Aluminum to Silicon

Part IV: Know Your Index, or Why Compute Markets are Tough to Trade and Easy to Manipulate

Ok so here's a classic puzzle in the world of statistics and housing markets: in City XYZ, the cheaper homes (let's say, in the suburbs) were up 10% last year, while the more expensive homes (downtown) were up an even spicier 15%. So how is it possible that the average price of homes was DOWN 21% year over year in this city?

The answer is mix shift. Look up Simpson's Paradox. Average sales price of homes could easily have declined year over year if the underlying mix of homes shifted from predominantly higher priced homes to predominantly lower priced homes (maybe a lot more of them transacted, or more were built). The same thing happens in compute: if a pile of cheaper, older H100 capacity happens to rent in a given hour, a GPU price index can drop even though nobody's price changed.

Every home got more valuable, yet the average sale price fellIllustrative City XYZ. Cheaper suburban homes went from 30% to 80% of sales.20252026$0k$200k$400k$600k$800k$300k$330kSuburbs+10% per home$600k$690kDowntown+15% per home$510k$402kAverage sale price-21% overall
Illustrative numbers

This illustrates one of many reasons why index construction is actually an insanely difficult exercise to do well. In compute markets, as in real estate, the choice of index can matter as much as the asset itself.

So, believe me when I say there's no single thing that is "The Price of a Home" in a given market, instead, there's actually two different ways to measure it: a repeat-sales index (Case-Shiller, for example) or a hedonic index (Zillow Zestimates). A repeat-sales index is what it sounds like: it tracks longitudinally single assets and tracks repeat transactions and how those prices move over time. A hedonic index uses every sale and statistically adjusts each price for the home's attributes (square footage, bedrooms, lot size, location, and so on), so that a 3-bedroom ranch and a 5-bedroom colonial can be translated into the same constant-quality unit.

No approach is perfect.

Repeat-sales indices are clean in the sense that they compare like with like, but a) they only ever see homes that actually trade, and b) they only count a home once it has traded twice. In a market where turnover has collapsed (today's market, for example), no actual person believes that transacted homes are representative of how actual housing stock is trending. Repeat-sales indices also overstate appreciation, especially when owners renovate between sales, since the index can't tell a new kitchen from a rising market.

The cleanest real-world example I know of is San Francisco coming out of the financial crisis: Case-Shiller, which only sees homes that trade, swung from down 32% year over year in January 2009 to up 18% by April 2010, largely because of which homes happened to be trading (foreclosure sales first, then a rush to beat the 2010 homebuyer tax credit), while Zillow's index of every home went from down 15% to roughly flat.

Same city, same houses, two very different storiesSan Francisco home prices, % change year over yearCase-Shiller (homes that sold)Zillow ZHVI (every home)-30%-20%-10%0%+10%+20%+30%2007200820092010201120122013Jan 2009: -32.1% vs. -14.7%Apr 2010: +17.9% vs. -1%Case-Shiller swung about 50 points in 15 months; Zillow moved about 14
Sources: S&P CoreLogic Case-Shiller San Francisco Home Price Index (via FRED) and Zillow Home Value Index for the San Francisco metro, both seasonally adjusted

Hedonic indices, on the other hand, use all the data and can price homes that never trade, but they're only as good as the model, and whoever chooses the attributes and the functional form is, in effect, choosing the answer. Some common blind spots: a) anything the model doesn't measure (condition, view, renovation quality) leaks straight into the "price change," b) the coefficients go stale as buyers' tastes shift (home offices after COVID, for example), and c) they still only see homes that sold, so they adjust for the mix of what traded but not for which homes trade at all.

Compute indices map onto these somewhat loosely. It turns out nobody runs a true repeat-sales index for compute yet. Meanwhile, Silicon Data's indices are hedonic in nature. Ornn is closer to a third type of index, a commodity price-reporting index (think of an oil benchmark or SOFR): rather than adjusting anything, it excludes whatever isn't comparable and averages the realized trades that are left, which makes it, in effect, a "filtered average sale price" index.

The anatomy of a compute index, from data to derivativesIndexWhat goes inHow it's builtWhere money settles on itSilicon DataSDH100RT (hedonic)Trades plus executablelistings, all contract termsBasis-adjusted, modelrecalibrated weeklyCME Group futuresvia NYMEX, slated for JanOrnnORNNH100 (trade avg.)Realized on-demandrentals, one global print1-hour volume-weighted,winsorized averageICE futures (pending),Kalshi, PolymarketRepeat-salesnobody runs one yetSame provider, sameconfiguration over timeBlocked by upgrades,delistings, thin dataNo venue yetBoth live indices are also published on Bloomberg
The two live compute indices and the one nobody has built, from inputs to the venues that settle on them

1. Who's publishing the reference indices, and how

The two names that matter most right now are Silicon Data and Ornn, largely because each is attached to a futures venue.

Silicon Data's H100 Rental Index launched in May 2025, and CME Group's proposed compute futures (listed via NYMEX, now slated for January) would settle on a monthly average of its H100 index. Methodologically it's the hedonic approach:

Ornn's Compute Price Index (OCPI), on the other hand, works like a commodity price-reporting index, and its indices are on Bloomberg (tickers ORNNH100 and ORNNB200), with ICE's proposed compute futures pending CFTC review:

In practice, the two have landed close to each other: in mid-November 2026, Silicon Data had the H100 at about $3.13 per GPU-hour and Ornn at about $3.06 (after settling at $3.29 just two days earlier). That sounds reassuring until you remember how differently they're built, which means the agreement tells you as much about the adjustment model as it does about the market.

It's like we always said internally on the finance team at Ramp: "He who controls the model, controls the narrative."

2. The same blind spots, in silicon

A transaction-average index like Ornn's is the City XYZ problem in silicon: it reflects real clearing prices, which is its strength, but it averages whatever happened to trade in the window, so a shift in which providers, configurations, or regions happen to be renting that hour moves the print even if no individual price changed. Imagine if a batch of early-2023 H100s traded in one hour and a batch of 2025 H100s traded in the next. This hardware is known to command quite significant differences in pricing on the market (As I like to tell my wife, "Not all first-edition Charizards are created equal." This is usually when she walks away...)

A hedonic index like Silicon Data's controls for that mix, but it then depends on the adjustment model.

A true repeat-sales index for compute, which would track the same provider's same configuration over time, probably would be the cleanest of all, except that the market is currently thin, offers get delisted, and providers upgrade their networking and software stacks between observations, which is the compute version of renovating the kitchen between sales.

None of this is a criticism of any particular provider, since every index involves these tradeoffs. It turns out this also makes each of these indices uniquely manipulable.

3. Assume they're being gamed

If there's one thing finance has taught me about benchmarks, it's that any number against which there are real dollars at stake will eventually attract people trying to move it. LIBOR was a contributor-submitted benchmark, and the panel banks submitting it were found to have manipulated it for years; therefore, any compute index built on contributed data (which describes most of them) deserves the same skepticism. Every construction method has its own soft spot: listings can be spoofed, trades can be washed, and hedonic attributes can be mislabeled. How much enforcement exists against bad actors right now? Hard to say. Furthermore, the markets are thin enough that a single large contributor can matter.

This isn't hypothetical. Kalshi has already listed compute contracts on H100, H200, B200, and RTX 5090 rental prices settling against Ornn's indices, and Polymarket has listed Ornn-settled H100 contracts too. That's a real step forward for price discovery, but it also means real money now settles against a single print, and there is every incentive for someone with P&L on the line to "curate" (that's a euphemism) where that print lands.

Five ways to move a compute indexNone require fabricating a trade outrightThe settlement printwhat contracts pay out onBanging the closeReal trades timed into thin hoursSelective reportingLeaving cheaper rentals outRelated-party tradesAffiliates trading off-marketMislabeled attributesGaming hedonic adjustmentsPhantom listingsOffers posted, never filledWinsorization, trade verification, and penalties blunt these, but don't eliminate them
Structural exposure, not an allegation against any index

The better precedent for transaction-based compute indices, though, is probably the 2013 FX fix scandal rather than LIBOR. Transaction-based indices like Ornn's are harder to fake, since only realized prices count. However, the forex traders who rigged the WM/Reuters fix didn't fabricate anything either: they coordinated real trades inside the short window when the benchmark was calculated. A compute index calculated from a rolling one-hour window in a thin market has the same exposure, and therefore anyone holding a futures position worth more than the cost of renting GPUs at off-market prices for an hour has an economic reason to move the print.

Anatomy of an hourly fixIllustrative. The thinner the settlement hour and the bigger the bets on it, the cheaper the print is to move.Spot volume in thesettlement houra handful of GPU rentalsMoney settlingon that printfutures and event contractsMoving the print pays when the P&L on contracts tied to itexceeds the cost of renting GPUs off-market for that one hour
Illustrative mechanism, not a measurement of any index

None of this means anyone's doing it today. The tell to watch is whether the volume of futures settling against these indices starts to dwarf the spot volume underneath them, because once the money riding on the index is much larger than the money it costs to move it, every one of these vectors gets cheap, and that's how benchmark manipulation has started in most markets before this one.

A quick word on winsorization. Winsorizing is a way to stop a few crazy numbers from dragging an average around without throwing them away. That's great protection against fat fingers and one-off junk prints, but it's much weaker against a determined participant. Pushing lots of trades a little higher, all inside the band, passes straight through; trading enough volume at your own price drags the band itself toward you, since the band is calculated from that same volume; and in a thin hour with only a handful of trades, the percentiles rest on so little data that one participant can effectively be the distribution.

Winsorization stops one bad print, not a determined traderIllustrative trades in one hourly window, $ per GPU-hour. Fair average: $2.85.winsorization band$2.5$2.7$2.9$3.1$3.3One fat-finger print at $20index prints $2.87A determined trader pushing inside the bandindex prints $2.92$20 print, clipped to the edge
Illustrative trades and band; Ornn doesn't publish where its bands sit

Token indices are the clearest example of this today, and in my view they're already distorted. Most of them are usage-weighted, which makes them the "Case-Shiller problem" in its purest form. Silicon Data's LLM Token Expenditure Index (Bloomberg: SDLLMTK), for instance, has slid to about $0.80 per million tokens as of mid-November 2026, roughly 60% below its May peak of about $2.10, and Silicon Data's own page notes that the index moves even when individual model prices are unchanged, because a usage-weighted index falls whenever usage shifts toward cheaper models, regardless of whether any individual model got cheaper. And usage has shifted dramatically, as providers flood routed volume with free or heavily subsidized "stealth" test models that regularly top OpenRouter's rankings before anyone knows who made them.

Some indices are built to filter that out; Ornn's Token Price Index, for example, excludes free tiers and thin-coverage days, but it's still volume-weighted within each lab and only sees traffic routed through OpenRouter, priced at posted rather than negotiated rates. Meanwhile, list prices per token actually rose on most model tiers over the past year, even as the cost per unit of capability fell sharply, so depending on whether you weight by usage, list price, or capability, you can tell three completely different stories about whether tokens are getting cheaper.

Three stories about the same tokensIs AI getting cheaper? Depends entirely on how you weight it.Usage-weighted spendSilicon Data SDLLMTK$0.80 / M tokensMid-Nov 2026: about 60%below its May peakList price per tokenposted rates by model tierUp on most tiersOctober 2025 toOctober 2026Cost per unit of capabilitysame benchmark, cheaper modelDown sharplyone tracker puts it atroughly 50x cheaper
Sources: Silicon Data (SDLLMTK), Synthorai LLM API price tracker

4. So What Now?

If you're lending against GPUs, hedging compute costs, or writing an index into a lease, you should assume someone is going to try to screw you at some point, which is of course the credit investor in me talking. Keep in mind the following:

  1. Know exactly which index you're referencing, whether it's built from listings, transactions, or a hedonic model, who contributes the data, how it's weighted, and how outliers are handled
  2. Of course, backtest. A classic ratings agency exercise is basically to figure out what happens in a P90, P95, or P99 scenario.
  3. Reference more than one index, since Silicon Data and Ornn measure different things, and a divergence between them is information rather than noise.
  4. Write fallbacks into the documents from day one, because the LIBOR transition showed how painful it is to retrofit a replacement benchmark into thousands of contracts after the fact. I spent more time than I'd like to admit negotiating LIBOR replacement language during the early days of Ramp, first against Goldman Sachs, then against Citi. In retrospect it was worth it.
  5. Prefer quality-adjusted measures for anything long-dated, which ties directly back to why physical hardware grading matters: a hedonic compute index is only as good as its attributes, and longitudinal physical condition is the attribute nobody is measuring yet.
Index diligence checklistBefore you lend, lease, or hedge against a compute indexCheckRed flagInstitutional standard[01]Construction"Proprietary" inputs and weightsInputs, weights, and outlier rules documented[02]Tail stressUnderwritten to the base caseBacktested to P90, P95, and P99 outcomes[03]Dual feedOne index, one source of truthTwo indices, with divergence treated as a signal[04]FallbacksNo replacement languageFallback benchmark written in from day one[05]Hardware gradePriced off SKU and age aloneQuality-adjusted, condition-aware marks
The five takeaways above, as a term-sheet checklist

Housing investors learned a long time ago that the index you quote is a choice, not a fact. Compute investors are about to learn the same thing.

Anyway this was a fun little detour. We're going to come back to the world of ABS and GPU debt deals in the next few blog installments. Need to circle back on a few topics within the world of capital markets infra etc.