The Index Was the Wedge. The Product Is the Risk Desk.

Our August 1 methodology study treated Silicon Data as one of four index administrators racing to become the settlement benchmark for GPU derivatives — the quote-based, breadth-first entrant against Ornn’s transaction prints, Kalshi’s market-implied curves, and Compute Desk’s physical-delivery indices. That framing was right for the settlement question and is now too small for the company. Since April, Silicon Data has shipped or dated a forward curve (April 20), an underwriting workbench for data-center economics (SiteIQ, July 24), a listed-futures launch date (August 11, for October 5), and a growing family of token-market products — while the index family itself expanded to seven GPU series across five chips, plus RAM and LLM tokens.

The right frame now is not “index provider.” It is the anatomy of a mature commodity-market data franchise — the thing Platts became for oil and Fastmarkets for battery metals — being assembled in advance of the market it serves. Every mature commodity complex runs on roughly the same information stack: a spot assessment, a term curve, a listed contract settling on the assessment, a quality/assay standard, asset- and logistics-level intelligence, and price reporting for the refined products downstream. Silicon Data has now shipped a recognizable version of every layer:

Three things in the inventory deserve emphasis before any application framework, because they are the load-bearing members.

First, the suite is two-sided. Silicon Data is the only provider in the complex publishing daily benchmarks for both the primary input to AI production (GPU rent, RAM) and its primary output (LLM tokens). Ornn publishes token indices too, but as adjacent products; Silicon Data has built a full downstream family — realized expenditure index, posted-price book, and demand-side telemetry. Two-sidedness is what turns a price service into a margin service, and margins are what risk desks actually manage. Section 3 develops this.

Second, the suite now reaches the asset layer. SiteIQ is not a price product at all — it is an underwriting model that converts a facility’s power, cooling, and location into deployable GPU capacity, break-even utilization, and payback math. Paired with Navigator’s second-hand and refurbished pricing — the only public window we know of into GPU residual values — Silicon Data is positioning to answer the collateral question, not just the rent question. That aims the suite squarely at lenders, exactly the constituency our August 1 piece flagged as holding the biggest unclaimed dataset in the market.

Third, the reference layer is about to acquire teeth. On October 5 the indices stop being observational and start settling money. That both validates the suite and stress-tests it: every methodological softness we documented on August 1 — quote-based inputs, proprietary weights, the December 2025 restatement — becomes a live P&L issue for anyone holding a position. The framework below therefore treats “use the index” and “trust the index” as separate decisions throughout.

The core of the suite is the classic three-piece price stack, and it is worth being precise about how the pieces relate, because the application framework depends on it.

The spot indices are daily assessments of on-demand rental rates, built from ~150,000 daily pricing records across a claimed 50–100 platforms — list prices and private-platform quotes, normalized for machine spec, term, cluster scale, platform performance, and geography, then aggregated with proprietary weights. The critical segmentation is the December 2025 neocloud/hyperscaler split: identical H100 silicon prints $2.74 on the neocloud series and $7.20 on the hyperscaler series this week. That 2.6× gap is not noise; it is counterparty tier, contract structure, and bundling priced into a “spot rate,” and it is the single most important number in this piece.

The Forward Curve (April 2026) publishes two distinct objects daily for H100, B200, and A100: a term-structure rate — what it costs today to lock capacity for 1–36 months — and a forward rate — the market’s implied expected spot price at future dates, derived from the term structure by no-arbitrage. The distinction matters operationally: the term rate is a price you can act on (it is the lease market); the forward rate is a model output (there is no liquid forward market validating it yet). Silicon Data’s own framing of the intended users — providers anchoring contracts, buyers timing procurement, financial institutions settling swaps and parametric insurance — is a fair map of where the curve is genuinely useful versus where it is a forecast wearing a curve’s clothing.

The CME futures complete the stack on October 5: Silicon Data H100 Rental Index Futures and B200 Rental Index Futures on NYMEX, each contract representing a month’s worth of GPU rent, cash-settled against the indices — settling, per our August 14 verification, on the neocloud series. Pete Keavey’s launch framing (“compute has become the currency of the AI age”) got the coverage, but Carmen Li’s was the operationally telling one: procurement teams “will now have a benchmark to check that against.” The first advertised use case is not hedging. It is quote verification — which tells you where Silicon Data thinks adoption actually starts.

Our inference-economics work this summer built the case that the natural unit of risk in AI is not the GPU-hour or the token but the spread between them — the compute analogue of a power plant’s spark spread or a refinery’s crack: revenue per million tokens out, minus GPU rent in, divided by the fleet’s realized tokens-per-GPU-hour heat rate. The obstacle to using that spread as an actual risk-management object was always data: nobody published both legs consistently.

Silicon Data now does. SDLLMTK gives a daily realized $/M-token blend (with frontier vs. open-weight sub-readings — $4.20 vs. $0.85 at the April reading, a 5× capability premium), Token Pricebook gives posted per-model input/output rates, and the GPU indices give the input leg. The token leg is also doing real signaling work in the market already: SDLLMTK is down roughly 20% from May, a move Bloomberg flagged in July as one of the AI trade’s key pricing-power signals deteriorating — while the H100 neocloud index drifted sideways. The output price is falling faster than the input price, which is margin compression measured in public data, for the first time.

For risk management this changes what is contractible. A model lab can now index its compute procurement to a published input benchmark while watching its competitive output price in the same portal. An inference reseller can measure gross-margin-at-market daily. And once the futures list, one leg of that spread becomes hedgeable on a cleared venue — an imperfect, one-legged hedge of a two-legged spread, but the first time any of it has been executable at all.

It also changes Silicon Data’s competitive position. Ornn can fight for the settlement benchmark on transaction purity. But a settlement print is a single number; a margin framework is a workflow. Vendors that own workflows are historically much harder to displace than vendors that own numbers.

The July 24 SiteIQ launch is the clearest statement of strategy in the whole suite. It is aimed, by Silicon Data’s own launch language, at investors, lenders, and financial analysts — before data-center operators, who are “planned.” The pitch is what the company calls the Megawatt Illusion: a facility’s headline power capacity is not deployable AI compute, because high-density workloads lose a large fraction of power to cooling and support systems, with the loss a function of climate and hardware stack. SiteIQ takes power spec, location, and hardware configuration and returns deployable GPU count, climate-adjusted cooling stress, monthly operating cost, GPU depreciation, break-even utilization, and site-level payback — with every assumption editable for stress-testing.

Read against the financing wave we covered on August 14 — NVIDIA’s MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR to mobilize $500B+ of third-party capital into AI infrastructure — SiteIQ is a tool built for the underwriting desks that capital will flow through. Its economics module quietly depends on the rest of the suite: revenue assumptions key off the rental indices and curve, depreciation off observed pricing by chip generation. That is the cross-sell, and also the epistemic risk: a lender using SiteIQ is underwriting with the same vendor’s price view on both the revenue line and the collateral line.

Two quieter asset-layer capabilities complete the picture. Navigator’s second-hand and refurbished GPU pricing is, as far as we can establish, the only recurring public-market window into GPU residual values — the exact quantity behind NVIDIA’s case-by-case residual-value support (the up-to-~25% backstop we analyzed as a written put on its own hardware’s depreciation) and behind every GPU-secured lender’s recovery assumption. Aircraft finance runs on exactly this data (Ascend, Avitas appraisals); compute finance now has a v1. And SiliconMark is the assay: with delivered performance on nominally identical H100s varying up to ~38%, a benchmark that certifies what a specific cluster actually delivers is the beginning of a quality-grading standard — the thing that lets a lease, an insurance contract, or eventually a delivery obligation specify not just “H100 hours” but “H100 hours of grade X.” Compute Desk is approaching the same problem from basis tables; SiliconMark approaches it from measurement.

The audit gives us the shelf. The framework question is which product comes off the shelf for whom. We organize it around seven risk exposures that recur across the AI value chain — price level, term/rollover, margin, residual value, utilization, quality, and site/capacity — and five participant classes. The matrix first, then the playbooks.

Lenders and structured financiers: the underwriting stack, with one eye open

The most natural institutional fit for the suite as now constituted. A GPU-secured lending desk can source its revenue-side assumption from the rental indices and term-structure rates rather than sponsor projections; mark collateral against Navigator’s second-hand prints instead of straight-line depreciation schedules; sanity-check facility-level capacity claims through SiteIQ before term sheets; and, from October, run a macro overlay — short H100/B200 futures strips against a book of fixed-rate GPU-backed paper, so that a broad rent decline that impairs every borrower simultaneously pays off on the hedge. For NVIDIA-adjacent structures specifically, the residual-value support we analyzed on August 14 (the ~25% backstop) can now be marked against an observable resale series rather than negotiated in an appraisal vacuum.

The eye that must stay open: the indices price rate, not occupancy. A borrower’s take-or-pay book, customer concentration, and realized utilization remain invisible to every product here — and roughly three-quarters of the market’s actual cash flow sits in exactly those private structures. The suite upgrades the lender’s market view from anecdote to benchmark; it does not yet touch the covenant package that matters most. (It is also worth remembering who is not yet a data contributor: no lender or lessor feeds any index. The first one to negotiate data-for-analytics with an administrator will get better marks than its competitors.)

VC and growth equity: the unit-economics observatory

Equity investors in AI companies are structurally long the spread in Section 3 — their portfolio companies buy compute and sell tokens (or products denominated in them). The suite’s value here is diligence and monitoring, not hedging: SDLLMTK’s trajectory is the single best public proxy for portfolio-wide gross-margin drift (its ~20% decline since May is a portfolio-level margin event that most boards have not yet discussed in those terms); Token Pricebook turns any pitch deck’s “cost per token” claim into a checkable number; Market Pulse’s adoption telemetry flags when a portfolio company’s chosen model family is losing share; and the frontier/open-weight split prices the exact commoditization risk every application-layer thesis depends on. The forward curve, meanwhile, disciplines burn models: a company whose plan assumes flat H100 pricing through 2027 can be tested against a term structure that actually slopes.

Inference-dependent corporates: procurement first, hedging second — and mind the basis

For a corporate whose P&L runs on inference — a customer-service platform, a coding-assistant vendor, an enterprise with a large internal AI budget — the suite’s first-order value is exactly what Carmen Li said at the CME announcement: a benchmark to check quotes against. Pricebook and Market Pulse do for model procurement what telecom benchmarking did for carrier contracts; the index-vs-quote gap becomes negotiating leverage; the term-structure rate answers the lock-versus-float question with market data instead of vendor pressure. A further step most corporates have not considered: indexing supply contracts — writing compute or token supply agreements that reset against SDH100RT or SDLLMTK, the way commercial power contracts reset against hub prices. That converts an opaque bilateral renegotiation into a transparent formula, and it is how young commodity markets actually adopt their benchmarks (indexation first, derivatives second).

Hedging the budget with the October futures is the advertised use case and the one to be most careful with. The contracts settle on neocloud H100/B200 rent. A corporate buying tokens from OpenAI, or capacity from a hyperscaler at the $7.20 tier, is hedging a 2.6×-different price series plus a model-pricing layer on top — a correlation position, not a hedge, as we argued in the listing-window piece. The honest sequencing for a corporate risk desk: benchmark procurement now, index contracts where possible, and treat futures as a partial hedge whose basis you measure before you size.

Neoclouds and operators: the natural short, and the assay premium

Operators are the participant class the futures were arguably built for: structurally long forward rental rates on an asset that depreciates on an 18–24 month silicon cycle. The playbook writes itself — sell futures against unsold forward capacity to lock the rate on the merchant tail of the fleet; use the term-structure rate as the public anchor for private lease negotiations; use SiteIQ’s break-even utilization as the floor beneath any discounting decision; time fleet refresh against Navigator’s resale curve rather than book schedules. SiliconMark adds a subtler lever: an operator whose clusters assay above nominal performance can document a quality premium over the index — the compute version of selling above benchmark on assay. The caveat mirrors the corporate one: an operator’s realized revenue depends on utilization and its own customer tier, so the futures hedge protects the rate component of revenue only. It is a real hedge, but of one factor.

Traders and funds: a two-sided data platform with a listed venue attached

For funds, the suite is less a risk tool than a market-data platform arriving before the market is crowded: 8 years of instance-level history (Navigator’s top tier) to build views on; five chip indices plus segment splits generating relative-value structure (H100 vs. B200 as a silicon-cycle spread; neocloud vs. hyperscaler as a credit/tier spread; GPU vs. token as the margin trade); a no-arbitrage curve to trade listed futures against; and cross-administrator basis — CME/Silicon Data quote-based settlement versus ICE/Ornn transaction-based settlement on economically similar underlyings — which is a pure methodology trade, the kind that paid for years in early power and freight markets. Kalshi’s event contracts (settling on Ornn) complete a triangle: three venues, three price-formation mechanisms, one underlying reality. Early cross-venue dislocations will be less about compute and more about whose measurement you believe — which is precisely why the methodology work matters to trading, not just to governance.

A sixth constituency the suite is quietly courting: insurers

Silicon Data’s own forward-curve materials name parametric insurance as an intended settlement use. The pieces are all present: a daily settlement-grade index (trigger), SiteIQ (exposure assessment at bind), SiliconMark (asset condition attestation), resale data (actual-cash-value basis), and SiliconCarbon for sustainability-linked terms. A parametric product paying out when SDH100RT trades below a strike for N consecutive days is a business-interruption analogue for operators that would be underwriteable today — the insurer’s question, as always in parametric design, is basis risk between the index trigger and insured loss, which returns us to the 2.6× problem and to utilization.

The main event of this piece is the Silicon Data framework, but a framework that never routes to a competitor is marketing. Against each use case above, here is where we would send a client somewhere else — drawing on the full four-provider comparison in our August 1 study.

First: the inputs are quotes, and the methodology is young. Everything in Section 5 inherits the findings of our August 1 study, none of which have been cured since: the indices are built from posted and quoted rates, not executed transactions; the aggregation statistic, weights, outlier rules, contributor list, and reference configuration are all undisclosed; there is no published rulebook, IOSCO statement, independent audit, or oversight committee. The December 2025 revision restated SDA100RT’s entire history by +35–40%, and routine provider additions have moved indices mid-single digits per event — composition sensitivity that matters enormously once contracts settle on these numbers. The CFTC self-certification for the October launch (not yet public as of this writing) is the forcing function: it should surface cash-market depth analysis and anti-manipulation practices for the first time. Until the rulebook exists, size positions as if the index can be restated, because it has been.

Second: the basis is structural, not incidental. The 2.6× neocloud/hyperscaler gap, ~7× same-SKU dispersion across configurations and regions, and up to ~38% delivered-performance variance on identical silicon mean that “the price of an H100-hour” is a family of prices. Silicon Data handles this more honestly than most — publishing the segment split is what revealed the 2.6× — but the hedging instruments settle on one member of the family. Every participant in Section 5 should compute their own basis to the settlement series before sizing any hedge; for hyperscaler-tier buyers that basis is larger than most commodity markets’ entire price.

Third: the vendor is concentrated and conflicted in ways clients should price, not ignore. One firm now supplies the index, the curve, the forecast, the underwriting model, and (via CME) the settlement print — a lender using SiteIQ against index-derived revenue assumptions has a single point of epistemic failure. CEO Carmen Li has also served as CEO of Compute Exchange, a GPU auction marketplace, since October 2025. Compute Exchange was co-founded by Don Wilson — founder and CEO of DRW, the trading firm that co-led Silicon Data’s $4.7M seed alongside Jump Trading Group — and has run spot auctions since early 2025, facilitating over $1B of compute supply. Nothing improper is in evidence, and the auction channel could one day give the indices the transaction leg they lack — but an administrator whose leadership also runs a trading venue in the underlying, funded by a major trading firm, is precisely the governance pattern IOSCO principles exist to referee. The absence of any published conflicts policy is, at this stage of maturity, the most fixable gap in the whole franchise.

The contextualization this piece set out to build reduces to one sentence: Silicon Data has stopped competing to be the price of compute and started competing to be the terminal on which compute risk is managed — the assessments, the curve, the listed contract, the underwriting model, the assay, the residual data, and both legs of the production margin, from one vendor, at a $998/month entry point. No competitor fields more than one of those layers. The framework in Section 5 is our attempt at the manual the suite ships without: lenders get the most complete toolkit (and the biggest blind spot), corporates get benchmarking and indexation before they get a clean hedge, operators get the natural short, funds get a two-sided data platform with a methodology basis trade attached, and insurers get trigger infrastructure that is quietly already built.

What would change our assessment, in order of information value: the CFTC self-certification filing for the October contracts (the first hard disclosure of the methodology's regulatory posture — and explicit confirmation of the settlement series); a published rulebook or restatement policy at or before listing; any move to add a transaction leg (watch the Compute Exchange channel); an Ornn methodology re-publication that resets the settlement-integrity comparison; the first lender or lessor data partnership anywhere in the complex; and any sign of a utilization benchmark — the product that would finish the map. October 5 is fifty days out. The listing doesn’t just launch a contract; it converts every question in this piece from analysis into P&L.