It has become fashionable to quote our token index as support for a bearish view on the AI trade. We have pushed back gently a few times, because nothing here is definitive and people should come to their own conclusions. It now seems worth saying a little more. In June we clarified what our LLM Token Index measures: what API users in our sample actually paid per million tokens. A usage-weighted price. Not token volume, not total spend. We noted it can be read, loosely, as revealed willingness to pay for frontier intelligence. We should have emphasized the conditions. The binding one is that the intelligence content of a token stays stable. Over a few weeks that is defensible. Across a couple of release cycles it clearly is not. While a token is simply the wrong unit for intelligence, measuring capability itself without bias is also extremely hard. [Btw we will struggle with this measurement problem all over the place as AI proliferates delivering non-market economic value.] The upshot is that the message from June landed. Maybe a little too well, because it created a new misread: that a falling index is necessarily bearish for the AI trade, since if token prices fall, model-layer margins must follow. Two things are being conflated. The deflation in our sample is real, and it is recent. It dates from the end of May. The index peaked at $2.07 on May 28 and sits at $1.00 as of Sept 21, down 52%. But the same index rose 67% from January into that peak, and few read the rise as bullish for lab margins. It is not bearish now. A usage-weighted price moves with the mix, in both directions. What the mix actually says: more work, at least within our sample, is being routed to cheap, fast models, and labs keep shipping more mid-tier variants. Most everyday tasks never needed a frontier model. That is partial equilibrium for the users we track, and our methodology note is explicit that this index alone cannot separate substitution from efficient agentic routing. Compute demand is a different question entirely, answered by different data series. Our H200 non-hyperscaler rental index has rerated through the summer: $2.87 average in June, $3.29 now, with a record $3.32 on Sept 19. B200 is $5.76, up 7% over the same stretch and 31% year to date. B300, our newest and thinnest series, is up 44% since inception in late April. H100 is off 7% from its August high, consistent with workloads migrating up the stack. Rents on the parts that are actually scarce are not signaling a demand stall. A world of mass agentic use is one where cheap tokens are nearly all of the count. Total token usage should keep growing far faster than the price is falling, with most of that growth coming from cheap, fast, and increasingly open models. The usage-weighted price can keep falling anyway. Our view, not a finding from this index: the highest value-add work still routes through frontier models, and that is where most of the economics will accrue. In any event, the implication for compute is more, not less. An astronomical number of tokens, most of them from cheap flash models, is what economy-wide AI proliferation should look like. It is not a demand stall.

GPU financing has become a core pain point in AI infrastructure. AI startups report that instead of reserving one year of compute, they're increasingly asked to reserve three while putting down 30-40% upfront. Compute providers aren't imposing these onerous terms arbitrarily: demand is outstripping supply, and building a data center is capital-intensive, with financing itself expensive to obtain. Part of why: financial markets still treat GPUs as fast-depreciating assets. Standard market practice still leans on 3-year straight-line schedules to assess GPU residual value, an assumption that flows straight through to the terms providers can offer, and the terms they in turn ask of their customers. We plotted in the quoted post the actual evolution of residual values for major GPU generations as a share of their original purchase prices. These are market-based fair value estimates, derived from the forward curves implied by observed GPU rental contracts. The A100 and H100 are both holding value well in excess of what 3 or 5-year straight-line schedules would imply. In fact, B200, whose supply is scarce, we estimate a current residual value well in excess of its original purchase price! We believe that the financial institutions and capital markets will eventually come view GPUs as the long-lived income-generating capital assets they are and allow them to be financed accordingly. Silicon Data is committed to bringing price transparency and indices to help improve the capital efficiency around GPU financing.

Our CEO @carmenli's X account was compromised earlier today. It has been secured and is back under her control. Posts published from the account between approximately 8:34 AM and 1:22 PM ET were not hers. This includes posts referring to a cryptocurrency token. Silicon Data has not issued, and will not issue, any cryptocurrency token. A token using our company name was created by the unauthorized party from that account. We have no involvement with it and receive nothing from it. Please disregard any communication sent from the account during that window. We are reporting the matter to the relevant platforms and are reviewing it internally.

🚨With CME launching futures on our indices this October, we’ve seen tremendous interest from market participants looking to use Silicon Data’s H100 and B200 indices for block trades, swaps, and other OTC transactions. To help accelerate adoption, we’re opening up our index APIs and offering complimentary licensing for firms using the indices as reference prices for OTC trading. If you’re trading or structuring compute products and want access, just ping us. We’ll provide the data access and a straightforward licensing agreement. Excited to support a deeper, more liquid compute trading ecosystem. Licensee may use Silicon Data’s H100 and B200 Indices solely as reference prices for bilateral OTC transactions, including block trades, swaps, forwards, pricing, settlement, and transaction-specific marks. The Indices may not be redistributed, resold, sublicensed, published, or used for broader risk management, portfolio valuation, benchmarking, regulatory reporting, or creation of other financial or data products without Silicon Data’s prior written consent. If interested, please reach out to contact@silicondata.com

GPU residual value update: - A100 $4,956 (-12.0% YTD) - H100 $20,308 (+3.6% YTD) - B200 $71,057 (14.4% YTD) A month on, the picture has not much changed. A100 ( launched in May 2020) has been roughly flat near $5k since late 2025. H100 and B200 have in fact both appreciated as the rising GPU rental income more than offset time decay! That is not a 2–3 year scrap curve! GPU financiability has increasingly become the central question for the AI buildout: credit, not chips or power, is what stalls smaller builds. Banks still often mark GPU residual to zero after three years of straight-line depreciation. Our estimates are going-concern value or what the GPU should be worth if it keeps running. "Zero after three years" is the wrong prior for that number. A six-year-old A100 still printing ~$5k is the living proof!

"We don't think anything is important unless the data tells us that way." Carmen Li @carmenli , CEO of @Silicon_Data and Compute Exchange, on how she turns wildly different GPU specs into one clean price index, no assumptions, just what the data actually shows. Full episode available now.

It turns out the LLM token index has not put in bottom. In fact, it put in a new low. In the last two weeks, both open and closed LLM token indices declined, however the decline of the closed models is the dominant factor. YTD, open LLM index has overtaken the closed LLM index.

Here’s our paper: And feel free to give SiliconMark a try: Our free tier allows you to benchmark and certify your GPUs at zero cost.

Excited and honored to participate as a panelist at the “Funding AI Compute” event in NYC on Sep 14th organized by @PrimaryVC. Thanks to @gabyllorenzi and @BSchech for the kind invitation and hosting! Look forward to many great discussions and learning from the fellow panelists!

I’ve had a lot of conversations with people who want to finance and trade compute as an infrastructure asset. One question I keep coming back to: once you’ve financed a GPU server for five to ten years, how do you actually know what you own—and what condition it’s in—throughout those five years? How do you independently verify which physical GPUs and components are actually there? How do you know those servers are being properly operated and maintained when they may sit in a data center thousands of miles away? And how do you know the equipment you financed on Day 1 is still performing as expected on Day 1,000? Physical inspection doesn’t scale. Self-reporting isn’t enough. You need third-party verification. That’s what we’re building with SiliconMark. We work with infrastructure providers to track machines down to component-level UUIDs—GPU, CPU and the broader system—and build a persistent identity and performance history for the asset. You can know what the machine is, its expected depreciation curve, how it is actually performing, its thermal behavior and quality history, with timestamped records over its lifecycle. And because our tests are open-sourced, the results are reproducible and independently verifiable. Think of it as a digital service record for compute infrastructure, maintained by an independent third party. For equipment financing, knowing the original purchase price isn’t enough. You need to continuously know what the asset is, that it exists, how it has been treated, how it is performing, and ultimately what it is worth. If GPUs are going to become a financeable and tradable institutional infrastructure asset class, this verification layer is a fundamental building block. Third-party verification is the trust layer between the physical GPU and the financial asset.

As a gauge of AI compute demand, we like looking across the full cross-section of GPUs and the term structure. But if we have to pick a single favorite metric, it’s not the H100 rental rate — it’s the A100. Released in May 2020, this “ancient” chip is still going strong and rented out at near-full capacity (as CoreWeave’s recent earnings call confirmed). A100s are now used predominantly for inference. The fact that A100 rental rates have held so steady — even as newer and far more powerful GPUs keep coming online — is a clear signal of the unrelenting strength and growth in inference demand.

Everybody saw it in the market. The @Silicon_Data indices all put in lows around then and since have been up and to the right. However not all GPU indices are equally well constructed. This graph highlights a serious discrepancy. I know which futures contract I will be using.

DeepSeek raises model prices 4 times. As we had observed last week, an interesting trend in AI model layer is that open models are becoming more expensive while closed frontier models are becoming cheaper with successive price changes from ChatGPT, Grok and Muse Spark not to mention price “per intelligence”.