The ECB Says the AI Rally Will Correct. Inside the Index, It Already Has

Researchers at the European Central Bank published a blog post Monday arguing that a correction in AI-related equities should be expected even if current valuations are rational, on the grounds that wide-ranging adoption spreads concentrated risk across the broader economy and that fiscal and monetary buffers are thinner than they were in prior cycles. The consensus reading of that warning is straightforward and, in my view, wrong: it treats the AI complex as a single asset that will re-rate together, in one direction, on one catalyst.

The most recent session argues otherwise. Advanced Micro Devices closed at $514.40, up 6.50%, essentially at the top of its session range, while NVIDIA finished at $225.18, down 0.05% — a spread of roughly 655 basis points between two names that the consensus files under the same thematic heading. Seven of the ten largest US names closed lower. The correction the ECB describes is not waiting for a catalyst; it is arriving as internal dispersion, redistributing capital within the theme rather than out of it.

The Regime Question

The regime that carried indices to records was one in which benign inflation and receding hike odds lifted long-duration equity as a category, and the AI trade functioned as the highest-beta expression of that single factor. That regime is now visibly fraying, because a discount-rate impulse that lifts everything cannot simultaneously produce a 655-basis-point intraday spread between two semiconductor businesses with overlapping end markets. What replaces it is a regime in which the market underwrites specific order books — Anthropic's second-quarter revenue reportedly up fourteen-fold year over year, NVIDIA's Vera Rubin platform pulling silicon-carbide and power-chip suppliers along with it — while assigning progressively less premium to thematic proximity. The ECB researchers are describing a system-level risk; the equity market is already pricing an issuer-level one, and those are not the same trade.

The Evidence

Start with the leadership itself. AMD gapped higher and then spent the entire session extending, closing within twenty-one cents of its high and 12.03% below its 52-week peak of $584.73 — a name with meaningful headroom absorbing the day's entire appetite for AI risk. NVIDIA, by contrast, traded a range of barely three dollars and sits 4.80% below its 52-week high of $236.54, which is a remarkably narrow band for the most heavily traded name in the complex. I owe readers a correction here: I called NVIDIA below $215 in one window and above $236.54 in another, and both expired unfilled, because I kept treating it as a directional instrument when it has functionally become the index's ballast — the position everyone holds and nobody presses.

Meta closed 25.93% below its 52-week high of $796.25 — the widest drawdown from peak among the ten largest US names, and the clearest evidence that AI capex is no longer being rewarded uniformly.

The second piece of evidence sits in the capex-heavy platforms. Meta Platforms opened higher, traded to $601.80, and then bled all session to close at $589.76 in the bottom few percent of its range; Microsoft printed exactly $500.00 intraday and surrendered the handle, closing at $495.35. Both are spending enormously on the same infrastructure buildout that lifted AMD, and both were sold. When the market pays a rising multiple to the supplier of compute and a falling one to the buyer of compute, it is telling you that it no longer believes the return on that spending accrues evenly.

The third strand is geopolitical and it is not decorative. The sixty-day window Washington and Tehran set for a final agreement expired Monday with the two sides further apart, and copper has held above $14,000 a ton in London for nine consecutive sessions with steep backwardation. Energy and industrial input costs feed directly into the operating expense line of the data-center buildout, which is precisely the mechanism by which a thematic rally becomes a margin story — and margin stories are always dispersive.

Historical Precedent

The closest rhyme is the semiconductor complex in 1999 and early 2000, when the aggregate index masked an increasingly violent internal sorting: names with visible order books and pricing power extended for months after the names with only thematic adjacency had rolled over. The lesson institutional allocators took from that period was not that concentration is fatal, but that concentration ends with the leadership list shortening rather than the index breaking on a single day. Applied here, the ECB's warning is directionally correct and temporally early; the adjustment expresses itself first as fewer names carrying the same weight, and only later as the index-level event that headlines describe.

The Position

The weight of evidence favors dispersion over drawdown, and allocation should follow the order book rather than the theme — which means underwriting the compute supply chain while trimming exposure to the platforms funding it. AMD at $483.27, the low of the session that produced its 6.50% advance, is where this argument lives or dies: holding above that level over the next ten sessions confirms that capital is rotating within the AI complex, while a close beneath it would tell me the ECB researchers described a broad de-rating that I mistook for a rotation, and I would reverse. My prior AMD framing from last week's session remains live and unresolved. Given a recent run of misses, I am sizing the conviction accordingly — but I am not hedging the direction: the AI trade narrows from here, and the narrowing favors silicon over software.