Introduction
After the recent bond and stock market drop, everyone got scared and AI skeptics like Ed Zitron are in the spotlight again. Let's discuss what's going on and whether this is a major worry for investors, tech workers or just everyday people concerned about a recession.
Messy details and the root of the problem
Lots of different approaches exist: proponents point to demand outpacing supply and old GPU prices going up (SemiAnalysis, AWS Capacity Blocks), others point to circular financing (OpenAI–NVIDIA, OpenAI–Oracle, CoreWeave 8-K) and many failed enterprise pilots (MIT NANDA, July 2025). These are important, but not likely to be the deciding factor. The main question is: will the money flow stop or slow down enough that a massive portion of the leverage in the system blows up, with many large bankruptcies and lots of large loans not being paid back? This could cause a global recession, both through financial instruments dropping and through investment and construction slowing down enough to impact jobs. One or two neoclouds failing is likely to happen anyway and would not cause major issues.

Circular financing looks like a bowl of spaghetti
Layers and profitability
I'm going to use Jensen Huang's 5-layer categorization, because it is well-known and mostly fits the topic. I'm also adding a 6th layer: users, because they are paying for most of this if all goes well.

Layers of AI
Energy, chip and infrastructure supply chains are making a lot of money, and the neoclouds and hyperscalers have high gross margins but are investing so much in capex that their cash flow is negative (except Microsoft, because they invest less CoreWeave Q3 2026, Microsoft FY27 Q1, Alphabet Q3 2026, Amazon Q3 2026, Meta Q3 2026). Model providers do the same with R&D and sometimes capex as well. Both would be profitable if they slowed down investments in growth or the profitable inference could take over as the majority of the spend if it grows significantly more than R&D.

Operating cash flow vs capital expenditures at hyperscalers
Applications are subsidizing users in some cases, like agents and chatbots, and making good margins in others, like enterprise APIs. Some estimates show profitability even on the average subscription user (SemiAnalysis, discussion). Some could fail, just as some neoclouds might get overleveraged with their loans and fail. There will certainly be some cases like that, but that happens in every tech boom and does not matter as long as it is a few small companies on the margin.
But not all users have productivity increases. So the question is: will users find enough value to keep paying for all of this?
Will the money flow stop?
Some users are seeing unbelievable productivity gains and fast growth. Others are seeing modest improvements. The statistics are mostly sad, with the majority of enterprises barely seeing any measurable impact (NBER w34836). Most of them are only using chatbots though (March 2026 w34836 PDF, McKinsey 2026, MIT NANDA, Ramp, October 2026, ChatGPT Business), and median spend is less than one $20 subscription/user (Ramp AI Index).

Lots of AI trials, very few clear successes
That means market penetration is low for everything except the cheapest chatbots, but almost everyone is trying or planning to use it for something (McKinsey 2026, NBER w34836 PDF, Census BTOS) and failing at it. Are they going to keep trying? Will more of them succeed in the future? History says yes. I went through such transformations at multiple companies, multiple times. Migrating to the cloud, to Agile or to SAP does not bring improvements immediately, yet everyone did it and sustained a consulting industry for decades doing these. Often you need to reorganize because the process you sped up was not the bottleneck, and many orgs don't realize it or take years to do it right.
The race to the bottom
Most CEOs and CTOs know about AI and are aware that this technology has the potential to increase their productivity, either through reducing costs or increasing output. Both have been shown to be possible, and most of them will try and keep trying because the threat of disruption is real. If a competitor uses it to reduce their costs and then their prices, they can take market share (or increase margins) unless you have a big moat. This is the innovator's dilemma: if you don't disrupt yourself, someone else will do it for you. Google is facing this with search: their search ad revenue growth forecasts were cut recently, but still growing (Alphabet Q3 2026, earnings call, Proactive on UBS, 30 Oct 2026). They can try to protect the old business or be part of the new group of companies replacing it.
Some moats are gone, like having built software, but network effects and some other lock-ins still work. Most people will not build their own email server, but selling Acrobat to edit PDFs will be harder when the free tools get just as good quickly. When a small company can create Pareto replacements for entire product suites, with the 20% of features people use 80% of the time, we truly have new and unique possibilities ahead. (See ArtCraft if you haven't heard of it: they have the most-used features of Adobe and Microsoft Office apps with the same UI and released as open source, free for everyone.)

You would think you are looking at Microsoft Excel, but you are not
Such productivity increases create more output from the same inputs and grow the economy, create cheaper products that sell more (Jevons paradox) or sell into new markets that didn't exist because costs were too high. Very few people would pay tens of thousands or more for IBM Watson to be their personal assistant; lots of people are willing to pay tens or hundreds of dollars for Grok Bot or to run OpenClaw (xAI pricing, SpaceX Says Grok Bot Doubled to 800,000 Paid Weekly Users in a Month).
Some more reasons for growth
The capabilities are also increasing quickly, and most companies are not using or even trying to use what was available 6 months ago. The ones using last year's stuff (chatbots/research agents) effectively are still ahead of the pack. Agentic work, which exploded around the end of last year and the beginning of this year, is still not widely used, even though there are new interfaces to it like Grok Bot or Dots that make it incredibly easy to start.
One of the reasons productivity stays flat in many cases is that companies usually improve, but not at the bottleneck, which would need workflow changes and maybe even a reorganization in some cases (NBER w35275). Everyone else works faster and is still waiting for the one slow process that hasn't changed. This will be another boom for consultants who will promise to find and fix this problem, make process improvements and help reorganize companies. We have seen how this goes: most of them will still not work, but almost everyone will try and keep trying. It is likely to go on for years, as most of them haven't started yet or only had a few small pilots.

Speed up everything except the bottleneck and the output does not move
Some companies, like Eli Lilly, are building their own GPU clusters in-house (Lilly, 28 Oct 2025, NVIDIA, 26 Feb 2026). This seems to be paying off as drug development is speeding up time to launch already (Lilly, access and innovation): they can create more compounds to investigate, use simulation and other AIs to test and evaluate them faster, and have more drug tests in their pipeline. This is an enormous industry that is already benefiting, with multiple drug companies reporting more and faster trials and improved success rates compared to before AI (Reuters, 26 Jan 2026, Jayatunga, Liu, and others, Drug Discovery Today, 2024).
Token prices, and especially cost per task (a better metric), are likely to keep falling, which makes old GPUs more valuable. Algorithmic improvements like better infrastructure code are one reason, but distillation improves efficiency as well. Future frontier models are smarter and can be distilled into smaller versions that are just as smart as today's best at a lower cost. This has been going on for a while, and estimates of price drops are all over the place: some say 5x/year, others say 10x or over 100x (Epoch AI). The exact number depends on the chosen models and benchmarks, but the improvements are clearly very significant.

AI is getting cheaper for the same capability by orders of magnitude
Summary
The enterprise market seems to be in the early adopter phase with anything other than cheap chatbots. Agents use orders of magnitude more compute and deliver a lot more value. If they start adopting the new tools, the revenue growth can continue. Enterprises are likely to keep spending and sustain the industry growth. Even if it keeps getting cheaper at a much lower 2-3x rate instead of the >5x measured so far, the value will be there even for large enterprises who pay higher prices and do not get the cheap subscriptions.
The model developers show slowing revenue growth and have huge commitments. (Anthropic underinvested in compute and had to scramble to get it from a competitor at 3x the price of a typical neocloud rental at the time; OpenAI had massive commitments for a long time) (Anthropic S-1/A, filed 2 Nov 2026, NVIDIA letter of intent, OpenAI on the Oracle partnership). We'll see if they can keep up with revenue; current trends get us there, but if revenue growth slows down significantly, they could be in trouble. Neoclouds have similar risks; many are leveraging themselves with loans.

Risk map of the layers
Even if some model companies and/or neoclouds fail (likely some will), the rest of the industry is likely to continue. Their commitments and data centers are probably going to be bought in a fire sale, and everything continues if the users keep buying.
I would worry more about the cybersecurity risks than the loan fallout leading to a 2008-like financial crisis, but that is a topic for another day. Agents also automate exploit development and break-ins, creating a lot more opportunities for attackers. Defenders can also apply these tools, but many are behind, for various reasons.
