Financing the AI boom: from cash flows to debt [pdf]
bis.org
bis.org
But like what happens if the government guarantees open ai's loans if the company is structurally unprofitable? Does the government create an operating subsidy?
The government intervened to prevent massive job losses, protect the domestic auto industry, and, in the 1979 case, preserve critical national security manufacturing, as Chrysler produced the M-1 Abrams tank.
Now, I suppose it is possible to imagine that the US government might bail out either or both of OpenAI or Anthropic (whether or not there's an ROI like there was with the Bank Bailout of the 2008 crisis) if the govt. deemed the technology critical to keep a fast pace on (I think we can say without doubt that requirement is satisfied) but, crucially, the government's calculus is that it is better to have these companies compete rather than bring the knowhow in-house.
Is that what you're thinking?
So it would look like the government taking ownership, letting investors lose their stake, and then operating as inference-only, which would turn a profit
We have absolutely 0 hard proof of this. We have a lot of wishful thinking but no hard numbers, audited numbers from any public entity.
I'd love to see them if they are available.
Do we have the balance sheet for OpenRouter & co?
Especially in this age where if you put AI in your company's mission statement you're drowned in money.
Let's hold off on calling something "cheap" until the external financing money runs out and the actual numbers are revealed AND audited.
> yes it's cheap.
When running toy models that do basically 0 of what regular people expect from state of the art LLMs, sure.
Running Apache is cheap. Running Google search isn't. They both serve web pages.
Running a local LLM isn't a mainstream normal thing to do, sure but saying it's "basically 0 of what regular people expect from state of the art LLMs" is lazily dismissing evidence because it contradicts your beliefs. It does work, and it works at a level somewhere above "basically zero" for nerds who are willing and able to set it up for themselves today. The comparison isn't Apache, it's ElasticSearch. It's not Apache cheap and simple, but it's also not Google Spanner expensive.
Again, you have no way of knowing this. During a bubble a myriad small companies nobody hears about get funded for millions and billions.
Also, plenty of startups max out the founder credit cards and then they go bankrupt.
Let's revisit this discussion and see if 10% of all the companies in OpenRouter are around in 2030.
> is lazily dismissing evidence because it contradicts your beliefs
No, it contradicts my experience. The models you can run on 128GB of VRAM/unified RAM (which is realistically the maximum a regular person can buy) are basically bad compared to Anthropic/OpenAI.
More than that hardware prices spike like crazy (and even if consumers could afford them, the hardware itself is basically a huge DYI project).
Let alone the fact that regular users need to run other things on their system so can't dedicate absolutely everything to the LLM.
Again, the financials of these businesses are at best unproven and at worst critically unsound.
I'm sure the industry will consolidate by 2030, as industry always does.
AI is currently a sunk cost to the US stock industry that is repeating the bad actor scenario. Not a single AI company is profitable and none of them produce deterministic nor cost effective solutions
Microsoft's statment of using AI to find the most resource intensive applications being ran highlights this. Task Manager does the same thing and does not need a server farm for training. It also uses MB of RAM vs GB.
If manufacturing had the same error rate in production as AI, those plants would of went out of business.
Both industries heavy use legal bribes, donations. Politicians will gladly bail them out to take ℅ of the cut in bribes.
Too big to fails are false claims to retain the bad actors that fund politicians. Bad actors need to fail so the good ones can properly operate.
Too big to fail is also allowing large corporations to skirt copyright laws. You or I seeding TB of copyright content would be thrown in jail.
Too big to fail is rebranding of legalizing corruption.
Are you saying that it is likely that at some point in the not too distant future, OpenAI and Anthropic will need bailout-size cash infusions from the US Government to continue existence and that the US government will do it and not face severe political consequences?
I just don't think that chain of events is likely. The current administration pays very close attention voter sentiment.
Didn't they say this themselves? https://edition.cnn.com/2025/11/06/tech/openai-backtracks-go...
> The current administration pays very close attention voter sentiment.
Is that how the Anti-Weaponization Fund came about?
OpenAI backtracked, according to the article you linked to.
> Is that how the Anti-Weaponization Fund came about?
It hasn’t come about has it? It was blocked. Besides, I don’t see the connection with the topic at hand.
In which alternative universe? The amount of bribery and self-enrichment is staggering. The treatment of the war is mind boggling. The actual political moves seem to be designed to punish disloyal republicans rather then win more votes.
Pushing against regulation supposed to prevent this happening again. They also resumed taking a lot of risks, just like before. Their managers got rich while doing the same decisions as before, because other people paid the price.
> What’s the argument for OpenAI being so inherently critical and interweaved with the rest of the economy
It does not need to be inherently critical and interweaved with the rest of the economy. It has to pay the bribe to the right president. The "military necessity" excuse will be used then.
AI companies are more and more interweaved with the economy because half the world is owning their stocks or has lent them money. Or, they have invested in companies that in turn have invested in AI. It is very similar to the situation before the credit crisis.
As the funding profile moves more towards debt, there's a bigger chance for problems as leverage tends to amplify the bad outcomes.
If Big AI (basically Mag 7) continue to issue debt, then there may be financial stability risks if it all blows up. The numbers are fine now, but the trend line is concerning (hence the BIS paper).
At this point anything less than "medium growth" will crash the economy. We'll have bigger problems if that happens (think 2000 or 2008)
Debt presumes future growth.
[1] a broad and poorly defined group of "we" - typically investors and tech-bro types.
Earnings did come, and the outcome of the internet include Google, Amazon, and others -- several of the ten biggest companies in the world.
In 2000, there was no sane way to predict who the success stories versus failures would be, timeliness, or otherwise.
AI will be a big change. We don't know how big, when, or who the losers and winners will be yet. Everything could be grossly overvalued or undervalued. We'll only know in hindsight.
Duolingo is such a company you would expect AI to help a lot. Surely AI could allow it to cut costs substantially. And yet, in the past year its stock is down 70% and in Q1 2026 profit has not seemed to increase compared to Q4 2025. In fact, other than Q3 of last year which had some tax shenanigans, their profit is relatively flat. Not a great look given that AI is highly disruptive to their product.
---
AI is actually insidious. Suppose you're in a competitive industry like Costco making 3% (net profit) margin. Suppose the average costco employee makes 60K. Then you come in and think it would be great to have an AI agent lets every employee ask questions of inventory to help customers. Surely if employees could use AI that could somehow make more money for Costco. Hypothetically let's say this ends up costing about the same as the basic subscription in terms of tokens. $20/employee/month Can't be that bad right?
$240 ÷ 0.03 = $8,000 (in other words, generate over 10% of their own salary in marginal additional net profit every year). Is Costco really going to generate 8K more per employee? Nope. And yet, firms like Costco who choose AI effectively just lower their own profit margins.
I want to add context for Duolingo because I think it's confounded by cultural conversation. Duolingo in some Q1 earnings call stated it was going to go AI-native. They were going to switch from contractors to AI. This led to a huge PR crisis, especially on TikTok, where they used to have a big online presence and GenZ influence. I'm not sure if they ever recovered the image they had after that.
They tried, and they got pushback from their consumer base.
Your Costco example makes a lot of sense to me. Right now at my company we are spending hundreds and thousands per employee, but it's not like everyone has an idea that is meant to be integrated into product. So everyone's just making more vaporware.
I do think this is the year the numbers and projections will start coming down to the ground. We can already see this with the fact that the leaders at the frontier labs have stopped talking about AGI and have started talking about token costs and just direct comparisons. It's no longer pie in the sky.
So far, there hasn't been a clear win for "software wrapping an existing LLM" - if the AI is good enough, then the users can go directly to the source.
That's a different and older kind of AI than chatbots, but it's not fundamentally different, but AI is the engine that makes their ad targeting so effective, and why they're some of the most profitable large companies on the planet.
It is. We are talking about LLMs here.
If you have a lemonade stand you might sell a cup for $1, but overall after paying yourself and for the cups, lemons, etc you might only get 3 cents each cup.
For costco revenue equals sales and membership.
What you are mentioning with salaries is not relevant.
What you are saying is again irrelevant. It is not about justification. Is about net margins…
https://arxiv.org/abs/2510.12049
> We quantify the short-term impact of Generative Artificial Intelligence (GenAI) on sales performance through a series of large-scale randomized field experiments involving millions of users and products at a leading cross-border online retail platform. Over 2023-2024, the platform integrated GenAI into seven business workflows spanning customer service, consumer-product matching, advertising, and seller services. We find that GenAI adoption increases sales in most workflows, with effects ranging from no detectable impact to 16.3%, depending on GenAI's marginal contribution relative to baseline firm practices. Across the four GenAI applications with positive sales effects, the implied annual incremental value is roughly $5 per consumer−an economically meaningful impact given the retailer's scale and the early stage of GenAI adoption. The gains operate primarily through higher conversion rates rather than larger cart values, consistent with GenAI improving the shopping experience by reducing search, information, communication, and personalization frictions. Importantly, these effects are not associated with worse post-purchase outcomes, as product return rates and customer ratings do not deteriorate. Finally, we document substantial demand-side heterogeneity, with larger gains for less experienced consumers. Our findings provide novel, large-scale causal evidence on how GenAI shapes sales productivity in online retail, highlighting both its immediate value and broader potential.
Impact on profitability itself is hard to determine due to caveats listed in the paper (which are important to read!) but offhand I would guess that incemental $5 margin per customer is much more than what their prompts cost.
So if AI is real then that‘s the cherry on top: people can now make an alternative to your ineffective messy app even easier.
For those kind of SaaS products with no moat LLMs could actually be a problem and definitely aren’t a good thing
AI announcement annoyed some people, the slop translate courses were, well, slop. That is the extend of the change.
It's a sign of how much the economy has grown that under "1% of GDP for a few years" now is far bigger than "over 10% of GDP for a few decades" was in the late 1800s.
https://news.ycombinator.com/item?id=44805979
Your estimate of current AI spending is also low. Hyperscaler capex alone is around 2% of US GDP, not including other costs (neoclouds, employee comp, etc.).
▫ Apollo Program: $257B, 14 years
▫ Interstate Highway System: $620B, 37 years
▫ AI data centers: $930B, 6 years and still accelerating
This is the pattern I’d expect for their IPOs too, give their current fantasy valuations.
[1] https://news.ycombinator.com/item?id=48870966 pgrust passes 100% of the Postgres regression tests
[2] https://news.ycombinator.com/item?id=48837877 Rewriting Bun in Rust
[3] https://news.ycombinator.com/item?id=48789325 My AI-built PHP engine in Rust passes 17% of PHP-src tests, renders WordPress (ekinertac.com)
AIUI, OP is saying that, with all these DCs with no load, we'll have excess electricity generation capacity that was built to support these DCs. That's the "cheap power" he is talking about, not necessarily "cheap computational power".
Also, a bunch of that money they spent will be on transmission infrastructure that is 100% useless if you aren’t moving that amount of power along exactly that route.
Companies like Bloom Energy are putting up power plants right next to DCs. If datacenter power demand then these companies will sell back to the grid creating an oversupply.
If the debt load is too high they're need to declare bankruptcy, at which point the bond holders will lose their capital, not the public.
For example, did macro investment in factory automation predict future productivity gains?