Ed Zitron 长文剖析 AI 泡沫如何靠年化收入等虚假数字支撑

If you liked this piece, you should subscribe to my premium newsletter, and you can subscribe on the following links: $70 a year, $18 a quarter, or $7 a month.

In return, you get a weekly premium newsletter that dives deep into the finances of the AI bubble. Over the past few months, I’ve published a two part guide to the world of AI debt, an analysis of the (dubious) impact of generative AI on the global economy, deep analyses of the biggest players in the AI infrastructure space like NVIDIA (part one and two) and Oracle (part one and two), and the role private credit is playing in funding AI data centers, and how private equity, reeling from years of successive failed software acquisitions, is doubling down on the AI bubble. 

On Friday, we’ll take a look at the JOBS Act — or, the Obama-era legislation that created the regulatory environment that would ultimately prove essential for the AI bubble to inflate - and why it might be time to reform or entirely repeal it.

If you want to get in touch — and especially if you have any juicy information about Anthropic, OpenAI, or any other companies in the AI bubble — hit me up on Signal at ezitron.76. I’m also on IB on your Bloomberg Terminal. 

Soundtrack: Cleopatrick - Hometown

"I think the like, state of being an entrepreneur is like, how well do you make decisions in an information-poor environment, right?" - Mark Zuckerberg

The entire AI bubble is inflated based on stupid, made up numbers that only exist to pump the bags of investors and make the AI startups look more successful than they are. It is a failure of journalism that they have been popularized, as they do not tell us very much about anything, other than that there’s a new, bigger number to get excited about.

Last week, the markets dropped on the news that OpenAI’s annualized recurring revenue (ARR) — which was reported as $70 billion by Axios and confirmed by the Financial Times — was actually $50 billion. The number in question was, according to the FT, inflated by “grossing up” the total to include sales through cloud partners, a discrepancy between how OpenAI and Anthropic make up a number that doesn’t really tell anyone anything:

Anthropic recognises gross revenue on sales through cloud partners. That is because Anthropic is the principal in the transaction, according to a person with knowledge of its accounting practices. The Claude maker’s revenue run rate was calculated by annualising the prior 28 days of consumption-based revenue, plus recurring monthly subscription revenue, they added.

OpenAI, meanwhile, counts only its own cut of proceeds from arrangements with partners such as Microsoft.

The $70bn figure reported in September arose from attempts by investors to compare Anthropic and OpenAI performance on a like-for-like basis, according to a person close to OpenAI. However, OpenAI described the $70bn as a “net” revenue metric when it was approached by the FT at that time.

Back in August, Anthropic had told investors that it had hit $65 billion in annualized run rate, but at no point defined what that meant, what period it referred to, what any of the underlying numbers are, or what “gross” versus “net” bollocks it was playing. Per Irrational Analysis the calculus might have been even sillier than that:

Last month, there was a whole kerfuffel in AI/semis/finance circles on Anthropic July ARR. Two numbers were going around. I don’t remember the numbers and frankly it does not matter. You will see.

One ARR number was the traditional “trailing 28 days * 13” number. Personally I hate this venture-capital clown metric but whatever a lot of people use this.

The traditional ARR number was bad and implied deceleration in growth. So the people massively long Anthropic came up with a new ARR number that was July 31st * 365 days.

This is the problem of saying what the revenues of a company might be based on numbers that are not its actual revenues. There is absolutely no value in using “annualized run rates” for anyone other than the company and its investors, and the fact that modern tech and business journalism continues to accept it for what are meant to be the largest and most-important companies in the world is a sign that something is very wrong with how we cover the tech industry.

Sidenote: When one “annualizes” a period of revenue, the intention is usually to take a time period that’s indicative of the consistent performance of the business and extrapolate from there what it might make on an (as you’ve guessed) annual basis. This calculation is useful only when connected to specific periods and forms of revenue — recurring subscriptions and consistent contracts rather than, in the case of the AI bubble, consumption-based revenues that are not ‘recurring’ and thus mean very little when annualized.

The other problem is that these run rates do not measure if revenues go up or down, because you’re choosing the best period possible.

I am not going to spend much time dwelling on gross versus net. It appears that, per Reuters, 47% of Anthropic’s 2025 revenues came from sales made by Google and Amazon, who take a 16% cut as opposed to the roughly 20% that Microsoft takes from OpenAI. It appears that Anthropic for whatever reason counts every dollar of sales as part of its annualized run rate figure and OpenAI does not, for reasons that escape me, as I simply don’t buy that it is “more ethical.” 

My evidence? The following quote from the FT, emphasis mine:

The $70bn figure reported in September arose from attempts by investors to compare Anthropic and OpenAI performance on a like-for-like basis, according to a person close to OpenAI. However, OpenAI described the $70bn as a “net” revenue metric when it was approached by the FT at that time.

This is the same OpenAI that tried to claim that a $34 billion loss was actually an $8 billion loss when the FT asked them about my reporting on its audited financials.

I must also say how utterly disappointed I was to see how Bloomberg credulously reported that OpenAI “expected to hit $70 billion in annualized revenue by end of 2026,” a transparent attempt to send the markets back up with a story that tells the world less than nothing. 

Wow, OpenAI, by the end of 2026, will increase a number that “represents a projection of [its] yearly sales based on a shorter period” by $20 billion! What is the period in question? Who knows! How is that number derived? Who cares! Number go up! Put the story in the bag! Your job is not to cover the tech industry, it is to celebrate it! Stop thinking!

Sidenote: another very obvious reason these companies use these numbers is SEO! If you try to Google OpenAI’s projected revenues, you’re guaranteed to get one of these stupid annualized figures. In reality, the company projects to make somewhere in the region of $30 billion in 2026. 

In other words, OpenAI’s made up number was actually lower than Anthropic’s made up number because of the way it made the number up, but by the end of the year, the made up number would be bigger than the other made up number. These made up numbers are reported by the largest and most well-respected news outlets of the world, who lend said made up numbers respectability despite the numbers in question being utter bullshit.

I’ll give you an example.

For the first half of 2026, OpenAI made $12.4 billion in ($5.7 billion in Q1 and $6.7 billion in Q2) revenue per The Information and the Wall Street Journal, while Anthropic, per the WSJ and the FT, made $15.5 billion ($4.6 billion in Q1 and $10.9 billion in Q2). We do not have consistent coverage of their underlying costs, but in Q1 (per The Information), OpenAI spent $8.6 billion on training and $3.5 billion on “cost of revenue,” with no reporting on its sales and marketing spend where it allocates the cost of serving free users or discounts. 

Anthropic said it had hit an annualized run rate (undefined period multiplied by undefined number) of $48 billion on May 28, or roughly 430% more than its entire revenue in the second quarter. In fact, if you added up every dollar that both Anthropic and OpenAI made in the first two quarters, you’d be $19.1 billion short of Anthropic’s May “run rate.”

We do not know how any of these numbers were actually conceived outside of they are somehow derived from a non-specific 28 day period. The information they convey is sparse, if it communicates anything at all. 

Reporting “run rates” is a marketing exercise, and if you disagree, please answer me the following: what does a “$65 billion annualized run rate” tell you about Anthropic, exactly? It’s MEANT to make you think it will make $65 billion a year in revenue, but that is not what it actually means. This number tells you nothing, and exists only to poison the well of information. 

Without an actual period of time — and I mean an exact one, such as “July 1 to July 31, 2026” — and the exact way it was calculated, this number effectively means nothing, and reporting it without this information only serves OpenAI or Anthropic. You can say “investors use it all the time” all you want. Investors are fucking stupid for using it. If the argument is “it shows how fast OpenAI is growing,” why don’t you find out the actual growth rate of the company? 

The fact these numbers moved markets is a reflection of the definitive — and accepted — ignorance built into the entire AI bubble, where despite trillions of dollars on the line, we know very little about the actual underlying numbers, and those numbers we do know are inflated using techniques that would get you shot by an accountant.

In fact, we don’t know most of the numbers that would actually tell us whether any of this is a good idea.

The AI Bubble Is Built On Bad Data, Because The Real Numbers Kinda Suck

Reporting on the true revenues of the AI industry feels less like financial analysis and more like legal discovery, and it’s fundamentally insane how little tangible, meaningful information that exists to feed the trillion-plus dollars in capital expenditures and endless full-court press in the media about how AI is the next big thing. 

We Do Not Know The Actual Revenues Of Most Of The AI Industry

We do not know Google, Amazon or Oracle’s AI revenues, and we have a single year — fiscal year 2026 — of Microsoft’s AI revenues based on an estimate from Bloomberg calculated from its disclosure of its OpenAI revenues in an SEC filing. We do not know how much each individual service makes, how much their true revenue shares are with the AI labs or, as I discussed a few weeks ago, how much actual, operational AI capacity exists as a result of the trillion-plus dollars in capital expenditures, though Bloomberg and I have both reported that Microsoft has around 2GW of it, even though it claimed to bring online a gigawatt in each of its last three fiscal quarters. 

We do not know the revenues or much about the economics of [include funding totals) Perplexity, Harvey, Instinct, Reflection, Cognition, Replit, Lovable, and Factory, all of whom have raised somewhere in the region of $15.8 billion in funding, and have a combined valuation of $215.8 billion. We do not and will never know the true revenues or costs of Cursor’s operations as a result of its acquisition by SpaceX, and as a reminder, annualized revenues are not actual indicators of the economic health — the revenues, costs or cash flows — of any of these companies. 

Perhaps this was permissible in the past because so much less capital was being annihilated in the name of startups, but things have escalated so much further than any other era, including the Dot Com bubble.

Look, the entire AI bubble has been inflated based on the implicit belief that there will be demand — from AI startups and elsewhere — for AI compute, to the point that Goldman Sachs says that hyperscalers need AI startups to make $425 billion in annual revenues just for the hyperscalers to break even…and we don’t even know, outside of Anthropic and OpenAI, how much any AI startup actually makes.

What we do have is “annualized run rates” pegged to non-specific time periods and non-specific numbers that tell us, for example, that Cognition hit $1 billion in late September, a number that means next to nothing. Even when The Information was able to report that Cognition would burn $800 million in 2026, it resorted to saying that it was “generating around $900 million in annualized revenue, or $75 million a month” as of August 27 2026 — adding that executives had “projected it would hit $1.5 billion in annualized revenue by end of year,” meaning that two journalists were able to successfully find out its actual expected burn, that it burned $200 million in Q2 2026…but not the actual revenues that the company generated on a quarterly basis.

Let’s be clear about something: the reason we’re not getting these numbers is that these companies are not actually making these amounts on an annual basis. If Cognition was actually going to hit anything close to $1 billion — and let’s be honest, it’s likely less than $500 million! — it’d leak it versus using these weird monthly snapshots.

It is astonishing that we, four years into the AI bubble, do not know the revenues of more than a handful of AI companies outside of OpenAI and Anthropic. All we have is a mishmash of annualized run rates, and at this point I have to wonder if it’s a failure of journalism, successful opsec, even more successful marketing, or that these companies don’t have great accounting. 

But really, go and look up basically every AI company. Perplexity? Nothing. Harvey? Nothing. Cursor? Years of reporting — endless thousands of words spent — and nary one report I can find, outside of these fucking annualized run rates. 

AI Startups Are The Dirty Little Secret Of The Bubble — Raising Billions of Dollars To Make Millions of Dollars, Making Up Around 80% of OpenAI and Anthropic’s Revenues By Spending 5x+ Of Sales On AI Tokens

Part of the reason, I imagine, is this chart from Ramp that I’ve brought up before:

From what I hear, the vast majority of those 1% of customers are AI startups that power their products using OpenAI and Anthropic’s models, predominantly using monthly subscriptions that allow users to burn way more in tokens than their subscriptions actually make.

For example, SemiAnalysis did a deep study into the amount of tokens you could burn on the various different AI subscriptions out there, finding that you could burn anywhere from $83 to $118 a month for $20 and from $1,264 to $1,331 a month for $200 on Cognition, dropping to $46 on a $20-a-month plan and $507 for $200-a-month on Cursor when using Anthropic’s models.

While OpenAI’s models are being dropped from Cursor, SemiAnalysis found you could burn $154 a month using OpenAI’s models on Cognition’s $20-a-month plan, and $1,736 a month on its $200-a-month plan.

The various boosters reading this and making noises like the Skesis from Dark Crystal will say “but these companies get discounts!” as if that makes up for the fact that customers are able to burn two to six times their actual subscription every month. They’ll also argue that “this is the gym model, and most customers don’t spend that much,” a statement that they have no proof for other than that it makes them feel better. 

The problem — and the reason I’d argue that the actual revenues of these companies are so shrouded in secrecy — is that the underlying economics do not match up when you consider that effectively every AI startup has to resell tokens at a discount. It doesn’t really matter how much revenue you make if you’re sending twice or thrice or five times that amount out the door to Anthropic and OpenAI, but saying a larger “annualized run rate” makes the amounts you’re raising far less worrying than what I imagine are particularly-shrimpy sales.

AI Startups Are An Extraction Machine That Turns Venture Capital Dollars Into Anthropic and OpenAI’s Revenues

Let’s take Harvey — which raised $550 million at a $15.5 billion valuation in September — as an example. As of early August, Harvey had hit $350 million in annualized revenue, up from $190 million in January. If we assume it uses the same “trailing 28 days” model as Anthropic, this means that it made $26.9 million in a 28-day period ending sometime in August, up from $14.6 million in a period that ended some time in January 2026.

Since December 2025, Harvey has raised a total of $910 million across three rounds, or roughly 33.8 times the estimated underlying revenues of its most-recent run rate, which increased by 45.7% in that period.

Bloomberg claimed in late September that Harvey’s Gross Margins had gone from 50% in January 2026 to negative 50% in June 2026 to “positive again” (I imagine in September) due to using open source models, but something about those numbers is deeply weird to me, mostly because I’ve never, ever heard of a company having had that happen. 

And none of that explains why a company that makes tens of millions of dollars in revenue needs to raise hundreds of millions of dollars every few months, unless, of course, most of those dollars are flowing immediately to Anthropic and OpenAI, which is, it turns out, exactly what happened:

The $15.6 billion legal startup Harvey built its business around training AI models like OpenAI’s GPT-4 to do specialized work for lawyers. Recently, however, soaring artificial intelligence costs have nudged it to rethink its dependence on the AI giants.

After a March update to its AI agents, Harvey’s customer usage spiked, but its gross margins dropped sharply — from about 50% at the beginning of the year to -50% by June, according to a person familiar with the matter.

This is the antithesis of any good business. In basically any industry, the more demand you have, the more money you make, but in AI, the more demand you get, the more money you lose, and your most engaged customers are also your least-profitable. 

However you might rationalize the underlying economics of a company like Harvey, the numbers we do know are extremely worrying — $900 million raised since December 2025, and, in my most-optimistic calculation, around $175 million (though likely less than $150 million) in revenue through the end of August, for a company that raised another $600 million across two other rounds in 2025. That’s $1.66 billion in about a year and a half to make an unprofitable company with less than $30 million in monthly revenue that is somehow worth $15.5 billion.

Harvey has categorically failed to find product-market fit, unless we qualify that as “literally finding a customer.” It’s very clear that its underlying product cannot support itself, and that the company is entirely-dependent on venture capital, raising at least $160 million every three to six months since February 2025. It is abundantly clear that Harvey’s customers — like every AI startup — do not actually want to pay what its services cost, which means it does not have product-market fit, nor what I’d call a “business.”

Then there’s Cognition, which has raised a combined $3 billion in 2026 alone (in May and September), with its run rate going from $492 million to $1 billion in that period. If we assume the same calculation as Anthropic uses, that means revenues were $37.8 million and $76.9 million respectively, and that its last two rounds of funding are roughly 39 times the size of its 28-day trailing revenues. It is unclear what has substantively changed about the company, other than that the $400 million it raised in September 2025 was insufficient to last.

I know there are going to be venture capitalists who argue that “these companies are just raising opportunistically,” but that doesn’t explain why every AI startup raises so much money in so little time, and how one reconciles that distance with any reality outside of “we can raise unlimited money forever.” 

And if they can’t do that, large chunks of OpenAI and Anthropic’s revenues disappear. 

Anthropic and OpenAI Are Only Growing Because Of AI Startups’ Perpetual Unsustainable Spending On Tokens — Once VC Dollars End, So Does Their Growth

Let me summarize all of this in simple terms, as it’s all terribly important.

• The Revenues
  • 80% of Anthropic and OpenAI’s revenues come from 1% of their customers, the vast majority of which are AI startups that spend anywhere from 300% to 1000% of their revenue on AI tokens.
• AI Startups Run Services at a Loss, Allowing Users to Burn Thousands of Dollars of Tokens for Hundreds of Dollars a Month, Requiring Constant Venture Funding
  • Customers of AI startups are repulsed by the actual cost of tokens, meaning that for every dollar of subscription revenue, AI startups must allow customers to burn anywhere from 2 times to 10 times their revenue in tokens. For example, Cognition’s $200-a-month users can spend upwards of $1200 a month in OpenAI or Anthropic tokens. 
  • Even with these “discounted” tokens, AI startups have very small revenues.
  • As a result, every AI startup runs at a loss. This necessitates raising massive amounts of capital every few months.
• Every New Model Release Artificially Boosts Revenues Through AI Startups
  • When a new model releases, every AI startup must immediately integrate it, because that’s what their customers expect, and because the customers in question are curious, they’re likely to use the new models more on release.
  • This means that Anthropic and OpenAI get a massive bump of new revenues — funded by AI startups that are allowing users to burn tokens without restraint — with every release.
• All Of This Is Only Made Possible By Perpetual Venture Funding, And Once It Expires, It Eats Anthropic and OpenAI’s Revenues Alive
  • All it takes for one of these companies to stop being able to spend tens or hundreds of millions of dollars a month on OpenAI and Anthropic’s services to meaningfully reduce their revenue.
  • It doesn’t really matter what big businesses are or are not using Claude or ChatGPT — the underlying revenues of these companies are held up by venture capital by proxy.

Every AI startup is effectively an extraction mechanism for OpenAI and Anthropic, reselling their tokens through their services at a massive discount because users are unwilling to pay the actual cost of using LLMs. 

Using annualized run rates allows everyone to feel like this isn’t a problem, because it’s far easier to rationalize that Cognition has raised $3 billion this year and is worth $48 billion if it has $1 billion in annualized run rate rather than $80 million a month in revenue. 

The obfuscation of the actual underlying financial of these companies is intentional, because it makes it harder to do the kinds of comparisons I’ve done today, 

If you disagree, start giving me actual revenue numbers for these companies instead of your magic elf number you made up or got told. If you’re an investor in these companies and have not been told the actual most recent quarterly revenue, there’s a reason — the real one sucks! 

If you know it, leak it if it’s so impressive. You know where to find me. Only one of us is being weird, and it’s the cloister of you who condone using bad information because you know the real number — the actual quarterly or annual revenue — is probably going to make people panic. 

These companies are also, at least in theory, the ones that pay for AI compute, which means that we also have a very, very poor understanding of the actual customers behind those hundreds of billions of NVIDIA GPUs.

In fact, that’s a great place to go next.

We Do Not Really Know How Much Money AI Data Centers Cost or Make, And The Data Sucks

People are going to get mad at me for this because they are cowards, but I posit that we have little understanding of the true price that people are paying for AI GPUs, or about the underlying economics of AI data centers writ large.

Yes, yes, I know, you’re going to tell me that SiliconData has an index, and that that index shows “GPU rental prices going up,” but these prices overwhelmingly favor spot pricing — as in what it costs to rent a GPU right now — versus what companies are paying on a long term basis.

It’s kind of like valuing a shipping company based on skyrocketing spot rates during a capacity shortage rather than the year (or years)-long rates that they regularly charge. You know, kind of like annualizing the data to make things look better.

SiliconData claims to create a “single standardized reading,” but does so using a mishmash of non-specific data that “standardizes for rental term length, cluster scale and interconnect” in a way that only ever seems to benefit the providers. 

The reason that SiliconData’s index has become so popular is that we otherwise lack many good measurements, which isn’t to say that its measurements are good, just that they’re the best we’ve got at the moment. While I think the company is doing its best with what it’s got, measuring GPU rates in general like this only seeks to confuse the market through abrupt shifts upwards and downwards based on what I imagine are larger contracts ending or beginning — not that we know, because SiliconData does not actually explain what’s changed or why it might have changed.

This is a huge issue, because NVIDIA has sold $535 billion in GPUs since 2022, and we still don’t know on any reliable basis how much their customers can rent them for. We don’t know with much certainty how much it costs to run one, or have a firm grasp on what an hourly comparison would be — as in how many dollars an hour it cost to run one — outside of analyst estimates (and my own model) that all seem to vary wildly and always end up at a place that says “what a great investment!”

There’s also some compelling evidence that the actual long-term rental rates for these things might not be as high as we think.

People Are Paying Much Less To Rent GPUs Than You Think

I realize some people might read what I’m saying and immediately want to pelt me with rocks, but I’m just telling you what I’ve found, and if I’m honest, I think I’m doing a better job than the vast majority of people covering this shit, in that I’m trying to actually show what people are paying versus aggregating various publicly-available sources of information to simmer out info-slop.

Again and again, I’ve found companies paying way, way less than the supposed going rate of compute.

Krea — B200s at $2.80-per-hour-per-GPU, vs. SiliconData rate of $5.94-per-hour

In late April, The Information ran a story about startups renting AI GPUs, specifically focusing on NVIDIA’s B200s, which SiliconData claims rent at $5.94 an hour, emphasis mine:

Six months ago, Krea signed a six-month contract to rent several hundred Nvidia Blackwell chips for $2.80 an hour per chip after several cloud providers competed aggressively for the startup’s business. But in the last month, when the startup began looking for more AI servers to train a model from scratch, some sales representatives at cloud providers wouldn’t pick up the phone, said co-founder and CEO Victor Perez.

[a few months later]…In the end, the founders reached an agreement to pay $3.70 an hour for another several hundred Blackwell chips in a one-year commitment—32% higher than their last deal. And even that price seemed cheap, given the chip prices he had seen from other providers.

To be clear, the B200 GPU, as part of an 8-GPU pod, works out to about $62,500 a GPU. At this rate, 32% of the first six years of its useful life would be absorbed by depreciation, assuming that it was sold in its entirety for every single hour of every single day of those years. 

OpenAI — Stargate Abilene — GB200s at $2.54-per-hour-per-GPU, vs. SiliconData rate of $5.94-per-GPU-per-hour (for B200)

Then there’s Oracle’s under-construction “Stargate Abilene” data center that landowner Lancium believes will make around $10 billion in annual revenue (see slide 5). At $10 billion a year and 450,000 GB200 GPUs, which works out to — if we assume a take-or-pay rate — to about $2.54-per-hour-per-GPU. If we assume that each of these is inside a GB200 NVL72 at about $3.5 million a piece, that’s about $1.61 per-GPU-per-hour of revenue for the first six years before you get to the COGS.

Sharon AI/ESDS Cloud — B300s at $2.80-per-hour-per-GPU, vs. SiliconData rate of $6.65-per-GPU-per-hour

Then there’s this contract between Australian neocloud Sharon AI and ESDS Cloud FZ-LLC, a Dubai-based subsidiary of Indian enterprise cloud company ESDS, renting out 8208 NVIDIA B300 GPUs at $3.30 a month over 60 months (five years). If we assume a cost of $540,000 per eight-GPU B300 cluster, that leaves us with about $1.54 in depreciation over six years, leaving us with about $1.869-per-hour-per-GPU of actual revenue for the first six years, in a contract that is five years long.

Canopy Wave — B300s at $4.3-per-hour-per-GPU for three years, $3.50-per-hour-per-GPU for two years (total five) vs. SiliconData rate of $6.65-per-GPU-per-hour (Average = $3.98-per-GPU-per-hour)

Here’s another. A contract I found between neocloud Kidz AI (which rents capacity from a company called Limestone, or, more-specifically, its subsidiary Catalyst Compute) and AI inference company Canopy Wave priced 256 B300 GPUs at $4.30 an hour (per GPU) for the first three years and $3.50 an hour for the last two of the five-year-long contract.

Your depreciation costs here work out to about $1.54-per-GPU-per-hour, meaning that you’re losing nearly half of your revenue for the first six years…of a five year long contract.

Sure hope there aren’t new kinds of GPUs — or an abundance of older GPUs — after these things end!

QumulusAI — B300s at $3.70-per-GPU-per-hour for three years vs. SiliconData rate of $6.65-per-GPU-per-hour 

QumulusAI is a former cryptocurrency mining company that converted into renting GPUs. In this contract, Qumulus is renting out 128 B300 nodes — so around 1,024 B300 GPUs — at $3.70-per-GPU-per-hour, sitting at about $1.284-per-GPU-per-hour of depreciation.

Again, sure hope there aren’t a bunch of data centers under construction that have the same chips as these!

WhiteFiber for BaseTen — B300s at $4.516-per-GPU-per-hour for three years vs. SiliconData rate of $6.65-per-GPU-per-hour

Buried deep in the earnings report of AI infrastructure company WhiteFiber sits a three-year-long contract with buzzy AI inference firm Baseten, which raised $1.5 billion in June:

In August 2026, we entered into a service order with a new customer, BaseTen Labs, Inc., to provide services utilizing a total of 1,392 B300 GPUs. The service order has an initial term of 36 months beginning on the services commencement date, after which it automatically renews for successive one-month periods unless terminated by either party. The service order represents an aggregate revenue opportunity of approximately $165.2 million. The deployment and revenue generation is scheduled to begin in November 2026.

This all works out to about $1.284-per-GPU-per-hour of depreciation on $4.51-per-GPU-per-hour, which is still more than two dollars less than what SiliconData quotes.

As mentioned, BaseTen just raised a ton of money, so it’s willing to pay through the nose, and even then it’s paying so much less than spot.

There’s also another contract with a customer in Iceland that works out to about $3.46-per-GPU-per-hour:

In August 2026, we entered into a new five-year service order with an existing customer to provide services utilizing a total of 576 Nvidia B300s GPUs in Iceland. The service order represents an aggregate revenue opportunity of approximately $87.5 million over its initial term, with additional potential upside through revenue sharing. Service under this agreement is targeted to commence in December 2026.

We Don’t Even Know How Many AI Data Centers Exist

As I said a few weeks ago, we do not really have any firm understanding of how much AI compute actually exists, to the point that the largest companies in the world (read: Microsoft) are actively misleading the public about how much they’re bringing online.

We also have no clear understanding of how much data center capacity is actually under construction, because nobody actually wants to tell us, and reports like this from Sightline Climate claiming 190GW “in planning” lump in every imaginable kind of data center with ones specializing in AI compute.

As a result, basically any AI compute capacity estimates are guesses. I’ve seen somebody say there’s “50GW of capacity,” I’ve got guys in my mentions saying that actually Bloomberg is wrong in saying that Microsoft has 2GW of capacity and the real number is 5GW based on nebulous “conversations” that only ever end up as a positive for the AI industry. 

We Don’t Really Know The Economics of an AI Data Center Either

In a vacuum of good information, people simply start making things up, which is why The Information published one of the more insane things I’ve read in a while, emphasis mine:

In a time of server scarcity, speed matters. The Louisiana campus plans to consume 330 MW of power to produce 245 MW in usable compute capacity. If that capacity can get online even one month faster than its current schedule calls for, Anthropic could generate an extra $1.3 billion or more in revenue, based on investors’ and analysts’ estimates. That assumes it uses all the capacity to serve up answers rather than develop new models.

I asked Okoye for an explanation, and he gave me one in the comments of the article:

• Facility rent: Hut 8’s lease for this 245 MW site is $7.0B over 15 years. That is about $39M a month.
• Compute: IREN’s 5-year, $9.7B contract with Microsoft covers about 200 MW of GB300 GPUs. That is about $9.7M per
• MW-year, or about $200M a month at 245 MW. It’s a GPU contract, so treat it as a benchmark for this TPU site, not its actual price.
• Lab revenue: the $1.3B is the midpoint of the $50M to $75M per MW-year estimates Ann cites [referring to a comment from the author Ann Davis Vaughan].
  • [Ed note: Those numbers that she cites come from another article where Ric Boada, the CEO of ten-month-old neocloud Volta, saying:]
    • He figures that the frontier AI labs are generating about $75 million in revenue per megawatt of use a year for inference output compared to an annualized cost to build and provide the infrastructure of $10 million to $15 million per megawatt a year. 
• Each layer is paid out of the one above it. Even at the low end, about $1.0B a month, lab revenue is roughly 5x the compute bill. That gap is why labs push datacenter developers so hard on timelines.

All of this hinges on some very wishy-washy numbers about the per-megawatt revenue of AI labs, which in turn hinge on wishy-washy numbers about how much compute labs have. For example, the New York Times reported that Anthropic would have 5GW of capacity by the end of 2026, up from 1.5GW at the end of 2025, according to statements made to its investors, but I can find no compelling evidence that there was even a gigawatt of compute for it to buy. 

If we assume these statements are accurate, then:

• Anthropic, which made $4.6 billion in revenue against (theoretically) 1.5GW of capacity, made about $3 million a megawatt in 2025.
• In 2026, assuming it makes $35 billion on 5GW of capacity, it would make $7 million a megawatt.

The peanut gallery will scream and oink at me for not working out how much is inference versus training, to which I say nobody fucking knows how much of its capacity is actually diverted to inference and I don’t even think they know how much capacity Anthropic has writ large!

We are basing the vast majority of words written and billions invested in the AI bubble on assumptions built on assumptions built on assumptions, several of which fall apart when you ask simple questions like “how does that work?” or “how did you get that number?” 

This would all be significantly less-worrying if it wasn’t for the fact that everything is built on the idea that AI is going to be the panacea for the entire tech industry! 

Put another way, NVIDIA plans to sell $670 billion or more GPUs next year on top of the more than $500 billion it’s already sold and $800 billion has been invested in AI startups mostly by people with little or no idea about the underlying economics of the data centers or their customers!

Nobody Knows What They’re Investing In

A few months ago, a dear friend of mine told me that “nobody knows what they’re investing in,” and it’s taken me a while to fully understand what he meant.

The entire world believes we’re in an AI-powered economic boom of data center capex and AI startup growth based on flimsy information sources that vary from fanciful to utterly hysterical, constantly guessing about how much capacity OpenAI or Anthropic or any of their hyperscale sugar daddies actually have operational, or how much a theoretical data center could charge under the best of circumstances to a customer that may or may not actually exist.

We are four fucking years into the AI bubble, have little or no understanding of the actual economics of AI, and that which we do have is extremely grim. Yet when you try and tell an AI booster (or many members of the media!) that Anthropic lost $8 billion and OpenAI lost $20.97 billion in 2025, their immediate response is to say that 2026 will be fine because of quarters with discounted compute costs and adjusted operating profits that don’t actually resolve the underlying problem that these companies have $1.3 trillion in compute commitments, with at least $413 billion of them being uncancellable.

The hysteria has grown so significant that Gavin Baker of investment firm Atreides (invested in both Anthropic and SpaceX) has simply chosen to suggest that “nobody is taking them seriously,” only to disappear into the ether when asked about which of the contracts he believes are “not taken seriously” and “by whom,” probably because he’s talking out of his ass. 

This argument is one you make when you are flat out of ideas, and simply hope that people will leave you alone if you sound authoritative enough, and is known in debate circles as the “nuh uh” technique. The reason Baker is deploying it is that he, like every AI booster, knows that while you can fool the press and investors with made up numbers, there’s very little that “annualized run rates” or “Anthropic ending the year with $100 billion in revenue” can do when Broadcom CEO Hock Tan, who recently likened AI investors selling their investments on the news of OpenAI's adjusted financials to degenerate gamblers, asks Dario Amodei to pay some of the $161.2 billion in non-cancellable contracts he signed.

We live in a climate of intense hysteria and profound confusion, made worse by a media ecosystem that currently appears more interested in sustaining narratives than investigating their underlying truths. The vast majority of hyperscaler AI revenues come from Anthropic and OpenAI’s compute spend that is dependent on near-perpetual funding, much like the vast majority of Anthropic and OpenAI’s revenues come from AI startups that depend on near-perpetual funding to continue their spend. 

And to make matters worse, we quite literally have no idea about the underlying financial condition of any other AI companies outside of the main AI labs, and no, annualized run rates don’t mean jack shit. 

This is extremely bad and extremely important to resolve. Anthropic and OpenAI’s growth depends on the AI startup ecosystem being able to spend increasing amounts on tokens every single month, without fail, for what appears to be the rest of time. Said startups’ ability to keep doing so depends on venture capital being willing to sling them hundreds of millions of dollars a year each, with each investment increasing the paper value of the company but adding little or no actual qualitative improvement other than their ability to keep playing at the Sam Altman Casino.

I genuinely don’t know why most of the tech and business media is ignoring this problem. Is it that it’s too difficult to comprehend? Is it that you, as a writer, truly think that all of this will work out? Is it that you think that Anthropic and OpenAI will just “negotiate out” of their contracts? If that’s the case, what happens to all the data center capacity? Is it that there’s advertiser or editorial pressure to be “positive on AI”? I truly want to know! 

Anyway, I’ve realized now that “nobody knows what they’re investing in” applies to basically everything. Most people writing about AI believe it’ll all work out because nothing bad has happened or they have a single statistic that affirms their worldview. Most people are still saying “it’s like Amazon Web Services!” or “it’s like Uber!” or “it’ll be like the Dot Com Bubble,” even though those are not correct or sensible or rational takes that actually mean much other than “money burn good actually,” and that thin level of rationality goes up to the highest rungs of finance and journalism to a degree that turns my fucking stomach.

Most coverage of AI lives in its own bubble where nothing bad has happened or could happen, and that every promise will come true. Every deranged, multi-faceted “model” that shows OpenAI or Anthropic will print money in the future is an attempt to get away from very, very simple mathematical problems. 

And, fundamentally, I’m just not sure anyone gives that much of a shit about being correct outside of sounding as close to the consensus as possible, which is why it’s been so profoundly strange hearing people say things recently that I’ve been saying for years.

You can shove your fingers or whatever other objects you’ve got handy in your ears and say “la la la” all you want, but Jefferies’ Chris Wood just said that “the most likely long-term outcome of the AI boom will be massive capital destruction in the US,” and Barclays just said the following (aside: that lab revenue number is completely wrong, but whatever):

…let's face facts; the combined disclosed lab revenues of circa $150bn and total Al revenues running at around $250bn is not going to cut it when the ecosystem needs well in excess of $1trn to run even the maintenance capex of this up front spend and I've yet been presented with a credible framework as to where this money is going to come from other than pie-in-the-sky TAM numbers whimsically inserted in prospectuses.

I want to force boosters to read this like the guy at the end of A Clockwork Orange. I want them to answer me, in detail, why this is actually fine. 

I already know the answer: they’d produce some sort of asinine number that shows that growth is parabolic, that there’s insatiable demand for AI compute, that Anthropic and OpenAI are the next Google and Facebook combined, that there’s not enough compute to go around, or another anecdote that is literally anything other than a meaningful response. 

There is little difference between how the various run rates, projections, “models” and “frameworks” used to justify future AI investment are used and the underlying tools of a religious con, except the only community you get here is made up of zyned-up twenty-somethings and the elder millennials desperate for their approval.

The AI bubble is inflated by bullshit numbers peddled by the ignorant, craven or easily-led. Your ability to distinguish between which of them you’re talking to is critical to not losing your shirt.

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

If you want to get in touch — and especially if you have any juicy information about Anthropic, OpenAI, or any other companies in the AI bubble — hit me up on Signal at ezitron.76. I’m also on IB on The Terminal.