I'm a little confused here. Cost of revenue is lower than revenue. That's good. R&D is the main contributor to losses here and this seems normal in an industry like this. For OpenAI specifically, I think this is problematic. They were the first movers but despite the large R&D they've lost so much ground to Anthropic despite Anthropic seemingly gifting them with weird PR self owns. But if we were to extrapolate this to the industry as a whole, this seems more positive than negative. Am I reading this incorrectly? Unless there's an assumption that R&D costs have to forever go up in order to increase revenue, I feel like this shows that the AI industry is actually on a path to profitability in the long term.
Whether it can physically be as all encompassing as it makes itself out to be or whether it will just be healthily profitable remains to be seen. Kind of like how Uber went from "We'll autonomously drive the world" to "Look, we deliver food, goods, and people to locations and we figured out how to do that in a way that makes profits. Also, ads".
I’m not sure how people are looking at numbers that show, even if we wipe off the enormous R&D expenditures, they are still in the red for inference + sales/marketing + admin and responding “this seems positive”.
It’s like being a sold a car and being told “well if you ignore the fact it has no engine it’s a good buy” yet it also has no wheels.
> Unless there's an assumption that R&D costs have to forever go up in order to increase revenue, I feel like this shows that the AI industry is actually on a path to profitability in the long term.
There are three futures right, I’ll rank them in order of fantasy -
1. Someone achieves AGI. At that point the economics of an individual company don’t even matter.
2. R&D costs do have to forever continue, because LLMs can be continually iteratively improved. Much like chip development, there is no end in sight, at least not on a near term timescale. If you are not continually at the frontier, customers will use a competitor or open/local alternatives.
3. LLMs reach a plateau of functionality. Further gains are minimal, quality reaches the apex of what the technology permits. In this scenario the hyperscalers have no business because open/local models will rapidly reach that same plateau as well.
The leaked numbers completely ignore how much of their compute is subsidized.
It also ignores how much of "R&D" is actually needed for the thing they offer to keep working. Looking at the thread everyone seems to be presuming "R&D" is all "training new models", but that is uncertain.
The people who are completely sold on the belief that AI providers are running at a profit believe him to be utterly, totally and completely wrong in every one of his predictions.
The people who are completely sold on the belief that AI providers are running at a loss they can never recover from believe him to be utterly, totally and completely correct in every one of his predictions.
The reality is that it's not his predictions that matter, but his data, which is almost always correct as of time of writing. If you ignore his opinions, the data presented on liabilities, spend, revenue, loans, commitments, etc across Coreweave, Stargate, Oracle and all of the usual AI companies is, as far as I can tell, correct.
IOW, when it comes to his opinions, it's all about your priors. His data is good, though.
> The reality is that it's not his predictions that matter, but his data, which is almost always correct as of time of writing. If you ignore his opinions, the data presented on liabilities, spend, revenue, loans, commitments, etc across Coreweave, Stargate, Oracle and all of the usual AI companies is, as far as I can tell, correct.
Yeah, I think that he does well with sources and data. I also think that his editorialising can be off-putting for lots of people. I kinda enjoy it, but accept that I have niche tastes.
Not only not understanding ARR, he simply doesn't do data analysis properly - he misses some few months and days in his calculation to prop up his point. This is a mistake chatgpt would have caught.
That's the trouble I have with ARR, because there's no standard, people engage in shenanigans. I do find the 5bn lifetime revenue versus their ARR figures pretty sketchy which is why I really want to see the S1.
Can you be more specific on his incorrect calculations please?
Wait. ARR has no precise definition but has a clear social understanding. It’s clear Ed doesn’t get that and ARR not having a clear definition doesn’t absolve him of the mistake. His misunderstanding was on a different axis.
The miscalculations are pretty clearly pointed out in the tweet I linked earlier.
> Wait. ARR has no precise definition but has a clear social understanding.
This is (historically) a recipe for fraud and badness. If ARR is important enough to be reported, then there should be a GAAP definition.
Do you use calendar month or four week rolling? Do you account for seasonality? How do you recognise revenue? (My sense is that Anthropic do sketchy things with credits, as the consumer ones last for like 180 days and then expire).
ARR is a really, really, really easy metric to make sound like whatever you want which is why I am sceptical of it.
EDIT: I looked at the tweet which is a screenshot of a supposed sheet that Ed built. Unless you have a source for the sheet then I'll need to assign this relatively low credibility (don't know the user, it's a screenshot with no link).
The user is someone I've followed for more than a decade:
> I’m a reporter who has written about technology, economics, and public policy for more than a decade. Before I launched Understanding AI, I wrote for the Washington Post, Vox.com, and Ars Technica. I have a master’s degree in computer science from Princeton.
> I’m working on Understanding AI full-time, and I have no outside investors or donors. Since I started it in 2023, paying subscribers have accounted for a large majority of my income (you can see full details on my source of income on my disclosure page). Their support allows me to work on the newsletter full-time.
Passes my credibility check because I've read a lot of his work, and he's been around the block a few times in journalism circles.
> But I’m a curious little critter and went ahead and added up all of the times that Anthropic had talked about its annualized revenue from 2025 onward, and the results — which you can find with links here! — and based on my calculations, just using published annualized revenues gets us to $4.837 billion.
It’s here in the blog.
> This is (historically) a recipe for fraud and badness. If ARR is important enough to be reported, then there should be a GAAP definition.
I don't think anyone believes the major AI providers are running at a profit? They are openly investing heavily into R&D and building out infrastructure, and according to these numbers way more than revenue. It wouldn't make sense for any of these companies to run at a profit right now as they're still aggressively expanding. The question is whether they will break even in the future, and capture a large enough market segment to sustain the business, allowing revenue to outgrow costs. If these numbers are real, revenue is already higher than COGS which is a really good signal for them.
I think the question is more about whether people believe this is a sound business in the long term, which imo isn't possible to tell based on these numbers yet.
> The people who are completely sold on the belief that AI providers are running at a loss they can never recover from believe him to be utterly, totally and completely correct in every one of his predictions.
It's funny, because you can both believe that these entities are bleeding money on every token and also believe that "financial engineering" will bail them out when they IPO despite this fact.
The fundamentals of running a business that sells products or services for more than the cost to produce them seem increasingly decoupled from the financial success of the company and its owners.
Up until this post, I thought he was someone with good financial insight, analytical chops, and business sense, stuck with an audience that thinks it's still 2023 and ChatGPT 3 is still the pinnacle of the technology, and that he therefore has to pander to in order to pay the bills.
After this supposedly being the reveal for his bubble-bursting massive revelation that will send the industry flying and lead to journalists kicking in his door for interview requests and exposés, I think... well, not that anymore. I thought "the frontier labs are losing money" was rather universally understood, and this really isn't even as bad as the stuff that's publicly visible; the fact that they keep raising hundreds of billions of dollars that they'll one day supposedly be required to show returns on?
> After this supposedly being the reveal for his bubble-bursting massive revelation that will send the industry flying and lead to journalists kicking in his door for interview requests and exposés
I mean, the fact that lots of expenses are not scaling with revenue (sales and marketing 5xed versus revenue 3xing) and that the losses are very very large is important. More importantly, these are audited figures which haven't been seen before.
Right, but this still isn't exactly new information. I don't think anyone was assuming that the labs are close to being profitable or that the losses wouldn't be rather large. The way this was announced was as if it was going to be a bombshell, but it just confirms what everyone (including the investors) was assuming anyway. Now if he had concrete numbers about whether inference at API pricing is profitable, that'd be a different thing (and it's what that hype bit was heavily implying since it's something he constantly keeps harping on, and rightfully so), but as it stands, nothing about these numbers says anything about whether this fundamentally has a road to profitability. It just says that this is a super high-risk high-reward investment, which isn't new information.
> As OpenAI’s worth rose, the increased value of those investor rights created a roughly $30bn charge, added the person. The charge is not expected to recur following the restructuring, they said.
> Stripping out the charge and other non-cash expenses, such as stock-based compensation of staff and computing credits from Microsoft, OpenAI’s losses were $8bn, according to the person.
As a long time FT subscriber, I'm happy you're using them as a source. The Zitron details were more useful to me though.
And none of my points have anything to do with the once off losses. I'm observing that a bunch of costs appear to be scaling with revenue or above revenue, which does not bode well for future profitability.
Also, as an aside, stripping out equity grants is really misleading for a private, high growth tech company.
The losses are scaling with revenue because increase in (expected) revenue increases valuation which increases compensation.
Once expectation stabilises these losses won’t happen because the valuation will remain constant.
A lot of people were paid really high equity grants simply because they started low. You can’t expect them to be paid the same amount each time.
FT themselves point this out and who you believe is up to you.
> The losses are scaling with revenue because increase in (expected) revenue increases valuation which increases compensation.
My original point around equity is that if you pay a substantial fraction of comp in this form, then leaving it out of expenses is pretty bizarre.
Is it your contention that the equity grants are the cause of their increasing losses?
I believe that this is probably not true at all, it's more likely to be S&M (salespeople scale as N not log(N) like engineering/product) particularly given that the product requires tuning for lots of companies (hence all the FDE hires).
More generally, the training costs seem to be increasing which is bad for their future profitability.
Also it should be obvious that you shouldn’t extrapolate stock based compensation in a scale up. People make a one time bounty but that is not recurring obviously.
> Before OpenAI’s switch late last year to become a public benefit corporation, investors in the company received convertible interest rights rather than conventional equity. Under US accounting rules, those interests were treated as liabilities and periodically revalued as the company’s valuation increased.
As OpenAI’s worth rose, the increased value of those investor rights created a roughly $30bn charge, added the person. The charge is not expected to recur following the restructuring, they said.
Stripping out the charge and other non-cash expenses, such as stock-based compensation of staff and computing credits from Microsoft, OpenAI’s losses were $8bn, according to the person.
I presume that this is what you're talking about, right?
That doesn't actually disagree with what I noted above using the (more detailed) figures from Ed's article. I noted that their revenue scaled by about 3x, while many costs (cost of revenue, sales & marketing, r&d) scaled by either equal (r&d) or greater than their revenue scaled. That's the point I was (apparently badly) making, nothing to do with the stock based compensation causing their losses. In any case, the loss was actually driven by treatment of the non-profit shares.
> Also it should be obvious that you shouldn’t extrapolate stock based compensation in a scale up. People make a one time bounty but that is not recurring obviously.
Correct, in some sense this is a once-off, however, most tech companies continue granting stock over time, so it's definitely worth including in actual margins. (This is a more general point that's not exclusive to Open AI).
The Uber comparison makes no sense. This is the opposite situation. Uber lost money on rides, OpenAI is (possibly) making money on inference. Uber used an R+D moonshot to autonomous driving to justify capturing an established industry without reducing costs meaningfully. OpenAI has a core product that risks becoming a commodity with open source models only 6 months behind.
Do you have a source for the claim that Uber was making money on rides during its decade of enormous unprofitability?
Its public stance was that growth was more important than profit. Why wouldn't they be subsidizing rides to fuel growth if that is their publicly stated goal?
And anyway, we got the Uber Files some years ago which made it explicit:
"In October 2014 in Madrid, the presentation shows, the hourly subsidy to drivers of $17.50 was almost twice the hourly fare it charged, which was only $9.10. In Berlin, the gross hourly fare Uber charged was $2.20, while the subsidy it paid out to drivers was $10.20 an hour. Uber burned through cash to “buy revenue”, in the words of the presentation."
The vast, vast amounts of money they spent on driver incentives city by city would seem to support the OPs claim (source: I was familiar with their spend on ads in the US approximately 10 years ago).
This is an impossible ask unless one works at Uber. I can tell you that i saw how much they were spending on ads back in 2016, and how long it continued and can assure you that they were 100% losing money back then.
Like, even now their margin is around 10% (they made 5bn on 50bn of revenue). Other software companies make a much, much, much better margin because Uber is basically not a real software business, it's an app attached to a low-margin delivery business.
Uber kept fares artificially low while simultaneously paying high bonuses to drivers to build a massive network. After burning through roughly $30+ billion over its first decade, Uber then pivoted its business model by raising rider fares, increasing restaurant fees on Uber Eats, and cutting driver pay.
"Cost of revenue" isn't the entire cost of running the company, (ie R&D, operations, sales, marketing, etc). It's just a cost they've associated with revenue IN ADDITION to the other costs I mentioned.
HSBC say they need to turn a 13b revenue to 200b by 2030 AND also find another 204b, in order to become profitable.
> It's just a cost they've associated with revenue
Its a little less arbitrary than that. Cost of Revenue/Cost of Sales/Cost of Goods Sold are clear, if you're following GAAP. To label these expenses as cost of revenue, they must meet the matching principle in that the expenses must be directly tied to the generation of specific revenue. If you didn't make that "sale" then that specific cost would not exist.
Other operating expenses come later on the income statement.
Total Revenue - Cost of Revenue = Gross Profit first, then you subtract OpEx from there for EBIT.
For OpenAI, I'd assume cost of revenue is almost directly inference costs + customer support & app dev.
How in the world could you read that article and think there is anything positive about OpenAI's prospects? We've been hearing for months that these companies need to make trillions of dollars in a handful of years, growing at record rates in order to break even and justify their massive outlay.
I tend not to focus on that future too much. I used to do so long ago. For example, how could Facebook possibly justify their losses while asking for such a big valuation? Same for Uber. Same for any number of big companies. And it turns out that growth in the future is impossible to predict accurately. Shopify is a good example where at the time the addressable market of online stores was tiny. But it turned out that Shopify created its own market which is huge today. Technology improvements have a way of creating new markets which far surpass today's total addressable market. Factor in currency depreciation and whatnot and sometimes, futures that looked impossible turn out to be possible.
Not saying anyone is wrong in pointing at the buildouts for AI and questioning its feasibility. Just making the argument for why I personally only look at operational costs and revenue because it's the only real-ish value I can look at and judge if a business can grow sustainably.
As a counter point, the red flag to all of this is R&D costs growing for each model release. If that continues and revenue cannot outstrip it, then these companies have a problem and it'll probably be that just 1-2 frontier labs can survive this once the dust settles.
I don't think Uber was a great ROI for investors though. It lost almost all the money they gave it in return for a business with entirely average profit margins (average across all industries, far lower than average for a SaaS app).
Since Uber's never paid dividends, ROI is easy to calculate.
At the end of its first day of trading (in May 2019), Uber's price was $41.57 per share, and it is currently $72–73 per share for a compound annual growth rate (CAGR) of roughly 8.2% per year.
In comparison, the BVP Nasdaq Emerging Cloud Index earned roughly 22–25% CAGR over the same time interval.
Whether it can physically be as all encompassing as it makes itself out to be or whether it will just be healthily profitable remains to be seen. Kind of like how Uber went from "We'll autonomously drive the world" to "Look, we deliver food, goods, and people to locations and we figured out how to do that in a way that makes profits. Also, ads".