Silicon Data

When it rains, it pours. Since Jul 2, our H100 GPU rental curve has risen yet again with even more providers hiking prices. Our perspective below on "meta selling compute" also seems to have been affirmed by today's news. Acc to our data, compute market continues to tighten.

@mattarderne Yes, coming soon

@Silicon_Data do you have any geographic breakdowns of this data?

When it rains, it pours. Since Jul 2, our H100 GPU rental curve has risen yet again with even more providers hiking prices. Our perspective below on "meta selling compute" also seems to have been affirmed by today's news. Acc to our data, compute market continues to tighten.

👀

Saw a recent post from @TechCrunch states “For better or worse, @nvidia value as a company is tied to the price of compute and that price is falling. Micron and its cohort are tied to the price of DRAM, and that price keeps rising.” Lower $/GPU spot doesn’t necessarily mean Nvidia is worth less. Throughout computing history, lower unit costs have expanded demand and grown the overall market. I’d also look beyond spot rental prices. AI infrastructure is a capital asset—its economics depend on lifetime cash flows, utilization, and residual value, not just today’s spot hourly rate. A healthy secondary market and strong residual values lower the cost of ownership and support continued investment. The question isn’t whether compute gets cheaper. It’s whether demand for compute grows even faster.

“Each LLM token, open or closed, has become measurably more intelligent or capable over the past 7 months. Our data suggests that users have been willing to pay more for that increase in intelligence or capability!”

Markets were jittery this week on news reports that Meta may be selling compute, raising concerns about excess supply. We thought we’d share some perspective using our rental term curves on why this news, if true, doesn’t have to be bearish for GPU rental prices. Last year, ahead of the agentic AI boom, Meta aggressively locked in a large amount of compute capacity. It quite possibly secured more than it ultimately needed. It was a smart strategic bet to secure scarce supply early, with the flexibility to either deploy it internally or sell excess into a tight market. They acquired valuable real options at the time. As recently as last November, the compute market looked very different. Spot and forward GPU rental rates were much lower, and the term curve was sharply backwardated. This is classic commodity behavior when the market anticipates new supply coming online and pressuring prices lower. Since then, the picture has changed dramatically. As shown in our H100 term rate curves below, the entire curve has both risen sharply in level and flattened significantly, moving out of its steep backwardation. In fact, rental rates have firmed further around the 1-year term over just the past week (Jun 25 – Jul 2), with multiple providers raising prices. For all the concern about a glut, the rental market is doing the opposite of pricing one in: rates are firming, not softening. It now makes perfect financial sense for Meta to shed some of its older secured capacity while continuing to invest in newer, more powerful clusters. The real option they purchased has appreciated meaningfully. At the same time, demand for their specific models and use cases may not have materialized as strongly or as quickly as anticipated. This looks like firm-level rebalancing rather than a signal about the broader market. Reallocating from legacy commitments toward frontier hardware is what a maturing market looks like: optimization, not weakness. Little in our data suggests the demand tailwinds from agentic AI and inference are softening. If anything, the term structure of GPU rental rates points to a market that’s tightening, not loosening. Our forward and term curves are updated daily at silicondata.com. Happy Fourth! 🇺🇸🎆

Eight months ago, when we started building the LLM Token Expenditure Index, the goal was simple: bring transparency to how much, as a society, we’re actually spending on AI per million token. I do not think it was that useful for me to tell you what you’re paying for one model—you already know your own bill. What I found much more interesting was the market as a whole: a volume-weighted view across models that shows our collective willingness to pay for AI. Eight months later, it’s exciting to see the index now sitting at the center of this conversation. This is exactly why we built it. Happy to chat.

Thanks to Bloomberg @JPBarnert @mikamsika @business for featuring and discussing our LLM Token Expenditure Index in their article today! See the quoted post for more details on our interpretation of the index. 😊 bloomberg.com/news/articles/…

At the start of the month, we observed (see QT) that "the token index has stagnated, which suggests that usage migration towards frontier models has slowed. Time will tell whether this is just a pause or an inflection in the trend as users move back towards open-weight models." Since then, we've seen the arrival of GLM 5.2 and a wave of reports of enterprises both big and small becoming more conscious of the quality-cost tradeoff between frontier models and open-weight models. So much so that a debate now rages on X over whether open-weight models pose existential or geopolitical risks to frontier models and the AI buildout altogether. We do not pretend to have answers to questions of such profound magnitude. What we can offer is what the data shows. Since Nov 2025, the token index has fluctuated: rising sharply into Jan, falling through Feb, recovering through May, and pulling back again in June. But through the ups and downs, the trend line has remained distinctly upward. The index, which tracks the expenditure-weighted price of LLM tokens across frontier and open-weight models, now sits meaningfully above where it started. Each LLM token, open or closed, has become measurably more intelligent or capable over the past 7 months. Our data suggests that users have been willing to pay more for that increase in intelligence or capability!

👀 Since last Nov or the start of agentic AI GPU rental curve has risen significantly across the term structure. The flattening of the rental curve means that long-term contracts no longer receive discounts as compute providers feel confident that future rents may go even higher.

Finally our estimate of the residual value for the H100 GPU is basically unchanged since the start of the year despite recent fluctuations in spot rental rates. Demand for GPUs remains strong as far as we can tell. Jevons effect dominates the substitution towards cheaper models.

First thing to notice is the much older A100 rental price continues to rise. The 2nd thing we want to highlight is that if we look at the reserved contracts and out further at 1 year or longer, there’s a distinct rise of the entire term curve since Nov 25 but unchanged since Mar.

There are some jitters in the market about demand for AI compute partly inspired by our Token Index and partly perhaps due to the decrease in the rental price for the H100 GPU. We too have observed a recent downtick in the H100, we don’t agree that compute demand has decreased.

SITUATION EXPLAINED: Why does AI compute need a futures market? We asked @carmenli, founder and CEO of Silicon Data and Compute Exchange: "We believe compute will be humans' largest resource, surpassing all traditional energy combined, so oil, gas, and electricity." "If you're a SaaS company, you only had people costs and then your margin. However, now going forward, everybody has to consume GPU or token. And this can be your largest cost component up to 80% or even 90%." "If 80-90% of your cost has 40-50% daily volatility, your margin is essentially uncontrollable." "If American Airlines tells you, 'I cannot hedge my oil prices,' your ticket price can go from $20 to $5k back to $200. How can you plan any trips? You can't." "Similarly, any companies that heavily rely on GPU or token have to have a way to manage their cost structure so they can have a healthy, manageable margin." @sodofi_

NEW ODD LOTS - BUILDING A MARKET FOR COMPUTE People love saying 'compute is the new oil. But to actually get a commodity market for GPUs, you need things like standardization, benchmark indices, etc. @TheStalwart and I talk @carmenli of @Silicon_Data podcasts.apple.com/us/podcast/car…

@TheStalwart @tracyalloway @carmenli @computeexchange And @Silicon_Data! 😊

NEW ODD LOTS: In a LIVE conversation, @tracyalloway and I talk to @carmenli about @computeexchange and the imminent launch on the CME of a compute futures market. podcasts.apple.com/us/podcast/car…

👀

It is pretty cool to be cited in @nikkei nikkei.com/article/DGXZQO… can’t say I understand the whole article 😂

No doubt @Silicon_Data has a way better grasp on this than me. But I don’t get why usage is always framed as “expensive SOTA models vs cheaper open weight ones.” What about cheaper, closed, non-SOTA models, which keep getting better and can ably handle many tasks?