Why Compute Prices Are Still Rising | Steve Hou of Silicon Data on GPU Rental Indices, Fed Meeting, and Anthropic S-1 (LIVE)

Anthropic

Across the shows: 68 speakers in other episodes, 63 bullish and 55 bearish calls, past 30 days

• Steve Hou · Silicon DataBullish

  Anthropic's enterprise adoption is extraordinary, and its Opus model is very popular, despite a reported Q3 growth slowdown.

  Steve cites the S-1 as showing extraordinary enterprise adoption of Anthropic models and says Opus 5.5 is very popular. He acknowledges that Anthropic's investor day revealed a slowdown in Q3 revenue growth. He points to Ramp data that he reads as a reacceleration, driven by Opus's popularity. He says it remains to be seen whether growth inflects back up.

  “the S1 is showing extraordinary amount of enterprise adoption of Anthropic models... Opus 5.5 is very popular.”at 7:04

• Steve Hou · Silicon DataBullish

  Anthropic's enterprise adoption and near-total retention make its customers very sticky.

  Hou cites Anthropic's very strong enterprise adoption and what he describes as 100% retention. He says that once a company is onboarded and the product is embedded inside it, customers can keep expanding usage, which makes the business extremely sticky. He notes his own experience: Claude's early problems were fixed, but he still feels a pull to stay with Anthropic. He presents this as evidence of durable enterprise position, while stressing that future growth pace is a separate question.

  “they have very strong enterprise adoption, and they have very strong retention, right? 100% retention... once you are in the enterprise... that's extremely sticky.”at 47:00

• Steve Hou · Silicon DataBullish

  Threat from cheap open models to Anthropic is overstated in the short run; its own cheaper models hold customers.

  Hou argues that competition from cheap open models over cost is probably overstated in the short run. Anthropic has its own cheaper models. Enterprises would have to adopt a post-trained Chinese model and work through compliance and safety issues, which is much harder than picking a cheaper model from the same vendor's dropdown. He thinks the bigger risk to Anthropic is internal, namely that OpenAI's model is so good that users never consider Anthropic's alternatives.

  “this idea of they're being threatened by open models, over cost, I think is probably in the short run overstated... It's just much easier to just click over to a slightly cheaper model on the dropdown menu”at 47:23

CoreWeave

Across the shows: 54 speakers in other episodes, 14 bullish and 53 bearish calls, past 30 days

• Steve Hou · Silicon DataBearish

  CoreWeave lags despite high rental rates because the market doubts that rising compute rental prices will be sustained.

  Asked why neoclouds like CoreWeave have underperformed despite high rental rates, Steve gives the same explanation as for memory and storage stocks. The market does not believe the rise in compute rental prices will last. Investors are pricing in the eventual end of these earnings. He says Silicon Data is launching compute futures so people can express this view and so suppliers can hedge the risk of compute prices falling. He reads the market's stance as broad skepticism.

  “And that's clearly what the market is saying, and they don't believe whatever that's happening in terms of the compute rental prices to the extent it's going up is going to, you know, be sustained.”at 23:52

SK Hynix

Across the shows: 2 speakers in other episodes, 2 bullish and 0 bearish calls, past 30 days

• Steve Hou · Silicon DataUncertain

  SK Hynix's low earnings multiple despite huge profits signals market skepticism that memory earnings can last.

  Steve cites SK Hynix's recent quarterly earnings, which he describes as enormous. He notes the company's earnings are more than those of the hyperscalers combined. Despite this, its earnings multiple is lower. He reads that as clear skepticism that such earnings can be sustained for years. The same pattern applies to memory and storage stocks, whose P/E ratios have fallen for a long time.

  “if you look at [SK Hynix's] earnings multiple, you know, it's lower. So clearly there is a skepticism that this could be sustained years out.”at 25:03

Marvell Technology

• Max Wiethe · Monetary Matters NetworkUncertain

  Marvell paying top dollar for Lumentum signals strong conviction in optics, though it could be a mistake.

  Max notes Lumentum has not fallen sharply from its June high, so a deal would be at close to the stock's trading level, and it is therefore top dollar. Given the market caps, he thinks the deal probably includes a cash component. He reads the price as a sign of conviction that optics will matter, while accepting it could be a huge mistake and that there is no guarantee an acquisition works.

  “Somebody has having conviction in the future being optics... It could be a huge mistake... It shows a lot of confidence to pay top dollar.”at 1:01:59

AI infrastructure & data centers

• Steve Hou · Silicon Data

  Data center operations are profitable today, but continuous new investment keeps cash flow negative, so the sector borrows.

  Steve argues that existing GPU and data center activity has positive margins, so the businesses are profitable on current operations. He says the picture is a rolling one, because operators keep making new capacity investments. At any snapshot, cash flow is therefore negative and the buildout is financed by borrowing. He frames the key question as how to assess that combination on a going basis.

  “On the one hand, the existing activity, the margins are all positive. On the other hand, you're making continuous new investments in new capacity... your cash flow is actually not positive, right? You're having to borrow to keep investing.”at 2:31

• Steve Hou · Silicon Data

  GPU rental rates are surging across chip generations, with the most advanced chips rising fastest, showing shortage and accelerating demand.

  Silicon Data's rental price series for H100, B200 and B300 are all rising, and in recent weeks the lines have gone almost vertical. The most advanced chips show the largest increases, which Steve links to them producing the most tokens. Across the chip range, on-demand rental prices are rising, which he reads as a shortage. He sees no sign of a demand slowdown and says recent charts show acceleration. He acknowledges the open question of whether demand comes from diverse enterprises or from AI-native and AI-adjacent companies, possibly in a circular loop, but says the observable data show strong demand.

  “They are all surging. And in most recent weeks, you know, they've been actually almost going vertical. And there's almost a sense in which the most advanced chips are seeing the greatest amount of rental rate increase.”at 5:13

• Steve Hou · Silicon Data

  Physical constraints on finding power and land, not lack of demand, are the main limit on data center buildout.

  When asked why executives don't just keep building, Steve says they are building more, but the constraint is physical. Operators need to secure power and land, and visibility on these is poor even over the next few months. Long-range capex projections are therefore uncertain, and the scale of the spending is becoming a fraction of GDP. He says even the most powerful executives cannot know these answers well.

  “right now the constraint is on the physical side of actually, can you actually find that power and land and whatnot?”at 14:01

• Steve Hou · Silicon Data

  Component inflation in memory and other parts means data center capex must rise in nominal terms just to build the same capacity.

  Steve agrees with Jack that component costs are rising sharply, including memory and other parts beyond chips. He argues that building the same amount of data center capacity now requires more spending in nominal terms. Capex growth therefore partly reflects cost inflation, not only more capacity. He ties this back to the question of whether real returns will show up.

  “The underlying components are seeing extraordinary inflation, you know, it's not just chips, but memory and other aspects... You've got to increase CapEx in nominal terms to keep up.”at 15:37

• Steve Hou · Silicon Data

  Data center credit spreads are widening because investors doubt AI investments will pay off, not because of higher rates.

  Steve argues that the bond market never priced in AI-driven productivity, let alone AGI-driven productivity. As investment sizes grow, doubts about whether the spending will pay off have pushed up spreads on AI-related debt. He says this is not because of the level of rates or a change in the underlying macro backdrop. He concludes that AI itself is the driver of credit spread moves.

  “the bond market never believed even once that AI was going to actually produce, you know, miraculous productivity.”at 22:12

• Steve Hou · Silicon Data

  The compute rental forward curve has risen at every maturity and turned upward-sloping, signalling expected higher future rental rates.

  Silicon Data's term curve of implied GPU rental rates has risen across all maturities since October 2025, and the curve now slopes upward, especially at longer tenors. Steve attributes this to cloud providers' growing unwillingness to lock into very long contracts. They want flexibility to raise prices as spot and short-term contracts roll. He compares it to airlines' jet fuel hedging trade-offs and notes that a storable commodity like oil is rarely in contango, while compute cannot be stored. He reads the upward slope as the market expecting rising future rental rates. He stops short of a definitive statement that the market is in contango.

  “the entire forward curve is risen in level, right? So every term length, every maturity forward has actually gone up. And in fact, the shape of the curve has also become, you know, upward sloping.”at 30:53

• Steve Hou · Silicon Data

  GPU residual values sit well above zero, far above accelerated depreciation schedules, even for older chips like A100.

  Silicon Data estimates going-concern fair value and residual value for the chips, assuming a maximum eight-year life. It discounts cash flows as observed, using market demand and implied rental rates, and applies a conservative decay after that. For the three chip generations shown, the residual values sit well above what an accelerated depreciation schedule would imply. The A100 still has value six years after mass availability. Steve concedes depreciation is happening, but argues new chips are not making older ones obsolete overnight, and rental rates for H100 have not fallen secularly despite new releases.

  “this is actually all lying well above what would have been implied, right, by such depreciation schedule.”at 35:05

AI pricing & unit economics

• Steve Hou · Silicon Data

  The falling token price index reflects cheaper models spreading as usage rises, not a deflationary race to the bottom.

  The usage-weighted token index has trended down since late May. Steve argues the bears read the chart as deflation and commoditization, but the same data fit healthy AI proliferation. Within the index, the proprietary and open-model sub-indices have not fallen. The overall decline comes from substitution away from expensive proprietary models toward cheaper open models. Total token usage has risen sharply, and this is consistent with people using cheaper models for more tasks.

  “You can tell a bearish story. I can also tell you that this is exactly what AI proliferation should look like, right? And prices of tokens coming down and people using the cheaper models to do a lot more stuff.”at 4:33

• Steve Hou · Silicon Data

  Token serving margins are clearly positive, and newer chips widen them; durability of margins is the open question.

  Hou says the chart being cited does not show token production costs above selling prices, since it uses dual axes with normalised figures. He argues newer chips are much better at producing tokens, so tokens are selling well above cost. He then says the real question is whether those margins persist. If the people earning large margins from running or owning GPUs are joined by more entrants, competition could compress the spread to zero, unless frontier models and existing operators have a durable moat. He calls that the trillion-dollar question.

  “newer chips are so much better at producing tokens that you're certainly producing, you know, those tokens are certainly selling well above cost.”at 41:08

Enterprise AI adoption & AI software

• Steve Hou · Silicon Data

  Companies actively adopting AI show faster growth and stronger profitability, per Ramp data, though the pace of payoff is uncertain.

  Steve says the macro-level ROI from AI would show up as higher productivity, profitability and revenue growth. He says some evidence of this exists. Ramp data show companies actively adopting AI are growing faster and are more profitable. He concedes that these effects may take a while to play out, and the speed of the payoff is the open question.

  “we do see some signs of that, right? Ramp has seen, has shown data, for example, on how companies that are actually actively adopting AI are also seeing faster growth and, you know, stronger profitability.”at 16:31

US rates & the Fed

• Steve Hou · Silicon Data

  Small rate moves won't halt AI buildout unless the Fed hikes to fight AI-driven inflation, which would slow it.

  Steve argues that if AI investments produce the returns people project, in a stable way, small moves in the discount rate will not deter investors. He says the scenario that matters is one where AI is strong enough to drive aggregate demand and inflation, so that it becomes an enemy of the Fed's dual mandate. In that case the Fed would raise rates until the economy slows, including the AI economy. That would change investment plans and hit the buildout. He also says the interest rate debate ties back to whether AI investments will generate sustainable profits.

  “I actually don't really think the interest rate itself is such going to stop anyone here per se. ... You're only in a case in which if the Fed really comes to see that the AI economy ... is creating inflation that it becomes directly an enemy”at 20:03

Frontier AI labs & model race

• Steve Hou · Silicon Data

  Agents make models interchangeable, so the moat may lie in accumulated user context rather than the model.

  Hou says agents make users almost model-agnostic. He uses agents without knowing which model powers them and cares only about whether the outcome is delivered. He still sees some moat, because over time an AI knows the user better, and the more it has learned, the harder it is to leave. He is explicit that he is not making a bullish or bearish case, and says the moat is an open question.

  “the funny thing about agents is that agents actually are making you almost model agnostic... All I care about is whether or not it's actually delivering an outcome.”at 53:46

AI chips & semiconductors

• Steve Hou · Silicon Data

  Price rises in memory and logic chips with strong pricing power will hit AI capex more than small components.

  Hou argues that what matters for a supplier's effect on costs is how large it is as a share of the bill of materials. Memory and logic chips are big items, so when their prices rise with strong pricing power, they move capex much more than a small component whose price rises because of a shortage. He sees the consolidation in optics as notable but says the bill-of-materials share is what matters.

  “how big you are as a share of the overall BOM... memory and logic chips are obviously big deals, and when they go up in price with very extremely strong pricing power, that's going to affect the CapEx much more”at 1:00:59