Insert token to continue, says AI. Yeah, about that...

OPINION It is too early to call peak 2026, but if we allow midterms, it's hard to beat Caveman. Caveman is a Claude Code skill that strips away non-essential linguistic components of the AI's output, making it communicate in a parody of a coding Neanderthal. Yes, this is a good way to use the products of an industry expected to spend a trillion dollars on capex this year. Ug fix API. Hyperscalers issue debt. Token minimization has been a set of techniques of interest for a while, something that's good to know yet a lot of faff if nobody's counting the cost. Well, they're counting it now, and Caveman is merely the latest and most surreal consequence. There are parallels with the changing importance of data compression in storage and transmission as the cost, capabilities, and demands on disks, networks, and memory evolved, and we should expect minimization to rapidly evolve to ease the expense of AI. It's a good indicator that in the area where LLMs have the most demonstrable effects on efficiency, in the production of hard, usable code, they cost too much, which has come as an unpleasant surprise to bean counters who had been told that AI = profit. The question of when that equation stops working for investors is one that's becoming increasingly bothersome, with the Bank for International Settlements noting that the capex carnival looks a lot like the history of canals, railways, and electrification, where new technologies encouraged Himalayan investments but produced only flatlands of profit. Oh, and global economic convulsions. The BIS is the bankers' bankers' banker, but we've had enough of experts, right? Unhelpfully, the structural failings of the hyperscaler AI economy will only be obvious in retrospect. There are just so many candidates. Will it be supply constraints of chips and power that quench promised growth? How about huge frontier model testing and training costs for products that go obsolete in months? Should they focus on consumers, enterprises, or some mixture of the two? There are a lot more, like how ready enterprises will be to make rapidly morphing AI integral to lines of business, as this study of OpenAI's financials entertainingly demonstrates. It came to no conclusions, not least because the work was completed just before OpenClaw made everyone aware that agents are a whole new world of tokens-a-gogo. And these guys are optimists. AI is not merely eating all the money; it is using that money to compete with everything else in tech for energy, components, datacenter utilization, attention, and resources within the hyperscalers that already provide so much essential infrastructure, among other grim side effects. This has pushed an annual inflation rate of 300-400 percent into the memory supply chain, damaging OEM economics in an already sticky market. Refresh cycles will slow, and important developments that may help AI adoption, such as on-prem and hybrid models, will become less attractive. There is no sign of an end to the competition for resources, nor how far into other industries the voracious appetite of AI will extend. Fans of Douglas Adams can be forgiven for becoming uneasy here. This is reminiscent of his economic theory of the Shoe Event Horizon, where a society becomes so fixated on new shoes that it becomes increasingly difficult to start any new retailer that isn't a shoe shop. At the Shoe Event Horizon, there is nothing but shoe shops, civilization collapses, and everyone evolves into birds to avoid even thinking about feet. Elon Musk is a big Adams fan. We cannot exclude the possibility that he sees this as an attractive macroeconomic theory, alongside the neurotic AIs and intergalactic levels of corruption. That may be infinitely improbable. What is not merely probable but demonstrable is that a new and poorly understood technology is having huge global economic effects while becoming increasingly desperate to generate cash - not to become profitable overall, but to demonstrate to investors that once the mad capex era is over, a sustainable and unassailable profit machine will remain. This is what matters most to keep the monster alive until it learns to feed. It has led to the tokenpocalypse and the reassessment by enterprises of what today's models are actually worth, and the realization among AI companies that revenue is not as elastic as they assumed. Which is why the world's most advanced and out-of-control technology is under threat from a caveman. Douglas Adams may have died 25 years ago, but he's still writing the script. (R)