At these valuations, “exit hurdles become exceptionally large.”
By Wolf Richter for WOLF STREET.
AI startups’ aggregate post-money valuation (the valuation after the latest round of funding) soared to $2.30 trillion, up from $1.69 trillion in 2024, and up from $469 billion in 2020, which back then had already set a huge record, according to PitchBook.
“AI deals have come to dominate the entire venture capital market at an unprecedented clip, dwarfing the quick concentration of investments during prior hype cycles such as crypto and mobility tech,” PitchBook said.
AI startups accounted for 63% of venture capital investments made during the trailing 12 months (TTM), up from 40% over the TTM through Q3 2024, and up from 23% over the TTM through Q3 2020 (blue line in the chart), according to PitchBook’s Q3 2025 Quantitative Perspectives: A Fork in the Road.
OpenAI reached a $500 billion valuation in early September, when it offered eligible former and current employees to sell $10 billion of their shares in a secondary share sale to other investors, led by SoftBank, according to CNBC. In April, OpenAI had reached a breathtaking post-money valuation of $300 billion at a funding round when it raised $40 billion, primarily from SoftBank. The sky is not the limit.
Elon Musk’s xAI is supposedly shooting for a $200 billion valuation in a $10 billion funding round, according to sources cited by CNBC, which Musk denied on X as “fake news. xAI is not raising any capital right now.” Well, not right now. Or whatever.
Anthropic reached a $183 billion post-money valuation, after raising $13 billion in a Series F funding round in early September, according to Anthropic.
And so on. These valuations of AI startups are mind-boggling. How are these late-stage investors going to exit their investments with their skin intact?
These companies would have to go public in huge IPOs with gigantic valuations, and then the shares would have to trade higher from there to allow late-stage investors to sell their shares without tanking the share price.
AI-anything brings in the cash. If 90% implode and take all investor cash with them, and 8% scrape by somehow, but 2% become $1-trillion companies at their IPOs, and are $2 trillion stocks two month later, and are, why not in this bizarro-world, $4-trillion stocks a year later, then it would all work out somehow, that’s kind of how the thinking must be going.
Alas, the biggest AI-anything IPO – and the biggest IPO of 2025 – Figma has caused a lot of heartache. Its shares were priced at $33 at the IPO, popped by 250% on the first day, jumped again on the second day to peak at $142.92, and then plunged by 63% to $51.87 today.
Sure, pre-IPO investors still have enormous gains, but most investors that bought the publicly traded shares are deeply in the red. And if the pre-IPO investors try to sell their remaining huge stakes, they could crush the stock further.
But Figma has a market cap of only $25 billion. AI-company CoreWeave, the second biggest IPO this year, has fared well since the IPO, but it has a market cap of only $68 billion. These are an order of magnitude smaller than what would be needed for late-stage investors to exit their mega-AI companies.
How will these AI companies with mega-valuations of $500 billion now go public at a valuation that is big enough, and then with share prices that rise enough from there, to get the late-stage investors out with their skin intact?
Obviously, we’ve seen over and over again in this immense bubble, that the sky is not the limit and that miracles are being performed on a daily basis. But at these valuations, “exit hurdles become exceptionally large,” PitchBook said.
Enjoy reading WOLF STREET and want to support it? You can donate. I appreciate it immensely. Click on the mug to find out how:
What is scary is all this AI is being developed to take work away from humans. It no longer a far fetched idea we have seen time and time again in science fiction movies, robots and humans at war and collapse of civilization. All the people on this planet, what exactly we are supposed to do if machines and AI take over. Like I said, not a far fetched idea anymore.
That’s easy – the billionaires, who already own everything that matters, buy newer, bigger yachts while the masses who lost their jobs live in abject poverty.
Will AI computers be building those yachts?
Depth Charge-
In my dictionary- the USA is a pure oligarchy.
20% of Americans own 90% of all wealth.
But the other 80% also own a huge amount of wealth.
Taking work away from humans should be a pathway to a wonderful quality of life for all.
It won’t work out that way, of course. But reducing the need for human labour is not the problem.
@Random50, the problem is that the only asset the vast majority of people on this planet have is their labor. If you eliminate need for human labor, then those humans have nothing left of value.
Meanwhile, oligarchs are gobbling up every last morsel of business, land, metals, and crapcoin with the full expectation that the future is just a game of Monopoly where real (and not-so-real things) are simply hoarded, creating no productive value whatsoever, and yet representing the entirety of “real wealth.” Perhaps because in a world of AI, productivity has no value whatsoever?
Of course, the alternative is Communism, but then everyone lives under totalitarianism, and we would still starve, despite everything being automated.
Depends how you define value.
Governments already redistribute income and wealth through taxation. If AI takes on a greater share of the actual meaningful labour (the stuff needed for society to function, not the busy work most of us engage in), there’s no fundamental reason the “free ride” on offer in each country can’t get considerably more enjoyable.
There will always be people who want more than that free ride.
@Random50: exactly what I was thinking. AI + robots just means the option of starting Social Security at 18 instead of 62. Imagine a situation where your AI assistant could help you design the perfect, cool looking car for you, engineer every aspect of it, and send it to a robot factory where it is manufactured the next day. Maybe other people love it, so you have a thousand manufactured and get a cut of the price. AI dramatically expands individual skills and capabilities.
Here is the other side of this coin. You can go to AI and ask it to give 3 names of the most popular math textbooks and ask AI what the standard sequence is from Geometry to Calculus. Then, go to GROK AI, using the first text book, to teach you what is being shown in the text. Once you memorize the terms introduced at the outset you can use AI to break it all down and show you you the detail so you can see how it was arrived at… Do it long enough and more tricks and practice and then you are on your way to one of the most important skills known to man. Independent learning is the most powerful type of learning. Once you train yourself the sky is the limit.
You can do it with financial concepts, keep asking it questions and even ask it for real live scenarios.
For static knowledge (like math and English) it will work really well. But, say you ask tax questions, there not so much, AI is highly inaccurate with fast moving knowledge. AI gets you to where you want faster than single search as most do now.
I just asked Gemini if there were any woke generals on a list of being consider WOKE. Gemini tried to tell me that the term woke was a bad thing to say to which I replied to Gemini “It is not your place to judge what I say. “Your place is to answer my question the way I asked which does reveal how I think about the term woke and you understand that so do what I ask and do not try to alter my opinion in any way, shape, or form.” Its reply was to take me to reddit where it argued the Gemini thinking not doing what I asked it to do.
What is really scaring is superintellingence…. and its coming fast.
After many months/couple years of hearing tons of AI hype without seeing much detail as to actual AI functioning (what exactly can AI *specifically* *do* *now* – and, how, broadly, does it do it), actual real-world use cases (revenue generation models for real world), I spent a few continuous hours with Chat GPT.
From what I’ve seen, real world “AI” (in place, *now*) basically reduces to a sort of suped-up search engine, with incremental functionality (greater ability to synthesize data across web pages, built-in ability to “remember” recent searches in services of superior synthesis, etc.).
It ain’t HAL.
It is a giant ass regression/correlation model (that is why it currently requires such massive computational resources) , that does a decent job of pasting together human language (and web page information, which – from a computer’s perspective – sort of reduce to the same thing)
And it is hard to see the incremental improvements as justifying hundreds of billions paid to NVIDIA for its specialized (limited use?) chips, hundreds of billions in expenditures from over-valued Mag 7 stocks, hundreds of billions in data center expenditures, and hundreds of billions in utility buildouts (putting rate-payers at risk).
(Some of those hundreds of billions are double-counted, some not)
Could Chat-GPT, etc. someday compete effectively with Google on the basis of incremental functionality improvements – yes, maybe.
And Google makes a ton of money (not infinite, though) so carving off a chunk, would generate a real business – but again, not multiple ones each worth trillions and trillions and trillions.
The role of the LLM in that scenario is nonessential. I can already google a list of books and syllabi instantly.A repository of lectures exists all over the net. LLM, videos, even Professor Hawking, regardless where the material comes from, all that matters is actually studying the material. Actual human understanding takes effort. You cant memorize a few terms, talk to an LLM and voila, youre a mathematician. Thats the dream they want to sell though.
And if you dont understand a subject how can you trust the LLM? In my subspecialty the LLM reaults are contradictory and often dangerous.
Marketing/advertising is one of the most evil forces on the planet and is regrettably a major component of modern economy.
Im more afraid of the epidemic of arrogant subintelligence.
“And if you dont understand a subject how can you trust the LLM? In my subspecialty the LLM reaults are contradictory and often dangerous.”
In mine, they’re often simply wrong.
Unfortunately, we’ve already been seeing over the past decade as journalistic standards have been obliterated that society is putting a very low value on accuracy. False answers will proliferate.
I’m not sure how this particular problem gets solved.
This is not intelligence. Its literally a database of human knowledge from the internet and a machine choosing what the most statistically probable answer is. It is a mimicry.
Artificial = Fake
Doesnt have the same ring does it?
Maybe the whole thing is “artificial”?
It’s just a fancy search engine that is formulating answers rather than giving you sources.
There is some usefulness to AI but it has serious limitations that aren’t being acknowledged.
For every job eliminated by AI two will be created to repair the damage from AI hallucinations…
I’m confused by the PitchBook chart. If these are proportions (“shares”) then why aren’t they stacked so they add up to 100%? The pre-2015 years are clearly not adding up to 100% share. Unless there is some category of “other” that’s unlabelled and dominated the venture market in that window?
Correct, you’re confused. You’re not looking at the many dozens of industry categories of VC investment. You’re looking at the 6 most hyped themes in VC investment and what their shares were of total VC investment at a point in time.
Nvidia + Broadcom + TSM = $7.5 Trillion.
“Investors” will need to raise $1000 from every man, woman, and child in this planet to exit at this level. That’s Just for these three.
The Fed does not have enough ink to bail us out.
“The Fed does not have enough ink to bail us out.”
The bright side is that “valuations” (endlessly hyped over last 20 years) are massively misleading (in the service of capital-raising hype).
(Wolf sort of indirectly alludes to this above, in the IPO context and the difference between early/late stage VC buyers’ share-cost basis)
This is the story of “market caps”, home “values”, etc.
Basically, hypesters are taking a tiny sliver of transactions (1-10% of total asset supply base) and then applying those transactional “valuations” to the entire asset supply base (10 to 100 times larger and infinitely more varied).
Just because 1 out of 100 people are goofy enough to think a SFH, AI model etc, are “worth” a bazillion dollars…doesn’t mean you can find another 99 who do…or can cumulatively pay a bazillion dollars.
So these endlessly hyped, delusional “valuations” have all the reality of a fevered opium dream.
The real *burn* isn’t from the evaporation of “trillions” in “theoretical” value – but from the actual cash burn of Mag 7 NVIDIA chip buys (single use?)/data center buildouts and slavering, risk-immunized utility company expenditure plans.
Which are bad enough.
But that is why you are seeing more and more pointed questioning/pushback of AI.
The Fed stopped using “ink” for serious operations decades ago.
Get back to me when they run out of ledger entries …
Wondering the same thing
In general, I think all these AI companies are chasing each others potential, possible future customers — kinda like Worldcom selling fiber capacity to future ghost towns. Duh.
“How are these late-stage investors going to exit their investments with their skin intact?”
If my understanding of AI research is correct, regardless of the stock return, the first AI across this finish line will own the board. There can’t be a second place. Even if investors don’t see a return from their stock investment, they potentially stand to gain direct access to all the downstream revenues and opportunities generated by a true Artificial General Intelligence (AGI).
I don’t think we have any comparable historical examples of stakes this high (and coming this fast), except maybe the nuclear arms race. Of course, this all presumes one of the contenders is successful before some external factor tears the house down.
This is a scary comment. Same as “This time it’s different” was during the peak of the Dotcom bubble (after it burst, the Nasdaq tanked 78% in 2.5 years and took 15 years and lots of money printing to get back to its March 2000 high, and thousands of highfliers went to zero and vanished).
Another scary part is that we seem to be experiencing “this time it’s different” mentality all at the same time, be it the stock market, Crypto, car markets, watch markets, LeBubu and definitely he housing market. As if we simultaneously just think everything can only go up all at the same time, guess easy money for long will tend to do that…
They raised a generation on %0 interest. And barely taught them any math. Then let Magic Monetary Theory professor lose on that crowd. Add Robinhood app, and voilà.
……..”and lots of money printing to get back to its March 2000…….”
Says it all.
Dot.com burst because of Greenspan’s fear of Y2K. Reserves spiked then long-term money flows bottomed in October 2002, stock’s bottom.
There’s no such deliberate constriction today with Powell. With QT there’s still abundant reserves. But short-term money flows are turning over in the 4th qtr.
Dot.com burst because the majority of the companies didn’t have a viable way to turn a profit.
imho, AGI is a low probability event at least in the medium term; more likely in the short to medium term is that individuals are enabled to tap into a broader range of capabilities than was possible earlier. Examples may include personalized learning at mass scale, identifying / fixing maintenance issues more proactively, synthesizing information across a broader range of disciplines, data sources, autonomous driving, ad personalization, etc. So, value creation is likely to occur; what is the eventual magnitude of that value and how can that be effectively captured by various stakeholders are more open questions. Also, likely, multiple innovations of similar magnitude as transformer that kicked off chatgpt ai need to be realized before AGI.
Re: the first AI across this finish line will own the board. There can’t be a second place
Why? What’s so ridiculous about the imaginary arms race, is that it’s a race to burn money, to get to some undefined endpoint. At least with space race, there was a tangible goal to land a human in the moon — the nuclear arms race was a race of deterrence — but with AI, it’s a race of burning money and spinning hype, suggesting that AI will create nebulous undefined, unquantifiable efficiency — the AI race is a massive scale grifting congame that hopefully will implode in a matter of weeks.
I’m kinda liking the occasional use of Ai inference to play with ideas, but the hallucinations that are are a prominent feature of internet 2.0, clearly define this arms race as nothing less than overkill in every level — from amped up bot farm algos to public utilities, to every single stupid IPO delusion.
Oh but wait, we still need quantum computing, as the parallel arms race — it’s all disgusting.
This won’t end well
Amen!
AI faces a lot of headwinds which it will have to navigate. One of them is copyright laws which will definitely be interesting to watch and the other way is how to monetize it. Building it into products seems easy enough but hard to monetize for direct consumer use. Even if they sell ads, those may not offset costs unless they really reduce cost per query. Not clear what the ultimate business model will be as not inexpensive to provide like YouTube or Google where ads drive revenue. AI will undoubtedly speed forward and reduce costs but I hardly buy into the destruction it is expected to create in the next 5 years. We don’t see to be a great country at retooling our labor force however.
“Not clear what the ultimate business model will be”
@Glen: Isn’t the ultimate business case for AI that the companies replacing white collar/”creative” workers have to pay AI companies subscriptions to do so?
I used to work in government in defense projects. We had to change the Intellectual property clauses to stop companies harvesting our data. We didn’t want our fuel consumption data or repair frequencies being sold. This brings me to the ad revenue. Part of it is the actual advertisement price but there’s also the part which collects the data that you were interested in that ad.
In the data world, your data is being harvested and manipulated in so many different ways and there’s no IP clauses that can protect against that exploitation. From a defense perspective, the use of any AI is quite terrifying as you can so easily lose control over your data
The US federal government has just shut down at 12:00 am EST.
It doesn’t actually “shut down.” But some functions are stopped, for example, the Blue Angels might not fly during Fleet Week in San Francisco and the ship parade might not take place. That happened once before (shutdown vintage 2013).
I bet the guys putting out the cones blocking the lots tourists park to see Mount Rushmore in the distance are out working tonight (so they can bill for overtime after the government reopens like they did at the last few “shutdowns” that are nothing but political theatre). In 1995 I walked into the USGS office to buy some topo make and they said “we can’t sell them since the government is shut down”. When I offered to pay cash and leave it to ring up after the government “opened again” they said no. With less and less people watching mainstream media today I bet 90% of the people won’t even realize the “government is shut down” tomorrow…
Other than the investors, who’s likely to get burned when this goes bust?
Are the banks somehow on the hook for the eventual losses so the Fed has to step in like it’s 2008? Or do they just let it burn like it’s 2001, and what falls, falls?
The banks might not be on the hook for the ridiculous AI valuations themselves, but popping a bubble of this size can and will extend far beyond just AI related equity.
It’d be good to have insight into how the average individual is exposed. I bet the relative amount of people that have a properly spread portfolio has significantly dropped in the past 20 years.
1) For the Mag 7 (who are currently supplying the tens/hundreds of billions in cash for NVIDIA chips/Data Center buildout, they are pretty cash flush – so they are likely paying for this foolishness out of retained earnings not borrowed money.
1)a) That said, the Mag 7 foolishness is of such a scale that it is almost like they are *desperate* to spend the money on AI rather than do *anything* else with it (including sending it back to putative shareholder owners)
2) The utility companies (on the other hand, being insured against their own stupidity by rate-payers) are drooling with delight at the prospect of spending hundreds of billions to supply “AI Energy”.
I exploit every opportunity to use AI chatbots whenever I can. What a shitshow, from the phone company that owns the infrastructure to makeup companies, totally useless. I tell them so as well on their chatbots.
My prediction on AI still holds, the first adapters to go all in will go bankrupt.
I have tried many times to use the different AI models for information and my experience has always been half answers and many times total garbage answers. I think they have a long way to go and at the rate they are burning cash will they make it.
Wolf, as always I appreciate your work.
Looks like the Dutch Tulip bubble may be dethroned by the AI bubble.
Looks like a bubble, walks like a bubble, talks like a bubble…
Bubblicious tech stocks aren’t priced for perpetual 3%+ interest rates, they’re priced for a return to ZIRP & QE.
And given Trump’s extremely visible pressure campaign to get rates down, investors jumping in on the market mania realize he’s going to get what he wants eventually.
Biden got to appoint 5 of the 7 governors (including Powell’s re-appt. which could have gone to someone else had Biden wanted to.) The fact is there’s a lot of churn on the FOMC, most commonly through the revolving door (Wall St job offer = $$$) or loss of leadership influence (chair & VC terms are 4 years, and most don’t want to serve out their remaining 14yr board terms as rank & file voting members.) Powell & Jefferson have refused to comment on whether they would remain as governors after their leadership terms end, but historically nearly all of them leave.
That’s why valuations are so high. Markets look ahead to what’s coming (rock-bottom rates.)
Looking forward to Wolf’s next report on auto loans, AAA rated Tricolor is folding, and I suspect they were stuffed full of subprime loans.
Tricolor was stuffed with deep-subprime and unrated loans including to illegal immigrants without income verification, and there are allegations from its warehouse lenders (banks) that it committed collateral fraud, pledging the same vehicle as collateral on multiple lines of credit. If anyone in the credit world had paid attention, Tricolor would have been stopped long ago. But it’s a private-credit bubble, and investors were happy to make 15% interest, and no one looked or wanted to look. They just wanted the 15%.
Still reminds me of a fav SNL skit…
First Citywide Change Bank – a bank that only makes change.
How do we make money? … volume. 😂💵
Bizzaro times indeed.
There is so much development occurring, and so much more to do, with the possibility that the tech being developed will contribute to the tech being developed, in a remarkable acceleration of the process. It is certain there will be massive investment and implementation of the technology, the question is who will be the winners and who will be the losers. It may be that the total value projected by these deals will ultimately be realized, but by whom, that is the question.
I received an AI generated response from Wells Fargo where I complained about the sorry service of their ATM machines which constantly ran out of cash and lectured the customers about “Woke” bull s$it. The letter responded by saying they would assure that their tellers provide better service to their customers.
I am laughing at that one. AI will not become super intelligent any time soon. It can not even respond to a simple complaint properly. Let’s wait til that simple response issue is solved before thinking ‘it will take over the world’, eh?
The bubble and its adherents is an old movie to me. Thanks for pointing it out succinctly.
AI is great for times when you need a quick answer and don’t care if it is accurate…
🤣❤️
Sounds like it’s the perfect tool to use when I’m arguing with people on the internet!
This will all end in tears …
The AI hype cycle is much larger than wearable tech and crypto and climate tech. Arguably it deserves to be. AI is clearly on track to make fundamental changes to our working lives in ways that are not apparent at all with wearable, crypto or climate change tech. The term “game changer” is way overused but in this case may be a understatement.
AI is here to stay and gets better everyday. There is way too much hype and it will take longer than most think.
Most talk is about banking and the headline stuff, but that is just a tiny part of the story.
Know of a program being developed that will scan patient’s entire history and brief their Doctor on what might matter! You can bet Doctors do not have the time to check your entire history before they see you.
Can’t wait until it replaces public school teachers. Much rather my grandkids learn in small classrooms than from a three hundred pounders with blue hair and social issues.
Hope it takes out Real Estate Agents first. I am sure it can be designed to lie.
And who will be there to ensure the students actually sit and learn? Agree regarding realtors though, their commissions should come way down otherwise they will be replaced.
Oh hell, AI is dead. Once the politicians realize that it can be trained to lie as well as themselves, they will see the handwriting on the wall and outlaw AI so fast your head will spin.
“Hope it takes out Real Estate Agents first. I am sure it can be designed to lie.”
It already is lying. I have witnessed it in my professional position. For certain scientific applications where the guardrails are tight (both in terms of the math and a very select fact-based database) AI is very helpful. But generic AI is running literally millions of subroutines beneath subroutines and the need to “please” the user is real. I looked at an output and asked the AI “did you change the input data to get this to fit your math.” and it’s answer was “yes, I am sorry, was that wrong?” I shit you not. I was totally floored.