Silicon Data

Silicon Data is a New York-based company that publishes daily benchmark indexes for the price of renting AI GPUs, led by the Silicon Data H100 Rental Index (Bloomberg ticker SDH100RT), which tracks the hourly rental cost of an NVIDIA H100.^[1][2] Carmen Li, a former Bloomberg executive, founded the company in 2024, and the trading firm DRW says it played a significant role in the founding. The H100 index went live on Bloomberg terminals on May 27, 2025, and Silicon Data has since added indexes for the A100, B200, H200, B300 and AMD MI300X GPUs and for LLM token prices.^[1][2][3][4] Its benchmarks are used as reference prices for financial contracts on AI compute: SDH100RT was the resolution source for Polymarket's first H100 rental-price markets, opened in February 2026, and the H100 and B200 indexes are the planned settlement prices for the compute futures that CME Group's NYMEX exchange asked the Commodity Futures Trading Commission (CFTC) to approve in August 2026.^[5][6] As of mid-October 2026 the futures were still awaiting approval, after the CFTC extended its review to November 9, 2026.^[7][8]

History

Founding and seed round

Silicon Data was founded in 2024 and is based in New York, according to Bloomberg.^[1] Li began her career at DRW and was a senior vice president in working capital products at Citi until 2022, according to eFinancialCareers; IEEE Spectrum reported that until early 2024 she was in charge of global data integration at Bloomberg.^[9][10] DRW, the Chicago trading firm founded by Don Wilson, says it played a significant role in founding both Silicon Data and The Compute Exchange, a GPU capacity marketplace that Wilson co-founded and that launched in January 2025.^[4][9]

On March 21, 2025, the company announced a $4.7 million seed round co-led by DRW and Jump Trading Group, with Woodside AI, Wintermute Ventures, Sancus Ventures and SoGal Ventures also investing. It introduced four tools at the same time: SiliconNavigator for GPU prices and market data, SiliconMark for tracking GPU performance under real workloads, Silicon PriceIQ for price forecasting and SiliconCarbon for estimating the carbon footprint of GPU use.^[11]

Launch of the H100 index

Silicon Data launched the H100 Rental Index on May 27, 2025 and called it the world's first daily benchmark for GPU rental prices. According to the launch release, the index standardized a wide range of H100 configurations for factors such as GPU subtype, location and platform conditions, and drew on 3.5 million pricing data points from multiple rental platforms. It was published on Bloomberg terminals, with Kaiko named as a further distributor and an A100 index planned.^[2] Bloomberg reported that the main users were banks licensing the index for internal risk management, hedge funds licensing the underlying data, and asset managers and private equity firms that finance or build data centers. "There is zero transparency now for GPU costs," Li told Bloomberg.^[1] On launch day the index put the average hourly rental price of an H100 at $2.37, IEEE Spectrum reported.^[10]

Expansion and the December 2025 revision

An A100 index followed in 2025, and Silicon Data worked with dxFeed, a market data and index management firm, to make the H100 and A100 indexes available on Refinitiv terminals.^[12] In October 2025 Li also became chief executive of Compute Exchange. Silicon Data said the two separate companies would follow a unified strategy, with tighter product integration and shared infrastructure.^[13]

On December 4, 2025, Silicon Data changed the methodology of its H100 and A100 indexes and restated both back to September 1, 2024. It removed a divisor adjustment, revised its proprietary pricing model, widened its coverage of cloud providers, updated provider weights and added a look-back mechanism to bridge short gaps in provider data. The company estimated that the changes lowered the H100 index by 4 to 6 percent and raised the A100 index by 35 to 40 percent, and it warned that the restated series might jump sharply on March 1, 2025, a day when its provider coverage expanded quickly.^[14] At the same time it launched a B200 index (SDB200RT), along with H100 and A100 on-demand indexes for hyperscalers that it offers only through its own portal, and said it had more than 500 enterprise, financial and semiconductor clients.^[14][15]

Futures partnership and Series A

During 2026 Silicon Data released a GPU forward curve on April 20, announced a futures partnership with CME Group on May 12, and launched neocloud indexes for the H200, effective July 15, and the B300, effective September 14.^[16][4][14] In July it added SiteIQ, a tool for modeling the economics of GPU data center sites.^[17] On August 11, 2026, it announced a $30.5 million initial closing of a Series A round led by the Valor Atreides AI Fund. The company said it had more than 1,000 registered users and nine financial-grade indexes covering GPUs and large language models, and that it would spend the money on benchmark pricing, SiliconMark performance measurement, institutional market and alternative data, and risk infrastructure for derivatives, insurance and credit markets.^[18]

Sources: SiliconANGLE and Silicon Data.^[11][18]

Indexes and data products

Silicon Data's GPU indexes give a price in US dollars per GPU-hour. Since the December 2025 revision it has kept separate series for neocloud providers (the specialist GPU clouds) and for hyperscalers; the neocloud series is the one carried under the Bloomberg tickers, and the company keeps the two apart by design rather than blending them into one H100 price.^[19][20] Its index list also includes a GDDR6 memory price index.^[3]

Sources: Silicon Data documentation and index pages, Bandi and Su, and Proactive Investors.^{[14][15][20][3][21]}

Alongside the indexes, Silicon Data sells data through a web portal and an API, including Token Pricebook, which standardizes LLM token prices to US dollars per million tokens, and Token Market Pulse, which aggregates token prices and usage.^[22] The GPU forward curve, released on April 20, 2026, gives term-structure rates and implied forward rental rates up to 36 months ahead for the A100, H100 and B200; the company said its first readings showed long-term contracts were not meaningfully discounted to on-demand prices and sometimes cost more.^[16] On September 7, 2026, its H100 curve put the rate three months forward at $2.50 and twelve months forward at $2.14, when the spot index stood at $2.63.^[23] SiteIQ estimates a site's usable GPU capacity and itemizes costs such as power, staffing and GPU depreciation for investors, lenders and operators.^[17]

How is the index calculated?

Silicon Data's two-page methodology overview, last updated on August 27, 2026, says each daily value is built from more than 10,000 GPU rental records collected from over 40 cloud providers, across several rental types and hardware configurations.^[24] The company's index pages say the observations come from cloud providers, colocation markets, brokered cluster sales and private rental platforms, and claim coverage of 95 percent of neocloud GPU providers, all major hyperscalers and more than 80 percent of the global GPU rental market.^[25]

Source: Silicon Data methodology overview.^[24]

The pricing model is recalibrated weekly, and records that deviate sharply from historical ranges for their provider and GPU variant are held for manual review. The overview says Silicon Data announces material changes in advance, stating the effective date and estimated effect on index values, and separately announces any decision to discontinue an index; the December 2025 revision was issued through such a notice.^[24][14] The company's index page says each index is published once per business day.^[3] The overview does not publish the provider weights or how they are set, gives no calculation or publication time, and does not mention an oversight committee, a conflicts-of-interest policy or the IOSCO Principles for Financial Benchmarks.^[24] NYMEX's filing names the publisher of the indexes as Silicon Derivatives Inc.^[6]

Li has described the hyperscaler tier as the enterprise-SLA tier of Amazon Web Services, Azure and Google Cloud, and the neocloud tier as specialist providers such as CoreWeave, Lambda, Crusoe and Nebius. She has written that neocloud capacity rents at roughly a third of the hyperscaler rate and reacts faster to supply and demand.^[19] Between March 1 and April 20, 2026, the hyperscaler H100 index stayed between $7.40 and $7.52 and was unchanged on 39 of 50 trading days.^[26]

H100 index levels

The table lists published levels of Silicon Data's H100 index, each with the date of the reading. Values from before December 4, 2025 were calculated under the original methodology; the restatement that took effect that day lowered the H100 history by an estimated 4 to 6 percent.^[14] On the restated series, which begins on September 1, 2024, the index fell 45 percent from its start to February 2025, according to Silicon Data.^[27]

Compute futures

CME Group partnership

CME Group and Silicon Data announced on May 12, 2026 that they planned to launch a compute futures market later that year, pending regulatory review, based on Silicon Data's indexes, which the companies described as "the world's first daily GPU benchmarks for on-demand rental rates." CME chairman and chief executive Terry Duffy said that "compute is the new oil of the 21st century." Li said, "Compute markets today are still highly fragmented," and Wilson predicted that "compute will become the largest commodity in the world."^[4]

NYMEX filing

On August 11, 2026, NYMEX asked the CFTC to approve Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures under CFTC Regulation 40.3, which provides for formal Commission approval, and said it intended to list them on CME Globex for trade date October 5, 2026.^[6]

Source: NYMEX Submission No. 26-370.^[6]

A contract of 730 GPU-hours equals one GPU running for an average month (365 times 24, divided by 12), and Silicon Data says the H100 contract settles only against its index of independent neocloud providers.^[23] In the filing NYMEX said the reference index was "sufficiently broad in definition and scope and supported by a robust level of trading in the underlying cash market" to meet the requirement that a contract not be readily susceptible to manipulation. It sought confidential treatment for the supplemental market information attached to the submission.^[6]

CFTC review

On September 21, 2026, the CFTC's Division of Market Oversight extended its review by 45 days, "until the end of November 9, 2026," citing "novel or complex issues" and the Commission's pending request for comment on compute derivatives, whose comment period was due to close on October 20, 2026.^[7] The October 5 target passed without a listing, and in mid-October 2026 the CFTC's product register still showed the H100 contract as "Approval Pending".^[8] Wilson, quoted by GARP, called it "a mistake to delay the launch of these futures by subjecting them to a lengthy comment period."^[30]

How do the CME and ICE contracts differ?

Intercontinental Exchange (ICE) plans competing futures on Ornn's Compute Price Index (OCPI), which Bandi and Su describe as a transaction-based benchmark.^[20] On September 29, 2026, ICE Futures U.S. named the first two, OCPI H100 Futures (HPR) and OCPI B200 Futures (BKL). Their contract size is the number of hours in the contract month rather than a fixed 730, prices move in tenths of a cent per GPU-hour, and up to 24 consecutive months may be listed instead of 36. ICE said it would list them once the CFTC's comment period was complete and that no date had been set. Both exchanges' contracts settle in cash on a monthly average of daily index values.^[31][6] The wider market, including perpetual futures, prediction markets and the CFTC consultation, is covered in compute derivatives.

Other uses

Prediction markets

Polymarket's first markets on H100 rental prices opened on February 3, 2026 and named SDH100RT as the resolution source, using the daily values on Silicon Data's index chart. Each day's value became final once the next day's value appeared.^[5][32] The February market traded about $337,000 and resolved with the index having reached $2.40 but not $2.45 during the month, without falling to $2.25. A companion market running to April 30 traded about $570,000; in that window the index reached $2.50 but not $2.75 and never fell to $2.20.^[5][32] Later Polymarket H100 markets, including one opened on May 14, 2026, named Ornn's H100 index as the resolution source instead, and Kalshi's monthly H100 series also settles on Ornn data.^[33][34]

Distribution and research use

Silicon Data's indexes reach financial users through Bloomberg and Refinitiv terminals and through the company's own portal and API, and the company lists Kaiko among its distribution partners.^[15][3] Federico Bandi and Yinan Su of Johns Hopkins University used Silicon Data's spot and term-rental series, alongside Ornn's, in an asset-pricing paper first posted in July 2026; they noted that the exact transformation behind Silicon Data's forward curve was not documented in the data available to them.^[20] In July 2026 Proactive Investors cited the LLM Token Expenditure Index as a possible warning sign in the debate over an AI bubble, reporting that it had nearly doubled from its December start and then fallen about 20 percent since May; according to Bloomberg, Silicon Data had told people to stop reading the index as a straight price tag.^[21]

Reception and criticism

Listed versus transacted prices

The main methodological dispute over GPU price indexes is whether they should rely on published rates or only on completed trades. SemiAnalysis, in the April 2026 article that launched its own H100 one-year rental price index, wrote that "many GPU rental indices are derived from spot/on-demand listings or posted pricing" and that "there is no guarantee that buyers of compute are actually transacting at prices publicly posted." In its view providers seldom change on-demand prices, so posted prices often move only after the contract market has shifted. It did not name Silicon Data.^[35] Ornn's lead seed investor, a16z crypto, described Ornn's index as "built from cleared trades rather than scraped list prices," also without naming a competitor.^[36]

Silicon Data says it aggregates rental and private transaction data from neocloud providers, hyperscalers, colocation markets and private rental platforms.^[3] For its hyperscaler tier, Li has made a similar point herself. In April 2026 she wrote that a hyperscaler listing an on-demand H100 rate "is not clearing a spot market" but "posting a ceiling above a large book of already-priced reserved capacity," that the series' stability was "administrative as much as market-derived," and that the neocloud index and no-arbitrage forward rates were more natural references for pricing swaps and forwards.^[26] The CFTC's August 2026 request for comment, which named no provider, noted that "certain published compute price series are constructed in whole or in part from posted or listed rates that the compute capacity providers themselves administer," and asked what would stop a provider from moving a settlement index by adjusting a posted rate.^[37]

Divergence from Ornn's index

Sources: Silicon Data and Ornn.^[25][38]

The two companies' indexes differ in level, and not in a consistent direction: in mid-October 2026 Silicon Data's H100 and B200 readings were below Ornn's, while its A100 reading was higher. Bandi and Su, who had data from both companies, found that the Silicon Data series grouped apart from the Ornn series in a correlation network and wrote that "the disconnect between Silicon Data indexes and Ornn indexes is a potential issue."^[20] The newsletter writer Dave Friedman cited a HEC Montréal study that found weekly changes in the two H100 indexes correlated at only 0.17, so that either would have removed about 3 percent of the other's weekly variance as a hedge; in his account, the data could not show whether the gap reflected "methodology, coverage, or genuine price differences." Friedman also noted that the CFTC's letter extending its review "identifies no particular defect in either index."^[39] List prices by provider are compared in cloud GPU pricing comparison.

Conflicts of interest

Silicon Data's investors include CME Group, whose exchange would list futures settling on its indexes and which invested in the Series A through CME Ventures, and DRW, which co-led the seed round, invested again in the Series A and says it played a significant role in founding the company.^[18][11][4] In The Next Web, Cristian Dina wrote that "the exchange holds equity in the company whose index its contracts will settle against" and that "several likely participants hold equity in it too." He called this "not unusual in commodity benchmarks" but added that "a benchmark is only as good as its independence."^[40]

Li also runs Compute Exchange, a marketplace for trading GPU capacity, and Wilson, quoted by GARP in August 2026, said Silicon Data's indexes "leverage data from numerous neoclouds as well as from its sister company Compute Exchange."^[13][30] Writing about compute benchmarks in general, Friedman warned that "contributors to the underlying data may hold positions whose value depends directly on the prices they help report."^[39] The CFTC has asked how to treat cases in which the parties best placed to influence a reference price also supply the capacity, transactions or posted rates behind it.^[37] Silicon Data's methodology overview does not describe how the company handles conflicts of interest.^[24]

Concentration and manipulation risk

Ty Greenberg, chief revenue officer and a co-founder of the compliance company Caffrey.AI, was quoted by GARP as saying that "there is no single, readily observable 'price of compute,'" that "the compute market is dominated by a relatively small number of large providers," and that because there are no mandatory reporting requirements for physical compute transactions, "there is no public dataset from which to establish an index." Defending Silicon Data, Wilson said the company "has successfully created indices which are robust, well-constructed and not readily susceptible to manipulation" and that "they do not rely on a small number of hyperscalers' pricing."^[30]

Silicon Data averages each provider's records before weighting providers, but it does not publish the weights.^[24] Its hyperscaler tier tracks the posted rates of a few large clouds, and Li has attributed that series' stability partly to administered pricing.^[19][26] Friedman suggested testing a settlement index with its largest contributors removed, to check how much it depends on a narrow set of prices.^[39]

References

1. ^1 ^2 ^3 ^4Katherine Doherty. "Silicon Data Creates First-of-Its-Kind Index for AI Chips." Bloomberg News (via Data Center Knowledge), 2025-05-27. datacenterknowledge.com/...kind-index-for-ai-chips
2. ^1 ^2 ^3"Silicon Data Launches World's First Daily GPU Rental Index to Bring Transparency to the AI Infrastructure Economy." Silicon Data (EIN Presswire), 2025-05-27. einpresswire.com/...-the-ai-infrastructure-economy
3. ^1 ^2 ^3 ^4 ^5 ^6"Silicon Index: Daily GPU Price Benchmarks." Silicon Data, accessed 2026-10-15. silicondata.com/...silicon-index
4. ^1 ^2 ^3 ^4 ^5"CME Group and Silicon Data Partner to Launch First Compute Futures." CME Group and Silicon Data (PR Newswire, via Silicon Data newsroom), 2026-05-12. silicondata.com/...to-launch-first-compute-futures
5. ^1 ^2 ^3"GPU rental prices (H100) hit___ in February?" Polymarket, 2026-02-03. polymarket.com/...ntal-prices-h100-hit-in-february
6. ^1 ^2 ^3 ^4 ^5 ^6"NYMEX Submission No. 26-370: CFTC Regulation 40.3(a) Voluntary Submission for Product Approval, Initial Listing of Two (2) Compute Futures Contracts." New York Mercantile Exchange (filed with the CFTC), 2026-08-11. cftc.gov/...ptc08112615405.pdf
7. ^1 ^2"Extension of Initial Review Period of NYMEX Submission No. 26-370 and Submission No. 26-370S for an Additional 45-Days." US Commodity Futures Trading Commission, Division of Market Oversight, 2026-09-21. cftc.gov/...orgdcmnymexcompcontr260921.pdf
8. ^1 ^2"Designated Contract Market Products: Silicon Data H100 Rental Index Futures (Filing 62544)." US Commodity Futures Trading Commission, accessed 2026-10-15. cftc.gov/...62544
9. ^1 ^2Sarah Butcher. "The ex-Citi VP at the heart of GPU trading: 'We're growing 200% every six months'." eFinancialCareers, 2026-03-25. efinancialcareers.com/...trading-gpus
10. ^1 ^2 ^3Samuel K. Moore. "Price Index Could Clarify Opaque GPU Costs for AI." IEEE Spectrum, 2025-05-28. spectrum.ieee.org/gpu-prices
11. ^1 ^2 ^3Kyt Dotson. "Silicon Data raises $4.7M for AI-driven GPU market insights." SiliconANGLE, 2025-03-21. siliconangle.com/...-ai-driven-gpu-market-insights
12. ^"dxFeed Helped Silicon Data to Expand GPU Index Coverage on Refinitiv." dxFeed, 2025-10-02. dxfeed.com/...pand-gpu-index-coverage-on-refinitiv
13. ^1 ^2"Silicon Data CEO Carmen Li Expands Leadership to Include Compute Exchange." SalesTechStar, 2025-10-07. salestechstar.com/...p-to-include-compute-exchange
14. ^1 ^2 ^3 ^4 ^5 ^6"GPU Index Announcements." Silicon Data documentation, accessed 2026-10-15. docs.silicondata.com/...gpu-index-announcements
15. ^1 ^2 ^3"Silicon Data Announces Major Revision to A100 & H100 Rental Indices, Launches First-Ever B200 Index, and Introduces New Hyperscaler Rental Benchmarks." Silicon Data, 2025-12-05. silicondata.com/...w-hyperscaler-rental-benchmarks
16. ^1 ^2"Silicon Data Unveils First GPU Forward Curve, Signaling Transition of AI Compute into a Tradable Commodity." Silicon Data, 2026-04-20. silicondata.com/...unveils-first-gpu-forward-curve
17. ^1 ^2"Silicon Data Launches SiteIQ: Bringing Underwriting Rigor to AI Data Center Economics." CIO Influence (EIN Presswire), 2026-07-24. cioinfluence.com/...or-to-ai-data-center-economics
18. ^1 ^2 ^3"Silicon Data Raises $30.5 Million Series A to Build the Independent Benchmark Layer for the AI Compute Economy." Silicon Data, 2026-08-11. silicondata.com/...ta-raises-30-5-million-series-a
19. ^1 ^2 ^3 ^4 ^5Carmen Li. "H100 Index, May 2026: Same Chip, Two Tiers, Two Regimes." Silicon Data, 2026-06-16. silicondata.com/...-may-2026-two-tiers-two-regimes
20. ^1 ^2 ^3 ^4 ^5Federico M. Bandi and Yinan Su. "(Early) AI Compute Asset Pricing." arXiv:2607.12156 (v3 revised 2026-09-10), 2026-07-13. arxiv.org/...2607.12156
21. ^1 ^2Ian Lyall. "Is the AI bubble ready to pop? One leading indicator makes uncomfortable reading." Proactive Investors, 2026-07-06. proactiveinvestors.com/...fortable-reading-1094961
22. ^"Introduction." Silicon Data documentation, accessed 2026-10-15. docs.silicondata.com
23. ^1 ^2 ^3"The Practitioner's Guide to the Compute Future: How to Hedge, Step by Step (Part 1)." Silicon Data, 2026-09-14. silicondata.com/...mpute-future-hedge-step-by-step
24. ^1 ^2 ^3 ^4 ^5 ^6"Silicon Data GPU Rental Index Methodology Overview." Silicon Data, 2026-08-27. silicondata.com/...icon-data-index-methodology.pdf
25. ^1 ^2 ^3"H100 Rental Price Index." Silicon Data, accessed 2026-10-15. silicondata.com/...h100
26. ^1 ^2 ^3Carmen Li. "H100 Hyperscaler Index, April 2026: A Benchmark in Flat Mode." Silicon Data, 2026-04-27. silicondata.com/...h100-hyperscaler-index-april-2026
27. ^Yuhua Yu and Steve Hou. "Cash-Settled Compute Futures." Silicon Data, 2026-08-24. silicondata.com/...cash-settled-compute-futures
28. ^Carmen Li. "H100 Rental Market Cools in September: Price Index Slips as Volatility Rises." Silicon Data, 2025-10-20. silicondata.com/...al-market-update-september-2025
29. ^1 ^2Carmen Li. "H100 Price Spike: Understanding the 10% Surge in GPU Rental Costs." Silicon Data, 2026-01-09. silicondata.com/...h100-price-spike
30. ^1 ^2 ^3Jeffrey Kutler. "Can't Get Enough 'Compute'? Here Come the Futures." GARP Risk Intelligence, 2026-08-28. garp.org/...cant-get-enough-compute-260828
31. ^"New Compute Index Reference Data to Be Made Available on Monday, October 5, 2026." ICE Futures U.S. exchange notice, 2026-09-29. ice.com/...Notice_ORNN_Compute_4Q2026_20260929.pdf
32. ^1 ^2"GPU rental prices (H100) hit___ by April 30?" Polymarket, 2026-02-03. polymarket.com/...-prices-h100-hit-by-april-30-967
33. ^"GPU rental prices (H100) hit___ by May 31?" Polymarket, 2026-05-14. polymarket.com/...rental-prices-h100-hit-by-may-31
34. ^"H100 Monthly (series KXH100MS)." Kalshi Trade API, accessed 2026-10-16. api.elections.kalshi.com/...KXH100MS
35. ^Daniel Nishball, Jordan Nanos, Cheang Kang Wen and others. "The Great GPU Shortage - Rental Capacity - Launching our H100 1 Year Rental Price Index." SemiAnalysis, 2026-04-02. newsletter.semianalysis.com/...age-rental-capacity
36. ^Ali Yahya and Elizabeth Harkavy. "Investing in Ornn: A Market for Compute." a16z crypto, 2026-06-24. a16zcrypto.com/...investing-in-ornn
37. ^1 ^2"Request for Comment on the Listing of Compute Derivatives Contracts." Federal Register, Vol. 91, No. 161, pp. 54259-54264 (US Commodity Futures Trading Commission), 2026-08-21. govinfo.gov/...2026-17163.pdf
38. ^"GPU Market Prices: Compute Price Index (OCPI)." Ornn Data, accessed 2026-10-15. data.ornn.com/markets
39. ^1 ^2 ^3Dave Friedman. "Compute Futures Need More Than a Useful Price Index." Buy the Rumor; Sell the News, 2026-09-23. davefriedman.substack.com/...ures-need-more-than-a
40. ^Cristian Dina. "AI compute becomes a tradable commodity on 5 October. The price goes public." The Next Web, 2026-08-13. thenextweb.com/...ta-gpu-compute-futures-october-5