The DeepSeek Effect

What Happened

In January 2025 DeepSeek released its R1 reasoning model, reaching capability comparable to top closed-source systems at a far lower training cost. Nvidia's stock fell sharply that day. It became the most-cited single example in the AI-bubble argument.

The Market's Answer Came Later, and Elsewhere

Reading that one day gets the direction wrong.

Nvidia's market capitalization stood at $3.29T at the end of 2024. Over the nine months after R1 it did not stall — on 2025-10-29 it became the first company ever to close above $5T. The market took a year to answer the question of whether efficiency gains would cut compute demand, and the answer was: no, and here is more money.

The cooling came afterwards and from elsewhere: $4.53T at the end of 2025, $4.73T at the end of July 2026. The real drawdown happened nowhere near the news blamed for it, and for unrelated reasons.

This is the point the site keeps making. A single point is noise; the trend is the evidence. An argument pivoting on how much something fell on one particular day rarely survives a longer time axis.

Neither Reading Has Been Falsified

• The bubble reading: if equivalent capability can be had for less compute, capex priced on the assumption that compute is capability loses its footing.
• Jevons' paradox: falling unit costs enlarge total demand, and whatever compute efficiency frees up is immediately consumed by more calls.

Both find support in the post-2025 data, because they answer the same question over different horizons. In the short run, falling costs did compress pricing power per unit of compute; over a longer window, growth in call volume pushed total demand back up. Nothing in the data has ruled either out.

Where the Efficiency Route Stands Now

The interesting signal is the direction of pricing: in August 2026 DeepSeek announced a substantial increase in its API prices. If low cost was a stage-appropriate pricing strategy rather than a structural cost advantage, its long-run implication for compute demand needs re-estimating.

Meanwhile China's efficiency route is attracting American-style valuations — Moonshot AI's valuation climbed to $50B within a month. Whether a technical path born of constraint returns to capital intensity once it succeeds is the question this topic follows from here.

How to Use This Page

The Nvidia series below is the only verified series on this site that spans both sides of R1, so you can see for yourself where that day sits in the whole curve. The signal list is filtered live from the feed under the models and chips categories.