Steve Hou, head of research at Silicon Data and a veteran quantitative investment strategist, says the fear of fast GPU depreciation has been "quite overblown". He argues that a six-year-old Nvidia A100 is still holding "pretty significant rental rates".
On Full Signal with Phil Rosen, in his fourth appearance on the show, Hou discussed the spread between token prices and GPU costs, what investors are pricing into AI stocks, and why he calls the bubble label a "tiresome, oversimplistic characterization."
The squeeze between token prices and GPU costs
Rosen said GPU rental prices are "right near a record high" and have been rising all year, even as token prices look like they have "fallen off a cliff since about May this year."
Hou said Silicon Data tracks GPU rental and token pricing as AI fundamentals. The figure to watch, he said, is the spread between what tokens sell for and the cost of renting a GPU, which he likened to the squeeze between CPI and PPI.
He said the token-price chart is a usage-weighted index, so a shift toward cheaper flash models would deflate the effective price of a million tokens. If price falls by half and demand rises 100 times, a lab still makes more money, a dynamic he described as Jevons paradox.
What the market is pricing in
Hou said worries about the labs' economics helped drive a local peak in AI-related stocks. He said the S&P 500 was only a couple of percentage points from the highs, while a median semiconductor name was maybe off 50% before the most recent rebound.
In his view, stocks are pricing the second derivative of growth, not a halt in the AI buildout. Spending is still growing, he said, but prices may be anticipating that it will not keep accelerating from 2028 onward.
Asked whether valuations look attractive, Hou declined to make a recommendation. He said fundamentals have improved a little, pointing to Anthropic's S-1, which he said showed good margins, a low burn rate, and a cash-flow-positive quarter.
Why depreciation fears look overblown
Hou said depreciation cuts against hyperscaler earnings. Using on-demand rental rates and market-implied long-term rental curves, he put a six-year-old A100, first announced in May 2020, at a 22% fair value, a chip he said should already be dead under three-year straight-line depreciation.
He put the H100 at about 50% to 60% of its original sale price, and estimated the B200's fair value at about 180%, "well over 100%". That, he said, is good news for Nvidia and its customers, because the chips can stay productive for a long time.
Why he calls the AI bubble label oversimplified
Rosen said nothing in Hou's research was necessarily flashing a bubble signal. Hou called the bubble framing a "tiresome, oversimplistic characterization." The usual fear, he said, is an overbuild of unused compute, like dark fiber in the dot-com era.
Hou said "that's certainly not the case" today, because "all the GPUs are being used." On-demand rental rates, he said, have been "going vertical" in recent weeks, "like the GPUs are melting." The question he takes more seriously is whether AI demand will diffuse beyond the AI industry itself.
Tune in to the full conversation with Steve Hou on YouTube, Spotify and Apple Podcasts.
For more on the AI bubble debate, read Sparkline Capital's Kai Wu on the AI Bubble, the K-Shaped Market and Disruption, and subscribe to the free Opening Bell Daily newsletter for a daily briefing on stocks, macro and Wall Street.
This article is for informational purposes only and is not investment advice. Views and stock opinions are those of Steve Hou and market figures are as stated by the guest at the time of recording.
