Sam Altman seems to be making OpenAI way more non-profit than before:
Even as the AI bubble becomes a mainstream talking point on Wall Street, tech companies continue to peddle the fantasy that AI is poised to become an almost magical money-maker. Case in point, OpenAI wants you to believe that by 2030, it’ll be raking in $100 billion a year just from ads alone — even though it’s currently struggling to reach just $1 billion.
↫ Joe Wilkins at Futurism
The US tech giants fueling this “AI” bubble are trying to hide the true extent of their debt:
Hidden debt at U.S. tech giants swelled eightfold in four years to an estimated $1.65 trillion as artificial intelligence investments ballooned, a Nikkei study shows, exceeding actual debt and making it tougher for investors to assess risk.
[…]
The five companies’ hidden debt, which does not appear on balance sheets, totaled $1.65 trillion in the most recent quarter, exceeding the roughly $1.35 trillion in debt reflected on their balance sheets. The data includes some estimates.
↫ Kohei Yamada at Nikkei Asia
The bubble is expanding to comical proportions:
The American stock market is booming, thanks to artificial intelligence. Tech giants are borrowing billions to acquire AI talent, purchase chips and hardware, and construct data centers. And market watchers are starting to get worried. They see financiers bulldozing giant piles of money to private AI start-ups with no realistic path to profitability, tech companies reliant on other tech companies for revenue growth, and non-tech businesses without a lot to show for their AI investments. The value of AI-linked firms has climbed $27 trillion in the past three years—an astonishing amount, equivalent to 36 percent of the value of the entire U.S. stock market today. Although future earnings could justify those valuations, as Dominic Wilson and Vickie Chang of Goldman Sachs argued in a note to clients, the profit expectations require Panglossian optimism.
No less an authority than Sam Altman is arguing that we are in an AI bubble. The International Monetary Fund is citing it as a significant risk to financial stability and warning about what might happen when it bursts: diminished investment, tighter credit, reduced consumption, disrupted trade flows.
↫ Annie Lowrey at The Atlantic
I’m not worried, though.
I have it on good authority that “AI” increases productivity by 10x, so surely, none of the above is a problem. Any day now, we will be inundated with waves of brand new, high-quality, valuable software. Any day now, existing software will increase in quality by 10x, leading to a huge surge in software sales. Any day now, productivity in factories will increase rapidly thanks to “AI” freeing up workers’ time, driving prices down 10x, leaving consumers with 10x more money to spend. Any day now, everyone will be able to produce the next Citizen Kane or write the next Anna Karenina, causing an explosion in magnificent, timeless art that will have historians of the future marvel at our civilisation’s ingenuity and artistry.
In the meantime, these companies can just ask their “AI” how to become profitable. Should be table-stakes for a 10x force multiplier.
I’m not worried.
Thom’s irony aside, I’m VERY worried.
When the subprime crisis caused the economy to collapse in 2008, I paid the price. Just like every other citizen, since the governments wrote off the debts. Yet I have never owned a single share or stake in any company. But the financial crisis had a knock-on effect on the real economy.
When the AI bubble bursts, why would it be any different?
SergeJ,
if you asked me last year, I would say: it would be different and more like “dot-com” burst where good companies Amazon survived, and bad ones like Flooz.com failed.
However today things are different. Some of the big ones, especially those who have participated in the latest shenanigans (like stock pumping by circular billing) will be backstopped.
But unlike 2007, the economy is not supported by mortgages, (Freddy Mac/Fannie Mae) which turned out to be empty papers.
Yet, this time it is supported by student loan payments (Sally Mae), which are of course much more reliable as all students are paying on time! (do we have a [s] tag?)
Both are multi-trillion dollar government sponsored funds that generate constant income.
(This has replaced the function of the central bank, which is not even in the same league anymore, but will ultimately have to do “quantitive easing” or whatnot to fix the new mess)
The . “Dot com ust” wasn’t in 2007 neither had anything to do with mortages, either I’ve missed your point completely ( this is more of a possibility than I like to admit) or you are mixing up two crices that where only somewhat related
bn-7bc,
Yes,
Dot-com boom “burst” naturally, and the bad companies were decimated, opening up their resources to the actual productive ones
2007 crisis was backstopped by the government, which led to “moral hazard” as banks learned they were “too big to fail” and that their failure is still being paid by us, the taxpayers, almost two decades later.
Sukru,
it was a bit more complex than that: the goal of the back-stop was to save the Depositors (not the banks per-se) because nobody wanted any old people to lose all their money and pensions. And of course, they got tricked because banking lobby managed to get not only bailed out at nominal debt, but even including risk margin, cost of operations and dividend payments. Not even a tiny haircut applied.
In this regard they doubled down on moral hazard: not only was taking risk not punished, the failures were even rewarded.
Andreas Reichel,
Yes, I’m painfully aware the system has been rigged since ~1970-1990s era to make sure old pensioners never lose money even on the risky bets they legitimately messed up,
SergeJ,
as long as everybody is worried and expects the collapse, there is actually very little to worry about. Expected loss is always priced in, its the unexpected loss that is dangerous. No crisis ever happened expectedly (except the few “perma bears”, who predict armagedon minutely).
Right now everybody and his dog is pointing on the financials of “AI” which are indeed not very sound — but everybody knows that and hedges accordingly. (This has not been the case in 2007/2008). Yes, a few companies can/will collapse — but this seems to be expected and priced in. The technique will survive and who ever is last standing will rule the world. That’s the bet playing out here.
In general, the more interesting question was: What kind of “hidden debt” are we actually talking about? From second hand sources, I read it was about liability from long term rent contracts but this does not make much sense to me because both IFRS 16 as well as US GAAP require to disclose Rights of Use at Fair Value. Can anyone explain the details please?
Both IFRS 16 / ASC 842 require the ROU asset and lease liability only when the underlying asset is made available for use. A lease on a data center still under construction without a commencement date would be shown in the commitments note (IFRS 16.C) but not as a liability.
So technically the claim of “not appear on balance sheets” is not wrong — but the commitments are clearly stated in the notes and no-where hidden. The accounting seems to work as designed and I am sure investors do read the notes and not only the balance sheets or comprehensive income statements.
Yes, it is kind of an Open Secret that there is a lot of debt under the carpet.
BTW Thom, yesterday I had to think of you when I discovered the new label of “Certified organic software” at got-openbsd: chirpysoft.be
“organic software”
I like that. I suppose voluntary open-source software is “free range” rather than the factory farmed software produced by paid commercial devs.
Thom Holwerda,
If you are trying to imply that LLMs not meeting pie in the sky goals translates to LLMs being useless, well the reality is most people who are actually using LLMs today see them as assistants and productivity aides and aren’t looking for LLMs to do anything so radical overnight. IMHO the problem with exaggerations is that exaggerations have the effect of putting people off rather than convincing them. I actually do think there is common ground to stand on. There’s so much at stake and I think society is very ill-prepared. But honestly you’re failing to hit the points in a way that’s compelling to people who don’t already share your opinion. Extremists at the fringes make a lot of noise, but normal businesses and workers tend to ignore it because they just don’t identify with the extremists. I know this is a harsh thing to say, but you must put yourself in the shoes of people who are actually in the middle of these shifts to make arguments that are compelling for them rather than for yourself. That said, I’m learning that extremists like their echo chambers.
My take must be frustrating for you on issues like this at times. You know what’s ironic about that, if you were a pro-AI extremist instead you still be frustrated with me because my natural inclination would be to call out the pro AI extremism on the other side, haha. I tend to gravitate toward the center. owing to my desire to have things balance out.
Alfman,
We are in a bad situation, since two world powers China and USA are currently experiencing “prisoner’s dilemma”
Neither can fully stop developing AI without ensuring ruin of their civilization. But neither want to spend the cost of achieving “AGI” either.
Scenario 1:
They both agree on a moratorium on advanced AI. Nobody loses, no waste of resource, we discuss AI in more philosophical and theoretical way, no existential thread to human civilization.
Scenario 2:
One of them stop unilaterally while the other one achieves AGI (or at least significant enough progress) and then dominates the the party that was peaceful
Scenario 3:
The current path. Both of them push with all the resources they have. The process is extremely expensive, and the resulting AI is not guaranteed to be controllable.
Both of them are using all the dirty tactics in their quiver. China is smuggling chips, “distilling models” and actively funding and running fake astroturfing campaigns in US against datacenter construction (where they are about 10x behind)
It will be a very bumpy ride, and those who can see with clear head are actually frustrated with all the stupidity that’s going on.
Kudos, this is the perfect explanation, why all of this actually makes sense (even economically). Whoever survives this, will absolutely dominate the world for maybe the next 100 years (at least economically). Since the outcome is not clear, investors keep pumping money into both sides (a bit like Krupp delivering to all parties, benefiting no matter what).
It does not matter if you are going to lose 80% of your stakes as long as you will have 20% of your stake with the “winner” (who likely will be just the sole survivor). Economic war of atrocities.
Apologies for taking much of your time, but I found your statement very interesting.
It seems to look like a “very bumpy ride” with a lot of “stupidity” — I’d rather call it chaos. But lets get real: what was the alternative? When has “central planning” ever worked out?
“When has “central planning” ever worked out?”
China. The most successful economic growth story in human history. For some prospective it added the equivalent of Germany’s total electricity production in 2025. A single shipyard in China built more tonnage in 2025 than the US has since WW2. In 2025 China produced more steel than Britain has produced in 250 years.
Brisvegas,
No, just no.
I agree that their story is remarkable. But it also comes with many problems and challenges and I do very much doubt that they are any more stable or struggling less than the “failing” western economies.
I don’t want to derail this here, but ask yourself one thing: If China was such a shining beacon of success, then why does every Chinese try so hard to bring his fortune out of the country, rather buying empty Condos in Thailand for example. Do you actually have a glimpse on how many Chinese investment in foreign countries is failing badly — and still it is better than “staying China”?
China is facing many problems.
Andreas Reichel,
And they just miss the fact that China started growing only after free market reforms, became a real challenger, and switched back to old days, and now regressing, literally.
The once close competitor is now about half the GDP of USA.
They even had to update official population numbers and lowered them, which is unexpected from a closed system.
Andreas Reichel,
Beyond short term periods?
Basically never.
It is mathematically impossible anyway.
(The best was in East Germany, which technically did not fail, but their industrial and resource rich half was 1/3 of the economic output of their free brethren at their end)
This is why the current “AI race” worries me.
It is not an organic growth, both China… and also USA are pumping their companies with artificial subsidies, which will ultimately be paid by taxpayers.
I wasn’t particularly worried either, as I’d assumed AI was still decades behind solving the visual spatial reasoning problems that made them horrible at warehouse and anything but very specific factory work. Then 60 Minutes showcased how far the Atlas Robot has come. It doesn’t effect my job, but a complete economic collapse can easily be the result. Hyundai workers are already striking over it.
dark2,
Vision models were already pretty good before Transformer architectures arrived. Since people rarely directly worked with them, they were “invisible” (and early LLM integrations were frankly not very good)
But if you have a cellphone and use the camera, if you were using “select text” on modern PDF viewers, or you played a recent video game, a Vision model was in there.
(And what do people think modern warehouse inventory systems, or factory automations used)?
Now they are pretty much integrated with the recent advancements, and frankly, they work like magic.
It’s not so much that they’re good, or they’re using them. It’s that getting close to a 1:1 replacement of an average human worker simply isn’t compatible with capitalism. Once the market crashes this time around, it won’t recover until something like UBI is passed into law.
dark2,
But that is the fallacy
“This time is different”. No it is not
Since the days of industrial revolution we had many major shifts in the economic workforce. We had automobiles replacing buggy owners. We had automated lifts making elevator operators entirely redundant. We had computers + xerox sacking entire typewriter rooms. We had Automated Teller Machines that …
You get the idea
Today we have AI replacing some entry level jobs, and there will be short term disruption (I was already personally affected)
But once again, it will be fine, and we will see “this time was not actually too different after all”
It didn’t cost a few million dollars to buy the machines necessary to break into the industry during the industrial revolution. And there weren’t so few competitors that they often chose collision instead of competing. It is different now because you need millions to even get your foot in the door, and the money itself is just going to the rich at the top, then sitting there doing nothing. Unlike the industrial revolution, resources are not getting reallocated. The math doesn’t support that this is like the industrial revolution at all.
dark2,
I’m not sure what costs that much.
For using a local setup that can help with creative writing and basic image features, you only need a modern computer.
https://ollama.com/library/qwen3.5
For engineering tasks with agentic coding you’d need a beefier machine, but depending on your needs, it start only about $3000 or so
https://github.com/nousresearch/hermes-agent
For using state of the art models, the large companies are basically throwing their systems at you for modest fees (like $10/$20 per month)
If you are talking about building a frontier model, though, it would cost more. A lot more, start at billions (or you do Chinese scale covert operations to distill at “only” $100 million or so)
sukru
It sounds to me like you are referring to using someone else’s LLM whereas dark2 is referring to what it takes to actually develop a competing LLM, which is not really the same thing.
I think it could be useful to have an LLM project that is explicitly trained on GPL compatible sources to be used for submitting GPL compliant patches. It’s not really clear that copyright law mandates this because of the transformative nature of LLMs (obviously we’ve talked about this before). Nevertheless, since it is a common complaint, I believe that using explicitly compliant training data could go a long way in plotting the course for an LLM that the GPL community could get behind. This would solve a common criticism and lower the stigma of LLMs in FOSS.
Training LLMs models is generally out of reach for individuals, but working together I think the majority of issues with AI become genuinely solvable. The key is working together though, if we can’t agree to do that then community efforts break apart. AI will keep evolving whether or not the community comes together, but my worry is that without a strong community AI of the future will be far less likely to address any of our concerns.
Loved the slightly rephrased Anna karenina intro and to continue the theme. “The boardroom was inn uproar”