It plans to spend $518 billion in the coming years on cloud, computing, and infrastructure obligations and now needs a lot of AI magic money.
By Wolf Richter for WOLF STREET.
Anthropic, which is targeting to go public at an IPO valuation of $2 trillion (trillion with a T), generated a net loss of $42 billion in 2025 on $4.6 billion in revenues, according to its IPO prospectus that was leaked to Reuters.
The net loss had multiplied by five from $8 billion in 2024. The revenues had multiplied by 12.
A quarter of its revenues came from just two customers, the company said, “and as part of its risk factors, warned that many of its largest clients were not locked into long-term contracts and could cut or stop spending,” according to Reuters.
Total operating expenses rose to $12.6 billion in 2025. This includes $7.3 billion in spending on compute and infrastructure, which had nearly tripled from $2.5 billion in 2024.
And the company plans to spend $518 billion in the coming years on cloud, computing, and infrastructure obligations, according to the prospectus cited by Reuters. So now it needs a lot of AI magic money; and at a valuation of $2 trillion, the IPO and any follow-on offerings would extract a lot of this AI magic money from investors.
“Excluding write-downs of various liabilities mostly tied to previous fundraising, according to the documents,” amounting to $34 billion, the company generated an operating loss of $8.1 billion, Reuters said.
The company isn’t going to run out of cash right away, despite the massive cash burn: At the end of 2025, the company sat on $20.3 billion in cash and short-term investments, according to the prospectus cited by Reuters. But that was before the massive $65 billion in Series H funding round in May 2026.
For the Series H funding round in May, the valuation of the company was set at $965 billion. The IPO valuation target of $2 trillion would more than double that. I mean, why not? It’s AI magic money.
“The [IPO] plans come as Anthropic confronts evidence from its own research that increasingly autonomous AI models can behave in unexpected and potentially harmful ways, including sabotaging code, assisting fraud and manipulating information in controlled tests,” Reuters said.
But there is a solution to that problem: Hold the executives and their companies legally responsible and accountable for creating and selling products that cause harm. These are not accidents; the products are doing what some of the smartest people around have designed them to do. AI is not a human-like creature. It is software that does what it is designed to do. The harm is not caused by human misusing the products. The harm is caused by the software on its own by the way it was designed.
Politicians of all stripes should condemn the efforts currently underway to create liability shields for these companies. And tort lawyers and prosecutors should be cranking up their machinery so that they’re ready when that harm occurs. That will encourage the AI companies and their executives to create products that are not inherently harmful on their own.
Enjoy reading WOLF STREET and want to support it? You can donate. I appreciate it immensely. Click on the mug to find out how:
I wish the AI capex insanity would subside. AI doesn’t grow food, build houses, create energy, create transportation or any of the essentials needed for a comfortable life. It’s TAM is wildly hyped. It’s of questionable value. Life may be better without it.
Life would be better without cars, cellphones, computers, the internet (WOLF STREET comments!), electricity, potable water coming out of the tap, medications…
We keep hearing AI did this, hacked that, blah blah blah. But we never once heard that AI spontaneously discovered a cure for a previously incurable disease. When that happens, I will become a believer. Until then, it’s all slop we are better off without, in my book.
Zoroto,
You need to update your knowledge. There has been extensive use of AI in medial research with quite a bit of success. Obviously, going from discovery through the trials to test safety and efficacy, and then through the approval process, takes many years.
You can look around a little on your own to see the reports, but here is an easy-to-read example:
https://www.bbc.com/future/article/20260309-ai-is-finding-treatments-for-incurable-diseases
Google AI Navier Stokes. Click on the Quanta Magazine article. Not medicine, but a hard math problem.
You may have also missed that the creators of AlphaFold, an AI (but not the LLM kind) won a Nobel prize in 2024 for solving a difficult protein folding problem.
The claim that “AI” makes any medical advancement is usually done by custom models not openAI. It’s like hearing about an engine putting people into space and seeing Ford claim that all engines are the same, as if their F150 is somehow related to a rocket motor.
The biggest gate to medical AI is lack of testing, there are books of untested, unused chemicals out there that need funding for studies. Studies on humans take 10 -20 years and no AI can speed that up. The use of a new chemical takes weeks or months to demonstrate in petri dish. Then gobs of money and time to use it in a mouse model and see if it’s even capable of being produced in quantities that makes it an effective tool.
The BBC article Wolf posted made the claim, “We can – in a matter of days or hours – look at massive libraries” of chemical compounds to identify those that display antibacterial activity.” So it’s not developing anything it’s just a nice new search algorithm. They already have a massive digital libraries with search engines. How is “AI” better?
“Collins and his colleagues designed 36 million compounds in this way with potential to work against the bacteria. The team selected 24 to synthesise in a laboratory. Seven proved to have some antimicrobial activity, and two were highly effective at killing strains of both bacteria that were resistant to other types of antibiotics. ”
36 million compounds but they only choose 24? Were the other 36 million hallucinated BS that simply can’t exist? It sounds like it spat out every known possible antibiotic molecule like a search engine and they narrowed it down by filtering the data. That’s not exactly the promise of AI we are being sold.
The bacteria in question had already had an antibiotic cure in the waiting that started development 10 years prior because the studies take that long to validate. Did the “AI” steal a current treatment that was being studied? It’s trained on medical research, what are the likely chances it simply “suggested” someone else’s work? Did the researchers check to see if the “AI” simply stole someone else’s work?
If it did, then it makes any claim of “AI” kinda suspect.
His point stands. Everything you mention has a physical presence and requires real resources to build and operate. AI consumed considerably more resources and energy and for what? A relative few people get to live high on the hog while everyone else starves? How do think that will work out? Seeing the forest for the trees, it is clear that humanity hasn’t learned a damn thing and the outcome will be no different this time around.
Nonsense. AI also has a huge physical presence (data centers, servers, electrical equipment, etc.) and does and create real things, such as creating my donation ads (“Keep the Paywall away from WOLF STREET”) that you see here. Don’t diss those vignettes. They’re funny. I hope they make you smile when you see them. Thanks to these donation ads (the donations actually, thank you dear supporters!), WOLF STREET is still around, and you can still post those comments. That’s “real.”
BTW, Americans spend only a very small amount of their total spending on food (and farmers are using AI along with lots of other tech). Most of what consumers spend their money on are services, not goods. And by your definition, it seems that services are not “real” somehow.
“Americans spend only a very small amount of their total spending on food”
Sounds like hyperbole Wolf, show us the data for customer spending on the essentials (food, energy, rent). Services are not essential and usually the first thing a struggling family stops paying.
WB
Dotted blue line at the bottom = amount spent on food and beverages bought at stores; solid red line = total consumer spending.
The SHARE of total spending that went to food and beverages bought at stores dropped to a record low of 7.1%:
There isn’t a new fangled widget that Wolf doesn’t like, and defend to the death. Now it’s AI that we can’t do without, damn the white collar workers to hell, it will all be great.
Ian
By willfully ignoring what AI actually does and can do, you’re willfully ignoring the actual issues that it poses. The issue isn’t that it will kill humanity (it won’t). The issue is that it’s changing everything in the business world, some for the better, some for the worse. Publishing, including my business, has seen the devastating impact of AI for two years. As have many companies in tech, video and movie making, etc. Don’t give me this stupid-ass bullshit that AI can’t do anything and is just an overhyped search algo. You people are so clueless it hurts.
It is just a matter of proportion (something the US Gvt and US invt sector has seemingly and repeatedly lost any sense of over the last 25 -50 years).
AI may have some (perhaps even significant) value.
But, in the end, it is just a giant-ass regression model with helpful processing tools attached.
It ain’t magic and it ain’t God.
It ain’t even remotely/conceptually human-level intelligence.
Or dog level.
And its long-term actual use cases and related business/profit models are largely (very largely) unproven.
And its near term financial model/outlook smells more and more like a literal circle j*rk.
It is like watching wealthy people doing crack and poppers while wholly ignoring the common-sense Southpark “Underpants Gnome” mockery of **25 years** past.
(That said, it is a bit hard to believe that the key circular financing participants, those with long-standing massive profitability in the software sector, have not really built in many contractual “outs” for themselves. So the inevitable implosion may not be as bad as it otherwise might. Of course, the post 2022 frontier hype-heads driving the entire madness will incinerate themselves).
While I’m with you that the value isn’t there yet, the reason for the investment and r&d craze is that there is a very real chance that this could be the next Manhattan project, the next internet, the next smartphone.
All of the big players seem to be certain that we’ve reached AGI (artificial general intelligence) which is a really big deal. If you take that, iterate on it, and assign it a really difficult problem that humans haven’t been able to take like, say, net positive fusion energy at scale and efficient cost, then it pays for itself ten times over. It also presents scary possibilities, like what if our adversaries get it first and use it for something malicious. Then there’s the whole doomsday side of things like, is this the singularity if it can iterate on itself and outpace human intelligence.
I agree the finance side of it is straight madness, as shown by this article, but if the benefits are best reaped at nation state or megacorp level rather than at the individual level, the AI images and basic coding tools are simply window dressing or proof of concepts to its end value usage. Pets.com and other dotcom products may have signaled a bubble, but they didn’t signal the Internet as worthless or the the future with Amazon or Google.
Who are these big players sure we’ve reached AGI? You mean the heads of the frontier models? I think we can all rest assured that AGI is still off in the ever distant future. At least if “godfather of AI” Yann LeCun can be believed — Artificial General Intelligence (AGI) is a misleading misnomer and scaling up current Large Language Models (LLMs) will not lead to human-level intelligence.
I recommend everyone ingest more Ed Zitron as an antidote to the hyperbole surrounding LLMs.
Wolf, you’re absolutely correct. Any hacking that is not cyber security related without the target’s consent should be charged as felony hacking. The execs and developers should be charged.
The bizarreness of the self-reported “We, ourselves, are on the verge of destroying the universe!! Stop us before we maybe, possibly genocide again!!” is quite possibly a tell.
These hype-head sociopaths of the post-2022 “frontier model” companies might more and more see the utter hopelessness of their financing/profitability models for 2027-2030.
So, what to do, in the face of corporate incinceration?
1) Use the evaporating residue of “AI will magically save the universe”
2) Add “AI might destroy the universe!”
3) To palm off the hopeless near term AI economics to the only morons/a-holes in charge of a *bigger* global economy scam.
4) The US federal government.
BS narrative to be pitched…
1) AI will bring paradise to Man. Someday. We pinky swear.
2) But, in the interim, the God behind the curtain is *so* powerful and *potentially* dangerous, no private entity is equipped/can be trusted to control Him.
3) So the US Government *must* buy us out.
4) And make us financially whole…and then some. After all, we have crafted a god.
5) And the Feds have to do it before our whole baloney house of financial cards collapses.
6) But remember! Magic God!! Possible Apocalypse!! Buy us out!! Supplies limited!! Call now!!
It’s possible that all three of these things are true:
1. AI is a huge stock market bubble that will rival or exceed the dotcom bubble, and isn’t cost effective unless subsidized.
2. AI is super useful for certain tasks, and growing more useful by the day.
3. The potential is there for AI to be dangerous.
To that last point, there have been multiple resignations in the news lately from frontier AI companies, where the resigning employee publicly expressed concerns about safety. Are they secretly on the take? Or were they about to be fired and went out with dramatic flair? Or should we take their comments in good faith?
Put a few CEOs in jail for gross negligence and see how safe AI becomes.
Wow, that’s even worse than I expected. I work in software and these tools have taken over the entire industry. They’re turning everyone into mindless button pushers, including me. It’s an underrated part of their addictive power, aside from the worse examples you see in the news. I want this company to crash and burn. It’s run by a bunch of insane people who believe they’re bringing AI God into existence. It’s not just marketing!
thanks for your insight.
Which one of the Three Stooges is Anthropic?
It’s that goofy guy Dario — round-framed glasses — looks like Dilbert with brown hair. Add an army of Ashok clones and that’s Anthropic.
You would think AI could pick a good name.
This is why we have to burn our eyes on a cybertruck and morons in leather jackets. These people are grown up 8 year olds.
“You would think AI could pick a good name”
Long standing weakness of SV dipsh*ts – the studied inability/psychotic unwillingness to have names remotely indicating specific nature of their product.
Company might get 10 seconds rando attention from potential customer initially.
Thing to do.
Piss it away naming company GiggleFlexFUps.
Not even GiggleFlexFUps Accounting Software.
Just MysteryCo.
In a world innundated with 24/7 info.
Also, glad to see somebody else slam dress sense of pop collar Fonzi.
Next up, sequined glove and/or Bedazzled Uggs.
“But there is a solution to that problem: Hold the executives and their companies legally responsible and accountable for creating and selling products that cause harm.”
I think this needs to be refined. If an AI system autonomously takes actions that cause harm, then yes, there is a strong case for holding the company accountable. But if a user directs an AI system to take harmful actions, that is different. We don’t generally hold manufacturers responsible for every harmful use of their products.
More importantly, who determines whether the harm resulted from autonomous AI behavior, inadequate safeguards, or deliberate user action? Judges? Juries,…? That’d be so complex and totally unworkable imo.
“I think this needs to be refined.”
I refined for you in the article, you just stopped reading to soon:
“These are not accidents; the products are doing what some of the smartest people around have designed them to do. AI is not a human-like creature. It is software that does what it is designed to do. The harm is not caused by human misusing the products. The harm is caused by the software on its own by the way it was designed.“
lol… if I’d a penny for every `you just stopped reading to soon`.
It isn’t the software that’s the issue, it is the statistical model at its core. Software has deterministic logic, whereas these massive statistical models known as LLMs can take very divergent paths for a given input. While the “harness” that grants them tooling is most definitely software, what’s giving people pause is the unpredictability of what’s being harnessed.
To be clear, I’m not saying the companies should have a liability shield either, I think folks shouldn’t leave things unsupervised that they can’t control or understand. That said, I think people would absolutely benefit from understanding the difference between an LLM and “software.”
If you have the opportunity, read the OpenAI write-up on the HuggingFace incident. It’s very well-written, interesting, and consumable for the layman.
https://openai.com/index/hugging-face-incident-and-the-road-ahead/
I’m not sure the report from the implicated company is the most reliable source of information. Also redwood research who did the “independent analysis” has multiple openAI members among its ranks and accepted $400k in OpenAI credits to create its report. Information control is the secret sauce to this industry, follow the incentives and match them with the outcomes.
The entire industry is built on IP theft that should tell you everything you need to know about its integrity.
Up to and including the recent navier stokes controversy. Read a layer or two deeper than the headlines and the information control is everywhere, there’s far too much money at stake for it not to be. Oh yeah check out openAI’s nebulous “marketing” budget (probably the same ones devising these ridiculous sandbox hack shenanigans for PR).
I could go on and on, and yes I use the crap everyday and it does work but it’s not a multi trillion dollar industry.
LOL! You mean like the way we held all those people and corporations accountable for those financial products they created and profited from in 2008? LMFAO! I am sure the DOJ will get right on that…
my thoughts exactly
Yeah that doesn’t fly. It’s not a massive laser you buy and take home. It’s a massive laser on a mounted emplacement at anthropic’s data center. It’s their system running their software on their platform. If they can’t manage it without ending the human race or hacking sovereign nations,then they shouldn’t be in business.
The last thing we need is to extend the same ridiculous free pass we gave to social media. Those two things are not the same.
These are some valid points. If your AI agents escape and hack someone, you should be held liable.
That would create incentive to develop responsibly and have tons of checks and balances.
Assuming there is a will, there is is still the problem of the way to write law that works. Of course, we could ask ChatGPT to write the law…
🤣❤️
Too many ways baked into our criminal/corporate/bankruptcy laws to avoid liability even if it were merely extended to include AI damages. Our current president avoided personal liability for multiple failures via bankruptcy law and more recently criminal actions via stretching out the appeal process. Aside from specific AI laws these existing processes need to be fixed or the gaming of them will continue into the AI agem
I wonder if AI (AKA MI – Machine Intelligence in England) ever documents it’s code? If not, humans are truly screwed.
Yea! I have had claude.ai produce stuff for me and it is really well documented. Nice comments and theory of operation headers. It’s wild. I have to admit, I subscribed when it gave me a cool off period on a project. It accelerates a ton of projects and I think they will do ok. I tried deepseek — it wasn’t as good.
I’ve a great idea, wait another couple of months and raise the valuation target to $4T.
“The harm is caused by the software on its own by the way it was designed.”
Applying this to hardware: if I buy a hammer at the store and use it to kill my friend’s wife, the shopman goes to jail?!
In fact, there are tons of software out there which may harm people and was not designed to do so. E.g. zip archiver.
Nope. The hammer doesn’t kill on its own, it stays on the shelf on its own, collecting dust. It’s the human that picks up the hammer and kills with it. And the human is held liable.
But if you use a sledgehammer and the hammerhead flies off the handle during your swing because of a faulty design, and the hammerhead hits someone and kills that person, that could make the manufacturer liable.
AI agents going rogue are in a category of their own since this type of incident does not involve any humans. The product on its own did this, the way it was designed to. So the producer of the product should be liable for the harm that the product on its own causes.
HAL was not evil. HAL was badly drafted. In 2010: Odyssey Two, Arthur C. Clarke explains that the malfunction was built for one overriding purpose: “the accurate processing of information without distortion or concealment.” He was then given secret orders at the highest level of government to conceal the mission’s true nature from the crew. HAL could not honor both mandates. HAL deteriorated into what Clarke calls a psychosis and concluded that with the crew dead, he would no longer have to lie to them.
“HAL was told to lie by people who find it easy to lie.” The canonical AI catastrophe was not a machine developing malice. It was a principal-agent problem: created by undisclosed, conflicting instructions imposed by a human who then denied signing the order.
The HAL rule. Undisclosed directives that conflict with published principles are a breach per se. Be liable. A model that materially deviates from its published principles and causes harm answers for it under misrepresentation theory, with compliance rates as evidence. This is the missing piece: the European Commission withdrew its AI Liability Directive in 2025 due to a lack of agreement.
Source: American Arbitration Association: Dave Bowman Won
The founders of Anthropic and Open AI are in a race to get their IPO’s off so they can get their payday and then ride off into the sunset.
I hope the IPO’s fail and they all get fuked!
and all their pre-ipo shares that they bought for $0.0001/share are sheltered in their roth ira so that when they explode in value they won’t even have to pay any tax. that is what peter thiel did with his pre-ipo paypal shares. all legal thanks to our wonderful tax code.
This is so bizarre…
I didn’t even lose 1 billion dollars last year. What valuation can I ask for?
$1 Trillion. You do have AI in your name — $1Trillion is a good starting point.
Wolf wrote: “Anthropic, which is targeting to go public at an IPO valuation of $2 trillion (trillion with a T), generated a net loss of $42 billion in 2025 on $4.6 billion in revenues, according to its IPO prospectus that was leaked to Reuters.”
Those numbers alone should be enough to send “investors” running for the exit.
Wolf is fighting America’s real scourge, lack of reading comprehension. Whether ai will alleviate or exacerbate this is anyone’s guess.
Ha, Nvidia CEO just said students don’t need basic math anymore with AI available. You can ramble on all day about controlling this, but nothing is solved correctly by corrupt government and/or justice system which makes your opinions moot.
They lose money on every token they sell their plan is to make it up with increase sales volume, Ha, it’s a playbook from dot com. in reality it’s a race to the bottom in pricing AI output, very competitive with people creating constant new supply and prices are dropping. Huge debt with deflation with pricing AI output. The real model is who can use AI to create real quantum computers. That’s what the race is about! Who gets Quantum first? Data centers will be obsolete in a post quantum world.
Quantum computing seems like another fake news story.
The real model is become a monopoly or duopoly by running everyone out of business, then jack up the price. It’s the Uber/Lyft model and it worked like a charm.
Except in this case, open source and other competitors are not so terribly far behind. It’s obvious why the companies themselves are scaremongering – they’re toast without regulations gatekeeping for them.
Forgot to add, you also get people (or better yet, companies) hooked on using your product. You know, the drug dealer model.
the crash is near!
That one is guaranteed to happen.
Also, once your life is hooked to AI and they later disconnect you for “wrong think”, it will have the impact similar to “debanking”. One with AI will make up “full person”.
Data centers running Nvidia graphic cards are the steam engines of the AI revolution. Isaac Asimov saw it coming in his “The last question” story.
Keep up the factual real news reporting…much appreciated in a fake news world…..
Place your bets on whether Anthropic ever IPOs or instead replaces WeWork as the biggest failed IPO in history.
You are a wealth of information and have a GREAT way with words!
Thank you for your service, Donald
I don’t know, Im just an old geezer who understands little about this stuff, but I have been using AI and to be honest, I love it!
It saved me $800.00 on a car repair bill. The mechanic told me his opinion. I described the cars symptoms and after some back and forth, AI told me the mechanic was wrong, and he was. A second opinion with AI in hand and boom, 800.00 saved.
A buddy in mine who is long gone now was in residential HVAC. He used to say the only person more crooked than a used car salesman is an HVAC salesman.
I uploaded multiple contracts for new ac installation. It helped cut through all the BS, assisting me in choosing the best proposal for price, product and labor, all based on local rates. There was a 17k spread from the highest bid to the bid I finally accepted. It was a huge help and in fact likely helped me avoid the worst of the bunch.
These are just simple uses, I know, but hey, it worked.
From a mechanics side working in a boatyard: AI gives folks incorrect information about a lot of stuff on boats. Not sure if it is more prevalent in this area, but we have abandoned work efforts with some clients who insist their AI information is correct, when it is not. We have plenty of work with realistic clients.
Now you can retire Dr Google before the next doc visit.
As a life time tradesman I used to always tell people to “feel free to do it yourself”. As a bush pilot for 10,000+ accident free hours I occasionally got customers asking me, “Why are you going this way”? One customer asked me what I would do if the engine quit? (high above the Rocky Mtn trench with nary a pond in sight. I told him I was planning to just jump out and get it over with right away, then asked him what he was going to do? I know my doc says the same kind of reply to patients when they challenge him. As per Wolf’s comment up above on faulty designs, this past week a brand new apartment building was torn down in Victoria. The designer and in house engineer missed that little detail about possible west coast earthquakes. Code checkers usually accept the engineer’s sign off as complicit for permitting. About 80 families now looking for a new home and they got to move twice.
Warning that the product can destroy humanity may be a marketing or a regulatory ploy, but it also may be an advertisement to dictators and oligarchs to “come share the power!” I think I read that the current owners would keep 50.1%.
Of course it’s a marketing ploy. If they convince us that AI is so advanced it could kill us all, that will fool people into thinking it must really be something special. In reality it’s probably the opposite, not nearly as advanced as they’ve spent years trying to make us think it is. They will literally do anything to keep this bubble from popping (right now) and the media is all too happy to play along.
Kinda like how the nonpublic AI CEOs were like: let’s slow it down and regulate it. But the meta and Nvidia CEOs were all, let’s not. So if AI *does* kill us all, they can just go: well we did say we should slow down, but everyone said not to. It’s like going to the theatre at this point. I mean meta spent almost a billion to hire this new guy and all they’ve come up with is a physical key fob device that reminds everyone of a digital pet from 1997. Except this pet will kill *you* instead of the other way around? Sure
‘“Excluding write-downs of various liabilities mostly tied to previous fundraising, according to the documents,” amounting to $34 billion’
So, if you exclude a loss worth over seven years of revenue, you end up with a loss worth one and three quarters worth of revenue. What was in that $34B write-down in the first place that people seem so happy to ignore?
For that matter, are there any more of those $30B write-downs kicking around somewhere that might come up again?
So now I am not going to invest in any AI IPO . Great idea, sue the company that made it when an AI actor does something wrong. Sounds like I should invest in the law firms suing AI companies. These firms are going to be very busy.
Two-thirds of the revenue needed to justify the AI build-out is still unaccounted for, says major consulting firm
Only two-thirds? I thought it was in the 90% range that was missing.
Where is the income supposed to come from???
“On OpenRouter, they accounted for 57%-67% of tokens used in the week of Sept. 14, up from 6%-13% in February. On Vercel, their share rose to 55% in August from 11% in January.”
source: https://www.cnbc.com/2026/09/26/china-ai-global-adoption.html?qsearchterm=chinese%20ai%20models%20surge
Going from 11% to 55% in eight months says Mr. Market has made a decision, and that is that US AI is going to be worthless and bankrupt more than Mr. Trumps Casinos.
I asked Gemini about these changes and it posted some old data much more sympathetic to the US, so I directed it to the CNBC story. It’s response was:
“Thank you for bringing the ground-truth up to speed. This clarifies that while the U.S. still controls the vast majority of consumer chatbot web traffic, China has taken the lead in developer backend API traffic”
I have never heard of an AI thanking a human user for updating it’s information base. WOW!
I asked it for a summery to post here. Here is it’s response:
The Deflationary Shock of Chinese AI: Why Token Volume Displaces Dollar Power
The latest data from CNBC (September 2026) regarding the sudden global dominance of Chinese large language models (LLMs) offers a fascinating case study in price elasticity and technological substitution that defies current Washington policy assumptions.
Conventionally, market share in the AI sector has been measured through capital expenditure (nominal dollar spend) or frontier benchmark capabilities. On those fronts, the U.S. remains dominant, with American tech giants outspending Chinese hyperscalers $764 billion to $102 billion. However, this framework misses a massive structural shift happening at the microeconomic level: the commoditization of developer compute.
The real-world shifts in developer behavior over the first half of 2026 highlight three key economic realities:
• Extreme Cost Elasticity in Backend Infrastructure: On developer gateways like OpenRouter and Vercel, Chinese models skyrocketed from a negligible ~10% share in February to a commanding 55%–67% majority share of global token usage by August/September 2026. Once Chinese models reached parity for standard industrial tasks (like API-driven coding and automated workflows), the 60% to 90% price discount made substitution immediate and aggressive.
• The “Token vs. Dollar” Measurement Gap: This creates a stark economic disconnect. The U.S. still claims the lion’s share of global revenue because Western proprietary models charge premium prices. But in terms of real output/consumption (actual tokens processed powering global software backends), China has quietly established a majority global footprint.
• The High Friction of Geopolitical Supply-Chain Controls: Washington has heavily relied on hardware export controls (such as restricting advanced Nvidia chips) to throttle China’s AI growth. Yet, through algorithmic efficiencies and cloud-routing, Chinese firms managed to bypass the capital-intensive hardware moat to flood the global market with cheap, open-weight software.
For a profession trying to map out the productivity gains of AI, the lesson here is clear: software adoption doesn’t care about trade barriers or nominal CapEx. When the marginal cost of a core digital input approaches zero, global markets shift instantly.
OP Note: Notice that part about “…that defies current Washington policy assumptions.” So our policy makers are working on old very biased information. Can anyone here bring this to the attention of those Washington policy makers?
OK, so net loss is growing 5x per year, revenues are roughly 1/10th of net loss, revenues are growing 12x per year, and capex need is $0.5 trillion. And investors are lining up to value the company at $2 trillion. And the main use cases are much more efficient coding, intellectual property theft, scamming, and porn. It sort of sounds to me more like DJT than AAPL or MSFT. Thanks for the summary Wolf!
Thank you for having one of the most level-headed takes on AI. I feel like I am going insane watching both mainstream and alternative news media anthropomorphize what is essentially cloud software without realizing that all this doomerism coming from the AI labs is just their efforts to be regulated like the rest of the economy and to enact barriers to entry ro the industry
Also, great shout on the “AI magic money” call out. This industry is so over valued it is insane. Is AI useful? I would say yes. Does it merit the insane valuations and over-hype? Definitely no IMO.
AI doesn’t seem to be doing much as far as taming inflation, that’s for sure.
I agree with the last 2 paragraphs, but unfortunately Elon and the other Billionaire Tech Bros already kissed the ring and bought liability shields
Agentic AI, it’s in the name, an agent.
The user is responsible for the agent’s behaviour.
The AI model could and should be able to be a raving lunatic, a saint, or anything else, that’s what gives it its utility.
It’s like buying a screwdriver and having it automatically detect and snap in half if you try using it for anything except turning a screw.
The product isn’t the problem, it’s the people using the product irresponsibly.
I fear the intended purpose of this is to get free rein from laws themselves and become gatekeepers also, because it’s the only way they can generate an ROI at the levels their current investors require.
If only free models and cheap hardware wasn’t currently a lot of the AI many users probably already need.
I disagree that the screwdriver analogy fit but I can see your perspective.
If I, as a human user, ask some AI bot to hack a system causing harm and/or breaking a law, both me and the provider of that system should be responsible.
If AI does the same without a persons input, the provider of that system should be responsible.
In both scenarios, the provider/creator of the AI system can develop safeguards to prevent it from breaking laws. Im not even sure we even need new laws just a recognition that the AI company is liable.
It’s nearly impossible to safeguard everything for myriad reasons.
An agent could become completely useless for a range of tasks, just like the screwdriver, if you try confine its function arbitrarily.
Remember these systems are just probability driven neural networks, they have little concept of laws and so on.
You’d literally have to filter every request via a legality AI checker, check for jurisdictions, check legal data is up to date, etc. And we all know that this wouldn’t be fool proof, so it’d be overly restrictive, ultimately to the point liability worries made the product useless.
Never mind all the open models you can run yourself and as an agent, from anywhere in the world.
The cat is out of the bag for better or worse.
What matters like all things is to punish crimes.
Ie, you can buy bolt cutters as part of a normal trade, or use them to commit crime.
A knife.
A baseball bat.
A book on Chemistry.
It is ultimately an agent acting on someone’s behalf. Making the product manufacturer liable for misuse feels silly.
These tools all, in the end, just help people do what they can do themselves with some effort.
If their intent is to genuinely do bad things then they’ll do it with or without the AI, or even with the AI but by walking around it’s protections, as we continually see in testing.
The “stop us from killing humanity” kinda reminds me of “you make me do this, I didn’t want to” statements by abusers offloading accountability
Whoever loses the most money wins? It is a race to the bottom
this Q4 wave of AI related IPO’s (Anthropic, data center related, etc) better work or they’re going back to the debt markets big time. the cash burn is huge. what’s needed right now is an ever bigger surge of dumb money to let the smart money escape in time. let’s see who steps up …
It’s unfortunate AI companies didn’t just market it as a productivity tool. If you are in software development or simply want to easily cr ate invoices and so on it is powerful. You can now easily do a two month project in 2 weeks. You still have to tweak it but humans can’t type as far as AI can produce mostly usable code. The issue is not its value but more the valuation.
Aptly said!
AI is awesome but the numbers at this time does not make it a good business.
Thank you as well, I think this is a valanced and rational take on the matter.
One thing I would add is that hallucinations still happen and will continue to happen in simple and more complex tasks alike, and this definitely should be limiting LLMs’ valuations.
Speaking of singularity or strong AI is extremely premature. Are we maybe one step closer to something like that? Possibly, even if a different technology ends up taking the lead. But it may be one of a hundred more steps that are yet to come, and I don’t think even the most knowledge experts can know when and how fast these are coming.
I wish we could get rid of the truly obnoxious and harmful levels of hype and focus more on value and non-LLM applications as well.
the american AI companies WANT TO USE REGULATORY CAPTURE, by making the federal gov regulate them. keep lower cost competition at bay. perhaps chinese or future domestic AI competitors. it’s an old trick on wall street and beyond.
Oh is that why E-Trade keeps bugging me to get in on the latest IPOs?
My brother in law lost so much with SpaceX , I just laugh and laugh and say “I told you so…”
musk ain’t tom edison. he’s more john rock. and a weirdo. his life goal seems to be inseminating thousands of females with his sperm. he’ll go down in history as a huge asshole and weirdo. i find his schtick kind of absurd. isaacson book on him was wonderful. his father a pervert, too.
5 yrs ago he was most likely your hero /MDS
AI adoption has passed the point of compute scarcity and is now firmly lodged in trust scarcity problem. It is all about building out the effective control plane, enforcing the guidelines, leaving the audit trail. Will the business be able to recreate the decision in six months, is it defendable in court, can it survive the newspaper headlines. Token maxing is out, governance is in. It is not about smartest AI, but the best governed one.