I watched fortunes bloom and wither in real-time. The AI capex narrative is no longer about rocket ships to the moon; it's about whether the engines can sustain the burn rate.
Over the past 48 hours, the data points have converged into a single signal: AI spending is slowing. Not collapsing. Not reversing. But the slope is flattening. And in a market where the top 20 stocks in the S&P 500 now command 50.8% of total index weight—a concentration without modern precedent, per JPMorgan—any deceleration in the AI narrative is a seismic event.
Speed is survival, but empathy is the signal. The market is not panicking yet. But it's asking a question that every engineer asks before a refactor: "Is this architecture sustainable?"
Hook: The Data That Broke the Narrative
On July 25, 2025, the Bank for International Settlements dropped a warning that should have been a headline on every terminal: "The spending frenzy could turn into a long-term investment bust." The BIS rarely uses that language. They are not traders. They are central bankers. When they say "long-term investment bust," they mean the structural underpinnings of the AI trade are being questioned at the highest level.
Within the same week, the American Bankers Association's July fund manager survey showed 45% of respondents now list "AI bubble" as the single biggest tail risk—up from 28% the month prior. For context, that's a 60% month-over-month increase in fear. Secondary inflation? Forgotten. The AI bubble has overtaken it as the market's primary anxiety.
But the most telling signal came from the rapid implosion of a fund run by a former OpenAI researcher. This fund, which had grown to $45 billion by leveraging concentrated positions in AI infrastructure stocks, collapsed to roughly $10 billion in a matter of weeks. It was eventually taken over by Citadel. The details are murky, but the narrative is clear: even the "insiders" are getting crushed.
Context: The Architecture of the Bet
To understand why this matters, we need to step back and look at the architecture of the bet. The AI capex narrative is not just about Nvidia GPUs. It's a multi-trillion-dollar infrastructure buildout that spans chips, storage, power, cooling, networking, and data centers.
Goldman Sachs estimates that by the end of 2026, annualized AI-related spending could exceed $800 billion. Morgan Stanley's projection is even more aggressive: by 2028, AI infrastructure investment could approach $3 trillion, with over 80% of that spending yet to occur.
These are not small numbers. They are roughly equivalent to the entire annual GDP of countries like Canada or South Korea. And they are being deployed by just five companies: the hyperscalers.
But here's the structural tension that the market is only now beginning to price: the spending is front-loaded, but the revenue is back-loaded. The thesis is that AI will generate massive returns—but the timeline is uncertain. And when you are spending $1 trillion over two years, the market wants to see revenue, not just R&D.
Core: The Earnings Quality Problem
The core of the current anxiety is not about whether AI is real. It's about whether the earnings growth that the market is currently pricing is sustainable.
Mac10, a quantitative research firm, has a compelling argument. They point out that the record-breaking forward earnings growth in the S&P 500 is not driven by organic operational improvements. Instead, it's being artificially inflated by a massive, one-time event: companies pouring unprecedented cash into AI infrastructure. This spending flows through the income statement as a current expense, which depresses earnings today. But the market is looking at the "adjusted" earnings, which strip out these expenses. The result is a phantom earnings growth that is not sustainable.
Goldman Sachs notes that roughly 64% of S&P 500 companies are beating consensus earnings estimates by at least one standard deviation. That sounds great. But in a market where the top 5 hyperscalers are deploying over $1 trillion in 2025-2026, how much of that beat is real operational outperformance, and how much is just the side effect of the capex narrative?
BlackRock, the world's largest asset manager, has pushed back. They argue that the current AI leaders generate real profits, have strong balance sheets, and are funding most of their investments through internal cash flow. This is not the dot-com bubble, they say, where companies had no revenue.
They are not wrong. But they are missing the point. The question is not whether the hyperscalers can afford to spend. The question is whether the spending will generate returns that justify the current valuation multiples.
From my experience monitoring DeFi summer in 2020, I saw a parallel pattern. Projects were subsidizing their TVL with liquidity mining rewards. The numbers looked great until the incentives stopped. The underlying metric—user retention, fee generation, sustainable yield—was never validated. The hyperscalers are not subsidizing users. But they are subsidizing the infrastructure. The question is: what happens when the capex cycle turns?
Contrarian: The Slowdown Is Not a Crash—It's a Calibration
Here is the contrarian angle that the market is missing: the AI capex slowdown might not be a bad thing. It could be a sign of maturity.
If the slowdown is driven by improving model efficiency—meaning we are getting more compute per dollar—then the unit economics of AI actually improve. The narrative that "AI needs infinite compute" is a scaling law assumption that may be breaking down. We are seeing evidence that model performance gains are plateauing on certain benchmarks. If that's true, then the marginal ROI of additional compute is declining. And that means the market is rationally adjusting its expectations.
In 2022, during the bear market, I hosted weekly "Code & Coffee" sessions to help developers understand the macro conditions. The biggest lesson I learned was that the market often overcorrects. When the narrative shifts from "infinite growth" to "inevitable collapse," the truth is usually somewhere in the middle.
Currently, AI infrastructure spending is still growing. The absolute numbers are still breaking records. The slowdown is in the growth rate, not the absolute level. This is a classic pattern in technology investment cycles: the fastest growth occurs early, then the slope flattens as the base gets larger.
Furthermore, the decline in GPU prices and the increasing availability of inference-as-a-service could actually democratize AI development. Smaller players, who cannot afford to build their own 100,000-GPU clusters, will benefit from the oversupply created by the hyperscalers. This is the same pattern we saw in the dot-com era: the infrastructure oversupply of the late 1990s led to the commodity internet of the 2000s, which enabled the application layer to explode.
The AI bubble, if it bursts, might not be the end of AI. It might be the beginning of its true utility.
Takeaway: The Next Watch
The next signal to watch is not another surprise earnings beat from Nvidia. It's the hyperscalers' capex guidance for 2026. If even one of the top five cuts its AI infrastructure budget, the entire trade will reprice.
But the deeper question is not about the market. It's about the technology. Code was the law, and I was its restless guardian. If the efficiency gains are real, then the slowdown is a healthy correction. If the slowdown is driven by a lack of demand, then we are in a different story entirely.
The market is asking: "Is the architecture sustainable?"
My answer, based on the data, is: "Not yet. But it's being calibrated."
Stability isn't the absence of volatility. It's the ability to survive it.