A chief executive left, and roughly 13% came off NVIDIA (NVDA). Nothing about that departure changes a wafer already in fab, a driver already shipped, or a cluster already racked — which means the selloff is doing work the headline isn't explaining.
The more useful signal came from the other side of the Pacific. Analysts cut Baidu (BIDU) earnings forecasts sharply, now modeling an EPS decline despite revenue growth, citing softer AI monetization and weaker advertising conditions. Revenue up, earnings down, AI revenue arriving slower than the spend that enabled it. That is the shape of the problem the chip complex spent the week pricing.
The Economics
The relevant unit in AI infrastructure isn't dollars per chip. It's the amortized cost of serving a million tokens, and it falls through two channels: a faster accelerator generation, or better utilization of the generation you already own.
Those channels have very different owners. New silicon is a vendor event. Utilization is a customer event — batching, quantization, model distillation, scheduling. When the roadmap's cadence becomes less certain, operators shift weight to the second channel by extending the depreciation schedule on installed fleets. The arithmetic is unforgiving in both directions: stretch a fleet's book life from three years to four and annual depreciation drops about 25%, which flatters cloud gross margin immediately. It also means the operator buys fewer new units in the interim.
A leadership change does not stop a tape-out already in flight. What it plausibly widens is the distribution of outcomes around the generation after that — and accelerator purchase commitments are written against roadmap confidence, not against last quarter's benchmark.
Who Benefits
The consensus answer is AMD (AMD) and the custom-ASIC franchise at Broadcom (AVGO). That trade is well understood and largely in the price; every buy-side model already has a share-shift line item for hyperscaler in-house silicon.
The less obvious beneficiary is the inference-serving software layer — the unglamorous tooling that raises tokens per second per installed GPU. Its value is inversely correlated with hardware supply confidence. When the next generation feels guaranteed, operators wait for it. When it doesn't, squeezing 30% more throughput out of an existing cluster becomes the cheapest capacity expansion available, and it lands in gross margin rather than capex.
Baidu's revisions are the cleaner data point of the week: revenue growing while earnings forecasts fall is what happens when AI capacity is bought faster than AI products are sold.
Note what Baidu's case does not support. One company's advertising softness is not a verdict on enterprise AI demand. It is one observation from a market with its own regulatory and competitive conditions, and it should be weighted as such.
The Second-Order Effect
Two things become possible at this point that weren't a month ago.
The first is political. Reporting suggests the leadership change gives Washington's China hawks an opening to push tighter technology controls, with the executive who personally argued against restricted-market restrictions no longer in the room. Export policy sets the addressable market for the highest-margin accelerators. If the stack bifurcates into two incompatible software ecosystems, the developer-portability moat that underwrites premium pricing gets tested in a way no benchmark result can answer.
The second is distributional, and it is quieter. Indian investors holding US-tracking ETFs absorbed the drawdown through vehicles trading at steep premiums to underlying value — structures where the local price can fall even when the US holdings don't. AI compute exposure has propagated into retail access products in markets where the plumbing introduces its own basis risk. That's a new transmission channel for a sector that until recently traded mostly between professionals.
Against all of this, Goldman Sachs still expects the global economy to grow 2.8% in 2027, with sustained AI investment named as a support. Both things can hold: aggregate AI capex continues, while the market re-rates who captures the margin on it.
The Leading Indicator
My recent calls on this complex have been wrong in a consistent direction. I leaned on the buildout — a Nasdaq Composite break below 25,600 on financing stress that never came, data center names that didn't clear my upside levels. The error was treating capex strength as a price forecast. It isn't; it's a volume forecast, and volume and margin have decoupled.
So the metric to track is the one that measures capture, not spend: AI-attributed revenue growth at the cloud and model layer relative to the prior quarter. Baidu's next disclosure is the near-term test. If AI revenue growth decelerates while infrastructure spending holds, the repricing that started this week has further to run.
On the stock, the thesis lives at two levels. If NVDA closes below $150.00 within 15 sessions, the market is pricing a monetization problem rather than a governance one, and the drawdown is structural. A close above $168.00 in that window refutes it — that would say the selloff was a personnel discount that the earnings line erases, and I would be wrong about the mechanism.
What changes my mind faster than price: a cloud or model-layer disclosure showing AI revenue growth accelerating alongside capex. That single data point would move the entire argument from margin compression back to a volume story.
Watch AI-attributed revenue growth in the next quarterly cycle. One number, and it settles this.