The Wall Street Journal's take is that panic selling in Nvidia, Broadcom, and other AI infrastructure names following DeepSeek's efficiency claims is disproportionate to the actual fundamental risk. The core bear argument — that cheaper, more efficient AI models reduce the need for massive GPU clusters — is real but not necessarily disqualifying for the hardware layer.

Nvidia reported $215.9B in revenue, up 65.5% year-over-year, with 71.1% gross margins and $4.90 diluted EPS. Broadcom posted $63.9B in revenue, up 23.9% YoY, with 67.8% gross margins. These are not the financials of companies whose moats are visibly crumbling.

The bull case rests on Jevons Paradox logic: cheaper inference tends to expand the total addressable market for AI compute rather than shrink it. If DeepSeek-style efficiency makes AI cheaper to run, enterprises will run more of it — and still need Nvidia and Broadcom silicon to do so. This is the thesis the WSJ is implicitly endorsing.

The bear case is more structural: if frontier model training itself becomes less compute-hungry, the hyperscalers that drive Nvidia's data center revenue could moderate capex plans. That's a slower-burn risk but a real one, and the market is right to at least price some probability of it.

What to watch: hyperscaler capex guidance in upcoming earnings (Microsoft, Google, Meta, Amazon), any commentary on H100/B200 order books, and whether Nvidia's next print maintains trajectory. The WSJ piece is a contrarian signal, but the resolution lives in the data, not the op-ed.