to have enough money to do this forever. And no, AI is not “too big to fail.”

Everybody makes fun of it. “AI” has become synonymous with generic, ugly, corporate slop. It’s a physical blight on the Earth, pumping horrifying toxins into minority neighborhoods and causing such noise that it makes people physically sick, and to make matters worse, some independent writers have made it their mission to cast doubt on these problems because they do not represent “the aggregate” of data centers.

Everyone trying to be the “rational” voice on data centers should know that they’re only helping make the AI industry stronger. If you’re anxious that people are being “unfair” about water use, you’re an active pawn of capital, and exist only to help pump the bags of NVIDIA and the billions of dollars of speculative investment going into these monstrosities. 

Without getting into the weeds, know that anyone talking about data center water use in terms of almonds or cattle is an actual industry plant. 

California does use a lot of water to make almonds — and also makes 100% of America and 80% of the world’s supply. Cattle and other livestock also take up a lot of water and land, but they also make food for people to eat. You can bicker about how much water a data center may or may not use, and you’re going to sound like a complete loser every second you do so, because you are fighting to make sure that the AI industry can build data centers for the largest companies in the world. 

Data centers are a monument to everything wrong with the world — horrifyingly large, loud, demanding of power and water and resources of all kinds. They create very few jobs, and those involved in their construction are usually from out-of-state. Their actual value to the world is largely tied up in their nebulous theoretical contribution to something an AI company does, and they get huge tax breaks, which means they don’t really contribute very much to many of the areas they’re put in. They are intentionally conflated with the smaller, useful data centers we’ve had in the past, all so that pedants can say “ehhmmm, you never had a problem with these before?”

I haven’t, because previous data centers haven’t been filled with GPUs or drawn more power than a small town, nor have they been rammed through by a combination of crony capitalism, tax breaks and endless debt.

And it’s fundamentally unclear why we need them! 

No, really, why do we need these fucking things? So Anthropic and OpenAI can do more of whatever it is they’re doing? Neither appears to be unable to serve customers — other than the lousy uptime of Claude — nor do they appear to improve their products based on the availability of compute. 

For such an offensively-large footprint — physically, fiscally and societally — nobody can really explain why the fuck we need all these things, other than the fact that they might make somebody money on a service that is best known for its huge mistakes and lack of profitability. 

As I’ve discussed, the demand isn’t there outside of these two companies, and the only reason anyone believes that it does is that the largest tech companies in the world have burned through every dollar they have to hide from you that they’re out of big ideas.

The AI industry fights like a bunch of losers because that’s what they are. They cannot win by telling the truth about their products, their infrastructure, the condition of their finances or their overall intentions. They cannot succeed without manipulation and deceit because they know, deep down, that their businesses don’t make sense and their actual products, described in the present tense, are impossible to justify what they’re asking for.

They require us to coddle them, to ignore their ruinous cost, avert our eyes when they hallucinate or delete somebody’s database, blame ourselves when they make mistakes and speak entirely in theoretical terms when we describe them because the present kind of fucking sucks. 

Absolutely nothing that the AI industry has created is worth even a fraction of the trillion-plus sunk into this industry, and at this point it’s very clear that these models cost about as much as a person and even then are neither capable of replacing one or profitable for the provider. 

The best shot the AI industry has is open source models that may only be getting better by distilling American models. At some point Anthropic or OpenAI is going to slow down and then stop making models entirely because it costs too much money to train models, and said costs are only increasing.

Even if GLM 5.2 is truly nearly as good as Opus 4.8, it did so by copying its outputs, which means that these models will likely only get as good as long as the foundation model companies keep training, which will only be possible if they can keep raising funding, which will become difficult if open source models eat their lunch in any meaningful way. 

Could Anthropic and OpenAI theoretically make better models in a vacuum? Sure! But they’re now going to have to slow-roll them, because Sam and Dario’s four or five-year-long scaremongering campaign has forced them into a situation where the US government demands oversight into their model releases at a time where the AI industry cannot afford to slow down. 

Their only option is to sit there and take it or, alternatively, admit that they’re making normal software, which will make the whole “let’s build a trillion dollars of data centers” thing a little harder to justify. 

This will also be a tougher sell to Masayoshi Son of SoftBank, who gave a truly demented presentation during the 46th annual SoftBank shareholder meeting, calling the company a “golden egg machine” that’s also a goose that lays eggs that are, at times, undervalued. 

Masayoshi Son has sunk $64 billion into OpenAI, and existentially tied a company with a quarter-of-a-trillion dollar market capitalization — the third largest on the Japanese stock market — to whether or not Sam Altman can turn a company that burned $20.9 billion in a single year into a company that makes more than $284 billion in annual revenue by 2030.

If you’re curious, the second-largest is Mitsubishi UFJ Financial Group, a massive Japanese bank with tens of billions of dollars invested in AI data centers, and the first is Kioxia, a memory and storage company that has seen massive revenues as a result of the massive demand for memory and storage for AI data centers. 

What do you think happens if AI data center capex slows? What do you think happens when it turns out there’s not enough demand for all those data centers? Even if MUFJ and SMBC (the second-largest Japanese bank, also heavily levered in AI) have sold off part of the risk, their counterparties are still part of the global banking system.

Anyway, SoftBank’s glorious, Geese-filled future depends upon OpenAI going public, and the New York Times just reported it’s likely pushed its IPO back to 2027, because bankers didn’t think it would get a trillion-dollar valuation, which is an absolute disaster considering its pre-money valuation (as in before the $122 billion it raised) was around $735 billion.

While it's partially blaming the floundering value of SpaceX, I think it’s possible (though I have no privileged knowledge to confirm it) that my story publishing its audited financials had something to do with it. 

One can present financial data in all manner of ways, and I have to wonder whether its S-1 might have differed in some way — perhaps how segments were broken down — to what I reported. Perhaps bankers saw the reaction to the numbers, the mess that is SpaceX, the weird state of the market, and said “yeah man you’re gonna be lucky to float at $700 billion.”

We may never know. 2027 may as well be in the year 3000 for how far away it is, and how much further OpenAI will have to drag itself to get there. 

While it “raised $122 billion” earlier in the year, it’s waiting for two more tranches of $20 billion a piece from NVIDIA and SoftBank, and will now straight up not get the $15 billion that Amazon conditioned on it either going public or reaching AGI. Considering that Mr. Altman can’t even con a bunch of bankers who were dumb enough to believe that SpaceX could 300x its AI revenue by 2030, it’s clear that the jig is up. 

Another worrying sign is that SoftBank was unable to raise a $6 billion margin loan with its entire OpenAI stake — likely valued, at least on paper, at over $100 billion — as collateral. This suggests banks have little faith in the company.

Some might believe that Anthropic has a better chance, and I’m just not sure there’s much that differentiates it from OpenAI anymore, other than how annoying Dario Amodei is and how much he appears to piss off the Trump administration. 

Anthropic is a large language model company that loses billions of dollars that has subsidized accounts that allow users to burn $8,000 a month in tokens for $200. To paraphrase and build upon something said by Cory Doctorow, if your business is only successful when you give away $40 for $1, that’s not a real business, it’s a way to feed venture capital dollars to hyperscalers and sell a bunch of people a product that doesn’t exist. 

Anyone still lazy enough to say “they’ll crank up the price” or use some hackneyed Amazon Web Services or Uber comparison is either deliberately ignorant (I explain here) or a loser like the rest of the AI industry. If you’re so confident about this shit, despite all the blaring warning signs, you need to start finding actual, real, tangible evidence, and you need it soon.

Every argument in favor of AI requires you to speak in the future tense and ignore your lying eyes. The AI industry will not allow you to discuss LLMs in terms of what they do today without reminding you that progress has been so rapid over the last few years and demanding you immediately acquiesce that something might be good in the future. 

Seriously, try and talk to somebody who loves AI sometime and criticize the tech and see how quickly they fall into the tropes of AWS losing money, AI models rapidly getting better (at benchmarks rigged in their favor because they can’t use a computer like you or me), about the “cost of intelligence going down” (when it’s actually going up), or any number of other tired tropes that mostly rely on you ignoring the present in favor of a billionaire’s dream of the future.

These are, as I have been saying, the acts of losers. This is what you do when you do not actually have a compelling story, cannot win by being straightforward or contrite, and have no way to prove yourself valid outside of appealing to cargo cults and doing financial engineering, except you’re such a loser that you’re not even doing it to commit fraud! You’re just writing PDFs so you get shares on Twitter. 

Forgive me for being so very brusque, but I have had to prove myself endless the last few years, and when I finally bring you the proof that OpenAI loses a bunch of money, you immediately jump for the first keys jingled above your head. If you truly love the AI industry so much, you should ask it for better proof! You should be enraged that OpenAI’s numbers are so shitty, and that you have to debase yourself by pretending they’re not! How utterly shameful! 

That’s loser shit! If you love large language models so much, go out and demand the people making them bring you the answers to my questions. Whenever I’m asked about how I might be wrong it mostly comes down to “but what if something that hasn’t happened happens?”  If your answer is “OpenAI will drive down the cost of its silicon using its “Jalapeño” chip from Broadcom,” you do not have shit! It’s still in early testing!

There is no future for the future these people are building. The demand does not exist for these data centers. It never has. It never will. You can give Baseten as much money as you want, you can talk about the exciting world of open source for hours, but there is not actually enough demand for this stuff unless it becomes something very different, very soon, in a very big way, that likely also involves it getting cheaper. 

Anthropic and OpenAI have $1.1 trillion in compute commitments that are contingent on their continued growth, at a time when their customers are protesting their costs, at a time when the market is clearly saying “you are not worth a trillion dollars.” 

What do you think changes that? 

The halo effect of AI has given way to a societal cynicism, even by the people that love it, who have a sort of vague reticent “I give up” vibe that I find exhausting to watch and will have a great deal of trouble forgetting once the bubble bursts. Even the people who claim to be excited are making jokes about Masayoshi Son and Sam Altman! 

Everything about AI has the stench of death and desperation, of losers pretending they’re winners who can only thrive in conditions that reward grifting, specious hype and forward-looking statements that vary from ridiculous to deliberately harmful.

It’s ugly, regressive, and when this era ends, I expect financial carnage and chaos that could have easily been avoided had so many people not so readily swallowed poison under the auspices of innovation.

Then again, some people might just be born to be regulated by the wallet inspector.

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 10,000 to 18,000 words, including vast, detailed analyses of the biggest events and companies in the AI bubble.