@adlrocha - Are we approaching a new AI winter?

An adoption crisis and the end of an AI cycle are aligning

I don’t know about you, but I feel like the mood around AI is shifting slightly. What once was the promise of the imminent disappearance of every knowledge worker in society so we could all enjoy our free time and hobbies, is cooling down and moving towards the realisation that this new technology in its current form still has limits. While it most certainly will eventually make some humans obsolete, it won’t be this early the promised Philosopher’s Stone that everyone was preaching.

I feel like we may be approaching the end of the LLMs and agents honeymoon. And please don’t get me wrong, I am not saying that this technology is not going to be useful, but that we are going to start seeing less optimism for a bit until we get to the next summer season.

This winter is not happening as a result of the technology plateauing or reaching a ceiling (which I don’t feel in a position to make an informed decision about) but due to a lack of adoption and technological diffusion. But coming from crypto I can tell you with some authority, winters are great to build without the distraction of the constant noise, and now is the best time to work on AI adoption.

> Special thanks to Pablo Grueso for helping shape and strengthen my opinion on the topic after our brief conversation about the matter in a car ride. Cheers!

---

What tokenmaxxing was

For about eighteen months, the dominant corporate theory of AI adoption was simple: the more your employees used AI, the better. Everyone needed to start adopting this new technology, become AI-native, and explore how their capabilities and output could be augmented. AI usage became the target metric. Companies built internal leaderboards, set token-consumption targets, and measured AI success the way they once measured digital transformation, by adoption rate, not by outcomes.

I am indeed referring to the infamous tokenmaxxing, i.e. feed the models as many tokens as possible, maximise throughput as this will maximise your production of results. The assumption baked into it was that more input would produce proportionally more output and consequently value. I still ask myself how this could be thought of as a good idea in the first place.

Amazon, for instance, ran a leaderboard called KiroRank, an internal ranking system that scored engineers by their activity on Kiro, the company’s AI developer platform. It seemed like a reasonable way to measure who was actually using the tools. What happened instead was predictable in retrospect: engineers assigned autonomous agents to run unnecessary tasks just to climb the rankings. Token consumption went up, but useful work didn’t follow (oh, surprise!). Amazon’s senior vice-president Dave Treadwell eventually told staff: “Please don’t use AI just for the sake of using AI. Use AI to help you solve customer problems, to help you solve business problems, to innovate”. However, the leaderboard wasn’t incentivising that behaviour, but the use of more and more tokens (fuck compacting your context). Obviously, the leaderboard was shut down.

Amazon replaced it with something more sensible: tracking whether engineers were regularly producing useful code with AI, not how many tokens they were burning through (a more subjective metric, harder to measure, but more aligned with the real output they were looking for).

This is not an AI problem, per se, but another example of the design of policies that do not have the goals and incentives in mind. But when AI was going to solve everything, more tokens could translate into more solutions. Turns out that AI may need to be conveniently steered in order to solve things, and strategy and domain knowledge doesn’t burn as many tokens and require humans to actually work.

First warning sign that we may have not figured out how to adopt this technology just yet.

---

Why it stopped making sense

The clearest data point on the ROI of the tokenmaxxing failure came from Uber. The company’s CTO revealed that Uber had burned through its entire 2026 Claude Code budget by April, four months into the year. The COO, Andrew Macdonald, then said publicly what a lot of people were already thinking: “That link is not there yet”, meaning the link between AI token consumption and features that users actually wanted. Uber pumped the brakes (again, oh, surprise!).

Michael Burry, who made his name betting against the 2008 housing market, described AI tokenmaxxing as a “crazy, rushed, temporary phase” driven by “quota-driven, leaderboard-driven, management-mandated overconsumption.” He drew explicit parallels to the late-1990s dot-com bubble and backed his view by purchasing put options on 1 million Nvidia shares. We’ll come to this comparison to the dot-com bubble in a few paragraphs.

Fortune’s analysis put it more formally: most companies are stuck at Stage 1 or 2 of AI adoption, i.e. basic implementation, workflow redesign. Real value requires business reinvention, the kind that most incumbents aren’t actually attempting. The companies eating their lunch are AI-native from the start.

This is not the same as saying AI doesn’t work. It’s saying that we still don’t know how to efficiently use and apply AI. This is why tokenmaxxing failed as a metric for AI adoption and proficiency in the company. Optimising for the measurement rather than the outcome, produces exactly what you’d expect: lots of activity, not much value.

If they gave me a penny every time that I’ve heard from friends and colleagues in the last few months the sentence: “since AI, I am working more than ever, it’s like there is no time to catch up with new developments”. And my question to them is always the same, “do you think that you’ve produced more value than before AI?” Spoiler alert: the responses differ greatly :) (related to this, this may be a good moment to read this post that I wrote a few months ago about how “we are not scared of AI, we are scared of irrelevance” if you haven’t already.

Yet another point in favour of my thesis about now being the best time to start adopting AI, but one where the public narrative will start feeling colder.

---

The convenient excuse

But AI adoption is not only becoming an obsession for some, but an excuse for others. Across 2025 and into 2026, a pattern emerged: companies announcing large-scale layoffs with AI as the stated rationale. Amazon (~30,000), UPS (~48,000), Oracle (~30,000), Microsoft (~23,000), Salesforce (~5,000). Roughly 80,000 jobs in 2026 alone, with 45+ CEOs citing AI as a driver (see this source I came across in my research).

Jack Dorsey cut Block from over 10,000 to just under 6,000 people and was explicit about it: “We’re not making this decision because we’re in trouble. Our business is strong... But something has changed.” Brian Armstrong at Coinbase put it similarly: “I’ve watched engineers use AI to ship in days what used to take a team weeks” (and then they have a massive outage). He announced the end of “pure managers” and described the goal as “rebuilding Coinbase as an intelligence, with humans around the edge aligning it.”

I don’t think these CEOs are lying. AI genuinely does change what’s possible with a smaller team. But AI is also functioning, in many of these cases, as a socially acceptable frame for a contraction that was coming regardless. Companies overhired during the low-interest-rate boom, and the correction was inevitable. AI provides a clean narrative, shifting the cause from “we made bad hiring decisions” to “the technology changed.” Both things can be true at once, but I am not buying the AI narrative just yet.

What I don’t expect (at least not yet) is a massive wave of net job destruction. The more likely near-term pattern is a contraction while people and organisations figure out what AI actually is, followed by an expansion once a new generation of AI-native professionals emerges who know how to use these tools with the right judgment. That’s historically how transformative technologies diffuse. It’s rarely a smooth upward curve, and it’s almost never the story that gets told during the first wave.

I am an AI-pilled myself, and I think AI is a different technology and will create a completely different revolution to the ones we are used to, but I feel it’s still early. Narrative-wise is great, but I can’t stress it enough, we still haven’t figured out the best way to use this technology.

---

The codebase regret

And here comes the reason that convinced me to write why I thought “winter is coming” as a result of us still figuring out how to use this technology. Some developers have started posting publicly, in increasing numbers, that they regret how heavily they relied on AI to build their codebases.

Dragos Nedelcu, a senior developer, wrote about generating roughly 150,000 lines of AI code across a production project. After a few months, he faced a mess: duplicated logic with almost no reusability, dead code everywhere, unit tests that didn’t assert anything meaningful, and cascading bugs that spread across seven or more files simultaneously. His conclusion was blunt: “It is faster to just start over instead of correcting hundreds of lines of messy AI-generated code.” Another engineer mass-deleted 14,000 lines of AI-generated code after seven months, the codebase shrank from 41,000 to 27,000 lines while keeping all the same features, and the bug rate fell 73%.

Let’s be honest, we’ve all faced this at some point in our relationship with coding agents (at least did).

Victor Taelin, who builds the HVM programming language and someone I’ve been following for years, documented his pain in real time on X. He used Opus to implement a new approach in a single day: 3,000 lines of C, 5x performance improvement. Then spent the next fifteen hours auditing it, finding what he called “retarded shit”: a case where the model had silently assumed HVM5 was supposed to handle under- and over-applied functions, implemented a massive system for that assumption, and never asked. None of it should have existed.

His conclusion shows another reason why I think that we may have not learnt how to use this technology effectively just yet: “I went from 0 to 95% in the first 5 hours. Yet, 15 hours later, it is still not 100%... if I have to read it all, review it all to ensure there is no retarded shit... what did I achieve by using AI, other than that dopamine anticipation?”

That last phrase, dopamine anticipation, is the most honest description of vibe-coding I’ve read. Oh, that beautiful AI slop slot machine that we’ve all become so addicted to.

Luis Ángel Alda, a Spanish developer, put the structural problem well: AI produces systems that are “locally correct, globally incoherent.” The models are good at optimising the next step. Architecture is precisely the opposite, a long-horizon, intuition-driven discipline that requires what he calls “feeling the software.” You accumulate that from years of building things, watching them fail, and rebuilding them. AI doesn’t have it. What it has is extremely good local pattern completion, which is useful for a large number of things and actively harmful when you need global coherence.

Again, none of this means AI is useless for coding. I use it constantly and it has genuinely changed how much I can ship. But there is a difference between using AI as a tool with judgment and using it as a substitute for judgment. The people posting these regret stories mostly did the latter, they handed over the architecture, not just the boilerplate.

The skill that actually matters here is knowing when to use it and when not to and this is what this winter will be about. That takes time and accumulated failures to develop. We’re all still learning it. I include myself. That “human/engineering taste” is still very much needed.

---

Why winters are needed

Every major technological wave has this shape, a honeymoon, a contraction, then a long slow diffusion that ends up being far more transformative than the original hype suggested, just on a completely different timeline.

Electricity took roughly forty years to fully diffuse through US manufacturing after Edison’s first commercial power station. The economists who studied this called it the “productivity paradox”. The technology was clearly working, but the aggregate numbers didn’t move for a generation. Businesses had to learn to reorganise around the new capability rather than just bolt it onto existing processes. That reorganisation is what unlocked the gains.

The internet had the same problem, but with an extra layer. The technology worked: email in the early 90s, the web in 1993, but for years people used it like they used old media. Banner ads that looked like billboards. Search engines that were just yellow pages. The capability was there; the mental model for how to actually use it wasn’t. Google didn’t invent the internet, but it gave people a way into it that matched how the thing actually worked. The interface had to catch up to the infrastructure before the value became visible.

AI is probably at that stage right now. Tokenmaxxing was the billboard phase, a genuinely new capability being forced into the shape of an old metric. The companies building internal KiroRank leaderboards weren’t stupid. They were doing what organisations always do when they don’t yet have the right mental model: they reached for the nearest familiar measurement. We’re waiting for the equivalent of a search bar, not a better model, but a better way of thinking about what the model is actually for.

AI will likely follow something like that path. Not because the technology stops working, but because the same absorption process takes time. The companies that figure out how to reorganise around AI, rather than just mandate token spending targets, will be the ones that capture the actual gains. Professionals who develop genuine “taste” about when to use it and when not to will become more valuable, not less. Software engineers aren’t going away, the role is shifting toward the people who can direct AI, write the spec, design the constraints, and catch architectural drift before it becomes a 150,000-line unmaintainable mess.

---

What the market is signalling

But how can we be approaching a winter if Anthropic closed a Series H at a post-money valuation of $965 billion, nearly a trillion dollars, and days later, it filed a confidential S-1 with the SEC? When Nvidia, whose actual earnings are real and growing fast, $215 billion in revenue in fiscal 2026, up 65%, trades at a forward P/E of around 25? Well, that sounds reasonable until you remember that 64% of its accounts receivable comes from three customers.

This is where the dot-com parallel becomes useful, not as a prediction but as a frame.

Amazon IPO’d in 1997 at 18 times sales. By late 1999, its stock had run to $107 on a wave of internet euphoria. Then it fell 92%, down to $7 by 2001. It took until roughly 2009 for the stock to recover its dot-com peak. What came in between? Nine years of grinding. AWS S3 launched in March 2006. EC2 followed in August. The thing that actually changed the world, web2.0, cloud computing, the infrastructure that every startup now runs on and became its biggest revenue source, arrived almost a decade after the IPO, and six years after the crash. Amazon was right about the internet. The timeline that the public narrative was painting, though, was just radically wrong.

And Amazon was one of the survivors. Pets.com IPO’d in February 2000 at a $290 million valuation, spent $27 million on advertising in its first year against $619,000 in revenue, and liquidated nine months later. Webvan raised $375 million in its 1999 IPO, first-day market cap of $8 billion, filed for bankruptcy in 2001 after losing over $800 million. In October 1999, the combined market cap of 199 internet stocks was $450 billion. Their total annual sales: $21 billion. Collective losses: $6.2 billion. The internet was going to change everything, and it did. Just not for most of the companies that raised money on that promise, and not on the timeline anyone expected. It took Google, and then the read-write web, and then smartphones, to unlock what the internet actually was. Not better technology, but a different mental model for what the technology was for.

None of this proves we’re in a bubble (because I honestly don’t think this is the case), but it smells like the end of the AI cycle. Anthropic’s $47 billion run-rate revenue is real. OpenAI’s ~$25 billion is real. The underlying technology is not Pets.com. But the pattern is a bit familiar, the extraordinary valuations are definitely concerning, IPO windows opening simultaneously, companies describing themselves in transformational rather than financial terms, that’s the pattern worth recognising. And we are seeing OpenAI, Anthropic, and even SpaceX (which has some AI components through xAI) rushing to IPO before the summer is over. The market feels quite heated.

Tokenmaxxing is AI’s billboard phase. The capability is real. The way most organisations are using it is just the old productivity metric with a new name. We haven’t had our Google moment yet, the interface, the workflow, the mental model that makes the value obvious and self-reinforcing. When it comes, the diffusion will accelerate fast. But we’re not there yet, and the market is priced as if we are.

Which brings us to what might actually be coming. Two cycles converging at the same time: an adoption winter, where the hype meets the actual difficulty of integration and the regret starts to accumulate, and a market correction, where valuations revert toward something connected to reality. Again, coming from crypto I know a bit about this.

AI is not crypto. The underlying technology is already being useful, and the enterprise revenues are not made of magical Internet money. But narrative compression works the same way regardless of fundamentals. If the adoption slowdown and the IPO hangover arrive in the same window, the optimistic story gets very quiet very quickly. Not because AI stopped working, but because the story ran too far ahead of the evidence, and the market has a way of correcting that, painfully and all at once. Humans tend to over-react in mass.

---

Is there something that I am not seeing?

All of this view of an upcoming winter comes from an outsider. I am an occasional Twitter reader, an actual user and practitioner of this technology, a follower (and lover) of the financial markets, and I am constantly talking with people in the space, but I am not based in SF, and do not work in a big frontier lab (unfortunately). The people who actually know whether the models are plateauing or on the edge of another leap are the researchers inside the labs, the ones running the experiments, watching the eval curves, seeing what the next generation of models can and can’t do before anyone else does, and how their technology is actually being adopted. Those and the executives and product people pushing for the diffusion could really assess how much of the mainstream AI narrative is real (Mythos, I am looking at you) and could fill the gap on our lack of adoption knowledge.

From the outside, the corporate behaviour and the market mechanics look like what end-of-cycle tends to look like. But so did the internet in 1996, and the real explosion came three years later. The labs might be rushing to close their IPOs because they’re worried. Or they might be doing it because they can see something coming that the market hasn’t priced in yet, and they’d rather have the capital before everyone else figures it out.

I genuinely don’t know which of those is true. And I’d be curious whether anyone closer to the inside does. But unless someone has additional data points to share, I think we can close here. See you next week.