Two chart updates in a two part 🧵 The LLM token index has not put in a bottom as we had previously speculated and has resumed its decline even as on-demand GPU rental indices (a signal of marginal demand and not where most of compute cost sits) continue to surge towards ATHs.

Everyone’s chasing the new vintage. Meanwhile, H100 rental rates rose 10.8% over the month, ahead of B200 and B300 in Silicon Data’s snapshot. GPUs aren’t fine wine, but apparently nobody told customers the older vintage was supposed to be worthless. $NVDA 🍷

In this week’s Reference Price newsletter we flagged an interesting “generational slope” in on-demand GPU rental rates. Over the past week the increases were ordered by chip generation: B300 +8.6% → B200 +4.7% → H200 +4.2% → H100 +2.5% → A100 +2.0%. To the extent there is a compute shortage, it appears concentrated at the newest chips. Long-context inference is increasingly memory-bound, which favors the chips with the most high-bandwidth memory — and that memory remains tight. The question is whether rents are starting to separate by capability rather than rising uniformly. Worth watching closely to see if the slope persists. Over the weekend @drwconvexity and @GavinSBaker discussed the “crunch spread” (relative moves in GPU rental rates vs. usage-weighted token prices). @deedydas also noted that the most advanced GPUs continue to deliver rising token throughput efficiency, which raises a related question: how much of the rental premium is justified by throughput gains versus pure scarcity of the newest hardware?

In the current macro environment, raising capital to fund AI compute buildout has become increasingly challenging. This is true even for large established players—Meta acknowledged as much in its latest 10-Q amid rising commitments and cash-flow pressure—and far more so for the smaller so-called “neoclouds.” This is occurring against a backdrop of rising compute rental rates across the entire term structure. Relentless demand for AI, visible in sharply higher prices, is precisely why so many are still racing to add capacity. Institutional-grade, liquid compute futures based on transparent and trustworthy benchmarks are therefore both important and useful for the further development of AI in America. Neoclouds that execute well and deliver reliable performance can lock in future revenues, reduce risk, and secure cheaper financing as a result. Silicon Data publishes daily market-implied forward rate curves for B200, H100, and A100. Once futures trading goes live, we will also have tradable futures curves. Current markets are displaying near-term backwardation (declining rates) and longer-term contango (rising rates).

“AI compute is the new oil—and needs a new market.” — @TheEconomist Cool to see @computeexchange featured as the marketplaces building that future! We are building a neutral, independent marketplace where buyers and sellers can discover prices, access verified supply, and transact through marketplaces, RFQs, and auctions. The compute market is coming. 🚀

Silicon Data CEO Carmen Li says the GPU forward curve has flipped to contango, so the market expects compute prices to keep rising over the next 36 months “So, the yellow line was last year, right? The horizontal line is the duration, and the dollar amount is the Y axis. So, last year was complete downward sloping curve for the yellow line, right? Meaning the longer term contract you signed, the cheaper, you know, was per GPU per hour." "And this year, starting early this year, the whole curve shift up and became contango. If you see the red line, right?" "So, translate to people not getting enough discount for longer term contract. Implies that people assume the price going to rise down the line." "And now you see the price, the whole curve keeps shifting up, got flattened, right? So, meaning people are now either get penalized or get, you know, awarded by signing long term contract." "This is reserved contract we observe. And if you look at implied forward rates, this is the implied forward rates, right? Because we don’t have a future trading. So, this is calculated based on our term structure. And we use similar methodology as Treasury futures curve." "So, you can see a market consensus view is the price could go up in the future, next 36 months.”

Renting Nvidia’s H100 costs 45% more than it did in December. It went on sale in 2022, and two newer generations have come out since. Yet, demand for compute is so far ahead of supply that older GPUs are getting more expensive.

Pretty cool to hear @CMEGroup call @silicondata “the most robust provider of data.” To all our data providers: thank you for trusting us! You chose a team serious about building the gold standard for compute indices. Your data is making a tremendous market impact! More to come! 🚀

Amazing panel AND an even better Q&A at @NVIDIA #GTCBerlin#! Loved sharing the stage with James Finlay (@nvidia ), Kevin Curtin (@jpmorgan), Joe Jackson (Apollo) & Thomas Galbraith (Barkr). From GPU pricing to financing and liquidity, AI compute is becoming an asset class. Thanks Nico Caprez for bringing us together! 🚀

GPU prices higher and token prices lower means the crunch is compressing. That means lower profit margins on the GPU rental to token production leg of the trade. Of course the profit margin on buying GPUs and renting them out is higher than ever (not exactly an indication of a slowdown in inference demand).

Another day, another round of on-demand GPU rental indices surging across the board and reaching ATHs. Even the A100 index, which recently saw some softness, seems to be inflecting a little.

Excited to announce our partnership with @nebiusai at @NVIDIAGTC Berlin! I love seeing infrastructure providers like Nebius embrace transparency. This is how we move compute toward a more open, data-driven marketplace. Check out @stevehou blog on our initial analysis: #SiliconData# #Nebius# #GTCBerlin# #AIInfrastructure# #ComputeEconomy# @dee_bosa

Our CEO @carmenli was featured in cover story in the @WSJ Markets & Finance section today. Thanks to @rudegeair and @candicechoi for covering Carmen's life journey and the AI compute market! [Link below]

What’s there to verify? Over a short period of time and a narrow sample anything you could happen. Over the summer based on our data rental price for B300 went up every day for months. All of compute was going up. Didn’t think it was worth making a big deal out of it to pump a new narrative like a new thing. What if it went down next week?

Allegedly “the most important chart” for understanding the current state of the AI economy. The model layer seems to be going through a period of commodification as non-frontier models saturate everyday use while frontier models compete fiercely. Compute layer sees massive increase in total demand for compute regardless of the source or use case.