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. My Hater’s Guides To 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 week, I’ll publish the second part to my ongoing series (“What If…We’re In An AI Bubble?”) about the factors and events that will cause the AI bubble to finally pop.
Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week.
AI is, as it stands, not economically viable for anybody involved other than the construction firms, NVIDIA, and the surrounding hardware companies benefitting from the irrational exuberance of a data center buildout that doesn’t appear to be happening at the speed we believed.
Every AI startup loses millions or billions of dollars a year, and nobody appears to have worked out a way to stop hemorrhaging cash. Hyperscalers have invested over $800 billion in the last three years, with plans to add another $700 billion or so in 2026 and another $1 trillion in 2027, meaning that they need to make at least three trillion dollars in AI specific revenue just to break even, and $6 trillion or more for AI to be anything other than a wash. I went into detail about this (albeit at a lower, pre-2026/2027 capex number) in a premium piece last year.
To give you some context, Microsoft made $281 billion, Meta $200 billion, Amazon $716 billion, and Google $402.8 billion in revenue in their most-recent fiscal years for every single product combined, for a total of $1.599 trillion. None of them will talk about their actual AI revenues. Yes, yes, I know Microsoft said that it had $37 billion in AI revenue run rate ($3.08 billion a month or so) and Amazon had $15 billion, or around $1.25 billion a month, but both of these are snapshots of single months that are meant to make it sound like they’re going to make that much in a year but in the end, you don’t actually know anything about how much money they’ve made from AI.
We do, however, now know that Microsoft has spent an approximate $100 billion on its OpenAI partnership after testimony from an executive during the otherwise-dull Musk-OpenAI trial, per Bloomberg:
That figure includes Microsoft’s original investments in OpenAI, as well as the costs of building infrastructure and hosting OpenAI’s computing, Microsoft deals executive Michael Wetter testified on Monday. It is cumulative through the current fiscal year which ends in June, he said.
This is a fascinating insight for a few reasons:
- Microsoft has spent a total of $293.8 billion in capex since the beginning of Fiscal Year 2023 (which began in the back half of 2022).
- This means that around 30% of Microsoft’s capex ($87 billion) went to building OpenAI’s infrastructure.
- Based on discussions with sources familiar with Azure architecture, this is the vast majority of Microsoft’s operational capacity.
At the end of 2025, OpenAI claimed that it had 1.9GW of capacity (likely referring to total power draw rather than the actual critical IT of the infrastructure at its disposal), which, per analyst estimates, ($42 to $44 million per megawatt) works out to around $79.8 billion. This claim was made around six months before the release of Microsoft’s most recent quarterly results.
In other words, Microsoft has spent 4 years sinking (either through spending or allocating the capex in advance) nearly $300 billion into…building OpenAI?
Okay, fine. Microsoft also has 20 million Microsoft 365 Copilot subscribers for an absolute maximum revenue of $7.2 billion…if every single one were paying $30 a month, which they are most assuredly not as Microsoft has been offering discounts on it for years.
Based on my reporting from last year, Microsoft made around $7.5 billion from OpenAI’s inference spend and $761 million from its revenue share in Fiscal Year 2025, a year when it invested (either spent or allocated) around $88.2 billion in capital expenditures.
I didn’t report it at the time, but I also had the numbers for all of Microsoft’s revenues for the first three quarters of Fiscal Year 2025 — a total of $8.9 billion of total AI revenue, with around $4.35 billion in revenues when you removed OpenAI’s inference. If we assume that Microsoft’s other AI services grew 10% quarter-over-quarter, I estimate that Microsoft likely made around $17.9 billion in AI revenue in FY2025, or a little under a fifth of its capex.
And let’s be clear: none of these numbers include the actual operating expenses.
Data centers, after all, need electricity to run, and AI data centers in particular need a lot of electricity. And some — though, admittedly, not many — people to handle the things like maintenance, repairs, and operations. And then there are things like taxes, insurance, and the other day-to-day costs that, when you add them all together, make a big, scary number.
You can argue that “actually GPUs are profitable to run” (I disagree!), but for any of this to make sense, four things have to happen:
- AI revenues have to explode.
- Capex has to stop being invested.
- GPUs need to be margin positive, including both their cost and the debt associated with operationalizing them.
- AI revenue has to stay consistent both before and after you stop spending that capex.
All four must be true. If AI revenues don’t explode, capex can stop, margins can be positive, and your best-case scenario is…you maybe broke even. If capex never stops being invested, you need revenues to explode dramatically — to the tune of effectively doubling Microsoft, Meta and Google’s entire businesses, and tripling Amazon Web Services’ annual revenue ($128 billion) — and for said revenues to be margin-positive, because if they’re not, eventually other healthy businesses will slow, leaving AI to tear a hole in overall margins. In all cases, AI revenue must stay consistent because, well, you need to get paid.
Sidenote: In all honesty, I have no idea how Meta makes this make sense, as it plans to invest at least $125 billion in capex in 2026 and has, to this point, not shown any actual, real growth in its revenue from AI, and no, those increases in conversion don’t mean actual revenue.
I also cannot find an economic scenario where this pays itself off.
Let’s assume that Anthropic is actually at $45 billion in annualized revenue (I believe it’s doing some very worrisome maths to get there), or around $3.75 billion a month. On an annualized basis, this would not be enough — assuming it had zero operating expenses (rather than losing billions) — to recover a single year of capital expenditures from Microsoft, Google, Meta, or Amazon from 2024 or 2023.
Even if OpenAI’s entire cloud spend ($50 billion) for 2026 went to Microsoft and it doubled its Microsoft 365 Copilot revenue (at full cost) to $14.4 billion, it estimates it will invest $190 billion in capital expenditures this year. Amazon’s $15 billion AI run rate, even if it doubled, wouldn’t put much of a dent in its $200 billion in investment plans. While we don’t know Google’s AI revenues, it plans to invest $185 billion in capex this year.
These AI revenues have to be completely fucking insane and they need to be that way extremely fucking soon, because otherwise the best they’ll be able to say is “our first few years of capex weren’t particularly useful but the stuff we built after it was,” which still works out to a few hundred billion dollars of waste.
Things get even worse when you realize that at least 70% of Microsoft, Google, and Amazon’s compute is dedicated to Anthropic and OpenAI, two companies that burn so many billions of dollars that Microsoft, Google and Amazon have already fed them a combined $54 billion in the last three years, with $28 billion of that coming in the last month and Anthropic due another $50 billion from Google and Amazon if certain performance obligations are met.
And there’s no real sign, outside of Anthropic and OpenAI’s compute spend (which is reliant on hyperscaler and venture capital money), of any real explosion in AI revenue. Per The Information (in a chart I love to share!), more than 50% of hyperscalers’ revenue backlogs comes from these companies:
If massive, incredible demand for AI existed, wouldn’t these remaining performance obligations be near the trillion mark? Wouldn’t there be other Anthropic or OpenAI sized chunks of revenue? There’s allegedly incredible, unstoppable, insatiable demand for compute. Why isn’t it lining up?
Let’s take a look at those RPOs!
- Microsoft’s RPOs jumped from $392 billion to $625 billion between Q1 and Q2 FY26 (or calendar year Q4 2025 and Q1 2026), driven by the $250 billion in “incremental Azure spend” from OpenAI (including already-existent commitments) locked up in October 2025 and the $30 billion promised as part of its deal with Anthropic from November 2025. Based on Microsoft’s own disclosures, without Anthropic and OpenAI’s additions, RPO would have been effectively flat, as evidenced by the fact that in Q3FY26, remaining performance obligations sat at $627 billion.
- Amazon’s RPOs jumped from $244 billion in Q4 2025 to $364 billion in Q1 2026, driven by its February 2026 $100 billion expansion of its $38 billion compute deal from November, and its extended partnership with Anthropic for 5GW of compute capacity unattached from any kind of dollar number.
- Google’s RPOs jumped from $242.8 billion in Q4 2025 to $467.6 billion in Q1 2026, driven by (per The Information) $200 billion in committed spend on TPUs and compute from Anthropic, meaning that it has expanded its future revenues by an unremarkable $24.8 billion when you remove Anthropic’s spend, when RPOs had previously jumped $85 billion between Q3 and Q4, likely driven by its compute deal from October 2025.
- It’s fair to assume a chunk of the remaining RPOs are from its deal to rent TPUs to Meta, announced in February 2026, which makes it likely that it accounts for the majority of the remaining $24.8 billion.
That was a lot of numbers, so let me make it simpler: outside of OpenAI and Anthropic, these three companies do not appear to be significantly increasing their revenues, and the only way to get that revenue is to feed money to one or both of these companies.
Put aside all the theoreticals and hypotheticals and metaphors and imaginary future scenarios and tell me: what, in the next year, are Microsoft, Google and Amazon going to do about this problem? How do they solve it?
If we assume the absolute best-case scenario, these companies are making a combined $70 billion in annual revenue on investments that now — including the money invested in the companies themselves — total over $900 billion. Doubling that won’t be enough. Tripling it won’t be enough. In fact, to pay this off, these companies will need to be making over $100 billion each in AI revenue in the next year, because otherwise there is no covering these losses.
And it all comes back to a very simple point: AI is too expensive. If the margins were good, they’d be sharing the margins. If the revenues were good, they’d be sharing the revenues (and no, run rates aren’t revenues). If the business was strong, it would be a separate category in their earnings.
But LLMs are too expensive! They cost too much to run, and said costs appear to increase linearly with revenues. The more a user uses a product, the more it costs the company to run it, and the more capacity they can take up. The only way to capture any growth is to buy and install GPUs, which in turn requires you to build somewhere to put them, which takes time and money.
I’m really struggling to see the argument in favor of continued capex investment. You’re more than $800 billion in the hole with, I estimate, less than half of that resulting in operational GPUs and capacity. Said capacity is mostly taken up by OpenAI and Anthropic, two companies that burn billions of dollars and do not appear to have an answer for how they might stop.
The more you build, the more your infrastructure becomes dependent on the continued existence of two perennially-unprofitable ultra-oafs, as your existent AI product lines are, at best, add-ons to products like Google Workspace or Microsoft 365, or further expansion of cloud compute capacity with lower margins and higher up-front costs than anything you’ve ever built.
Every quarter is an opportunity to put yourself another $30 billion or so in the hole, all in the hopes that, I assume, OpenAI or Anthropic will pay you $100 billion or $200 billion over the course of a few years, because nobody else in the entire universe is spending that much on compute. You are not recovering these investments without either a massive new product line that doesn’t exist today or three or four Anthropic or OpenAI-sized compute contracts.
Put another way, Amazon needs another AWS ($128 billion a year), Microsoft another Azure ($75 billion a year, including OpenAI’s 2025 compute spend) and Google a business line at least half the size of search (around $200 billion a year). These businesses have grown to this size by providing extranormally large amounts of value from the very moment they were created and impenetrable monopolies — and while there are quite literally other cloud providers that can physically provide the infrastructure to OpenAI and Anthropic (Oracle is trying to compete and may die as a result), the actual “monopoly” here is “being able to deploy hundreds of billions of dollars.” Anthropic proved this when it took 300MW of compute from Elon Musk.
Sidenote: I have absolutely no idea what Meta does, and my chaos bet is that it starts renting out its compute to Anthropic or OpenAI when things get rough. Perhaps it does some sort of incestuous deal where Meta gets equity. I really have no idea here! It’s a crazy and stupid company run by a moron.
In Oracle’s case, as I’ve explained at length, it has to successfully build 7.1GW of capacity, have that capacity actually be margin-positive (doubtful!), and then actually get paid for it by the time it’s built in, oh, I dunno, 2032?
Sadly, I have bad news about Oracle, Microsoft, Amazon, and Google’s largest customers.
Here’s a fun game: ask an AI booster how OpenAI or Anthropic becomes profitable!
Here’s what they’ll say:
- Silicon will get cheaper.
- They’ll start selling services.
- They’re profitable on inference.
I must be abundantly clear that nobody has any proof that anyone is profitable on inference, but we have plenty of proof they’re not. They’ll likely cite known liar Sam Altman saying OpenAI is profitable on inference from a party from August 2025, or Dario Amodei saying (in a sentence around “stylized facts” that are “not exact” and are specifically “a toy model” and specifically not about Anthropic) “the inference has some gross margin that’s more than 50%.”
Here’s a really simple way to dispute this: Coatue said that Anthropic’s revenues were 85% API calls in 2025. If it’s profitable on inference, how is it still losing money? You’re gonna say “training,” but that doesn’t actually answer the question: if Anthropic’s process of providing tokens to its models is profitable, how is it losing so much money? Why offer a subscription platform at all?
As I’ll get to, Anthropic has companies paying massive amounts for tokens — hundreds of millions a year in some cases — that’s all inference. Why are you bothering with these stinky, nasty monthly subscriptions?
The “inference is profitable” argument is a bedtime story told to people that can’t reconcile the logic of a company that allows people to burn between $8 and $13.50 of every dollar of their subscription revenue.
Otherwise, you have to reconcile with the fact that both Anthropic and OpenAI are both incinerating money and have no real path to any kind of sustainability other than, well, not doing that.
One very, very specific counter-argument people make is that open source models are cheap, and can somehow be compared to OpenAI and Anthropic’s, despite the fact that we have no idea what the actual parameters of Sonnet, GPT, Opus, or any other of their models actually are.
What we do know is that both of these companies lose billions of dollars.
What we do know is that OpenAI, per The Information, plans to burn $852 billion through the end of 2030, and that as of March 6, 2026 (per CFO Krishna Rao’s sworn affidavit), Anthropic made “exceeding” (sigh) $5 billion in revenue and spent $10 billion on inference and training.
Anthropic has done a great deal of work to obfuscate how much it actually makes or spends, but I think it’s likely it burns even more than OpenAI, given the fact that it’s had to raise $75 billion in the last 6 months (assuming its new $30 billion round closes), and that’s not including an additional $30 billion from Google and Amazon if certain unknown milestones are hit.
Then there’s the issue of those RPOs. Anthropic is now on the hook for $200 billion to Google, $100 billion to Amazon and $30 billion to Microsoft, I assume over the course of the next three or four years.
So let’s lay this out.
Anthropic — based on its own affidavit from March — appears to have spent $3 to make $1 of revenue on a compute basis, and that’s before you include any and all other costs like staff or electricity or the vocal coach that Dario Amodei uses to add that bass to his voice.
Additionally, it needs $330 billion to pay its cloud obligations to Amazon, Google, and Microsoft over the next four years. I’d estimate it needs $5 billion a year for its compute deal xAI (so $20 billion over the total period) and an estimated $30 billion to cover its deal with CoreWeave. That brings us to a total of $380 billion.
It’s hard to estimate the actual costs associated with running Anthropic because so much of the reporting no longer makes sense as a result of that affidavit. Nevertheless, I think it’s fair to assume it will need at least $20 billion of operating expenses across that four year period.
We don’t even need to play in the realm of “what might Anthropic or OpenAI’s revenues be?” to understand the problem here. Both companies aggressively burn money, and neither of them have any answer as to how they might stop. Numerous reports about how Anthropic will turn “cash flow positive” in either 2027 or 2028 are fantastical, illogical, entirely driven by ridiculous projections, and should never have been reported as anything other than an attempt by companies to mislead their investors. In both cases, reporters should’ve had more asterisks on those numbers than Q*Bert reading Frank’s lines from Blue Velvet.
And we have plenty of evidence that they’re losing more money over time. In January 2026, The Information reported that Anthropic’s gross margins were 40% in 2025 — 10% lower than its “optimistic” projections, specifically attributed to “…the costs of running Anthropic models from paying customers, in a process known as inference, on servers from Google and Amazon,” adding that those costs were “23% higher than the company anticipated.”
In February, The Information ran another story saying that OpenAI’s gross margins fell from 40% in 2024 to 33% in 2025, a full 13% lower than its projected margins of 46%, all because (and I quote) “…the company having to buy more expensive compute at the last minute in response to higher than expected demand for its chatbots and models.”
You know, exactly what Anthropic has had to do.
This is what I’ve referred to as the knife-catching problem for compute demand — you either don’t order enough compute and have to rush to buy some last-minute as demand intensifies, or you order too much, and, well, to quote Dario Amodei:
Basically I’m saying, “In 2027, how much compute do I get?” I could assume that the revenue will continue growing 10x a year, so it’ll be $100 billion at the end of 2026 and $1 trillion at the end of 2027. Actually it would be $5 trillion dollars of compute because it would be $1 trillion a year for five years. I could buy $1 trillion of compute that starts at the end of 2027. If my revenue is not $1 trillion dollars, if it’s even $800 billion, there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt if I buy that much compute.
And right now, as I’ve covered, there’s not enough compute being built to keep up with Anthropic or OpenAI’s voracious demands, meaning that they will both be bartering to buy whatever’s available at whatever price it’s available at. This naturally will savage their already-negative margins…
…and then what?
No, really, and then what? One of you fucking AI boosters, answer me, how does this actual reverse course? Because even if Anthropic were making $100 billion in annual revenue, it would probably be losing $300 billion or more to get there. The fact it had to raise $30 billion in February, $15 billion in April, and now $30 billion more in May all while allegedly pulling in more than $3 billion a month in revenue suggests that its COGS are fucking horrendous, and its growth is coming at a terrible financial cost.
Let’s say that Anthropic keeps growing and (as The Information suggests) hits $100 billion in annualized revenue (around $8.3 billion a month). How, exactly, does it afford to make that much money? Because right now it’s (allegedly) about to hit $45 billion in annualized revenue, and needs so much money that it’s absorbing (along with OpenAI) the majority of venture capital raised this year, and very clearly does not have any path to bring its costs down.
The answer is simple: it can’t! There is no mechanism to do so. More compute does not make OpenAI or Anthropic’s services cheaper to offer. There is no magical silicon coming that will make any of this more affordable, and no, Anthropic is not “profitable on inference,” because if it were, that massive revenue growth would have leveled out its margins rather than require it to raise a little less than the combined value of every Major League Baseball team, or more if you add the other $50 billion that Amazon and Google have promised based on secretly-held performance obligations.
The same goes for OpenAI, which “raised” $122 billion (around $45 to $50 billion in real cash, with the rest either paid in installments or on it IPOing or reaching (sigh) AGI) in February and is now already considering raising more.
Somebody might counter-argue that this is companies raising as a means of boosting their valuations, I think that’s a very convenient way of looking at two extremely problematic companies.
I should also ask why neither of them appear to be seriously considering going public. While both were rumoured earlier in the year to be planning to do so in 2026, both appear poised to raise more private capital.
I think the answer is simple: their CFOs know that doing so would reveal their actual margins, which are hot dogshit with sprinkles on top.
Nobody has a sensible or logical response here.
Which leads us right to our next point!
Organizations Cannot Afford To Keep Blowing Through Their AI Budgets Millions Of Dollars At A Time Without ROI
One important detail to keep in mind here is that as of a month or two ago, Anthropic moved all enterprise customers to token-based-billing, which will begin, I believe, a true stress-test of the true “value” of AI as costs skyrocket.
Just last week I ran the first of a two (or three, potentially) part premium series called “What If We’re In An AI Bubble?” and touched on the gruesome subject of whether organizations could afford to pay for AI long-term:
Per Laura Bratton of The Information, Uber, ServiceNow and multiple other organizations are blowing through their yearly API token budgets in a matter of months, and are currently in the “cope” stage, with Kellie Romack, CIO of ServiceNow, saying the following about a conversation with CFO Gina Mastantuono:
Romack said she recently met with ServiceNow Chief Financial Officer Gina Mastantuono to figure out how to contain costs so employees can keep using their Claude Enterprise accounts for the rest of the year.
“It’s a really hard problem,” Romack said. “She’s worried, I’m worried, and we’re working together to go figure this out.”
Let’s focus on that phrase “…can keep using their Claude Enterprise accounts for the rest of the year,” because it’s important. A public company with a CEO that previously boosted the metaverse and now has profound AI psychosis is saying that it isn’t sure whether it can continue to justify paying for Anthropic’s models through the rest of the year without containing its costs.
Earlier in the week, carnival barker and Salesforce CEO Marc Benioff said his company would spend $300 million on Anthropic tokens in 2026, and as I discussed in my premium from Friday, unrestrained AI spending is inflating the revenues of Anthropic and OpenAI in a way that isn’t sustainable for anybody involved:
For example, sources with direct familiarity with Stripe’s internal cost have told me that its technical staff (a little over 5000 people) are burning an average of $94,000 a day (around $2.8 million a month) in tokens, primarily on Anthropic’s coding models. Stripe’s EBITDA revenue was around $1.6 billion in 2024,so $33.6 million a year isn’t necessarily life-threatening, but if we assume an average salary of $150,000 per member of technical staff, that puts its raw headcount costs at around at least $765 million, making AI costs sit at roughly 4.392% of headcount.
As I said, this is one of the more-normal examples. Goldman Sachs reported a few weeks ago that AI costs are approaching 10% of total headcount costs, and “…could be on track to be on par with headcount costs in the next several quarters based on current trajectories.”
The problem is simple: nobody actually knows how much AI is going to cost them in any given quarter. This means that the current token spend you’re seeing is entirely experimental, which is why organizations keep burning through their tokens so fast.
This massive growth in spend is what underpins the “massive” (I have serious questions about its accounting) growth in Anthropic’s revenue. Executives have, across the board, given their engineers free reign to burn as many tokens as they’d like, and while I severely doubt that Anthropic actually hit $50 billion in annualized revenue outside of not-quite-fraudulent non-GAAP measurements, I believe its revenue growth has come from an artificial boost from a tech industry searching for a reason to pay somebody money.
To be very clear about what I mean, I think there is currently an AI token binge across both Anthropic and OpenAI. Enterprises do not know the actual value of AI, and do not know how much they should actually be budgeting, which is why Uber and others are running through their token budgets but not, it seems, spending less. We’re currently in an abundance phase — one where nobody is truly thinking about the costs outside of their fear of missing out — but there’s this nasty undercurrent of “wait, how much does that cost?” followed by “oh, fuck, well…you know I love AI but…”
Put another way, the current spend on AI tokens is not something that’s indicative of lasting, reliable revenue.
In some cases, the pressure to use AI for everything is turning companies’ software stacks into slop.
Zillow Spent Over $1 Million On AI Tokens Through Q1 2026, And Is On Course To Spend $7m to $10m In The Entire Year (20%+ of Its 2025 Profits)
Things are worse elsewhere. Something is wrong at Zillow. Something about LLMs has done something to its technical leadership, something that makes them talk strange and send weird slide decks with confusing, slop-ridden sentences.
The real estate tech firm spent over $1 million on AI services in the first quarter of 2026, and in April it spent $749,000 in tokens across Cursor and Anthropic’s services, as well as through AWS Bedrock. As of the end of the month, it was nearly 75% of the way through its annual Cursor token budget of $1.1 million.
As of the middle of May, its total AI spend had already crested over $300,000, and its Cursor budget sat dangerously close to the edge at 85%.
This is particularly-concerning when you consider that Zillow’s net income for Q1 2026 was $46 million, and ranged from $2 million to $10 million each quarter of 2025.
Zillow is currently on course to spend at least $7 million on AI in 2026, and at its current pace might hit as much as $10 million, which would amount to a little less than 50% of its 2025 net income ($23 million).
You’re probably wondering how Zillow manages to spend so much on AI, and the answer — as I’ll get into in next week’s free newsletter — is that its technical executives appear to have AI psychosis, saying that the short-term goal is for “software engineers to never open a code editor again.”
The reality is chaos. In a slide deck that I’ll discuss later, Zillow revealed that while engineering resources have largely stayed the same, outputs requiring human review have increased by nearly 50%. Meanwhile, code deployments and pull requests increased by 39%, and software reviewer load increased by 29,000 hours each month, creating a massive burden on the 1,500 or so engineers working at the company.
In simpler terms, that’s about 19 hours of extra work per engineer that’s literally just looking at extra code written by LLMs.
On Blind, the anonymous social network for tech workers, Zillow workers complain about Zillow’s code “slowly becoming AI slop,” with “much more code getting approved without guardrails or input due to people not being able to keep up the other’s velocity or just not caring anymore.”
One worker claimed that “the slop is job security,” adding that they “don’t want the output to be good or documentation to be clean [as] management will replace [them] with offshore/nearshore/AI agents at the slightest whiff of evidence that the slop cannon is self sustaining.”
Another said that they felt “lost in the agentic world,” and that they “didn’t have full grasp of where we are going or what [their] role is,” with a “lot of overlap in what people are doing.” Another said that “people are burning tokens just to hit internal AI adoption targets,” adding that “this is what happens when leadership ties metrics to usage instead of outcomes,” saying that it “literally subsidized busywork.”
This is all part of what an internal slide deck viewed by this publication called “AI-Native Engineering,” promising a “path to an agentic Zillow” and “faster outcomes for customers,” though customers are never mentioned in any other slide.
The deck — pumped full of AI-generated text — talks about “generic AI being a commodity,” saying that “Zillow-aware AI is a competitive advantage,” and at no point explains what that means. It encourages engineers to go from “AI-Assisted” to “AI-Native,” with “systems enabling org-wide leverage,” with engineers moving from being “soloists” — individual developers with AI tools — to “conductors” that orchestrate AI agents, to “composers” that “define systems AI can safely play,” adding that “2026 is the transition from conductor to composer.”
Yet the strangest part is named “2027: A Tuesday,” discussing a theoretical day in the office for whoever is left at the company.
- It’s an example of a typical working day. At 8:30AM, the engineer notes that confirmation rates in Dallas dropped 3% overnight.
- ‘Dallas inventory spiked; buyers went from 3 showings to 7. The agent shows the pattern: we’re hitting the same buyer 7 times in 24 hours with “tour confirmed” pings. They’re overwhelmed; they’re muting us.’
- The line before this says: “I don’t open the codebase — I open the spec and eval dashboard.”
- Half an hour later, the engineer changes the spec, which is then tested against previous data, showing an improvement.
- “The PM and I review diffs, check guardrails, approve.”
- Diffs are “differences” — essentially comparing two versions of the same document to see which lines have been changed.
- The code is then rolled out.
- At 11AM, the senior engineer mentors a junior engineer:
- ‘A junior engineer’s rescheduling agent is failing evals. I ask one question: “What happens if the buyer picks a slot the seller just blocked while the agent is negotiating?” We identify the race condition and add a constraint: “Always re-check availability at confirmation time.” She updates the spec and evals. The agent passes.’
- It is absolutely adorable they’re pretending that they’ll have junior engineers if this hellscape vision comes to life.
This theoretical example is, apparently, a process that would take weeks, but now takes under two hours.”
Zillow intends, based on this deck, to sacrifice everything to AI — code review, vulnerability fixes, policy checks, deployments, testing, and basically having agents take over everything, no matter how small, like having an agent do dependency updates and security hotfixes that could be handled with a simple shell script.
To quote Zillow:
Agent capabilities exist across the entire software development lifecycle-from ticket to production-with humans steering and approving rather than executing each step.
In practice, sources at Zillow tell me that there has been no actual movement toward this vision. Software engineers still open IDEs and review code manually, with one describing Zillow’s “vision” as “nonsense,” adding that “you can’t just throw buzzwords on a slide deck and change how all the engineers do their jobs.”
As for why token burn is so high, sources tell me that engineers are actively encouraged to use AI for everything, as much as possible, writing PRD (product requirement documents) in AI, then using the AI to make stuff based on the PRD, then doing a deck with AI, then writing emails with AI, using AI to brainstorm, or create weird, esoteric automations, with some managers pushing workers to have one personal AI “goal” to aspire to.
Zillow’s agentic “vision” is apparently a remit from the C-suite.
It’s hard to tell if this is AI psychosis or just classic Business Idiot bullshit.
Perhaps it’s a little of both.
Every organization I’ve talked to has exceeded or is nearing the edge of their annual token budget barely five months into the year, which means that everybody has suddenly given themselves an extra few million dollars’ worth of operating expenses for reasons that escape effectively everybody I’ve talked to.
Every engineer tells me the same thing: “I’m being made to do this, I don’t want to do this, my managers do not seem to understand, my bosses seem to understand even less than my managers, and if I don’t use AI somebody is going to fire me.”
Put another way, CEOs and CTOs are screeching at their underlings to “use AI as much as possible” to “find its incredible benefits” without anybody really knowing what those are and how much it’ll cost to get there.
Anthropic Is Deliberately Obfuscating Enterprise Data, Driving Up Costs In The Process
This might be because Anthropic obfuscates the data that might tell customers the real costs.
Per Laura Bratton at The Information,
One reason Anthropic costs are tough to predict, ServiceNow chief digital information officer Kellie Romack told me, is that Anthropic doesn’t automatically show customers the kind of granular data that allows them to see which of its users consume which tools; how much they use the tools, and how they’re using them. Software firms such as ServiceNow, SAP, Microsoft and Workday offer such “telemetry” data to their customers, she said.
Bratton’s article has numerous quotes from executives saying that Anthropic lacks transparency and granularity into the ways that tokens are being burned across an organization, in a way that I think sounds very, very suspicious, particularly when you add the following:
Anthropic also doesn’t offer so-called service-level agreements with customers that define how well the product will perform and customer-service response times that the customers should expect, Romack and Mehta said. Such agreements are standard in the software industry.
While I’m not accusing Anthropic of anything untoward, massive, multi-million dollar contracts that involve individuals burning thousands or tens of thousands of dollars’ worth of tokens with no service level agreement, transparency or true granularity into the burn is a perfect setup for a company — not saying it’s Anthropic! — to do something dastardly with those numbers.
While an individual might be able to monitor their own personal usage, in an organization of hundreds or thousands of engineers, who’s to know if, say, the particular token burn is consistent across every member of the company, or that those costs are actually matching up with what the user is doing?
This is a company ostensibly worth $900 billion dollars acting with disregard for the basic measurement of “how much did this cost, and how did it cost so much?”
Every Single AI Token Budget Is Bullshit Because You Can’t Measure How Many Tokens A Task Will Take
And in the end, how do you even measure it at scale? Say you’ve got 1,500 engineers, and they’re spending a combined $1 million tokens a month. How the fuck do you actually measure the return on investment for that spend?
How many tokens does it take to do one thing? Is it consistent across every model? Is it consistent across every employee? Are you even measuring how many tokens a task costs? Because if you’re not, that token budget is basically throwing a dart blindfolded.
Okay, now you’ve measured a task, did you make sure to measure it multiple times? Because LLMs can randomly do things differently even with the same prompt and same Claude.MD file and same strictures and same data sources. You’re gonna need at least 10 samples of each task, and you’re gonna need to make sure somebody who actually knows what they’re doing can measure them, because if you get a dimwit, they’re going to say it can do something it can’t.
Unless, of course, you can’t actually measure how many tokens a particular task can take with much accuracy, in which case every single AI token budget is bullshit. And each model does things differently depending on many different variables, some of them a result of the user, some of them a result of the AI labs themselves.
Alright, well, maybe you just need KPIs — measurements you can aspire toward, and by pursuing them you can start working out how much it costs to do stuff.
Wait, which metric works there exactly?
- You can’t say “burn as many tokens as possible,” because employees will — as happened at both Amazon and Meta — deliberately create ways to burn more tokens using scripts and automations.
- You can’t say “use AI every day,” because even if they do so, that doesn’t actually set up a success criteria.
- You can’t tell software engineers to try and “ship more software,” because that, again, emphasizes doing more, not making good stuff, and leads to an increase in velocity rather than how good the stuff is.
- You can’t say “pull requests” or any other metric a software engineer can manipulate, because in 100% of the situations where you give a software engineer a number to hit they will focus entirely on hitting that number.
In fact, it’s pretty hard to measure anything like “efficiency” or “productivity” in any business, because every metric connected to them can be gamed, leaving managers and executives with the problematic situation where they have to start learning how things work so they can see if they’re good.
Before AI, this wasn’t as much of a problem, in the sense that inefficiencies and wasted hours weren’t directly connected to a chatbot that is specifically designed to burn money. Managers and executives could come up with whatever deranged, self-gratifying office bullshit they pleased, wasting hours of people’s time in the process, but doing so didn’t immediately connect to a massive, ever-increasing cost.
AI is a perfect storm of failed concepts and organizations, and the apex of the Era of the Business Idiot, an epoch where we’re ruled by people so thoroughly disconnected from the actual workforce that it was inevitable that a technology would be created specifically to grift them.
LLMs are dangerous for many, many reasons, but the under-discussed one is how well they play to a certain kind of executive imbecile. Generative AI is — to quote Mo Bitar — really good at doing an impression of work, much like most managers and c-suite executives, and even if it’s completely incapable of doing something, it’ll absolutely say it can and tell you you’re amazing for suggesting it.
And that’s why Business Idiots love it.
Where regular human beings would say annoying things like “that’s not possible within that timeline” or “we don’t have the resources to do it,” AI will say “of course, right away!” and burn as many tokens as possible.
When it makes mistakes, it’ll apologize — as it should because it failed you — but then promise to do better next time, all while costing so much less, at least in theory, than a regular, stinky human being.
It’ll create a PRD of a theoretical software project with the confident and vigor that you need to take it immediately to a software engineer and say “build this immediately,” and when the software engineer tells you a bunch of bullshit about it not being possible, it’ll spit out several convincing-sounding responses. Fuck, why even bother talking to that engineer at all? Claude Code can mock up a prototype that you can then shove in their fucking face before you fire them for not using AI to do it themselves.
Any executive-level fuckwit you’ve met in your life now has a seemingly-powerful tool that can burp up mimicry of open source software and, if you constantly prompt it, eventually get something half-functional onto some sort of web server. When you face bugs, it’ll try and fix them, sometimes also “fixing” (adding or deleting code) from elsewhere to be helpful, like when Cursor using Anthropic’s Claude Opus 4.6 model deleted an entire production database and all its backups. It will never, ever say no, even if it’s incapable, even if it has no thoughts, even if what you are asking is equal parts impossible and unreasonable in both its timescale and scope.
A Business Idiot, given his druthers, can sit there and fuck around and make an LLM spit out something that makes him feel like he’s coding, which in turn makes him feel that you, a lazy and stupid engineer, could do even more with the power of AI. It doesn’t matter that it costs an absolute shit-ton of money, or that there’s no way to measure its efficacy. The Lion does not concern himself with things like “efficacy” or “productivity,” and the Lion is increasingly tired of your whining! The Lion doesn’t even understand what it is you do every day other than not doing what The Lion is asking for!
You laugh, but this is genuinely how the majority of managers and executives think and act, and now they have a special chatbot that can fart out functional-enough prototypes to convince a Business Idiot they can do anything, because executives and managers do not regularly do much work and thus have no idea what it looks like other than when they look over your shoulder, which is why they wanted you back in the office!
Organizations aren’t burning millions or hundreds of millions of dollars a year on AI because it’s good, they’re doing it because they are run by people who do not know what the fuck they’re doing.
In a sane world, randomly adding a massive, ever-expanding operating expense to your business with the express intent of — to quote IT firm Workato’s CIO, “eating the costs while employees experiment” — would have the board blow up your house. In our world, one dominated by disconnected, self-involved and massively-overpaid dullards, many businesses pushing their workers to use AI are doing so because the other guy is doing it, with about as much strategy and forethought as one would expect from somebody who spends 90% of their life reading emails, going to meetings, or going to lunch.
The majority of those I see trumpeting the so-called benefits of AI do not appear to do anything of note. I have yet to see one so-called multi-agent orchestrator engineer psychopath ship something remarkable or impressive or even functional. I have yet to see any AI-obsessed boss write or create or author or do anything I can remember. I don’t see any of these fuckwits running a company on their own outside of those who have learned to sell stuff to other AI psychosis victims or executive midwits of varying size.
And why oh why is it always the language of inevitability and possessiveness? Nobody who’s this insistent, aggressive and violative with their language of “it’s here and if you don’t adopt it you’re stupid and dead” has ever been right about anything. Nobody this desperate, insistent and forceful has ever had good intentions, good vibes or brought good omens — they are always bearers of some kind of con.
Most technology is sold on elevating and ascending human beings. AI cheapens every interaction by creating a work-shaped product from a person that doesn’t respect you enough to give you work that’s barely fit for a human because it wasn’t made for one.
You must accept becoming a dogshit dealer that loves accepting and receiving low quality goods. You must celebrate intentionless and decaying slop, and defend it and the machine that made it with your entire being. You must sully yourself — treat its unexceptional, sloppy and unreliable outputs as signs of sentience, or at least the proof that digital sentience is possible. You must defend horrible, abrasive, ugly, loud monoliths of steel full of $50,000 graphics cards. You must say they are necessary, and you must aggressively antagonize those who do not.
Every time you defend generative AI you defend a machine of capital that has burned $1 trillion and created one of the most-wasteful products in history. If people disagree with you, you must attempt to harm them somehow — ostracize them, mock them, attack them, denigrate them. You will justify this as moral, because you have been manipulated by a technology built and sold by two of the greatest grifters of all time — Dario Amodei and Sam Altman.
Anything less is opposition to an industry with all the trappings of authoritarianism down to the media toadies, the propaganda and the seizure of land in the name of a nebulous “greater good.”
But man, these men got people good.
Sam Altman helped propagate a technology perfect for conning people with potential, a larger extrapolation of Altman’s own life of taking dogshit — Loopt, for example! — and parlaying it into larger opportunities. It can make a really half-hearted demo of a lot of things, and that’s good enough to sell to Business Idiot.
Dario Amodei took this grift and perfected it. Anthropic is a company purpose-built to con people into giving it by money by making people feel smart. LLMs can do work-shaped stuff, sometimes, as long as you debase yourself to accept mediocre and often-broken stuff that you have to keep a vigilant eye on, and either use a subsided product that loses Anthropic money or pay a shit ton of money as an enterprise to Anthropic and they still lose money.
These companies were only capable of growing in an economy dominated by the gullible and work-shy. Only a capitalist culture dominated by people who don’t actually do or know stuff have let this get so far. Nobody wants this, nobody wanted it since the beginning, it was forced upon everyone, and to pretend otherwise is laughable and offensive. The amount of people who use this shit a bit and become convinced that we’re mere years from it costing over a trillion dollars to somehow making trillions of dollars and being an entirely different and good product should be aware that they are being manipulated. The more you feel compelled to defend AI the more scrutiny you must show it.
I am not your enemy! If you think that I am, you are on the side of a corporation or a product. You can try it, like it, and I don’t really care, but the second I see you trying to be condescending or judgmental or aggressive toward another person for not agreeing with your product choices I immediately feel suspicious. Can’t you see how these people act? Can’t you see how strange it is to defend a thing you pay money for that has terrible economics? If it wasn’t the “in” thing, being an AI person would be considered really weird. I look forward to the day it is. I hope you guys like having the stuff you said since 2022 repeated back to you! I’ve been saving it all. Time is running out for a graceful bow, and you better act quick!
If you feel self conscious while other people dunk on AI, that’s weird! I see people say they don’t like Macs all the time. Who gives a fuck! I’m not going to go to the mat for Tim Cook. People can make their own decisions.
Those comparing AI to AOL mailing CDs to people should feel ashamed of themselves. This is like if every single time you opened a magazine an AOL CD flew at your head, your boss told you he would replace you with a modem if you didn’t go online, and the news constantly ran segments called “I didn’t receive an email: father forgets son forever because he wasn’t online” or panels with “Internet experts” who said “I am on the Internet superhighway right now, and I’m certain that within 10 years AOL Time Warner will be able to email myself to my dad.”
Imagine if Shingy was a billionaire and went on TV every day in 1999 and told you “the world must get ready, because you’re about to get a ICQ message from The Lord.”
Generative AI was purpose-built to grift an economy run by executives and managers who don’t actually do any work. Its success has been driven by a remarkable, society-wide ignorance in the management sect, and its continued proliferation is only possible through the media’s continued trust and faith in the idea that CEOs are busy because they’re actually doing work.
Yet even a Business Idiot eventually realizes that too much money is being spent, and the first one of these dimwits to cut their token budget will send the rest of them running for the doors.
We should lock them. We should make everybody who obsessed over theoretical ideas about what AI can or will do ashamed for their intellectual deceit or constant ignorance.
At the end of the AI era, the only thing that will change the rot at the heart of our economy is the acceptance that the majority of companies are run by lazy, self-involved and ignorant fuckwits, and accountability for those who refused to scrutinize them.
