Riccardo Boccardi is a writer from Montalcino, Italy, known for creating Racconti Crestati, a vast collection of micro-stories each just under 600 characters long. Aside from being an impressive achievement in its own right, Riccardo’s stories have also inspired more than a thousand illustrations and several musical jingles by other artists. A selection of stories has turned into two published books as well as university teaching materials.

Can you tell us a bit more about Racconti Crestati and what inspired you to create this interesting project?

It all started in 2017 with the creation of a simple Facebook page. The idea was to share short stories, as long as a single post, meant to be read perhaps on the subway or while waiting for a bus. They had to suit a new form of communication: fast, immediate, and capable of capturing attention in just a few seconds. I named the project “Racconti Crestati” (Crested Stories) and chose a Mohawk hairstyle as its logo. The hairstyle symbolized the principle behind these stories: they needed to be direct, sharp, and full of character. Just like that haircut, all the superfluous was stripped away, leaving only what was essential. At first, I wrote a story every day, but the pace soon became unsustainable, so I switched to one “crested” story per week. Since then, these texts have been my companions: through them, I capture thoughts, visions, and fantasies, giving life to tales that are often grotesque or surreal.

As a writer, what draws you to the short story, or more like, micro-story form?

At its core, it’s about explosiveness. Everything unfolds in just a few breaths, even though each story can continue to echo in the reader’s mind long afterward. In my writing, I always try to suggest a “before” and an “after” surrounding the events I narrate: this way, time stretches, and the characters take on a fuller life in the reader’s imagination. Thanks to their brevity, micro-stories allow me to explore a wide variety of themes and literary genres, from the grotesque to noir, from romance to science fiction. Sometimes a single line of dialogue is enough to spark an entire story. Other times, the tales emerge from playful linguistic experiments, such as tautograms or lipograms.

It seems clear from your project that you enjoy working with constraints—what motivated you to set the limit of just under 600 characters for each story? How do these limitations foster creativity in your writing?

At first, when I was drafting my texts, I hadn’t set a precise length—let alone the strict one I follow today. Over the weeks, however, two guiding numbers emerged: 597 and 598. In fact, each of my stories must meet one of these character counts (excluding spaces, title included). Within this exact framework, my imagination moves freely, and my creativity is honed. In such a small space, every sentence, every word, every punctuation mark must be carefully chosen to create the right rhythm and evoke the intended atmosphere. What might seem like a limitation is, in fact, my greatest freedom: it allows me to cross boundaries I otherwise could not.

What is your creative process for crafting each of these concise stories? How does your writing routine look, and how has it contributed to your creative output?

Stories can emerge from a variety of sources: an everyday event pushed to the extreme, a person you meet who becomes a character, the musicality of a single word that becomes the story’s pivot, or a dormant memory that resurfaces and transforms. I don’t follow a strict routine: I set my goal in Ulysses (a maximum of 600 characters, excluding spaces) and then develop the story. Usually, I already know the ending and write the story backward. Other times, I begin with a theme or a striking image. Some “crested” stories are born quickly, while others require weeks of gestation before they come to life.

You have been using Ulysses since its very first release, which was in 2003, and have written over 570 micro-stories with it. What makes Ulysses such a perfect writing companion for you?

The first time I used Ulysses, I didn’t fully grasp its structure or potential. It was revolutionary: minimal, essential, almost stark in its simplicity. Everything existed in a shared, undefined space—no separate files, no folders, no formatting tools. Only through regular use did I come to appreciate the hidden power beneath its simplicity, and the advantages it offered both in the creative process and in managing texts structurally. Since then, it has become my daily writing companion. The ability to organize hundreds of texts with clarity and security is unmatched. The only thing I miss is the icon from the earliest versions: that “U”, part bard’s lyre, part Minotaur’s horns.

Writing over 570 micro-stories surely changes how you perceive narrative over time. How has this influenced your understanding of storytelling or your approach to new creative ideas?

Essentially, it has allowed me to sharpen my imagination, explore a wide variety of themes, and blend the comic with the dramatic, and vice versa. The natural science knowledge I gained at university has found new life in this literary form, which preserves the rigor and organizational skills of science. Metamorphoses, symbioses, geological events, and peculiar creatures often take center stage in my stories. Condensing, optimizing, and synthesizing are integral to the way I think, which is why the ultra-short form of writing suits me so well. Yet it remains a valuable exercise: when practiced consistently, it helps uncover new narrative and linguistic solutions, making each text more impactful in its brevity.

Even though all micro romanzi (micro-stories) are written by you, there is a strong encouragement of collaboration in this project. What is it like working together with illustrators, performers, and musicians? Do you feel that such interpretations elevate your stories, and, in turn, is your writing also influenced by other art forms, such as music, film, or visual arts?

Yes, the project has had a collaborative and voluntary spirit from the very beginning. It started with readings by Italian crime writers, then moved to illustrations—now numbering around 1,300—followed by performances from actors and voice-over artists, and finally musical jingles. Many of these collaborations are international, stretching from China to Chile, from Spain to Russia. The magic of it all lies in the fact that each creation springs from just a handful of words—a few sentences capable of connecting continents. Without a doubt, art feeds art, and I am constantly amazed at how illustrators interpret the stories in ways I had never imagined. As for influences on my writing, I believe every experience adds a new layer to the creative process: cinema, music, TV series, books, comics, or even simply lounging on the sofa.

It is clear that Racconti Crestati spread widely in the art world, crossing into many fields beyond writing. Was that your plan from the start, or did it grow in that direction over time? And what impact do you wish to create with this project?

The original idea was simply to bring new “interesting” content to a social network that, even then, was filled with images and trivialities. Collaborations began shortly after, sparked by simple requests sent through Facebook. As you can see, generosity is far from dead…

As for the future of Crested Stories, I don’t have any specific plans at the moment, apart from continuing to write and nurturing artistic collaborations.

This is a digital-first project, with all stories freely available on the project’s website (www.racconticrestati.com), but it has also grown to include printed editions with two published books. Can you tell us how Racconti Crestati evolved from digital shorts to printed books? And how has reader feedback, especially from online platforms, influenced the direction of the project?

Two collections of micro-stories have been published: one in 2019 (“Racconti Crestati”, Primamedia Editore) and one this year (“Ebdomadari Crestati”, Scatole Parlanti). The first contains 196 stories, the second 200. Both include a previously unpublished story that never appeared online. Transitioning from the web to print required careful selection and refinement of the texts: when something is printed, it demands the utmost attention and precision. “Scripta manent”, as my Latin ancestors used to say. Although short stories are not widely appreciated by the general public in Italy, both publishers embraced these brief tales enthusiastically—perhaps because of their unusual conciseness and themes. Reader feedback hasn’t shaped the project, but today, fans of the “crested” stories can enjoy them online and also have a physical copy on their bedside table.

One of the books was also used as teaching material in a storytelling course at the University of Turin. How did this come about? Could you share your thoughts on what micro-stories can teach us about storytelling? What insights have you gathered during this project that you’d like to share with aspiring short story writers?

From the very beginning, my stories received the appreciation of Alessandro Perissinotto, an Italian crime writer and storytelling professor at the University of Turin. When my first printed collection was released, Alessandro—who later wrote the preface for the second anthology—selected Crested Stories as recommended reading for his course. It was a tremendous honor, and I still take this opportunity to thank him.

I believe micro-stories offer the chance to experiment with a variety of narrative techniques: ellipses, plot twists, reversals, temporal contractions or expansions (as perceived by the reader), and the evocative power of words. My advice for anyone venturing into short-form writing is simple: write. Often, people are drawn only to the idea of being a writer, without realizing how demanding, solitary, and sometimes frustrating it can be to turn one’s imagination into words. The message is clear: write even when you don’t feel like it. With a bit of luck, inspiration will follow.

Lastly, after many years of writing micro-stories, how do you see the future of Racconti Crestati? Are there new formats, collaborations, or directions you’d like to explore next?

I believe this project will continue to accompany me for several more years: by now, these short stories have become an integral part of the way I view the world, its people, and its events. Often, when observing a face, a gesture, or a situation, I can already see one of my quirky stories taking shape. Looking ahead, I remain open to all kinds of artistic collaborations, without restrictions of genre or form. I have also experimented with longer stories, earning positive reviews and recognition in national competitions. My current goal is still the publication of a novel… but if that doesn’t come to pass, I can always weave together a thousand micro-stories.

If you’d like to read more from Riccardo, make sure to check out his personal website or his project's website and Facebook page.

Ich weiß nicht, wie Leute auf solche absurden Ideen kommen, aber irgendwas werden sie sich dabei schon denken, denke ich mir – und höre natürlich mal rein. Die Arbeit soll ja schließlich nicht umsonst gewesen sein.

I applied my „cursed“ recipe to Pyramid Song and combined it with the old Betty Boop cartoon Poor Cinderella. But here is the strange part. With this specific video, YouTube automatically generated a transcript (sometimes it does, sometimes it doesn’t) over the reversed and pitch shifted version of the song. And surprisingly, it showed no more or less that Thom is seemingly singing “Jesus is my God” in Russian. Any Slavic speaker of any Slavic language would understand it.

I know about The Wizard of Oz theory, and during those Kid A and Amnesiac sessions they were heavily experimenting with reversed recordings (we all know how Like Spinning Plates was made), so this finding is quite intriguing to me.

Also, black-eyed angels appear in the video, and that was not my intention when I created it.

(Direktlink, via Zwentner)

In Madrid on a residential block, BURR Studio has turned a disused fur shop into Benito – a modern, efficient home for two that balances openness with a clever spacial layout. The architects approached the small space as both a technical challenge and a design opportunity, crafting a layered interior that makes the most of its unusual split-level plan and spatial quirks inherited from the building’s commercial past.

The building’s distinctive “L”-shaped section defines its layout and atmosphere. The lower band, at street level, rises to an airy height of roughly 13 feet, creating a generous volume for the main living area. Above it, the mezzanine level sits about 5 feet 11 inches higher, forming a raised platform that allows cars to pass below on their way to a shared basement garage. This clever structural arrangement divides the home into two parallel zones: one public and open to the street, and another more private and retreat-like above.

At the intersection of these two areas stands the project’s defining feature – a clean white staircase that serves as both sculptural focal point and practical storage hub. Beneath its geometric form, BURR Studio hid a compact bathroom and a series of stepped compartments that follow the slope of the stairs. These concealed storage units house essentials like kitchen appliances, pantry items, and cleaning supplies, maintaining a seamless, uncluttered appearance. The compact bathroom tucked below feels almost like something aboard a train or airplane – efficient and just the right size.

Circulation flows through the home in a U-shaped loop, beginning at the entrance and wrapping through the living areas before returning to the front facade at the upper level. Large openings at both ends of the home promote natural cross ventilation and draw daylight inside. Each level also opens to outdoor spaces: the lower connects to a communal patio, while the upper level spills onto a more private terrace designed.

A restrained color palette enhances the project’s sense of calm. Smooth white surfaces bounce light, making the modest footprint feel more expansive. Against this backdrop, a central wooden structure introduces warmth and texture upstairs – an integrated system of wardrobes and bookshelves, along with a private full bathroom, separates the bedroom area from the living room.

To manage uneven ceiling heights and conceal shared building services, the architects introduced a half-barrel vault above the public-facing spaces. This soft, curving form lends a sense of visual unity to the compact dwelling.

To learn more about Benito and BURR Studio, visit burr.studio.

Photography by Maru Serrano.

On Saturday, millions of people across the U.S. attended “No Kings” protests—a slogan born in response to President Donald Trump’s self-aggrandizing social media posts where he’s called himself a king, including with AI-generated images of himself in a crown, and his continuous stretching of executive power. While Americans were out in the street, the president was posting.

In an AI-generated video originally posted on X by a genAI shitposter, Trump, wearing a crown, takes off in a fighter jet to the song “Danger Zone” like he’s in Top Gun. Flying over protestors in American cities, Pilot King Trump bombs people with gallons of chunky brown liquid. It’s poop, ok? It’s shit. It’s diarrhea, and in reposting it, it’s clear enough to me that Trump is fantasizing about doing a carpet-bomb dookie on the people he put his hand on a bible and swore to serve nine months ago. The first protestor seen in the video is a real person, Harry Sisson, a liberal social media influencer. 

But this was not clear, it seems, to many other journalists. Most national news outlets seem scared to call it how they see, and how everyone sees it: as Trump dropping turd bombs on America, instead opting for euphemisms. Some of the best have included: 

• The Hill called it “brown liquid” and “what looked like feces”
• The Guardian deemed it “brown sludge” and “bursts of brown matter”
• More “brown liquid” from the New York Times
• NBC News got close with “what appeared to be feces”  
• A CNN contributor’s “analysis” said Trump was “appearing to dump raw sewage”
• Axios’ helpful context: “suspect brown substances falling from the sky” 
• ABC News opted to cut the video before the AI poop even started falling

The New York Post, never one to waste a prime alliteration opportunity, didn’t disappoint: “Trump’s fighter jet was shown dropping masses of manure.”

I can understand some of these venerated news establishments might be skittish about using a word like “poop” in their headlines, and I can also concede that I haven’t had an editor tell me I can’t use a bad word in a headline in a long, long time. I can imagine the logic: we can’t “prove” it’s meant to be shit, so we can’t call it shit. But there’s nothing in these outlets’ style guides that has kept them from saying “poop” in the headline in the past: “Women Poop,” the New York Times once proclaimed. Axios writes extensively and frequently about dog poop. CNN’s analysis extends to poop often.

Along with the above concessions, I can also promise I don’t feel that passionately about getting poop on anyone else’s homepages. But we are in an era where the highest office in the country is disseminating imagery that isn’t just fake and stupid, but actively hostile to the people living in this country. When I first saw someone talking about the Trump Poop Bomber video—on Reels, of all places—I thought it must be someone doing satire about what they imagined Trump would post about the protests. I had to search for it to find out if he really did, and what I found was the above: trusted sources of truth and information too scared to call fake poop fake poop. It’s not about poop, it’s about being able to accurately describe what we see, an essential skill when everything online is increasingly built to enrage, trick, or mislead us. AI continues to be the aesthetic of fascism: fast, easy, ugly. When we lose the ability to say what it is, we’re losing a lot more than the chance to pun on poop.

Add to this the fact that no one in Trump’s circle will say what we can all plainly see, either: that the president hates the people. “The president uses social media to make the point. You can argue he’s probably the most effective person who’s ever used social media for that,” Speaker of the House Mike Johnson said at a news conference this morning. “He is using satire to make a point. He is not calling for the murder of his political opponents.” Johnson did not say what that point was, however.

So, I originally planned for this to be on my premium newsletter, but decided it was better to publish on my free one so that you could all enjoy it. If you liked it, please consider subscribing to support my work. Here’s $10 off the first year of annual.

I’ve also recorded an episode about this on my podcast Better Offline (RSS feed, Apple, Spotify, iHeartRadio), it’s a little different but both handle the same information, just subscribe and it'll pop up. 

Over the last two years I have written again and again about the ruinous costs of running generative AI services, and today I’m coming to you with real proof.

Based on discussions with sources with direct knowledge of their AWS billing, I am able to disclose the amounts that AI firms are spending, specifically Anthropic and AI coding company Cursor, its largest customer.

I can exclusively reveal today Anthropic’s spending on Amazon Web Services for the entirety of 2024, and for every month in 2025 up until September, and that that Anthropic’s spend on compute far exceeds that previously reported. 

Furthermore, I can confirm that through September, Anthropic has spent more than 100% of its estimated revenue (based on reporting in the last year) on Amazon Web Services, spending $2.66 billion on compute on an estimated $2.55 billion in revenue.

Additionally, Cursor’s Amazon Web Services bills more than doubled from $6.2 million in May 2025 to $12.6 million in June 2025, exacerbating a cash crunch that began when Anthropic introduced Priority Service Tiers, an aggressive rent-seeking measure that begun what I call the Subprime AI Crisis, where model providers begin jacking up the prices on their previously subsidized rates.

Although Cursor obtains the majority of its compute from Anthropic — with AWS contributing a relatively small amount, and likely also taking care of other parts of its business — the data seen reveals an overall direction of travel, where the costs of compute only keep on going up. 

Let’s get to it.

Some Initial Important Details

• I do not have all the answers! I am going to do my best to go through the information I’ve obtained and give you a thorough review and analysis. This information provides a revealing — though incomplete — insight into the costs of running Anthropic and Cursor, but does not include other costs, like salaries and compute obtained from other providers. I cannot tell you (and do not have insight into) Anthropic’s actual private moves. Any conclusions or speculation I make in this article will be based on my interpretations of the information I’ve received, as well as other publicly-available information.
• I have used estimates of Anthropic’s revenue based on reporting across the last ten months. Any estimates I make are detailed and they are brief. 
• These costs are inclusive of every product bought on Amazon Web Services, including EC2, storage and database services (as well as literally everything else they pay for).
• Anthropic works with both Amazon Web Services and Google Cloud for compute. I do not have any information about its Google Cloud spend.
  • The reason I bring this up is that Anthropic’s revenue is already being eaten up by its AWS spend. It’s likely billions more in the hole from Google Cloud and other operational expenses.
• I have confirmed with sources that every single number I give around Anthropic and Cursor’s AWS spend is the final cash paid to Amazon after any discounts or credits.
• While I cannot disclose the identity of my source, I am 100% confident in these numbers, and have verified their veracity with other sources.

Anthropic’s Compute Costs Are Likely Much Higher Than Reported — $1.35 Billion in 2024 on AWS Alone

In February of this year, The information reported that Anthropic burned $5.6 billion in 2024, and made somewhere between $400 million and $600 million in revenue:

It’s not publicly known how much revenue Anthropic generated in 2024, although its monthly revenue rose to about $80 million by the end of the year, compared to around $8 million at the start. That suggests full-year revenue in the $400 million to $600 million range.

…Anthropic told investors it expects to burn $3 billion this year, substantially less than last year, when it burned $5.6 billion. Last year’s cash burn was nearly $3 billion more than Anthropic had previously projected. That’s likely due to the fact that more than half of the cash burn came from a one-off payment to access the data centers that power its technology, according to one of the people who viewed the pitch.

While I don’t know about prepayment for services, I can confirm from a source with direct knowledge of billing that Anthropic spent $1.35 billion on Amazon Web Services in 2024, and has already spent $2.66 billion on Amazon Web Services through the end of September.

Assuming that Anthropic made $600 million in revenue, this means that Anthropic spent $6.2 billion in 2024, leaving $4.85 billion in costs unaccounted for. 

The Information’s piece also brings up another point:

The costs to develop AI models accounted for a major portion of Anthropic’s expenses last year. The company spent $1.5 billion on servers for training AI models. OpenAI was on track to spend as much as $3 billion on training costs last year, though that figure includes additional expenses like paying for data.

Before I go any further, I want to be clear that The Information’s reporting is sound, and I trust that their source (I have no idea who they are or what information was provided) was operating in good faith with good data.

However, Anthropic is telling people it spent $1.5 billion on just training when it has an Amazon Web Services bill of $1.35 billion, which heavily suggests that its actual compute costs are significantly higher than we thought, because, to quote SemiAnalysis, “a large share of Anthropic’s spending is going to Google Cloud.” 

I am guessing, because I do not know, but with $4.85 billion of other expenses to account for, it’s reasonable to believe Anthropic spent an amount similar to its AWS spend on Google Cloud. I do not have any information to confirm this, but given the discrepancies mentioned above, this is an explanation that makes sense.

I also will add that there is some sort of undisclosed cut that Amazon gets of Anthropic’s revenue, though it’s unclear how much. According to The Information, “Anthropic previously told some investors it paid a substantially higher percentage to Amazon [than OpenAI’s 20% revenue share with Microsoft] when companies purchase Anthropic models through Amazon.”

I cannot confirm whether a similar revenue share agreement exists between Anthropic and Google.

This also makes me wonder exactly where Anthropic’s money is going.

Where Is Anthropic’s Money Going?

Anthropic has, based on what I can find, raised $32 billion in the last two years, starting out 2023 with a $4 billion investment from Amazon from September 2023 (bringing the total to $37.5 billion), where Amazon was named its “primary cloud provider” nearly eight months after Anthropic announced Google was Anthropic’s “cloud provider.,” which Google responded to a month later by investing another $2 billion on October 27 2023, “involving a $500 million upfront investment and an additional $1.5 billion to be invested over time,” bringing its total funding from 2023 to $6 billion.

In 2024, it would raise several more rounds — one in January for $750 million, another in March for $884.1 million, another in May for $452.3 million, and another $4 billion from Amazon in November 2024, which also saw it name AWS as Anthropic’s “primary cloud and training partner,” bringing its 2024 funding total to $6 billion.

In 2025 so far, it’s raised a $1 billion round from Google, a $3.5 billion venture round in March, opened a $2.5 billion credit facility in May, and completed a $13 billion venture round in September, valuing the company at $183 billion. This brings its total 2025 funding to $20 billion. 

While I do not have Anthropic’s 2023 numbers, its spend on AWS in 2024 — around $1.35 billion — leaves (as I’ve mentioned) $4.85 billion in costs that are unaccounted for. The Information reports that costs for Anthropic’s 521 research and development staff reached $160 million in 2024, leaving 394 other employees unaccounted for (for 915 employees total), and also adding that Anthropic expects its headcount to increase to 1900 people by the end of 2025.

The Information also adds that Anthropic “expects to stop burning cash in 2027.”

This leaves two unanswered questions:

• Where is the rest of Anthropic’s money going?
• How will it “stop burning cash” when its operational costs explode as its revenue increases?

An optimist might argue that Anthropic is just growing its pile of cash so it’s got a warchest to burn through in the future, but I have my doubts. In a memo revealed by WIRED, Anthropic CEO Dario Amodei stated that “if [Anthropic wanted] to stay on the frontier, [it would] gain a very large benefit from having access to this capital,” with “this capital” referring to money from the Middle East. 

Anthropic and Amodei’s sudden willingness to take large swaths of capital from the Gulf States does not suggest that it’s not at least a little desperate for capital, especially given Anthropic has, according to Bloomberg, “recently held early funding talks with Abu Dhabi-based investment firm MGX” a month after raising $13 billion.

In my opinion — and this is just my gut instinct — I believe that it is either significantly more expensive to run Anthropic than we know, or Anthropic’s leaked (and stated) revenue numbers are worse than we believe. I do not know one way or another, and will only report what I know.

How Much Did Anthropic and Cursor Spend On Amazon Web Services In 2025?

So, I’m going to do this a little differently than you’d expect, in that I’m going to lay out how much these companies spent, and draw throughlines from that spend to its reported revenue numbers and product announcements or events that may have caused its compute costs to increase.

I’ve only got Cursor’s numbers from January through September 2025, but I have Anthropic’s AWS spend for both the entirety of 2024 and through September 2025.

What Does “Annualized” Mean?

So, this term is one of the most abused terms in the world of software, but in this case, I am sticking to the idea that it means “month times 12.” So, if a company made $10m in January, you would say that its annualized revenue is $120m. Obviously, there’s a lot of (when you think about it, really obvious) problems with this kind of reporting — and thus, you only ever see it when it comes to pre-IPO firms — but that’s besides the point.

I give you this explanation because, when contrasting Anthropic’s AWS spend with its revenues, I’ve had to work back from whatever annualized revenues were reported for that month. 

Anthropic’s Amazon Web Services Spend In 2024 - $1.359 Billion - Estimated Revenue $400 Million to $600 Million

Anthropic’s 2024 revenues are a little bit of a mystery, but, as mentioned above, The Information says it might be between $400 million and $600 million.

Here’s its monthly AWS spend. 

• January 2024 - $52.9 million
• February 2024 - $60.9 million
• March 2024 - $74.3 million
• April 2024 - $101.1 million
• May 2024 - $100.1 million
• June 2024 - $101.8 million
• July 2024 - $118.9 million
• August 2024 - $128.8 million
• September 2024 - $127.8 million
• October 2024 - $169.6 million
• November 2024 - $146.5 million
• December 2024 - $176.1 million

Analysis: Anthropic Spent At Least 200% of Its 2024 Revenue On Amazon Web Services In 2024

I’m gonna be nice here and say that Anthropic made $600 million in 2024 — the higher end of The Information’s reporting — meaning that it spent around 226% of its revenue ($1.359 billion) on Amazon Web Services.

[Editor's note: this copy originally had incorrect maths on the %. Fixed now.]

Anthropic’s Amazon Web Services Spend In 2025 Through September 2025 - $2.66 Billion - Estimated Revenue Through September $2.55 Billion - 104% Of Revenue Spent on AWS

Thanks to my own analysis and reporting from outlets like The Information and Reuters, we have a pretty good idea of Anthropic’s revenues for much of the year. That said, July, August, and September get a little weirder, because we’re relying on “almosts” and “approachings,” as I’ll explain as we go.

I’m also gonna do an analysis on a month-by-month basis, because it’s necessary to evaluate these numbers in context. 

January 2025 - $188.5 million In AWS Spend, $72.91 or $83 Million In Revenue - 227% Of Revenue Spent on AWS

In this month, Anthropic’s reported revenue was somewhere from $875 million to $1 billion annualized, meaning either $72.91 million or $83 million for the month of January.

February 2025 - $181.2 million in AWS Spend, $116 Million In Revenue - 156% Of Revenue Spent On AWS - 181% Of Revenue Spent On AWS

In February, as reported by The Information, Anthropic hit $1.4 billion annualized revenue, or around $116 million each month.

March 2025 - $240.3 million in AWS Spend - $166 Million In Revenue - 144% Of Revenue Spent On AWS - Launch of Claude Sonnet 3.7 & Claude Code Research Preview (February 24)

In March, as reported by Reuters, Anthropic hit $2 billion in annualized revenue, or $166 million in revenue.

Because February is a short month, and the launch took place on February 24 2025, I’m considering the launches of Claude 3.7 Sonnet and Claude Code’s research preview to be a cost burden in the month of March.

And man, what a burden! Costs increased by $59.1 million, primarily across compute categories, but with a large ($2 million since January) increase in monthly costs for S3 storage.

April 2025 - $221.6 million in AWS Spend - $204 Million In Revenue - 108% Of Revenue Spent On AWS

I estimate, based on a 22.4% compound growth rate, that Anthropic hit around $2.44 billion in annualized revenue in April, or $204 million in revenue.

Interestingly, this was the month where Anthropic launched its $100 and $200 dollar a month “Max” plans, and it doesn’t seem to have dramatically increased its costs. Then again, Max is also the gateway to things like Claude Code, which I’ll get to shortly.

May 2025 - $286.7 million in AWS Spend - $250 Million In Revenue - 114% Of Revenue Spent On AWS - Sonnet 4, Opus 4, General Availability Of Claude Code (May 22) Service Tiers (May 30)

In May, as reported by CNBC, Anthropic hit $3 billion in annualized revenue, or $250 million in monthly average revenue.

This was a big month for Anthropic, with two huge launches on May 22 2025 — its new, “more powerful” models Claude Sonnet and Opus 4, as well as the general availability of its AI coding environment Claude Code.

Eight days later, on May 30 2025, a page on Anthropic's API documentation appeared for the first time: "Service Tiers":

Different tiers of service allow you to balance availability, performance, and predictable costs based on your application’s needs.

We offer three service tiers:

- Priority Tier: Best for workflows deployed in production where time, availability, and predictable pricing are important

Standard: Best for bursty traffic, or for when you’re trying a new idea

Batch: Best for asynchronous workflows which can wait or benefit from being outside your normal capacity

Accessing the priority tier requires you to make an up-front commitment to Anthropic, and said commitment is based on a number of months (1, 3, 6 or 12) and the number of input and output tokens you estimate you will use each minute. 

What’s a Priority Tier? Why Is It Significant?

As I’ll get into in my June analysis, Anthropic’s Service Tiers exist specifically for it to “guarantee” your company won’t face rate limits or any other service interruptions, requiring a minimum spend, minimum token throughput, and for you to pay higher rates when writing to the cache — which is, as I’ll explain, a big part of running an AI coding product like Cursor.

Now, the jump in costs — $65.1 million or so between April and May — likely comes as a result of the final training for Sonnet and Opus 4, as well as, I imagine, some sort of testing to make sure Claude Code was ready to go.

June 2025 - $321.4 million in AWS Spend - $333 Million In Revenue - 96.5% Of Revenue Spent On AWS - Anthropic Cashes In On Service Tier Tolls That Add An Increased Charge For Prompt Caching, Directly Targeting Companies Like Cursor

In June, as reported by The Information, Anthropic hit $4 billion in annualized revenue, or $333 million.

Anthropic’s revenue spiked by $83 million this month, and so did its costs by $34.7 million. 

Anthropic Started The Subprime AI Crisis In June 2025, Increasing Costs On Its Largest Customer, Doubling Its AWS Spend In A Month

I have, for a while, talked about the Subprime AI Crisis, where big tech and companies like Anthropic, after offering subsidized pricing to entice in customers, raise the rates on their customers to start covering more of their costs, leading to a cascade where businesses are forced to raise their prices to handle their new, exploding costs.

And I was god damn right. Or, at least, it sure looks like I am. I’m hedging, forgive me. I cannot say for certain, but I see a pattern. 

It’s likely the June 2025 spike in revenue came from the introduction of service tiers, which specifically target prompt caching, increasing the amount of tokens you’re charged for as an enterprise customer based on the term of the contract, and your forecast usage.

Per my reporting in July: 

You see, Anthropic specifically notes on its "service tiers" page that requests at the priority tier are "prioritized over all other requests to Anthropic," a rent-seeking measure that effectively means a company must either:

- Commit to at least a month, though likely 3-12 months of specific levels of input and output tokens a minute, based on what they believe they will use in the future, regardless of whether they do.

- Accept that access to Anthropic models will be slower at some point, in some way that Anthropic can't guarantee.Furthermore, the way that Anthropic is charging almost feels intentionally built to fuck over any coding startup that would use its service. Per the service tier page, Anthropic charges 1.25 for every time you write a token to the cache with a 5 minute TTL — or 2 tokens if you have a 1 hour TTL — and a longer cache is effectively essential for any background task where an agent will be working for more than 5 minutes, such as restructuring a particularly complex series of code, you know, the exact things that Cursor is well-known and marketed to do.

Furthermore, the longer something is in the cache, the better autocomplete suggestions for your code will be. It's also important to remember you're, at some point, caching the prompts themselves — so the instructions of what you want Cursor to do, meaning that the more complex the operation, the more expensive it'll now be for Cursor to provide the service with reasonable uptime.

Cursor, as Anthropic’s largest client (the second largest being Github Copilot), represents a material part of its revenue, and its surging popularity meant it was sending more and more revenue Anthropic’s way.  Anysphere, the company that develops Cursor, hit $500 million annualized revenue ($41.6 million) by the end of May, which Anthropic chose to celebrate by increasing its costs.

On June 16 2025, Cursor launched a $200-a-month “Ultra” plan, as well as dramatic changes to its $20-a-month Pro pricing that, instead of offering 500 “fast” responses using models from Anthropic and OpenAI, now effectively provided you with “at least” whatever you paid a month (so $20-a-month got at least $20 of credit), massively increasing the costs for users, with one calling the changes a “rug pull” after spending $71 in a single day.

As I’ll get to later in the piece, Cursor’s costs exploded from $6.19 million in May 2025 to $12.67 million in June 2025, and I believe this is a direct result of Anthropic’s sudden and aggressive cost increases. 

Similarly, Replit, another AI coding startup, moved to “Effort-Based Pricing” on June 18 2025. I have not got any information around its AWS spend.

I’ll get into this a bit later, but I find this whole situation disgusting.

July 2025 $323.2 million in AWS Spend - $416 Million In Revenue - 77.7% Of Revenue Spent On AWS

In July, as reported by Bloomberg, Anthropic hit $5 billion in annualized revenue, or $416 million.

While July wasn’t a huge month for announcements, it was allegedly the month that Claude Code was generating “nearly $400 million in annualized revenue,” or $33.3 million (according to The Information, who says Anthropic was “approaching” $5 billion in annualized revenue - which likely means LESS than that - but I’m going to go with the full $5 billion annualized for sake of fairness. 

There’s roughly an $83 million bump in Anthropic’s revenue between June and July 2025, and I think Claude Code and its new rates are a big part of it. What’s fascinating is that cloud costs didn’t increase too much — by only $1.8 million, to be specific.

August 2025 - $383.7 million in AWS Spend - $416 Million In Revenue - 92% Of Revenue Spent On AWS

In August, according to Anthropic, its run-rate “reached over $5 billion,” or in or around $416 million. I am not giving it anything more than $5 billion, especially considering in July Bloomberg’s reporting said “about $5 billion.”

Costs grew by $60.5 this month, potentially due to the launch of Claude Opus 4.1, Anthropic’s more aggressively expensive model, though revenues do not appear to have grown much along the way.

Yet what’s very interesting is that Anthropic — starting August 28 — launched weekly rate limits on its Claude Pro and Max plans. I wonder why?

September 2025 - $518.9 million in AWS Spend - $583 Million In Revenue - 88.9% Of Revenue Spent On AWS

Oh fuck! Look at that massive cost explosion!

Anyway, according to Reuters, Anthropic’s run rate is “approaching $7 billion” in October, and for the sake of fairness, I am going to just say it has $7 billion annualized, though I believe this number to be lower. “Approaching” can mean a lot of different things — $6.1 billion, $6.5 billion — and because I already anticipate a lot of accusations of “FUD,” I’m going to err on the side of generosity.

If we assume a $6.5 billion annualized rate, that would make this month’s revenue $541.6 million, or 95.8% of its AWS spend.  

Nevertheless, Anthropic’s costs exploded in the space of a month by $135.2 million (35%) - likely due to the fact that users, as I reported in mid-July, were costing it thousands or tens of thousands of dollars in compute, a problem it still faces to this day, with VibeRank showing a user currently spending $51,291 in a calendar month on a $200-a-month subscription.

If there were other costs, they likely had something to do with the training runs for the launches of Sonnet 4.5 on September 29 2025 and Haiku 4.5 in October 2025.

Anthropic’s Monthly AWS Costs Have Increased By 174% Since January - And With Its Potential Google Cloud Spend and Massive Staff, Anthropic Is Burning Billions In 2025

While these costs only speak to one part of its cloud stack — Anthropic has an unknowable amount of cloud spend on Google Cloud, and the data I have only covers AWS — it is simply remarkable how much this company spends on AWS, and how rapidly its costs seem to escalate as it grows.

Though things improved slightly over time — in that Anthropic is no longer burning over 200% of its revenue on AWS alone — these costs have still dramatically escalated, and done so in an aggressive and arbitrary manner. 

Anthropic’s AWS Costs Increase Linearly With Revenue, Consuming The Majority Of Each Dollar Anthropic Makes - As A Reminder, It Also Spends Hundreds Of Millions Or Billions On Google Cloud Too

So, I wanted to visualize this part of the story, because I think it’s important to see the various different scenarios.

An Estimate of Anthropic’s Potential Cloud Compute Spend Through September

THE NUMBERS I AM USING ARE ESTIMATES CALCULATED BASED ON 25%, 50% and 100% OF THE AMOUNTS THAT ANTHROPIC HAS SPENT ON AMAZON WEB SERVICES THROUGH SEPTEMBER. 

I apologize for all the noise, I just want it to be crystal clear what you see next.  

As you can see, all it takes is for Anthropic to spend (I am estimating) around 25% of its Amazon Web Services bills (for a total of around $3.33 billion in compute costs through the end of September) to savage any and all revenue ($2.55 billion) it’s making. 

Assuming Anthropic spends half of its  AWS spend on Google Cloud, this number climbs to $3.99 billion, and if you assume - and to be clear, this is an estimate - that it spends around the same on both Google Cloud and AWS, Anthropic has spent $5.3 billion on compute through the end of September.

I can’t tell you which it is, just that we know for certain that Anthropic is spending money on Google Cloud, and because Google owns 14% of the company — rivalling estimates saying Amazon owns around 15-19% — it’s fair to assume that there’s a significant spend.

Anthropic’s Costs Are Out Of Control, Consistently And Aggressively Outpacing Revenue - And Amazon’s Revenue from Anthropic Of $2.66 Billion Is 2.5% Of Its 2025 Capex

I have sat with these numbers for a great deal of time, and I can’t find any evidence that Anthropic has any path to profitability outside of aggressively increasing the prices on their customers to the point that its services will become untenable for consumers and enterprise customers alike.

As you can see from these estimated and reported revenues, Anthropic’s AWS costs appear to increase in a near-linear fashion with its revenues, meaning that the current pricing — including rent-seeking measures like Priority Service Tiers — isn’t working to meet the burden of its costs.

We do not know its Google Cloud spend, but I’d be shocked if it was anything less than 50% of its AWS bill. If that’s the case, Anthropic is in real trouble - the cost of the services underlying its business increase the more money they make.

It’s becoming increasingly apparent that Large Language Models are not a profitable business. While I cannot speak to Amazon Web Services’ actual costs, it’s making $2.66 billion from Anthropic, which is the second largest foundation model company in the world. 

Is that really worth $105 billion in capital expenditures? Is that really worth building a giant 1200 acre data center in Indiana with 2.2GW of electricity?

What’s the plan, exactly? Let Anthropic burn money for the foreseeable future until it dies, and then pick up the pieces? Wait until Wall Street gets mad at you and then pull the plug?

Who knows. 

But let’s change gears and talk about Cursor — Anthropic’s largest client and, at this point, a victim of circumstance.

Cursor’s Amazon Web Services Spend In 2025 Through September 2025 - $69.99 Million

An Important Note About Cursor’s Compute Spend

Amazon sells Anthropic’s models through Amazon Bedrock, and I believe that AI startups are compelled to spend some of their AI model compute costs through Amazon Web Services. Cursor also sends money directly to Anthropic and OpenAI, meaning that these costs are only one piece of its overall compute costs. In any case, it’s very clear that Cursor buys some degree of its Anthropic model spend through Amazon.

I’ll also add that Tom Dotan of Newcomer reported a few months ago that an investor told him that “Cursor is spending 100% of its revenue on Anthropic.”

Unlike Anthropic, we lack thorough reporting of the month-by-month breakdown of Cursor’s revenues. I will, however, mention them in the month I have them.

For the sake of readability — and because we really don’t have much information on Cursor’s revenues beyond a few months — I’m going to stick to a bullet point list. 

Another Note About Cursor’s AWS Spend - It Likely Funnels Some Model Spend Through AWS, But The Majority Goes Directly To Providers Like Anthropic

As discussed above, Cursor announced (along with their price change and $200-a-month plan) several multi-year partnerships with xAI, Anthropic, OpenAI and Google, suggesting that it has direct agreements with Anthropic itself versus one with AWS to guarantee “this volume of compute at a predictable price.” 

Based on its spend with AWS, I do not see a strong “minimum” spend that would suggest that they have a similar deal with Amazon — likely because Amazon handles more than its infrastructure than just compute, but incentivizes it to spend on Anthropic’s models through AWS by offering discounts, something I’ve confirmed with a source. 

In any case, here’s what Cursor spent on AWS.

• January 2025 - $1.459 million
  • This, apparently, is the month that Cursor hit $100 million annualized revenue — or $8.3 million, meaning it spent 17.5% of its revenue on AWS.
• February 2025 - $2.47 million
• March 2025 - $4.39 million
• April 2025 - $4.74 million
  • Cursor hit $200 million annualized ($16.6 million) at the end of March 2025, according to The Information, working out to spending 28% of its revenue on AWS.  
• May 2025 - $6.19 million
• June 2025 - $12.67 million
  • So, Bloomberg reported that Cursor hit $500 million on June 5 2025, along with raising a $900 million funding round. Great news! Turns out it’d need to start handing a lot of that to Anthropic.
  • This was, as I’ve discussed above, the month when Anthropic forced it to adopt “Service Tiers”. I go into detail about the situation here, but the long and short of it is that Anthropic increased the amount of tokens you burned by writing stuff to the cache (think of it like RAM in a computer), and AI coding startups are very cache heavy, meaning that Cursor immediately took on what I believed would be massive new costs. As I discuss in what I just linked, this led Cursor to aggressively change its product, thereby vastly increasing its customers’ costs if they wanted to use the same service.
  • That same month, Cursor’s AWS costs — which I believe are the minority of its cloud compute costs — exploded by 104% (or by $6.48 million), and never returned to their previous levels.
  • It’s conceivable that this surge is due to the compute-heavy nature of the latest Claude 4 models released that month — or, perhaps, Cursor sending more of its users to other models that it runs on Bedrock. 
• July 2025 - $15.5 million
  • As you can see, Cursor’s costs continue to balloon in July, and I am guessing it’s because of the Service Tiers situation — which, I believe, indirectly resulted in Cursor pushing more users to models that it runs on Amazon’s infrastructure.
• August 2025 - $9.67 million
  • So, I can only guess as to why there was a drop here. User churn? It could be the launch of GPT-5 on Cursor, which gave users a week of free access to OpenAI’s new models.
  • What’s also interesting is that this was the month when Cursor announced that its previously free “auto” model (where Cursor would select the best available premium model or its own model) would now bill at “competitive token rates,” by which I mean it went from charging nothing to $1.25 per million input and $6 per million output tokens. This change would take effect on September 15 2025.
  • On August 10 2025, Tom Dotan of Newcomer reported that Cursor was “well above” $500 million in annualized revenue based on commentary from two sources.
• September 2025 - $12.91 million
  • Per the above, this is the month when Cursor started charging for its “auto” model.

What Anthropic May Have Done To Cursor Is Disgusting - And Is A Preview Of What’s To Come For AI Startups

When I wrote that Anthropic and OpenAI had begun the Subprime AI Crisis back in July, I assumed that the increase in costs was burdensome, but having the information from its AWS bills, it seems that Anthropic’s actions directly caused Cursor’s costs to explode by over 100%. 

While I can’t definitively say “this is exactly what did it,” the timelines match up exactly, the costs have never come down, Amazon offers provisioned throughput, and, more than likely, Cursor needs to keep a standard of uptime similar to that of Anthropic’s own direct API access.

If this is what happened, it’s deeply shameful. 

Cursor, Anthropic’s largest customer, in the very same month it hit $500 million in annualized revenue, immediately had its AWS and Anthropic-related costs explode to the point that it had to dramatically reduce the value of its product just as it hit the apex of its revenue growth. 

Anthropic Timed Its Rent-Seeking Service Tier Price Increases on Cursor With The Launch Of A Competitive Product - Which Is What’s Coming To Any AI Startup That Builds On Top Of Its Products

It’s very difficult to see Service Tiers as anything other than an aggressive rent-seeking maneuver.

Yet another undiscussed part of the story is that the launch of Claude 4 Opus and Sonnet — and the subsequent launch of Service Tiers — coincided with the launch of Claude Code, a product that directly competes with Cursor, without the burden of having to pay itself for the cost of models or, indeed, having to deal with its own “Service Tiers.”

Anthropic may have increased the prices on its largest client at the time it was launching a competitor, and I believe that this is what awaits any product built on top of OpenAI or Anthropic’s models. 

The Subprime AI Crisis Is Real, And It Can Hurt You

I realize this has been a long, number-stuffed article, but the long-and-short of it is simple: Anthropic is burning all of its revenue on compute, and Anthropic will willingly increase the prices on its customers if it’ll help it burn less money, even though that doesn’t seem to be working.

What I believe happened to Cursor will likely happen to every AI-native company, because in a very real sense, Anthropic’s products are a wrapper for its own models, except it only has to pay the (unprofitable) costs of running them on Amazon Web Services and Google Cloud.

As a result, both OpenAI and Anthropic can (and may very well!) devour the market of any company that builds on top of their models. 

OpenAI may have given Cursor free access to its GPT-5 models in August, but a month later on September 15 2025 it debuted massive upgrades to its competitive “Codex” platform. 

Any product built on top of an AI model that shows any kind of success can be cloned immediately by OpenAI and Anthropic, and I believe that we’re going to see multiple price increases on AI-native companies in the next few months. After all, OpenAI already has its own priority processing product, which it launched shortly after Anthropic’s in June.

The ultimate problem is that there really are no winners in this situation. If Anthropic kills Cursor through aggressive rent-seeking, that directly eats into its own revenues. If Anthropic lets Cursor succeed, that’s revenue, but it’s also clearly unprofitable revenue. Everybody loses, but nobody loses more than Cursor’s (and other AI companies’) customers. 

Anthropic Is In Real Trouble - And The Current Cost Of Doing Business Is Unsustainable, Meaning Prices Must Increase

I’ve come away from this piece with a feeling of dread.

Anthropic’s costs are out of control, and as things get more desperate, it appears to be lashing out at its customers, both companies like Cursor and Claude Code customers facing weekly rate limits on their more-powerful models who are chided for using a product they pay for. Again, I cannot say for certain, but the spike in costs is clear, and it feels like more than a coincidence to me. 

There is no period of time that I can see in the just under two years of data I’ve been party to that suggests that Anthropic has any means of — or any success doing — cost-cutting, and the only thing this company seems capable of doing is increasing the amount of money it burns on a monthly basis. 

Based on what I have been party to, the more successful Anthropic becomes, the more its services cost. The cost of inference is clearly increasing for customers, but based on its escalating monthly costs, the cost of inference appears to be high for Anthropic too, though it’s impossible to tell how much of its compute is based on training versus running inference.

In any case, these costs seem to increase with the amount of money Anthropic makes, meaning that the current pricing of both subscriptions and API access seems unprofitable, and must increase dramatically — from my calculations, a 100% price increase might work, but good luck retaining every single customer and their customers too! — for this company to ever become sustainable. 

I don’t think that people would pay those prices. If anything, I think what we’re seeing in these numbers is a company bleeding out from costs that escalate the more that its user base grows. This is just my opinion, of course. 

I’m tired of watching these companies burn billions of dollars to destroy our environment and steal from everybody. I’m tired that so many people have tried to pretend there’s a justification for burning billions of dollars every year, clinging to empty tropes about how this is just like Uber or Amazon Web Services, when Anthropic has built something far more mediocre. 

Mr. Amodei, I am sure you will read this piece, and I can make time to chat in person on my show Better Offline. Perhaps this Friday? I even have some studio time on the books. 

Hello readers! This premium edition features a generous free intro because I like to try and get some of the info out there, but the real indepth stuff is below the cut. Nevertheless, I deeply appreciate anyone subscribing.

On Monday I will have my biggest scoop ever, and it'll go out on the free newsletter because of its scale. This is possible because of people supporting me on the premium. Thanks so much for reading.

One of the only consistent critiques of my work is that I’m angry, irate, that I am taking myself too seriously, that I’m swearing too much, and that my arguments would be “better received” if I “calmed down.”

Fuck that.

Look at where being timid or deferential has got us. Broadcom and OpenAI have announced another 10GW of custom chips and supposed capacity which will supposedly get fully deployed by the end of 2029, and still the media neutrally reports these things as not simply doable, but rational.

To be clear, building a gigawatt of data center capacity costs at least $32.5 billion (though Jensen Huang says the computing hardware alone costs $50 billion, which excludes the buildings themselves and the supporting power infrastructure, and Barclays Bank says $50 billion to $60 billion) and takes two and a half years. 

In fact, fuck it — I’m updating my priors. Let’s say it’s a nice, round $50 billion per gigawatt of data center capacity. $32.5 billion is what it cost to build Stargate Abilene, but that estimate was based on Crusoe’s 1.2GW of compute for OpenAI being part of a $15 billion joint venture, which meant a gigawatt of compute runs about $12.5 billion, and Abilene’s 8 buildings are meant to hold 50,000 NVIDIA GB200 GPUs and their associated networking infrastructure, so let’s say a gigawatt is around 333,333 Blackwell GPUs at $60,000 a piece, so about $20 billion a gigawatt.

However, this mathematics assumed that every cost associated would be paid by the Joint Venture. Lancium, the owner of the land that is allegedly building the power infrastructure, has now raised over a billion dollars.

This maths also didn’t include the cost of the associated networking infrastructure around the GB200s. So, guess what? We’re doing $50 billion now. 

OpenAI has now promised 33GW of capacity across AMD, NVIDIA, Broadcom and the seven data centers built under Stargate, though one of those — in Lordstown, Ohio — is not actually a data center, with my source being “SoftBank,” speaking to WKBN in Lordstown Ohio, which said it will “not be a full-blown data center,” and instead be “at the center of cutting-edge technology that will encompass storage containers that will hold the infrastructure for AI and data storage.”

This wasn’t hard to find, by the way! I googled “SoftBank Lordstown” and up it came, ready for me to read with my eyes.

Putting all of that aside, I think it’s time that everybody started taking this situation far more seriously, by which I mean acknowledging the sheer recklessness and naked market manipulation taking place. 

But let’s make it really simple, and write out what’s meant to happen in the next year:

• In the second half of 2026, OpenAI and Broadcom will tape out and successfully complete an AI inference chip, then manufacture enough of them to fill a 1GW data center.
  • That data center will be built in an as-yet-unknown location, and will have at least 1GW of power, but more realistically it will need 1.2GW to 1.3GW of power, because for every 1GW of IT load, you need extra power capacity in reserve for the hottest day of the year, when the cooling system works hardest and power transmission losses are highest. . 
  • OpenAI does not appear to have a site for this data center, and thus has not broken ground on it.
• In the second half of 2026, AMD and OpenAI will begin “the first 1 gigawatt deployment of AMD Instinct MI450 GPUs.” 
  • This will take place in an as-yet-unnamed data center location, which to be completed by that time would have needed to start construction and early procurement of power at least a year ago, if not more. 
• In the second half of 2026, OpenAI and NVIDIA will deploy the first gigawatt of NVIDIA’s Vera Rubin GPU systems as part of their $100 billion deal.
  • These GPUs will be deployed in a data center of some sort, which remains unnamed, but for them to meet this timeline they will need to have started construction at least a year ago.

In my most conservative estimate, these data centers will cost over $100 billion, and to be clear, a lot of that money needs to already be in OpenAI’s hands to get the data centers built. Or, some other dupe has to a.) have the money, and b.) be willing to front it. 

All of this is a fucking joke. I’m sorry, I know some of you will read this, cowering from your screen like a B-movie vampire that just saw a crucifix, but it is a joke, and it is a fucking stupid joke, the only thing stupider being that any number of respectable media outlets are saying these things like they’ll actually happen.

There is not enough time to build these things. If there was enough time, there wouldn’t be enough money. If there was enough money, there wouldn’t be enough transformers, electrical-grade steel, or specialised talent to run the power to the data centers. Fuck! Piss! Shit! Swearing doesn’t change the fact that I’m right — none of what OpenAI, NVIDIA, Broadcom, and AMD are saying is possible, and it’s fair to ask why they’re saying it.

I mean, we know. Number must go up, deal must go through, and Jensen Huang wouldn’t go on CNBC and say “yeah man if I’m honest I’ve got no fucking clue how Sam Altman is going to pay me, other than with the $10 billion I’m handing him in a month. Anyway, NVIDIA’s accounts receivables keep increasing every quarter for a normal reason, don’t worry about it.” 

But in all seriousness, we now have three publicly-traded tech firms that have all agreed to join Sam Altman’s No IT Loads Refused Cash Dump, all promising to do things on insane timelines that they — as executives of giant hardware manufacturers, or human beings with warm bodies and pulses and sciatica — all must know are impossible to meet. 

What is the media meant to do? What are we, as regular people, meant to do? These stocks keep pumping based on completely nonsensical ideas, and we’re all meant to sit around pretending things are normal and good. They’re not! At some point somebody’s going to start paying people actual, real dollars at a scale that OpenAI has never truly had to reckon with.

In this piece, I’m going to spell out in no uncertain terms exactly what OpenAI has to do in the next year to fulfil its destiny — having a bunch of capacity that cost ungodly amounts of money to serve demand that never arrives.

Yes, yes, I know, you’re going to tell me that OpenAI has 800 million weekly active users, and putting aside the fact that OpenAI’s own research (see page 10, footnote 20) says it double-counts users who are logged out if they’re use different devices, OpenAI is saying it wants to build 250 gigawatts of capacity by 2033, which will cost it $10 trillion dollars, or one-third of the entire US economy last year.

Who the fuck for? 

One thing that’s important to note: In February, Goldman Sachs estimated that the global data center capacity was around 55GW. In essence, OpenAI says it wants to add five times that capacity — something that has grown organically over the past thirty or so years — by itself, and in eight years. 

And yes, it’ll cost one-third of America’s output in 2024. This is not a sensible proposition. 

Even if you think that OpenAI’s growth is impressive — it went from 700 million to 800 million weekly active users in the last two months — that is not the kind of growth that says “build capacity assuming that literally every single human being on Earth uses this all the time.” 

As an aside: Altman is already lying about his available capacity. According to an internal Slack note seen by Alex Heath of Sources, Altman claims that OpenAI started the year with “around” 230 megawatts of capacity and is “now on track to exit 2025 north of 2GW of operational capacity.” Unless I’m much mistaken OpenAI doesn’t have any capacity of its own — and according to Mr. Altman, it’s somehow built or acquired 1.7GW of capacity from somewhere without disclosing it.

For context, 1.7GW is the equivalent of every data center in the UK that was operational last year. 

Where is this coming from? Is this CoreWeave? It only has — at most — 900MW of capacity by the end of 2025. Where’d all the extra capacity come from? Who knows! It isn’t Stargate Abilene that’s for sure — they’ve only got one operational building and 200MW of power, meaning they can only really support 130MW of IT loads, because of that pesky reserve I mentioned earlier. 

Anyway, what exactly is OpenAI doing? Why does it need all this capacity? Even if it  hits its $13 billion revenue projection for this year (it’s only at $5.3 billion or so as of the end of August, and for OpenAI to hit its targets it’ll need to make $1.5bn+ a month very soon), does it really think it’s going to effectively 10x the entire company from here? What possible sign is there of that happening other than a conga-line of different executives willing to stake their reputations on blatant lies peddled by a man best known for needing, at any given moment, another billion dollars. 

According to The Information, OpenAI spent $6.7 billion on research and development in the first six months of 2025, and according to Epoch AI, most of the $5 billion it spent on research and development in 2024 was spent on research, experimental, or derisking runs (basically running tests before doing the final testing run) and models it would never release, with only $480 million going to training actual models that people will use. 

I should also add that GPT 4.5 was a dud, and even Altman called it giant, expensive, and said it “wouldn’t crush benchmarks.”

I’m sorry, but what exactly is it that OpenAI has released in the last year-and-a-half that was worth burning $11.7 billion for? GPT 5? That was a huge letdown! Sora 2? The giant plagiarism machine that it’s already had to neuter?

What is it that any of you believe that OpenAI is going to do with these fictional data centers? 

Why Does ChatGPT Need $10 Trillion Of Data Centers?

The problem with ChatGPT isn’t just that it hallucinates — it’s that you can’t really say exactly what it can do, because you can’t really trust that it can do anything. Sure, it’ll get a few things right a lot of the time, but what task is it able to do every time that you actually need? 

Say the answer is “something that took me an hour now takes me five minutes.” Cool! How many of those do you get? Again, OpenAI wants to build 250 gigawatts of data centers, and will need around ten trillion dollars to do it. “It’s going to be really good” is no longer enough.

And no, I’m sorry, they are not building AGI. He just told Politico a few weeks ago that if we didn’t have “models that are extraordinarily capable and do things that we ourselves cannot do” by 2030 he would be “very surprised.” 

Wow! What a stunning and confident statement. Let’s give this guy the ten trillion dollars he needs! And he’s gonna need it soon if he wants to build 250 gigawatts of capacity by 2033.

But let’s get a little more specific.

Based on my calculations, in the next six months, OpenAI needs at least $50 billion to build a gigawatt of data centers for Broadcom — and to hit its goal of 10 gigawatts of data centers by end of 2029, at least another $200 billion in the next 12 months, not including at least $50 billion to build a gigawatt of data centers for NVIDIA, $40 billion to pay for its 2026 compute, at least $50 billion to buy chips and build a gigawatt of data centers for AMD, at least $500 million to build its consumer device (and they can’t seem to work out what to build), and at least a billion dollars to hand off to ARM for a CPU to go with the new chips from Broadcom.

That’s $391.5 billion dollars! That’s $23.5 billion more than the $368 billion of global venture capital raised in 2024! That’s nearly 11 times Uber’s total ($35.8 billion) lifetime funding, or 5.7 times the $67.6 billion in capital expenditures that Amazon spent building Amazon Web Services! 

On top of all of this are OpenAI’s other costs. According to The Information, OpenAI spent $2 billion alone on Sales and Marketing in the first half of 2025, and likely spends billions of dollars on salaries, meaning that it’ll likely need at least another $10 billion on top. As this is a vague cost, I’m going with a rounded $400 billion number, though I believe it’s actually going to be more.

And to be clear, to complete these deals by the end of 2026, OpenAI needs large swaths of this money by February 2026. 

OpenAI Needs Over $400 Billion In The Next 12 Months To Complete Any Of These Deals — And Sam Altman Doesn’t Have Enough Time To Build Any Of it

I know, I know, you’re going to say that OpenAI will simply “raise debt” and “work it out,” but OpenAI has less than a year to do that, because OpenAI has promised in its own announcements that all of these things would happen by the end of December 2026, and even if they’re going to happen in 2027, data centers require actual money to begin construction, and Broadcom, NVIDIA and AMD are going to actually require cash for those chips before they ship them.

Even if OpenAI finds multiple consortiums of paypigs to take on the tens of billions of dollars of data center funding, there are limits, and based on OpenAI’s aggressive (and insane) timelines, they will need to raise multiple different versions of the largest known data center deals of all time, multiple times a year, every single year. 

Say that happens. OpenAI will still need to pay those compute contracts with Oracle, CoreWeave, Microsoft (I believe its Azure credits have run out) and Google (via CoreWeave) with actual, real cash — $40 billion dollars worth — when it’s already burning $9.2 billion in the first half of 2026 on compute against revenues of $4.3 billion. OpenAI will still need to pay its staff, its storage, its sales and marketing department that cost it $2 billion in the first half of 2026, all while converting its non-profit into a for-profit by the end of the year, or it loses $20 billion in funding from SoftBank.

Also, if it doesn’t convert to a for-profit by October 2026, its $6.6 billion funding round from 2024 converts to debt.

The Global Financial System Cannot Afford OpenAI

The burden that OpenAI is putting on the financial system is remarkable, and actively dangerous. It would absorb, at this rate, the capital expenditures of multiple hyperscalers, requiring multiple $30 billion debt financing operations a year, and for it to hit its goal of 250 gigawatts by the end of 2033, it will likely have to have outpaced the capital expenditures of any other company in the world.

OpenAI is an out-of-control monstrosity that is going to harm every party that depends upon it completing its plans. For it to succeed, it will have to absorb over a trillion dollars a year — and for it to hit its target, it will likely have to eclipse the $1.7 trillion in global private equity deal volume in 2024, and become a significant part of global trade ($33 trillion in 2025).

There isn’t enough money to do this without diverting most of the money that exists to doing it, and even if that were to happen, there isn’t enough time to do any of the stuff that has been promised in anything approaching the timelines promised, because OpenAI is making this up as it goes along and somehow everybody is believing it. 

At some point, OpenAI is going to have to actually do the things it has promised to do, and the global financial system is incapable of supporting them.

And to be clear, OpenAI cannot really do any of the things it’s promised.

Just take a look at the Oracle deal!

• Oracle needs 4.5 gigawatts of IT load capacity to provide OpenAI the compute for its $300 billion, five-year-long deal.
  • Despite Oracle CEO Greg Magouyrk saying “of course OpenAI can pay $60 billion a year,” OpenAI cannot actually afford to pay $60 billion a year. It’s on course to lose 
  • Even if it could, Oracle needs 4.5GW of capacity. Stargate Abilene is meant to be completed by the end of 2026 (six months behind schedule), but (as I reported last week) only appears to have 200MW of the 1.5+GW of actual, real power it needs right now, and won’t have enough by the end of the year.
  • Even if Abilene was completed on time, Oracle only has one other data center location planned — a 1.4GW data center plot in Shackelford, Texas that has only just begun construction, and will only have a single building by the second half of 2026.

None of this bullshit is happening, and it’s time to be honest about what’s actually going on.

OpenAI is not building “the AI industry,” as this is capacity for one company that burns billions of dollars and has absolutely no path to profitability. 

This is a giant, selfish waste of money and time, one that will collapse the second that somebody’s confidence wavers.

I realize that it’s tempting to write “Sam Altman is building a giant data center empire,” but what Sam Altman is actually doing is lying. He is lying to everybody. 

He is saying that he will build 250GW of data centers in the space of eight years, an impossible feat, requiring more money than anybody would ever give him in volumes and intervals that are impossible for anybody to raise. 

Sam Altman’s singular talent is finding people willing to believe his shit or join him in an economy-supporting confidence game, and the recklessness of continuing to do so will only harm retail investors — regular people beguiled by the bullshit machine and bullshit masters making billions promising they’ll make trillions.

To prove it, I’m going to write down everything that will need to take place in the next twelve months for this to happen, and illustrate the timelines of everything involved.