Lightnews — Scholar-powered news

Ed Zitron

@edzitron.com

170K followers 2.7K following 25K posts

British, But In Las Vegas and NYC
ezitron.76 Sig
Newsletter - wheresyoured.at
https://linktr.ee/betteroffline - podcast w/ iheartradio
Chosen by god, perfected by science
CEO at EZPR.com - Award-Winning Tech PR

Posts Replies Media Videos

Ed Zitron

@edzitron.com

The ones that are sat unpowered in Santa Clara are *48MW projects*. How, exactly, are we going to get the 200, 500 or 1.4GW ones up and running? I don't think this is a regulatory issue!

In Santa Clara, California, where the world’s biggest supplier of artificial-intelligence chips is based, Digital Realty Trust Inc. applied in 2019 to build a data center. Roughly six years later, the development remains an empty shell awaiting full energization. Stack Infrastructure, which was acquired earlier this year by Blue Owl Capital Inc., has a nearby 48-megawatt project that’s also vacant, while the city-owned utility, Silicon Valley Power, struggles to upgrade its capacity.

November 10, 2025 at 6:20 PM

Ed Zitron

@edzitron.com

Me, two weeks ago
www.wheresyoured.at/big-tech-2tr...

As we speak, big tech is hoarding masses of GPUs, and by extension, masses of depreciating assets that will without fail lead to a massive impairment loss, which will come directly out of their earnings, and when they’ve spent tens or hundreds of billions of dollars on GPUs, these impairments will be extremely harsh, and come directly out of earnings.

So, on top of spending hundreds of billions of dollars in capital expenditures, big tech is also conjuring a giant anvil above their heads. GPUs start losing value the moment you buy them, and by purchasing hundreds of billions of expensive, ever-rotting assets, big tech guarantees their revenues will take a hit sometime in the future, and likely multiple times as each generation ages.

Let me put it in really simple terms, and use a hypothetical example.

Ten companies bought $10bn of GPUs each to sell AI services, each requiring billions of dollars of data center infrastructure to go on top of them to make them run. As a result, one of the companies — let’s call them BallWeave — has had to take on an absolute shit ton of debt to buy them. BallWeave eventually runs into solvency issues, and has to start selling off masses of GPUs, massively reducing the market value of these GPUs by 30%.

This causes the other 9 companies to have to take an immediate impairment — a $3bn haircut directly off of their balance sheet. There is no spreading it out.

This is a problem, in part because $3bn is a lot of money, and in part because that’s more than the revenue being generated from AI. The impairment has an immediate negative effect on their assets, while also directly eating into their income, because this shit is almost certainly being financed.
This hypothetical scenario is quite real, and it can hurt you, and will hurt any and all companies buying masses of GPUs.

One last thing: Remember how I said that there are many ways to calculate depreciation over the useful lifetime of an object? Well, the depreciation rate is always the difference between the initial value of the item and its salvage value — which means how much a company can expect to make when they dispose of the item, either by reselling it or literally scrapping it for parts. So, for the sake of illustration, if a GPU costs $50,000 and has a useful life of five years, after which it has a nominal salvage value of $10,000, the amount said company could mark down as depreciation over that lifespan is $40,000.

This is simply me talking out loud, but I worry that there’s no real way to accurately predict what that salvage value of an enterprise GPU will be. It’s contingent on a bunch of factors — namely, how long these cards last in real-world conditions, where they’re experiencing near-constant utilization, as well the pace and the quality of Nvidia’s future releases.

It’s bad, and only compounds the fact that they’ve already created a giant, ugly, unsustainable hole on top of their earnings.

November 10, 2025 at 5:09 PM

Ed Zitron

@edzitron.com

Turns out Clammy Sam Altman actually did request help federal assistance for building data centers in a letter sent last week to the US government, specifically saying that the CHIPS act investment credit should apply to power, GPUs and data center construction.
cdn.openai.com/pdf/21b88bb5...

Expand tax credits that accelerate American manufacturing. The Administration has already
taken critical steps to strengthen American manufacturing by extending the Advanced
Manufacturing Investment Credit (AMIC) for semiconductor fabrication. OSTP should now
double down on this approach and work with Congress to further extend eligibility to the
semiconductor manufacturing supply chain; grid components like transformers and specialized
steel for their production; AI server production; and AI data centers. Broadening coverage of the
AMIC will lower the effective cost of capital, de-risk early investment, and unlock private capital
to help alleviate bottlenecks and accelerate the AI build in the US

November 7, 2025 at 9:17 PM

Ed Zitron

@edzitron.com

I cover this in the piece! OpenAI doesn't have capex - everybody else is paying for the land, the data centers, the GPUs, etc. OpenAI doesn't have assets other than real estate. They even buy other companies in stock.

What About Compute Contract Pre-Payments?
So, as we’re limiting this analysis to the period from 2023 through H1 2025, we’re limited to three vendors:

Oracle for Stargate Abilene, with the initial $30bn a year deal reported in early June 2025.
CoreWeave’s $11.6bn (at the time) compute deal.
Google, which is paying CoreWeave to provide OpenAI’s compute.
There is a chance that OpenAI had to pre-pay some chunk of its CoreWeave deal — five years at $11.6bn could foreseeably be $2.3bn (1/5th of the deal), and given that servicing OpenAI’s needs would require significant upfront expenses from CoreWeave (in the form of new facilities and hardware), it’s not inconceivable that it would ask Altman to reach into his wallet.

Then again, CoreWeave ended Q2 2025 with around $2bn in cash on a loss of $605m, and mentioned that its “weighted-average prepayment across all active contracts was 15% to 25% of the total contract value,” adding that “the level of prepayments…may fluctuate over time” as CoreWeave scaled its operations. To be clear, 15% of $11.6bn is $1.74bn, and 25% is $2.9bn.

Those are high enough costs that I cannot believe we wouldn’t see them in the reporting around OpenAI’s H1 2025 costs.

I also do not believe Oracle is actually making OpenAI pay in advance. The Information reported in April 2025 that OpenAI can walk away from the deal for Stargate Abilene “...if Oracle and its construction partners [don’t] complete the project quickly enough,’ which heavily suggests that OpenAI hasn’t had to pay them a dime yet. While Stargate is partially operational, that only happened in September 2025, and again, any prepayment would be kind of obvious.

While the common refrain is that OpenAI is “investing in data centers,” I see no actual evidence that OpenAI is paying for a single one.

Oracle is handling the GPU purchases for Stargate Abilene, and a joint venture is paying for the construction. Stargate Norway is handled and funded by Nscale and Aker, Stargate UAE is funded by the UAE itself, Vantage Data Centers is raising $38bn in debt to build more Stargate data centers too — and I see no evidence that OpenAI is having to dip into its own funds. OpenAI bought Jony Ive’s company in stock, Statsig in stock, and while we don’t know how much it paid for Software Applications Incorporated or Roi, I’d bet that was paid in stock too. OpenAI was meant to put $19bn into Stargate, but I can find no evidence that it did so, or even that SoftBank did so.

In fact, I can’t find any evidence of OpenAI spending any money on anything other than itself, outside of a $50m non-profit fund, which…is still part of OpenAI.

November 7, 2025 at 6:12 PM

Ed Zitron

@edzitron.com

Reports say that OpenAI will need to raise another $70bn in 2027.

Based on my calculations, I believe OpenAI needs to raise $100bn in 2026 just to handle its rate of spending, and potentially more based on $1.4tr+ in deals it has agreed to.

What a mess.
www.wheresyoured.at/where-is-ope...

OpenAI Will Need To Raise Over $100bn In 2026 To Survive At Its Current Rate Of Spend
The Information recently reported that OpenAI would raise $90bn in 2027, and I don’t believe this company can make it to July 2026 without raising at least another $30bn, especially because it now has the burden of its deal with Amazon Web Services, on top of CoreWeave, on top of Microsoft, on top of Google, and on top of Oracle via Stargate Abilene.

As discussed previously, these amounts work out to around $11bn a month by the year 2027, a year when this company believes it will make $60bn in revenue.

I do not see how it is possible in OpenAi’s current state to handle any of these costs, even with the supposed $20bn arriving in December from SoftBank. In its H1 2025 form — with effectively a single vendor compute contract at a discounted rate and little capital expenditure burn (as OpenAI’s partners shoulder the entirety of those costs) — OpenAI is on course to burn far in excess of the supposed $8bn or $9bn that it has leaked to the press.

One day, I imagine we’ll find out what it is that’s really going on at OpenAI.

And when we do, I’ll be taking detailed notes on those who doubted me.

November 7, 2025 at 5:35 PM

Ed Zitron

@edzitron.com

OpenAI CFO Sarah Friar claims that OpenAI could have "healthy gross margins" if it "wasn't seeking to invest so aggressively." I believe the truth is that OpenAI's on path to getting more out of their models is more inference - which will make margins *worse.*
www.wheresyoured.at/where-is-ope...

OpenAI Can’t Fix Its Cost Problems With More GPUs, As Inference May Be One Of Its Largest Costs And Modern GPUs Aren’t Making Inference Cheaper
OpenAI CFO Sarah Friar said at that WSJ conference that “OpenAI could reach profitability on “very healthy” gross margins in its enterprise and consumer businesses quickly if it weren’t seeking to invest so aggressively,” a statement that makes absolutely no sense to me based on my exhaustive analysis of everything OpenAI has ever said or done.

“More data centers” do not allow OpenAI to reduce its costs, and, indeed, there is no proof that OpenAI or NVIDIA has worked out a way to reduce the cost of inference at all. As David Rosenthal — employee #4 at NVIDIA — said a few weeks ago, “the key feature of each successive generation of AI chips is that it is more efficient at using power…but that doesn't mean they use less power, they use more but less per operation.”

OpenAI is filling eight buildings full of Blackwell GPUs that have negative 100% gross margins, and I see no reason to believe that the cost OpenAI pays is somehow profitable for the operator as a result.

Remember, OpenAI pays a discounted rate on Microsoft Azure. It’s highly unlikely Oracle is charging it less than Microsoft, and much less probable given that the costs are so incredibly high.

In fact, I don’t even know what OpenAI hopes to achieve with these data centers, as for the most part, OpenAI doesn’t own any of this infrastructure outside of the still-theoretical 26 Gigawatts of data centers it’s building for NVIDIA, Broadcom and AMD.

I also do not believe that new GPUs are making inference cheaper, and if they were, we’d certainly have heard about it from somewhere, or had literally any proof of any kind.

OpenAI Needs More Inference To Get More Out Of Models, Using Reasoning To Eek Out Incremental Improvements As It Runs Out Of Other Ideas
We have, at this point, reached the end of training-based scaling laws for Large Language Models, and I believe that OpenAI is buying more compute as a means of forcing more reasoning models into its users’ hands for incremental increases in model efficacy, because training models isn’t going to do it.

In simpler terms, OpenAI is going to try (or is already trying) to force its models to spend more time on test-time compute — “thinking” (these models don’t think like humans, it’s just going over the steps of an output) for more time in the hopes that doing so will produce a better output.

This is the kind of lazy, entitled crap that OpenAI is best known for. Instead of accepting that we might be at the end of the Large Language Model era, I believe this company intends to get as much compute capacity as possible with the goal of running user queries through compute-intensive reasoning models in the hopes that something new happens.

I don’t think it will.

November 7, 2025 at 5:35 PM

Ed Zitron

@edzitron.com

My theory is that OpenAI's inference costs are much higher than expected. Every new GPU is marketed for inference, Stargate Abilene is full of inference-focused Blackwell GPUs, and OpenAI's latest releases have all been very, very inference intensive.
www.wheresyoured.at/where-is-ope...

If Inference Costs Are Higher, Things Are Only Going To Get Worse For OpenAI
So, OpenAI's reported burnrate is around $13.7 billion, leaving us with a gap of $5.3 billion, and that's not including the cloud credits or line of credit of $4bn.

Of all OpenAI's costs, I can't think of any others that might be secretly higher than we knew. Staff? Nah, I believe 3000 people would be expensive, but at least consistent with what OpenAI has leaked. Training? I guess, but is that really $5.3 billion of burn more? What would they be doing? Why would OpenAI be spending that much more on training? We have tangible reporting about the costs of its training runs - $500m a run for GPT-5, for example - and I can imagine while those costs are high, I don't think they're out of control.

You know what might be out of control? Inference - the cost of serving the outputs from models.

While I can’t say conclusively, this makes a hell of a lot more sense than there just being billions of dollars left unaccounted for. I refuse to believe that OpenAI is simply sitting on piles of secret cash, as am I refusing to believe that it has magically brought down its expenses.

If this is the case, OpenAI is in trouble, because the cost of inference is out of control, and I believe the reason it wants more compute is because running ChatGPT is simply too compute intensive, regardless of whether it makes money.

80-90% of AI Compute Spend Is Inference, Every New GPU Is Inference Focused - Suggesting That Inference Is The Primary Reason That OpenAI Wants More Data Centers
Fun fact! According to the MIT Technology Review, 80% to 90% of computing power is used for inference, despite the commonly-held belief that much of it is used for training.

Don’t believe me? Let’s read some marketing collateral!

NVIDIA GB200 (Blackwell): “GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design. It boasts a 72-GPU NVIDIA NVLink™ domain that acts as a single, massive GPU and delivers 30x faster real-time trillion-parameter large language model (LLM) inference.”
This page also includes numbers like “LLM inference 30x vs. NVIDIA H100 GPU,” versus 4x for LLM training on the H100.
NVIDIA GB200 NVL72 (Blackwell): Title: NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference.
NVIDIA GB300 NVL72 (Blackwell): “purpose-built for test-time scaling inference and AI reasoning tasks.”
AMD MI450 “Helios” Rack: “...a generational leap that enables trillion parameter training and large scale AI inference.
Per CNBC about the Broadcom chip designed with OpenAI: “OpenAI and Broadcom have been collaborating for 18 months on a new line of co-designed chips optimized for inference and networked through Broadcom’s Ethernet stack.”
“NVIDIA Unveils Rubin CPX: A New Class of GPU Designed for Massive-Context Inference”: The word inference is used 5 times in this blog, and the word training is not used once.
Blog on NVIDIA titled “NVIDIA Rubin CPX Accelerates Inference Performance and Efficiency for 1M+ Token Context Workloads,” the word inference is used 26 times, the word training is not used once.

“NVIDIA Unveils Rubin CPX: A New Class of GPU Designed for Massive-Context Inference”: The word inference is used 5 times in this blog, and the word training is not used once.
Blog on NVIDIA titled “NVIDIA Rubin CPX Accelerates Inference Performance and Efficiency for 1M+ Token Context Workloads,” the word inference is used 26 times, the word training is not used once.
Jensen Huang in a conversation with investor Bill Gurley: “inference is about to go up by a billion times.”
Jensen Huang in a presentation at GTC Paris: “The number of people using inference has gone from 8 million to 800 million — 100x in just a couple of years.”
SemiAnalysis in March 2025: Vera Rubin is “built for reasoning,” and GTC, NVIDIA’s conference, was addressing the new “test-time scaling” era (IE: reasoning-based inference), specifically stating that this followed “post-training scaling” on the bizarre chart NVIDIA used in Jensen Huang’s speech:

OpenAI wants to expand its data center infrastructure so that it can continue its inference sprawl as the world’s largest AI compute customer, all while obfuscating its costs.

November 7, 2025 at 5:35 PM

Ed Zitron

@edzitron.com

Based on my analysis, OpenAI burned $19 billion of their $28.6 billion in cash from 2023 through 1H2025, meaning there's at least $4.1 billion in cash unaccounted for, as my analysis includes *ALL* reported costs that OpenAI has incurred. Where is the money?
www.wheresyoured.at/where-is-ope...

OpenAI Burned An Estimated $19bn In Cash Between 2023 and H1 2025 — Where Did It Go?
OpenAI had $9.6bn in cash at the end of June.

So, Microsoft’s funding in 2023 was $2bn in cash (and $8bn in credits, but I’m not counting those), add in the $6.6bn of cash from 2024 (i’m not including the line of credit, as that would be a “cash equivalent), $10bn from SoftBank and other investors in June 2025, and we assume $2bn in revenue in 2023, $3.7bn in revenue in 2024, and $4.3bn in the first half of 2025 (for a total of $28.6bn in cash), this means that OpenAI burned $19bn in cash from the beginning of 2023 through the end of June 2025.

Yet the reported losses are much smaller.

For the sake of argument, I’m also going to leave OpenAI’s $4bn line of credit out of the calculation.

In 2024, OpenAI allegedly burned $5bn. Through H1 2025, it had a “net loss” of $13.5bn but an “operating loss” of $6.7bn, and the painfully-worded accounting leaked to The Information is done in such a way as to make it unclear how much of that is cash, with the hopes that you believe it’s not cash burn.

$19bn in burned cash doesn’t make sense based on any reported numbers.

Even if we assume OpenAI burned $2bn in cash in 2023, $5bn in cash in 2024 and $6.7bn in cash in H1 2025, that's $13.7 billion in burn - yet OpenAI has only $9.6 billion in cash at the end of H1 2025.

That's a $4.1 billion difference in burnrate. $4.1 billion is a pretty big number.

That's a $4.1 billion difference in burnrate. $4.1 billion is a pretty big number.

I guess there could be prepayments as part of its deals with Broadcom or CoreWeave, but even then, The Information had another tidbit:

OpenAI also recorded around $670 million of equity it gave to unnamed entities for providing the company with servers in the first half of 2025, compared to $1 billion in such equity compensation last year. While it’s not clear which cloud provider or providers received the equity, the decision could help OpenAI minimize its cash burn as it projected it would spend $16 billion to rent servers this year.
So, on top of paying cash for compute, OpenAI has also been paying for it in stock.

If OpenAI “burned” $5bn in 2024, and had an “operating loss” of $6.7bn, how did it end up with $9.6bn in cash at the end of June?

With the revenue it’s (apparently) making, OpenAI should have significantly more cash than it has.

Unless, of course, what has been leaked isn’t reflective of OpenAI’s actual cash position.

November 7, 2025 at 5:35 PM

Ed Zitron

@edzitron.com

In an analysis of OpenAI's costs, revenues and funding from 2023 to 1H 2025, it raised $28.6bn in cash ($10bn rev included), and had $13.7bn in cash losses.

Reports say OpenAI ended 1H 2025 with $9.6bn. It should have $14.9bn in cash. There's a $4.1bn+ hole.
www.wheresyoured.at/where-is-ope...

A Brief History of OpenAI’s Revenues and Costs
If there are other articles reporting OpenAI's revenues or costs, please do get in touch. But I don't think anybody has ever done a full run-down. Behold!

On February 9, 2024, The Financial Times said that “OpenAI’s revenues have surpassed $2bn on an annualised basis,” with the company apparently hitting it “In December 2023,” according to “two people with knowledge of its finances.”
There was no mention of any comment related to OpenAI in this article.
On June 12, 2024, The Information reported that “OpenAI has more than doubled its annualized revenue to $3.4 billion in the past six months or so, OpenAI CEO Sam Altman has told staff.” This amounts to around $283m in revenue a month at the end of calendar year Q2 2024.
This appears to be sourced from something that Sam Altman told staff.
Curiously, this is the only time I’ve seen OpenAI comment on one of these stories, with “...an OpenAI spokeswoman [saying] the financial details were “inaccurate,” without elaborating.” I have read every single one of these and I cannot find another instance of this happening. This makes me believe, along with other information from sources, that these numbers are particularly accurate.
This article also confirms that OpenAI gets a “roughly” 20% revenue share cut from Microsoft selling its models.
On July 24, 2024, The Information published that OpenAI “could lose as much as $5 billion” in 2024, “according to an analysis by The Information, based on previously undisclosed internal financial data and people involved in the business.”
Spokespeople for Microsoft and OpenAI did not have a comment.
On September 27. 2024, the New York Times reported that “OpenAI’s monthly revenue hit $300 million in August…and the company expects about $3.7 billion in annual sales [in 2024],” according to a financial professional’s review of documents.
That’s $300m in revenue in the month of August 2024. Details are important. OpenAI declined to comment.

On October 9, 2024, The Information reported that “OpenAI burned $340 million in the first half of 2024” and that its “cash burn was lower than previously thought,” based on “OpenAI financial documents viewed by The Information.”
These documents also “project [OpenAI’s] total losses between 2023 and 2028 to be $44 billion” not including stock compensation, and, laughably, that OpenAI would “make $14 billion in profit [based on the analysis] in 2029.”
OpenAI also, in this document, said that “model training costs could rise sharply in the next couple of years, to as high as $9.5 billion a year in 2026.”
This same piece included projections that OpenAI would make $12bn in 2025, $26bn in 2026, $44bn in 2027, $69bn in 2028, and $100bn in 2029.
There was no mention of any comment related to OpenAI in this article.
On February 21, 2025, The Information reported that OpenAI forecasted that “one-third of [its] revenue growth would come from spending by SoftBank to use OpenAI’s products across its companies,” with SoftBank spending $3bn per year. It took until November 5, 2025 for the joint venture involving this spend to actually form.
This is sourced to disclosures by OpenAI to its investors.
This piece also asserts that $3bn from SoftBank would get OpenAI to $12.7bn in revenue for 2025.
There was no mention of any comment related to OpenAI in this article.
In this piece, The Information reports that “...[OpenAI’s] spending on Microsoft-owned data centers should ramp up significantly in the next few years, more than doubling from $13 billion this year to $28 billion in 2028.”
On March 26, 2025, Bloomberg reported that OpenAI “expects to more than triple its revenue to $12.7 billion in 2025,” and that it generated “$3.7 billion in annual revenue last year,” according to a person familiar with the matter who “spoke on condition of anonymity to discuss internal communications.
In the piece, Bloomberg states that “OpenAI does not expect to be cash-flow positive until 2029, t…

On June 9, 2025, an OpenAI spokesperson told CNBC that it had hit “$10 billion annual recurring revenue,” excluding licensing revenue from OpenAI’s 20% revenue share and “large, one-time deals.” $10bn annualized revenue works out to around $833, a month.
This news also revealed that OpenAI had reached $5.5bn in annualized revenue in December 2024, or $458m.
I have some fucking questions.
Why is OpenAI deliberately excluding that cut? Is it because Microsoft didn’t authorize it?
What exactly is the time period that “annualized” covers?
What are these “large, one-time deals” and why wouldn’t you include them in revenue?
On September 5, 2025, The Information reported that “OpenAI projected its cash burn this year through 2029 will rise even higher than previously thought, to a total of $115 billion.”
None of this information is sourced in a way that I can see. It’s the strangest thing. The only mention of anything like that is to say that one thing — that OpenAI would consider renting out servers to other companies — was “according to recent comments by OpenAI executives.”
While I can’t speak to how this information got the reporter, it feels like this was OpenAI. There is no mention of asking anyone for comment.
Nevertheless, this piece says that:
OpenAI’s cash burn would be around $9bn in 2025 (though the copy says $8bn, the “Cash Crunch” graph says $9bn), around $17bn in 2026, around $35bn in 2027, around $47bn in 2028, around $8bn in 2029, and then, inexplicably, have a positive cash flow of $38bn. Huh?
The piece also mentions that cash burn would be “more than $10 billion higher” than previous cash burn projections from OpenAI, without linking to or sourcing where they received that information (though it’s potentially part of their source’s info).
This piece also vastly increases revenues, saying that OpenAI will make $13bn in revenue in 2025, $30bn in 2026, $60bn in 2027, $100bn in 2028, $145bn in 2029, and $200bn in 2030.

On September 10, 2025 — five days after a story saying OpenAI would burn $35bn in 2027 — Oracle and OpenAI announced a $300bn, five-year-long cloud deal. I believe this was intentional as this deal is not reflected in the cash burn above.
On September 19, 2025, The Information reported that OpenAI “planned to spend about $100 billion renting backup servers from cloud providers over the next five years,” according to what executives have told shareholders.
According to The Information, “including the backup servers, OpenAI is planning to spend about $85 billion a year on server rentals over the next five years, on average, according to the company’s private projections.”
Yet deeper in the piece, The Information added that “OpenAI plans to spend $450 billion renting servers through 2030, including $100 billion in backup servers it says are ‘monetizeable’.”
There is no mentioned request for comment.
On September 29, 2025, The Information reported that OpenAI “generated around $4.3 billion in revenue in the first half of 2025, about 16% more than it generated all of last year, according to financial disclosures to shareholders.”
OpenAI “burned $2.5 billion” in the first half of 2025 (assume all following statements are from that time period).
OpenAI spent $2bn on sales and marketing.
OpenAI spent $6.2bn on research and development, which The Information believes is due to research compute costs.
OpenAI “looked on track to meet its full-year revenue projection of $13 billion and $8.5 billion of cash burn.”
There is no mentioned request for comment.
On October 31, 2025, The Wall Street Journal reported that Microsoft had mentioned in its recent earnings that it had taken a $4.1bn loss on its investment in OpenAI, implying a more than $12bn quarterly loss at the company due to Microsoft’s 32.% stake in that quarter (which has now dropped to 27%).
The reason I exhaustively wrote this out is that when you lay it all out, a pattern develops: OpenAI is bullshitting its projec…

November 7, 2025 at 5:35 PM

Ed Zitron

@edzitron.com

OpenAI is has now signed over a trillion dollars of cloud deals it can't afford. Based on these costs, OpenAI will spend $11 billion a month by 2027. This is more than the operating expenses of Meta, a profitable company with 78k employees.
www.wheresyoured.at/where-is-ope...

Though there are likely losses related to stock-based compensation, this only represented a cost of $2.5bn in the first half of 2025. The Information also reported that OpenAI “spent more than $2.5 billion on its cost of revenue,” suggesting inference costs of…around that?

I don’t know. I really don’t know.

But what I do know is that Sam Altman has made obscene, ridiculous promises to seemingly every company in the tech industry. He promised $38bn over seven years to Amazon Web Services (starting immediately), $300bn to Oracle over five years (starting in 2027), $22.4bn to CoreWeave over five years (started October 2025), and pledged that OpenAI will spend $250bn on Microsoft Azure, which I assume wraps in the already-reported $28bn it will spend on Azure in 2028.

Based on these calculations, and OpenAI’s already-reported costs, I calculate that by 2027, OpenAI will be spending over $11bn a month on compute costs alone.

Or, perhaps more. Last night, in a lengthy thread that was roundly mocked for its defensive tone and Dostoyevskian wordcount, Sam Altman said that OpenAI is “looking at [compute] commitments of about $1.4 trillion over the next 8 years.” Now, note the non-committal language there (“looking at”), but if we assume that it goes through with all of these, we’re looking at an average monthly spend of $14.58bn ($1.4tn divided by 96 months).

For some context, Meta’s last quarterly (not monthly) operating expenses were $30bn — up from $27bn in the previous quarter — which works out to around $10bn a month for a company that regularly makes more than $15bn a quarter in net income (on revenues of over $51bn, though its latest quarter involved a $15bn or so hit to net income due to tariffs).

Similarly, Microsoft had operating expenses of $39bn in its last quarter — around $13bn a month — on net income of over $27bn (and revenues of over $77bn).

For some context, Meta has over 78,000 employees, and Microsoft over 200,000. In August, The Verge reporte…

November 7, 2025 at 5:35 PM

Ed Zitron

@edzitron.com

November 7, 2025 at 3:33 PM

Ed Zitron

@edzitron.com

Had a lot of fun with this one

What is the thing! It's like vague posting being like, I really shouldn't be alone right now. You know, I just had such a bad night.
I wish someone would text me and tell me they're buying a GPU. I don't know, man.
You're the CEO of the largest company on the stock market.

November 7, 2025 at 2:30 PM

Ed Zitron

@edzitron.com

I believe the culprit for the missing cash burn is OpenAI's inference spend, something I have taken great pains to explain in this piece. This is the culmination of two years of collecting pieces of a puzzle that I am still yet to complete.

edzitronswheresyouredatghostio.outpost.pub/public/promo...

A Few Theories About Where OpenAI’s Cash Is Going
The following are hypotheses.

OpenAI Is Spending Far More On Inference Than We Have Been Led To Believe
Let’s look at the 2024 OpenAI cost breakdown from The Information:

Similarly, The Information reported the following of OpenAI’s 1H 2025 burn:

OpenAI also spent more than $2.5 billion on its cost of revenue in the first half. Those expenses are largely related to renting Nvidia-powered servers from Microsoft to run ChatGPT and other products.

I’m sorry - if this is inference (I’m confident it is), I…just don’t believe that’s how much OpenAI spent, based on both its reported 2024 spend and the maths I’ve run above.

Putting aside the unprofitability side, If OpenAI spent $2 billion on inference in 2024 to make $3.7 billion in revenue a year, how did it spend $2.5 billion on inference to make $4.3 billion - reducing their costs from $1.85 per dollar to $1.72 per dollar - while also blowing way, way more cash?

Putting aside the unprofitability side, If OpenAI spent $2 billion on inference in 2024 to make $3.7 billion in revenue a year, how did it spend $2.5 billion on inference to make $4.3 billion - reducing their costs from $1.85 per dollar to $1.72 per dollar - while also blowing way, way more cash?

If your argument is that “the cost of inference came down,” I would like an explanation as to how. Since September 2024, OpenAI launched its o1 reasoning model and ChatGPT’s advanced voice mode, its SearchGPT product in October 2024, its $200-a-month “pro” subscription that loses OpenAI money because users use it too much and a new version of its generative video model Sora in December 2024, its o3-mini reasoning model in January 2025, its Deep Research product in February 2025 (which Artificial Analysis found cost roughly $30 per API call, which is reflective of the user’s costs which are likely lower than OpenAI’s), along with GPT 4.5 - which Sam Altman described as a “giant, expensive model” and required OpenAI to add “thousands of GPUs,” ChatGPT 4o’s image generation product in March 2025 - which was so onerous on OpenAI’s infrastructure that Sam Altman said that it was rate limiting the product as “GPUs were melting,” o3 and o4-mini reasoning models in April 2025, and o3-pro in June 10 2025.

So, you’re telling me that after months of releasing the most compute-intensive models known to man, ones that have tested the limits of OpenAI’s GPU capacity, that their cost of revenue has decreased? That their inference costs have only doubled?

I think the truth is far simpler: I believe OpenAI’s cost of inference is much, much higher than anyone letting on.

If we assume that OpenAI burned $2 billion in 2023, $5 billion in 2024 and $6.7 billion through 1H 2025, it should have $14.9 billion left.

The only way that gap makes sense is if OpenAI’s inference costs are much higher than we have been led to believe.

Think about it. OpenAI has been flush with cash, and billions …

November 6, 2025 at 9:24 PM

Ed Zitron

@edzitron.com

This piece is the single-most detailed and thorough analysis of OpenAI's revenue, funding and costs in existence. It is a one-of-a-kind evaluation of a company with conflicting leaks and what I believe is an astonishingly higher burn rate.

edzitronswheresyouredatghostio.outpost.pub/public/promo...

On February 9 2024, The Financial Times said that “OpenAI’s revenues have surpassed $2bn on an annualised basis,” with the company apparently hitting it “In December 2023,” according to “two people with knowledge of its finances.”
There was no mention of any comment related to OpenAI in this article.
On June 12 2024, The Information reported that “OpenAI has more than doubled its annualized revenue to $3.4 billion in the past six months or so, OpenAI CEO Sam Altman has told staff.” This amounts to around $283 million in revenue a month at the end of Calendar Year Q2 2024.
This appears to be sourced from something that Sam Altman told staff.
Curiously, this is the only time I’ve seen OpenAI comment on one of these stories, with “...an OpenAI spokeswoman [saying] the financial details were “inaccurate,” without elaborating.” I have read every single one of these and I cannot find another instance of this happening. This makes me believe, along with other information from sources, that these numbers are particularly accurate.
This article also confirms that OpenAI gets a “roughly” revenue share cut from Microsoft selling its models.
On July 24 2024, The Information published that OpenAI “could lose as much as $5 billion” in 2024, “according to an analysis by The Information, based on previously undisclosed internal financial data and people involved in the business.”
Spokespeople for Microsoft and OpenAI did not have a comment.
On September 27 2024, the New York Times reported that “OpenAI’s monthly revenue hit $300 million in August…and the company expects about $3.7 billion in annual sales [in 2024],” according to a financial professional’s review of documents.
That’s $300 million in revenue in the month of August 2024. Details are important. OpenAI declined to comment.
On October 9 2024, The Information reported that “OpenAI burned $340 million in the first half of 2024” and that its “cash burn was lower than previously thought,” based on “OpenAI financial document…

On March 26 2025, Bloomberg reported that OpenAI “expects to more than triple its revenue to $12.7 billion in 2025,” and that it generated “$3.7 billion in annual revenue last year,” according to a person familiar with the matter who “spoke on condition of anonymity to discuss internal communications.
In the piece, Bloomberg states that “OpenAI does not expect to be cash-flow positive until 2029, the person said, a year when it projects revenue will top $125 billion.”
Per Bloomberg’s story, “OpenAI [did] not expect to be cash-flow positive until 2029, the person said, a year when it [projected] revenue will top $125 billion.”
OpenAI declined to comment.
On June 9 2025, an OpenAI spokesperson told CNBC that it had hit “$10 billion annual recurring revenue,” excluding licensing revenue from OpenAI’s 20% revenue share and “large, one-time deals.” $10 billion annualized revenue works out to around $833 million a month.
This news also revealed that OpenAI had been at $5.5 billion in annualized revenue in December 2024, or $458 million.
I have some fucking questions.
Why is OpenAI deliberately excluding that cut? Is it because Microsoft didn’t authorize it?
What exactly is the time period that “annualized” covers?
What are these “large, one-time deals” and why wouldn’t you include them in revenue?
On September 5 2025, The Information reported that “OpenAI projected its cash burn this year through 2029 will rise even higher than previously thought, to a total of $115 billion.”
None of this information is sourced in a way that I can see. It’s the strangest thing. The only mention of anything like that is to say that one thing - that OpenAI would consider renting out servers to other companies - was “according to recent comments by OpenAI executives.”
While I can’t speak to how this information got the reporter, it feels like this was OpenAI. There is no mention of asking anyone for comment.
Nevertheless, this piece says that:
OpenAI’s cash burn would be around $9 billion i…

On September 10 - five days after a story saying OpenAI would burn $35 billion in 2027 - Oracle and OpenAI announced a $300 billion, five-year-long cloud deal. I believe this was intentional as this deal is not reflected in the cash burn above.
On September 19 2025, The Information reported that OpenAI “planned to spend about $100 billion renting backup servers from cloud providers over the next five years,” according to what executives have told shareholders.
According to The Information, “including the backup servers, OpenAI is planning to spend about $85 billion a year on server rentals over the next five years, on average, according to the company’s private projections.”
Yet deeper in the piece, The Information added that “OpenAI plans to spend $450 billion renting servers through 2030, including $100 billion in backup servers it says are ‘monetizeable’.”
There is no mentioned request for comment.
On September 29 2025, The Information reported that OpenAI “generated around $4.3 billion in revenue in the first half of 2025, about 16% more than it generated all of last year, according to financial disclosures to shareholders.”
OpenAI “burned $2.5 billion” in the first half of 2025 (assume all following statements are from that time period).
OpenAI spent $2 billion on sales and marketing.
OpenAI spent $6.2 billion on research and development, which The Information believes is to due to research compute costs.
OpenAI “looked on track to meet its full-year revenue projection of $13 billion and $8.5 billion of cash burn.”
There is no mentioned request for comment.
On October 31 2025, The Wall Street Journal reported that Microsoft had mentioned in its recent earnings that it had taken a $4.1 billion loss on their investment in OpenAI, implying a more than $12 billion quarterly loss at the company due to Microsoft’s 32.% stake in that quarter (which has now dropped to 27%).

November 6, 2025 at 9:24 PM

Ed Zitron

@edzitron.com

Tomorrow's Premium: Based on analysis of years of reporting, OpenAI has burned $23 billion since 2023, ending 1H 2025 with $9.6bn cash. OpenAI appears to have $9bn+ more in burn than has previously been reported.

Here's $10 off annual.

edzitronswheresyouredatghostio.outpost.pub/public/promo...

OpenAI Burned An Estimated $19 Billion In Cash Between 2023 and 1H 2025 - Where Did It Go?
OpenAI had $9.6 billion in cash at the end of June.

So, Microsoft’s funding in 2023 was $2 billion in cash (and $8 billion in credits, but I’m not counting those), add in the $6.6 billion of cash from 2024 (i’m not including the line of credit, as that would be a “cash equivalent), $10 billion from SoftBank and other investors in June 2025, and we assume $2 billion in revenue in 2023, $3.7 billion in revenue in 2024, and $4.3 billion in revenue in the first half of 2025 (for a total of $28.6 billion in cash), this means that OpenAI burned $19 billion in cash from the beginning of 2023 through the end of June 2025.

Yet the reported losses are much, much smaller.

For the sake of argument, I’m also going to leave OpenAI’s $4 billion line of credit out of the calculation.

In 2024, OpenAI allegedly burned $5 billion. Through 1H 2025, it had a “net loss” of $13.5 billion but an “operating loss” of $6.7 billion, and the painfully-worded accounting leaked to The Information is done in such a way as to make it unclear how much of that is cash, with the hopes that you believe it’s not cash burn.

$19 billion in burned cash doesn’t make sense based on any reported numbers.

Even if we assume OpenAI burned $2 billion in cash in 2023, $5 billion in cash in 2024 and $6.7 billion in cash in 1H 2025, that’s $14.9 billion in missing burn to get us to the $28.6 billion total.

I guess there could be prepayments as part of its deals with Broadcom or CoreWeave, but even then, The Information had another tidbit:

OpenAI also recorded around $670 million of equity it gave to unnamed entities for providing the company with servers in the first half of 2025, compared to $1 billion in such equity compensation last year. While it’s not clear which cloud provider or providers received the equity, the decision could help OpenAI minimize its cash burn as it projected it would spend $16 billion to ren…

November 6, 2025 at 9:24 PM

Ed Zitron

@edzitron.com

AI boosters are beginning to crash out

Capitalism is good. Capitalism is right, capitalism works. Capitalism clarifies, cuts through, and captures the essence of the evolutionary spirit. Capitalism will win.
If you think central planning China can compete with buggy, unoptimized Al chips that crash all the time guzzling power against American ingenuity, the smartest Al researchers in the world, and hundreds of billions of annual capex spending from the most profitable companies in history, good luck with that.

November 6, 2025 at 5:04 PM

Ed Zitron

@edzitron.com

To be clear this is a misread of the situation. Was Friar testing the waters? Sure. Did she think the waters would be fully of angry fish? No
archive.ph/2025.11.06-0...

In addition to OpenAI’s deals with chipmakers, the ChatGPT maker is also eying a broad mix of financing vehicles to fund its infrastructure efforts. Friar said OpenAI is “looking for an ecosystem of banks [and] private equity” to support its ambitious plans. She also hinted at a role for the US government to “backstop the guarantee that allows the financing to happen,” but did not elaborate on how this would work.
Asked for clarification, an OpenAI spokesperson said Friar was speaking in the context of the broader AI industry. There are currently no immediate plans about OpenAI pursuing a federal backstop, the spokesperson said.

November 6, 2025 at 6:14 AM

Ed Zitron

@edzitron.com

What the hell is this statement from NVIDIA? What am I meant to do with this info?

NVIDIA Newsroom Y @nvidianewsroom
•••
X.com
A statement from NVIDIA CEO Jensen Huang.
Statement from NVIDIA
CEO Jensen Huang
As I have long said, China is nanoseconds behind America in Al. It's vital that America wins by racing ahead and winning developers worldwide.
November 5, 2025
• NVIDIA
6:56 PM • 11/5/25 • 260K Views

November 6, 2025 at 5:35 AM

Ed Zitron

@edzitron.com

They’re calling it “the most well-taken comment in history”

Sarah Friar
in
• Following
CFO OpenAl | Co-Launcher @ Ladies Who Launch | Board Member @ Wa...
1m •
I want to clarify my comments earlier today. OpenAl is not seeking a government backstop for our infrastructure commitments. I used the word "backstop" and it muddied the point. As the full clip of my answer shows, I was making the point that American strength in technology will come from building real industrial capacity which requires the private sector and government playing their part. As I said, the US government has been incredibly forward-leaning and has really understood that Al is a national strategic asset.

November 6, 2025 at 5:31 AM

Ed Zitron

@edzitron.com

OpenAI has lost the crucial “dunce” vote.

Matthew Yglesias & @mattygle.... 6h S ...
Come on
SARAH FRIAR
OPENAI CFO
OpenAI Wants Federal Backstop for New Investments
Sarah Friar, the CFO of OpenAl, says the company wants a federal guarantee to make it easier to finance massive investments in Al chips for data centers. Friar spoke at WSJ's Tech Live event in California. Photo: Nikki Ritcher for WSJ
By Wall Street Journal
2 hours ago
+ 2:09

November 6, 2025 at 5:01 AM

Ed Zitron

@edzitron.com

Does this change your read of the situation? Not arguing, I’m genuinely trying to understand it.

That's because Microsoft handles its OpenAI stake using an accounting approach known as the equity method, in which it simply reports its share of losses or income at the AI company. OpenAI recently restructured to become a for-profit entity, and Microsoft will have a 27.5% stake going forward.

November 6, 2025 at 12:50 AM

Ed Zitron

@edzitron.com

Why is the CFO of OpenAI claiming that something disclosed in Microsoft’s earnings (was originally reported as $11.5bn) - that OpenAI lost $12bn in q3 CY2025 - was “way overestimated”? This is disputing the accounting of a $4tr public company!

HT @gerritd.bsky.social

www.wsj.com/livecoverage...