Can the AI Boom Survive a Margin Squeeze?

Anthropic

Across the shows: 43 speakers in other episodes, 24 bullish and 36 bearish calls, past 30 days

• Steve Hou · Silicon DataBearish

  Slowing Anthropic growth and cheaper models may put its high IPO valuation in question.

  Anthropic went from little revenue to about $60bn annualised run-rate very quickly, but by some measures its revenue growth has flattened or slowed since the token-price peak. Structural competitive pressure from cheaper models could keep growth below what people expected. Steve says this will create questions around the company's upcoming IPO and its relatively high private valuation, which he puts at roughly $1.5-2tn for Anthropic and OpenAI. He hedges that the slowdown is only tentative evidence.

  “[Anthropic] has gone in almost no time at all from having very little revenue to $60,000,000,000 of annualized revenue run rate... There are signs that the growth rate is indeed slowing... and that will put into question some of the valuations”at 13:29

• Steve Hou · Silicon DataNeutral

  Anthropic's share pledge for AI-affected people looks like a behavioral, indulgence-style move.

  Steve finds it ironic that Anthropic has long voiced worry about the dangers AI poses to society while being at the forefront of inventing the technology. He suggests it is almost as though it has a product and wants to show its danger. He also sees a more pragmatic logic: other firms are developing this technology regardless, so Anthropic wants to help society deal with the consequences. He says there are few details on how the 5% share program will work and cannot judge what it will do. He compares it to buying indulgences.

  “So it's almost as though like, oh, we have this product. Look, look, you know, look at the danger it's actually, you know, posing.”at 44:40

OpenAI

Across the shows: 21 speakers in other episodes, 10 bullish and 16 bearish calls, past 30 days

• Steve Hou · Silicon DataBearish

  OpenAI's upcoming IPO and relatively high valuation face questions from cheaper-model pressure.

  Steve argues that the shift toward cheaper, good-enough models and slowing growth at the frontier labs will create questions around the upcoming IPOs of Anthropic and OpenAI. He says the valuations, roughly $1.5-2tn, look relatively high for non-public companies whose growth may not match earlier expectations. This is a hedged view, since he says he suspects the pressure will surface and the economics of who captures rent from frontier models remain open.

  “I suspect this is going to create some questions surrounding their upcoming IPOs and... Relatively high valuations.”at 12:53

AI bubble & tech valuations

• Steve Hou · Silicon Data

  The AI market topped mid-year; ROI doubts are growing, and the cycle is not early.

  Steve says every capex cycle goes through boom and bust, so the question is timing. He argues the market already topped in the middle of the year. Questions about ROI, how the investment produces a return, and whether frontier labs can justify valuations that rest on share sales to fund compute commitments are getting louder. If good answers are not forthcoming, the whole cyclical narrative could be challenged. He concludes the group is not early in this trade.

  “The question is a matter of timeline... The market has already topped in the local fashion in the middle of the year, right? ... these questions are growing louder, and we're not early in this trade.”at 37:13

• Steve Hou · Silicon Data

  Market is in a feverish phase with smart money already de-risking; an AI air pocket is real.

  Steve argues the market is probably in a pretty feverish moment, with some smarter money already having taken its chips off the table or having them taken off for it. He notes that a bubble's end does not come in a straight line, so violent counter-rallies are possible even if this is the top. He says the risk of an air pocket is very much real and that people are already living through it in different forms. He also says a melt-up is possible if things go right, so the bear and bull scenarios are not mutually exclusive, and he admits he cannot time them.

  “we are probably in a pretty feverish, you know, moment of the market that some of the smarter money has already either taken chips off the table or having their chips taken off the table for them.”at 38:58

AI pricing & unit economics

• Steve Hou · Silicon Data

  Token prices falling while GPU rental rates rise, squeezing AI model economics.

  Silicon Data's token index, the expenditure-weighted effective price paid for models, has fallen since about the end of May to roughly half its mid-year level, while GPU rental rates have risen rapidly and continuously. Usage is shifting to cheaper models that handle everyday tasks, so volume grows while per-token prices fall. Steve also questions how AI will be monetised: metered usage is unproven, Meta plans free or low-cost access, and distilled cheaper models may let others capture the value that frontier labs created.

  “since about the end of May, the token index or the expenditure weighted effective price paid for the models has been coming down pretty significantly, almost, you know, half of where it was back in the middle of the year. In the meantime, you know, rental rates for these chips have been going up, you know, rapidly continuously.”at 9:36

AI chips & semiconductors

• Steve Hou · Silicon Data

  Memory and storage demand has gone exponential from AI conversations, driving memory stock gains.

  Steve links the explosive rise in memory and storage stocks (SanDisk, Micron, SK Hynix, Samsung) to AI's need for memory. Models must keep track of long conversational context, and each conversation generates far more data that must be stored. He argues that this has made demand for memory and storage grow exponentially, which explains why Korea has drawn intense global attention.

  “part of the reason why we have seen such an explosive growth in the stock prices of memory and storage stocks such as SanDisk or Micron... is exactly that, right? We suddenly have a exercise where the demand for memory and demand for storage of data has just gone exponential.”at 22:32

AI infrastructure & data centers

• Steve Hou · Silicon Data

  Data center supply lags AI demand because of power and site constraints; the shortage is forecast.

  Steve argues that agentic AI, which runs continuously, will drive much more compute demand, and that demand multiplied across developed economies is not matched by supply growth. He says there is not enough physical power and not enough sites for large data centres, so builders are forced into smaller facilities in more corners of the world, facing energy shortages and local political pushback. He calls it a forecast shortage rather than a current one, but a reasonable forecast given recent trends.

  “To that extent, yes, there is a shortage in data centres... Nevertheless, I think it's a reasonable forecast, given the recent trend.”at 25:01

AI: overall outlook

• Steve Hou · Silicon Data

  AI is a fundamentally transformative technology; the bear view that it is vaporware is wrong.

  Steve rejects the bearish position that AI is vaporware, a con job, and that nobody wants it, calling that fundamentally untrue. He says AI is a fundamentally transformative technology that will reshape society. He gives a mechanism for it: products with good product-market fit can drive accelerating adoption, because people will pay for things that give them utility. He expects bullish believers to be proven right eventually, while acknowledging the timing and adoption risks he raised earlier.

  “But I do think this is a fundamentally transformative technology. I do think it's going to reshape society.”at 40:03

Enterprise AI adoption & AI software

• Steve Hou · Silicon Data

  AI demand may arrive slower than bulls expect, given bureaucratic hurdles, risking an air pocket.

  Steve's main concern is timing rather than whether demand exists. Bullish San Francisco voices extrapolate from coding, where programmers adopt AI quickly, to the whole economy. He argues they ignore bureaucratic hurdles to enterprise adoption that could slow things down. If supply arrives in 2027 and 2028 while enterprise adoption is slower, and consumer adoption also fails to find the product-market fit that the iPhone did, demand could fall short. He says this risk of an air pocket is very much real.

  “rather the demand could show up slower than some of the most bullish people in SF seem to think. They seem to think that because it works for coding, all the programmers adopt AI rapidly overnight”at 41:05