A Little Kyoto in AI

A debate that seems to mesmerize Wall Street revolves around the ROI of AI and whether we are in a bubble that will end badly like the bubbles before it. Jane Street took a $15 billion hit in July, driven partly by its stake in a former OpenAI researcher’s hedge fund, and there have been critical pieces on AI’s circular financing. This has reinforced the bubble view. But I was exploring Kyoto and its ancient temples this past week. And I found that embedded in its history were parallels and lessons well suited to providing a framework to think about AI and whether it’s a bubble. The first parallel is that it’s difficult to predict the intersection between supply and demand. When Kyoto was founded in 794, it was laid out as a perfectly symmetrical grid modelled on the Tang Dynasty split into east and west halves. Each was to have an equivalent population. However, the Emperor underestimated the infrastructure issues in the west, which sat on low, poorly drained land. It never attracted settlement at scale and reverted to fields and marsh within a couple of centuries, while the east grew fast. The same argument can be made for AI. The market was excited by Claude Code’s commit trajectory, Accenture’s training 30,000 employees on Claude and Fractal’s 112% net revenue retention amongst other supporting evidence for AI’s demand. However, DeepSeek and Chinese open-source models soaked up the token volume because they are cheaper. There are $500 billion of third-party SPVs structured by Nvidia via MOUs, CoreWeave’s $8.5 billion non-recourse facility, and Apollo’s xAI vehicle all suggest that AI companies can’t fund their capex from operating cash flows or plain vanilla credit. DeepSeek raised its prices as it had to price in its cost to serve. After centuries, eventually the west part of Kyoto ceased to be marshland, but only after the returns made sense. Another lesson from a Kyoto template illustrates continuity without sameness. Kiyomizu-dera was burned down more than ten times and rebuilt. But it was not rebuilt with the same materials. The design intent, the proportions, and the ritual function were faithfully re-established. However, the substrate was different. The same framework applies to AI. The models are being retrained, updated or run on different hardware with AI companies trying to make their own chips, approaches being tried to reduce memory demand and other hardware innovations. It seems unlikely that semiconductor companies will continue to enjoy such high margins. But the models retain their character and continue to evolve even though the underlying substrates are in flux. Then there is domain knowledge, where AI companies are paying experts for their knowledge to train their models. Responses are graded, and reinforcement learning flows into the model. But tacit knowledge resists codification. The iemoto system in Kyoto’s tea ceremonies, the families who make Buddhist altar fittings, the model master who assembles realistic shinkansen trains, all pass down judgments that were not written down as rules. It’s a reminder that expertise cannot be reduced to an explicit rulebook, no matter how detailed. The people who say robots will automate most tasks are underestimating the amount of data that will continue to be missing. AI engineering systems are becoming more complex as well. They are adding capabilities, options and output that may be overkill for many of the tasks. The dry gardens at Ryoan-ji, with its fifteen stones, raked gravel, and not much else, illustrates the concept of meaningful empty space. Restraint can be a design choice, and not a limitation, something that Steve Jobs at Apple took to heart.
Finally, Kyoto was spared bombing in World War II because of its cultural and historical significance. Secretary of War Henry Stimson wanted to preserve its ancient temples, even though the analytical guidance of his military planners thought it was a strong target for America’s first nuclear bomb. A human, contingent call about what was worth preserving mattered. As we move into autonomous weapon systems, or AI customer service, this ability for humans to make contingent calls will go away. But it will eventually lead to a backlash since there are institutions and artifacts that survive because a human decides they are worth the costs. None of these settles whether AI is a bubble; it reframes the question. Kyoto's west side wasn't a planning failure so much as a timing mismatch between infrastructure built to a blueprint and demand that showed up on its own schedule, a couple of centuries later. The AI capex build has a similar feel. The open question isn't whether the demand eventually arrives (DeepSeek's repricing suggests it already has, at the margin), but whether the SPV-financed layer of the stack can survive the wait without being forced back into the ground first. End Note Walking around old streets and modern department stores in Kyoto, Osaka, Busan and Seoul, people still like to shop, meet people and explore new ideas. The experience is curated. China is much the same. People still value interconnections and social spaces. Everything runs on time, leading to an observation that the more advanced a country is in AI, the more that time is wasted waiting for a train. Asia does use AI extensively on the back end, from behind-the-scenes logistics, product design, charge points, taxi efficiency and restaurant simplification, it’s very much an ‘integrate with humans’ vibe rather than let’s try to replace humans. Maybe this is why no one really talks about AI, even though it’s in the background. This brief is coming from an airport lounge on holiday. I’ll go back to more market-based analysis next week.