Market structure · Index & instrument design · Trading · Risk
A research and development practice for markets that are still being built.
What should the reference price measure, how should the contract settle, how is it collateralised, and who actually uses it once it lists. Those four questions decide whether a new market works — and in compute, tokenized collateral and event contracts, all four are still open. The research is published in full, and the analytics that produced it ship alongside.
Where the work concentrates
Three threads, and they keep meeting.
Each thread is a single argument built across many pieces rather than a folder of posts — and the interesting findings sit on the seams. The compute forward curve was extracted from event-contract ladders. Compute credit is financed on the tokenized collateral rails the settlement work tracks. The one place a cross-venue basis has no index basis inside it is a prediction market pricing GPUs.
Energy & Compute
Compute is congealed electricity. Almost none of it is priced that way yet.
The first new major commodity in decades is being financialised out of order — futures listed before benchmarks are governed, indices published before their denominators are measured, credit extended before any of it can be marked. Twenty-one pieces built as one stack: what a GPU-hour physically is, where its electricity is priced, who defines the benchmark, how paper converts into racks, what the tokens on the other side are worth, and who is financing it.
AI Invest
Series in progressThe one market here where the unsettled question is whether a machine may hold discretion at all.
The SEC withdrew its only AI-specific rule proposal, the banking agencies replaced the model-risk standard with guidance that explicitly excludes generative and agentic systems, and no regulator has specified a required control set for an agent that trades. Five parts published — the autonomy ladder, what nine platforms’ filings actually say, where the return really shows up, the enforcement layer nobody has built, and the number you cannot trust — plus the diligence tool they produced, free to download.
Newest first
Published work.
Every piece runs on primary data — settlement records, contract texts, regulatory filings, measured traces — and ships with an interactive tool, so the assumptions are drivable rather than buried. Modelled figures are labelled as such, and sample limits are stated before a critic finds them.
Open Source Is Not an Audit Trail
NATIVX open-sourced a complete compute-benchmark engine under Apache-2.0 — six providers, six aggregation methodologies, an ablation stress test, a provenance ledger, checksummed point-in-time snapshots, twenty-six passing tests. Every line is inspectable, which makes it the right subject for the one test that separates a benchmark from a number on a screen: take a published value, take the constituents published beside it, apply the stated methodology — does the number come back? It does not. Ten of the thirteen published indices equal a single vendor’s on-demand list price to the cent. Two match no observation anywhere. The twenty-five values in the historical series the dashboard charts as settlement prices have no constituents at all, and the checksums presented as integrity seals are not hashes of anything present. The forward starts 46% above the index it forecasts and falls at a quarter of the rate of the platform’s own published history. Transparency is necessary and nowhere near sufficient — what is actually for sale is governed reproducibility.
Same Algos, Same Book. Not the Same Margin.
Three announcements inside thirty-six hours. CME entered the wind market. FalconX and Kemet put Kalshi event contracts into the same execution and risk layer institutions already use for options, perps and spot — same algos, same book, same risk model. And Kalshi named The Weather Company its settlement source for a climate and weather vertical growing 500% year over year toward $1.1 billion annualized. The middle one is the genuine leap: workflow was the binding constraint and now it mostly isn't. But a risk layer is worth what the balance sheet behind it can carry, and the risk arriving in that book isn't digital-asset risk at all — it's weather and commodity risk wearing an event wrapper, sharing a factor with none of the crypto book. Its natural offsets sit at CME, at Nodal, in reinsurance trusts. Three collateral pools, no bridge. A dealer hedging a client across two of them funds both legs gross. The hedged book pays for its own prudence twice.
Five Indices, One Price.
Five providers now publish a compute price, and they are all publishing the same one: the flat on-demand rent of what is commercially the residual market. Underneath sit four series that decide whether anything can be settled — an attested meter, a term curve from disclosed prints, a graded token price, a settlement-grade residual — and not one exists in a form a contract could reference. They fail differently, which matters: residual value exists in three incompatible constructions at once, term structure exists but not from disclosed prints, and utilization is the single flat absence, with no series gated or free and no methodology for one. There are six working models of benchmark administration; the four compute has adopted are precisely the four that need no contributor to agree to anything. The two it lacks are the two that require negotiating with someone who has something to lose. Which is the real deficiency, and it isn't data. Appendix C asks exchanges to verify their index provider minimizes the 'opportunity or incentive to manipulate' — and incentive isn't addressable by policy. August produced the first two exceptions, and they bought different things: Compute Desk's benchmarks are now administered by GX Benchmarks, an FCA-regulated arm of General Index with no position in compute, which buys independence but pushes the test down onto contributor concentration; NATIVX open-sourced FSKU, which buys reproducibility but takes posted rates in and puts a modelled forward out. Appendix C also suggests administrators publish the names of their sources. Still not one does — FSKU names its sources because they are public price sheets that cost nothing to name. The insurance industry ran this experiment: ISO in 1971, independent board in 1995, for-profit in 1997, Verisk in 2009. That order is the lesson — along with the antitrust case nobody mentions.
The Compute Crack Spread.
The futures listing October 5 settle on rent, which barely moves — under 1% a week, and the hyperscaler series unchanged on 39 of 50 trading days. The price that does float, what a token sells for across a hundred providers repriced in hours, has no index at all. So we built it. A refiner hedges the crack, not crude and gasoline separately; a generator hedges the spark spread. The compute version is C = T·p − r, and the market heat rate H* = r/p is what a fleet must produce to break even at a venue's price. From a live endpoint book: two thirds of the active market clusters at 3.5–3.6M tokens per GPU-hour, which a saturated H100 covers three times over. At 10 requests per second almost nobody clears. Utilization sets the sign, not price. And the CFTC's August 19 objections — opaque bilateral price formation, undisclosed transaction prices — are true of the rent leg and false of the token leg.
The Contract Has a Date. It Needs a Dealer.
CME lists GPU futures October 5. The question nobody has answered in public is which commercial balance sheet carries a floating exposure to a GPU-hour that these contracts can offset. The answer, from the filings and the calls: none does. Every participant with a real exposure prices it fixed — two-to-five-year take-or-pay, term debt sized against contracted revenue, the merchant tail routed away from the senior tranche. Zero named hedgers on the public record, and CoreWeave's Q2 call never says futures, hedging or index. Nor does the filing say which Silicon Data tier settles, though the tiers differ threefold. History says what comes next, and it runs backwards from here: WTI deepened because banks warehoused bilateral risk and laid the residual off on the screen; Henry Hub deepened because the cash market was unbundled first. Compute has neither a floating physical leg nor a dealer — but it has something oil never did at this stage. The lenders already own the merchant tail.
The Long End Has Two Sellers Now.
The AI build-out is the largest private borrower of duration in history, and it arrived exactly when the Treasury needs the same buyers most. Take the debt-crisis argument as a conditional and ask the question the macro literature skips: if it happens, how does it reach a data center? Not the way you'd think. Meta's 40-year is fixed at 5.75% — a +150bp move costs the bondholder 18–20% and makes Meta a winner. GPU debt floats against SOFR, which a term-premium shock with the Fed on hold doesn't move. The exposed parts are narrower: new projects, where debt service per MW rises 13–32%; the 2027–29 refinancing wall, where spread beats base rate; and tenant credit. Then the worked case, which reorganizes the question entirely. A 100 MW merchant project breaches 1.0x coverage in year four or five under every regime — including today's, with no shock at all — because 22% annual rental deflation meets a five-year amortizer. Rates take the equity from 16% to negative but don't cause the breach. It dies of deflation. Rates are the second derivative.
The Other Side of the Spark Spread.
OpenAI is hiring a power trader — the first role at a frontier lab that reads like a merchant desk rather than a procurement seat. A generator is a call option on the spread between power and fuel. A lab has the same spread with an extra leg and the sign reversed, and the second heat rate is not physics: it is utilization. At $45/MWh, power is $0.069 per GPU-hour — 2.7% of a rented H100 and 0.8% of saturated token revenue — while the same chip's inference spread swings from −$0.29 to +$6.15 purely on how hard it is run. Which means the seat is not hedging margin at all. It is hedging absolute dollars and the tail: roughly $700M a year per $10/MWh at 10 GW, against summer evening hours that print $205/MWh where overnight prints $21. A generator sells those hours. A flat load buys every one.
The Other Side of the Block.
Cantor just opened institutional block trading in event contracts to roughly 3,000 clients. Cantor is the pipe, Kalshi is the venue, and Susquehanna is the balance sheet — the firm that quotes a single price for the full size and owns the risk until it resolves. That third role is where the economics live, and it splits the business in two. An FOMC block back-to-backs almost exactly in fed funds futures: 2,000,000 contracts at 63¢ against ~1,920 ZQ locks about $20,000 in either state, and the binary disappears. A 500,000-contract Super Bowl future cannot be laid off at any price worth paying, and the two cents of skew stops being a dealing spread and becomes an underwriting premium. So the moat is not the pipe — that gets competed within months — it is the willingness to warehouse a risk with no hedge, which can only be bought by losing money first. On June 10, one Knicks tip-in took $22.4 million out of the market-making community in a night.
The compute risk premium got its first print.
Two Johns Hopkins economists have put the first number on what the long side of compute gets paid — high single to double digits, paid by the fleets to whoever will carry their price risk. The framework is imported wholesale from electricity forward markets and it is the right import: nothing can store a GPU-hour, so cost-of-carry is dead on arrival and hedging pressure sets the sign of the premium instead. Their sharpest conceptual contribution is a name for something the corpus lacked — the physical access wedge, the gap between a reserved rental (which bundles a capacity-locking option) and a cash-settled future (which carries insurance and nothing else). The framework is right. The map is already out of date, and four corrections follow — starting with the market the paper calls unlaunched, which has been trading since May on curves that carry no wedge at all.
The contest can't be hedged. The financing can.
The BIS wrote down a model in which the AI build-out is rationally oversized by 40–50% — not a mania, a contest, where over-investing is each player's best response to the others — put roughly even odds on a bust, and priced the damage at $378 billion. The practitioner's question the paper deliberately skips: can anyone hedge this? The answer falls out of the paper's own logic. In a contest, commitment is the strategy and a hedge is the opposite of commitment, so a hyperscaler shorting compute futures against its own build-out would be visibly undoing the signal its capex exists to send. The racers are unhedgeable by choice — and empirically they sell protection rather than buy it. The demand migrates one layer out, to everyone who financed them without a ticket to the prize, and lands in a market roughly 120 times too small to absorb it. Four loss channels mapped to four instrument families, with honest residuals where nothing settles.
The speculation sleeve.
Half of Gen Z investors moved money meant for investing into sportsbooks last year, and a quarter now call it part of their long-term financial strategy. The industry's answer — stop betting, buy index funds — is losing to a product that pays out Sunday night, and fifteen years of research says voluntary restraint tooling barely moves net losses. So the question becomes structural rather than moral: if a generation insists on holding binary risk, what is the least destructive venue for it? Prediction markets are the only one where the house edge is a fee schedule rather than a business model — around 1.75¢ on a 50¢ contract against a 10%+ sportsbook hold. The honest half gets equal weight: settled-contract data shows takers losing ~32% of stake, because retail crosses the spread for longshots that pay ~2% of the time. Cheap structure, expensive behavior — and closing that gap is the entire design problem.
The index was the wedge. The product is the risk desk.
Silicon Data entered the compute race as one index among four. It is now an eleven-product platform covering spot, the curve, the listed hedge, asset underwriting, the assay, the resale market — and, uniquely, both legs of the AI production margin. That is not an index provider; it is the anatomy of a commodity-market data franchise, assembled in advance of the market it serves. This audits the suite, then maps seven risk exposures to products across five participant classes — with an honest routing table naming where competitors win, because a framework that never routes elsewhere is marketing. Three caveats carry through everything: the inputs are quotes with no published rulebook, the 2.6× tier basis means the futures settle on one member of a family of prices, and one vendor now supplies the index, the model and the settlement print. The finding is the gap: nobody, anywhere, publishes observed utilization — so rates can hold firm while fleets idle.
The contract got a date.
The first US compute futures — H100 and B200 rental-index contracts, each a month's worth of GPU rent — list on NYMEX on October 5, pending regulatory review. An index is an average over dispersion, so the piece builds the ladder the contracts must navigate before asking how settlement works: 1.4× same-chip performance variance normalized away inside a proprietary model, a 2.0× grade spread that becomes tradable on day one, a 2.6× gap between neo-cloud and hyperscaler H100 that is published as sub-indices but sits outside the settlement sample, and up to 7× across venues on identical open weights. That third rung is the design choice nobody is discussing: hedging a $7.19 hyperscaler bill with a $2.73 instrument is a correlation bet, not a hedge. Plus Carmen Li on the record from "very TBD" to a listing date eight weeks later, and a three-bin account of whose index documentation is actually public.
Compute became collateral. NVIDIA sold the floor.
Six of the largest alternative-asset managers in the world signed MOUs with NVIDIA to mobilize over $500 billion of third-party capital into vehicles that issue debt secured by GPU compute itself. But that half-trillion is money NVIDIA does not supply, promised over time, under memoranda that bind nobody until final agreements are executed. Read the release for what NVIDIA commits itself to and exactly one figure survives: residual-value support on up to approximately 25% of an opportunity, case-by-case — and what that describes is a written put on the depreciation of NVIDIA's own hardware, the instrument aircraft manufacturers have sold for decades to move metal through leasing channels. Half a trillion dollars of paper, and the mark it needs does not exist yet. Though the oldest objection to compute derivatives — where is the natural short — just acquired six logos.
The metal keeps calendar time. The contract ticks.
Kalshi didn’t list an asset. It listed a clock. In 2002 Emanuel Derman argued that short-horizon speculators don’t perceive risk in calendar time — they count trading opportunities, and capital chases an asset’s temperature: volatility times the square root of trading frequency. A quarter-hourly grid quotes the exchange rate between those two clocks in public, 96 times a day. But the grid is not the heat. Intrinsic time is endogenous; the exchange’s ruler is flat and imposed, which is exactly why it measures the terrain. What the product does create is a third clock — settlement time, 96 synchronized forced resolutions a day — and the temperature it raises belongs to the trader, not the metal: 1.75¢ feels like nothing at 2:15pm and compounds into everything by December.
One roof, four registrations, seventy-seven questions.
The vertically integrated exchange asked the CFTC for legitimacy for three years. On July 30 it got it — attached to an invoice. The proposal doesn’t break up the exchange-clearinghouse-broker-trading-desk model; it writes it into the CFR and sets terms. The sharpest is a mandated disadvantage: an exchange’s affiliated market maker may keep quoting, but its orders are filled last at every price level regardless of time priority — a tax landing exactly where affiliated liquidity claims its value. Mapped across thirteen corporate families, the bill is startlingly uneven: heavy on the two venues whose liquidity is affiliated, near-zero for the one that outsourced clearing and the one that outsourced liquidity, and close to costless for CME, whose own voluntary practice is what’s being codified.
Perp the S&P. Sue the format.
On August 17 a perpetual-style S&P 500 contract lists on a CFTC-regulated US exchange at up to 20x, with a filing pending to let traders post a stablecoin as margin — eight weeks after the benchmark's owner sued the regulator to have the entire format declared swaps rather than futures. Three fights are running at once: what a future legally is, who may license the index, and whether a margin obligation that never sleeps can be met during the 58 hours a week Fedwire is closed. CME's own legal theory carries a boomerang — win the case, and S&P DJI may be free to license the S&P 500 'non-future' to everyone. Meanwhile the S&P cash market prints 32.5 of 168 hours, so for most of the week a near-continuous S&P perp anchors its funding to something other than the thing it tracks.
The coin can't pay yield. The fund behind it can.
In seven weeks the largest cash managers launched three purpose-built stablecoin-reserve money market funds. Only BlackRock's is tokenized, and that difference is the story: the reserve asset now lives on Solana, Ethereum and Tempo — the same rails as the coins it collateralizes. GENIUS enumerates what a reserve may hold and simultaneously bars the coin from paying what that reserve earns, which is exactly why the yield migrates one layer down to the fund. Four days before the launch, the firm that keeps the register for BUIDL, Apollo, KKR and VanEck added its fifth registration, an SEC adviser licence — completing the first fully regulated vertical stack and leaving a private registrar performing Cede & Co.'s function without Cede & Co.'s oversight regime. Closes with five ways it reads as less than it looks.
Soft target, hard target: the event-perp feedback loop
Can intraday crypto event contracts and the perpetual complex bend each other's prices? Yes — but not equally, and not in the direction most people assume. Both sit on one spectrum: how manipulation-resistant their settlement print is. Kalshi averages 60 seconds of a per-second multi-exchange composite; Polymarket's short-dated markets ran on a near point-in-time Binance-heavy oracle until it became a 30–60 second TWAP on 7 August; perp liquidation marks are medians of capped multi-exchange indices, engineered so a single venue cannot move them. So the popular framing — move the prediction market to trigger liquidations — fails the arithmetic before it starts. The reverse works, and published evidence documents it. The contribution here is the leg that literature does not cover: liquidation prices are public on-chain, so a trader need not move the composite the whole distance — push spot into a visible cluster and other traders' forced selling carries the print through the strike.
Markets went continuous. Central-bank money didn’t.
CLARITY is stuck on the Senate floor over an ethics provision, and the sector is treating that as the story. It isn’t — the institutional stack was built by agencies without it, from stablecoin trust charters to a tokenized ETF posted as margin at CME. What is actually pulling stablecoins and tokenized collateral onto institutional balance sheets is the Federal Reserve’s operating calendar: US markets now trade 58 hours a week with no central-bank settlement rail open, and the Fed’s own 2028–29 expansion closes only 22 of them. Saturday never opens. Float grew 11% while adjusted volume grew 125%, which explains the yield war; the collateral map explains the rest — including the clearinghouse that considered tokenized collateral and said no.
Volta absorbs, Trillium distributes
In the same week, two announcements opened compute-market access to participants without investment-grade balance sheets — and they are close to opposites, routinely filed under the same headline. Volta is a purpose-built balance sheet that absorbs the credit and utilization risk unrated AI startups cannot carry: the startup sheds the five-year take-or-pay, and the take-or-pay moves rather than disappears — onto a stack carrying a reported $10B anchor against $300M of equity, on silicon no index prices. Trillium runs the pipe the other way, packaging a platform's prepaid compute credits — a wasting asset with a redemption rate and an expiration window — into marketplace notes whose current documents say plainly they are not secured. The exposure ladder each creates, the valuation gap both share, and the hedge program each side would need.
The number you cannot trust
Forty-five independent strategy variants are enough to manufacture a Sharpe ratio of 1.0 out of five years of pure noise. An afternoon's work, a thirty-year-old result, and almost universally ignored. Part IV specified what an AI system may do; this asks whether it should be there at all. In April 2026 the agencies rescinded SR 11-7 — the document every AI governance policy in US financial services was written against — and replaced it with guidance that excludes generative and agentic AI by name and is voluntary by design. Deflated Sharpe arithmetic, an even-handed reading of a genuinely contested replication literature, evaluation design for a system with no track record, the judge problem, and what drift means when a pinned model name is not a pinned model.
The token leg has a price screen. Nobody has built the index.
A March 2026 paper designed a complete futures contract for AI inference tokens — standard grade, volume-weighted settlement index, margin rules, circuit breakers — and never named a venue that could produce the index. The venue exists: it clears roughly 25 trillion tokens a week, publishes a live multi-venue order book with no login, and its public API is almost line for line the feed the paper requires. We pulled it — 56 quotes across five model books, 101 venues on the screen and 100 by that afternoon — then tried to build the index, with a widget that lets you set the administrator's choices and watch the print move. It breaks in four specific places, starting with a 7× output-price spread on identical open weights.
The shield
Interactive Brokers has the richest pre-trade risk control surface available to a retail investor, and in July 2026 it opened an MCP endpoint to every major AI client — where the agent may analyse, monitor and draft, but not execute. Robinhood, Alpaca, Coinbase and eToro do let an agent trade, and between them this piece could not find a single documented position cap, notional cap, rate limit, loss limit, drawdown breaker or restricted-instrument list. The broker with the controls does not allow execution; the brokers that allow execution do not have the controls, and nobody has connected the two. Part IV specifies the deterministic layer that belongs between them — how an investment policy statement compiles into constraints an agent cannot argue with — on the template the Market Access Rule established after Knight Capital lost $460M in 45 minutes.
Twenty minutes a week
A large RIA reported its internal AI assistant saves five thousand hours a year. Divided across the advisor base and the working year, that is twenty minutes per advisor per week — which does not become a new client, it becomes a slightly less rushed Tuesday. Efficiency here is almost always quoted as hours per firm per year, the one unit that makes a small number look like a department. The advisor stack audience by audience: the jagged frontier by category, the data layer as a gate rather than a nice-to-have, the supervision and recordkeeping obligations that attach the day a tool is switched on, build-versus-buy by firm size, and vendor risk reframed — absorption was the wrong thing to fear; the real risk is a $100-a-seat notetaker quietly becoming your system of record.
Read the filings first
Part I promised a hands-on evaluation of the retail AI advisors. Pulling their Form ADVs, Form CRSs and advisory contracts first changed the exercise, because the gap between the marketing and the paperwork turned out to be the finding. A platform marketing $40B of assets on platform reports zero regulatory AUM on its current brochure — the same figure the SEC litigated in its first AI-washing order, when it stood at $6B. A platform whose brochure disclaims individualized advice takes full discretion for 49bps. One adviser registration is simply inactive. None of the seven AI-native platforms splits risk capacity from risk tolerance, a distinction both pre-AI robos make. Nine platforms scored on documented investor-protection posture — explicitly not advice quality — plus the full hands-on protocol, and why terms-of-service prohibitions on automated access leave it unrunnable by anyone outside the firms.
Two venues, one index
Kalshi and Polymarket both list GPU rental price contracts, and both settle against the same Ornn dashboard URL — named in both contract texts, verified directly rather than inferred. That makes the spread the first live compute spread with no index basis inside it. It is still unreadable: matching the contract structure collapses the headline gap from $2.17 to $0.10, smaller than the bid-ask on one venue alone, and both ladders violate their own internal arbitrage conditions before any cross-venue comparison begins. Remove the disclosure problem entirely and the basis is still uninterpretable — index disclosure is necessary, and visibly not sufficient.
The inference spark spread
On 30 July a frontier lab cut one model tier's published price by 80%, three weeks after launch. Over the past year the neocloud H100 rate that tier is served on roughly doubled off its trough. Those two facts are in direct tension, and the tension has an exact form — a spark spread, the structure a power trader uses to price a generator's margin as the power price less the gas price times the heat rate. Written that way, inference reduces to a single unobservable term, tokens per GPU-hour, and the arithmetic inverts to a threshold: the serving efficiency a rate card requires in order to clear. Nobody publishes that heat rate, so we measured it — 1,024 vLLM serving runs on Llama-3.1-70B across four H100s, joined sample-by-sample to the NVML power traces recorded alongside them. The July cards raise the required efficiency 5×, and the sign convention makes labs and resellers natural counterparties on the token leg.
Verifiable is not replicable
The administrator catalogue closed on a stark finding: across the four providers positioned to bear settlement weight on CFTC-track compute contracts, not one publishes a rulebook, an IOSCO statement, an audit, or an oversight committee. A fifth has now surfaced from entirely outside that perimeter — a per-model GPU-hour benchmark on Solana, run by a protocol that is at once the rental marketplace, the derivatives venue, the index administrator, and the issuer of the token those derivatives settle in. It discloses more of its construction than any of the four inside the perimeter, including the only quantified bound on settlement circularity among the five. And it still cannot be reproduced by anyone who settles against it.
From 60/40 to autonomous agents — AI in wealth management, Part I
Part I of a new series. Portfolio construction has moved through three regimes, and the technology now arriving serves three very different buyers with three very different tolerances for autonomy. This installment builds the taxonomy the rest of the series runs on — the autonomy ladder, from copilot through advisor and delegate to autonomous agent — maps the current product landscape onto it with an interactive explorer, and closes on the question that will decide institutional adoption: the control problem of proving an AI portfolio process stays inside mandate.
Hedging the inference book
Inference providers — the Perplexity and Poe class of business — run a commodity retailer's book: fixed-price subscriptions downstream, floating token costs upstream. Two instrument families now exist to hedge it, and they settle in different units: token forwards in $/M-token against the routed API share, GPU compute futures in $/GPU-hour against the self-hosted share. The spread between the units is the inference-efficiency curve. A worked example evaluates and optimizes the mix — 91% variance reduction for the two-unit program against 57% for GPU futures alone — with an interactive optimizer and the case for capped structures over fixed strips.
How compute index providers actually calculate price
Four providers are positioning to become the settlement benchmark for GPU compute — posted rates at scale (Silicon Data), invoice-verified prints (Ornn), market-implied forwards with H100 conspicuously excluded (Kalshi), and an undisclosed physical-delivery reference (Compute Desk). The component-level catalogue: what each publishes, how each price is calculated as far as public record allows, who feeds the data, and the finding that unites them — zero published rulebooks, no utilization input anywhere, and a Dec 2025 restatement that moved one index's entire history +40%. The CFTC filings expected this half are the forcing function; this piece fixes the baseline they will be measured against.
Hedging the GPU balance sheet
A specialty compute lender is short one price twice: borrower revenue is the GPU rental rate, and the collateral is the same rate capitalised. Program design worked against the live venue landscape — cash-settled core with the EFP option held for the recovery state, benchmark-tier standardization with configuration basis charged rather than hedged, and a five-check index validation program that requires no administrator cooperation. The power leg is grounded in ERCOT and Dominion settlement data, with an interactive dashboard of the two risks a flat annual hedge cannot see.
Data centers are not a flat load
Utility filings and hedge desks model data-center demand as a flat 7×24 block. Measured H100 facility profiles are not flat at any utilization below saturation — and joined against 401,379 hourly DA settlements across eight ERCOT and PJM locations, including a fresh pull of the DOM zone, the load-weighted power price runs up to 9.4% above flat-block. The premium peaks at 40% utilization — the ramp phase every new facility passes through — and flips sign between an ERCOT summer and an ERCOT winter. The shape basis in every flat hedge, quantified, with a live join engine across all eight locations.
The parlay is a correlation market
Every combo price minus the product of its legs is an implied correlation — printed thousands of times a day, quoted by nobody. The partition rule applied to the markets that already trade: this week's Fed ladder against the BTC price ladders, a World Cup lookback in which jointly impossible finals conditionals met in the final, and $100M a week of parlay flow read as the correlation tier that arrived at retail before any benchmark exists to hedge it.
Implied compute forward curves
Kalshi's compute contracts do not settle against Kalshi. They settle against Ornn's Compute Price Index — and that one fact is what turns a prediction market into a price curve. Each rung of the ladder is a digital option on an externally administered benchmark, a strip of digitals is a discretised probability distribution, and a distribution has a mean. This works the construction end to end: ladder to density to forward, the convenience yield it exposes against provider term sheets, and a two-factor hedge calibrated rather than assumed. Then the governance question underneath it — one administrator now carries four instruments plus the deepest event-contract ladder in the market, and has published no calculation.
Hedging corporate event risk inside a hierarchy
Liquidity cannot survive being spread across ten thousand single-name event contracts — so concentrate it in a small number of parametric benchmarks and sell the idiosyncratic residual back as basis. That is how industry loss warranties and CDS indices already work, and this piece applies the pattern to corporate event risk across six families: pre-release content compromise and mass-tort litigation built out in depth, then M&A deal-break, cyber, recall, and approval risk. Underneath it: the CFTC event-contract regime including the unlawful-activity prong, ILW and cat-bond structure, index-versus-single-name liquidity, and market-scoring-rule sizing.
The denominator problem — what a GPU-hour actually costs in kilowatt-hours
Every compute-to-energy conversion in circulation multiplies nameplate TDP by an assumed PUE — including four dashboards on this site. The National Laboratory of the Rockies measured it instead, at 0.1-second resolution across training, fine-tuning and inference, and published the traces, the facility series and the tooling under CC-BY. Rebuilt on that data: real workload duty factors run 0.795–0.880, so nameplate overstates device energy by 12–21% (not the 4.5% gpu-burn gap everyone quotes); facilities peak at 73–80% of rated IT power, and peak sizing versus energy volume take different derates that must not be multiplied together. An energy-normalized compute index doesn't remove assumptions — it relocates them from power geography into workload mix.
Kalshi files metals perpetuals — and leaves the ratio leg on the table
Kalshi filed with the CFTC to list gold, silver, and platinum perpetual futures — its first expansion beyond crypto, on a 45-day review clock, filed onto CME's home turf while CME's lawsuit against the perpetual approval is live. The contract it didn't file is the interesting one: a Gold/Silver Ratio perpetual would complete a no-arbitrage triangle with the two legs Kalshi just filed — reference prices, funding anchor, and arbitrage discipline all from contracts inside the building. The margining case (ratio vol vs two gross legs, with the stressed-correlation caveat), the volume case (the FX-cross precedent), and the funding identity that glues the triangle together. Interactive throughout.
The Street just got a second settlement layer
Ondo's Oasis Pro Markets received FINRA authorization to offer tokenized NMS equities, ETFs, funds, and IPO allocations to US investors — a registered claim structure anchored at the transfer agent, not an offshore wrapper. The interesting part isn't trading access, it's collateral mobility: a long equity position that is simultaneously a tradable asset and a programmable collateral object. What that means for prime brokerage — financing disintermediation, the wrapper-basis haircut, sec lending, netting, the dual-book seam — plus an RWA tokenization timeline and the token-vs-SSF-vs-offshore-perp comparison.
Paper becomes racks — ComputeConnect and the first compute EFP
Architect and Compute Desk are building the first compute exchange-for-physical network: CFTC-regulated futures on H100/H200/B200/B300 rental indexes that convert into real GPU capacity via Compute Clear, with published basis tables by SKU, memory configuration, and location. EFP mechanics are the connective tissue between paper and physical in every mature commodity — the prerequisite for genuine hedger participation rather than purely speculative flow. Read through the crude, gold, gas, and metals precedents, with an EFP lifecycle explorer, basis-table simulator, and convergence lab. Companion piece: ICE × NATIVX's energy-normalized COIL contract, stress-tested against 374,550 hourly power prints.
Compute per unit of energy — ICE × NATIVX and the COIL Index
ICE's second compute-futures act lists GPU compute on NATIVX's energy-normalized COIL Index alongside the gas and power complex it already clears — the first contract design fusing the compute and power risk stacks. Stress-tested against 374,550 hourly ERCOT/PJM settlements: the energy content of a GPU-hour, the normalization-vs-shape arb, and the compute heat rate.
The market with no exit — CXMT's pre-IPO perpetual at a 526% access premium
Hyperliquid's xyz:CXMT perp extends the SpaceX pre-IPO playbook to a restricted foreign equity — offshore synthetic price discovery for an asset US and Chinese investors cannot directly access. No borrow, no delivery, no cash leg: a 526% gap that normally invites arbitrage, with the obvious trade structurally unavailable. Funding rates, mark price, and the oracle handoff to STAR Market × USD/CNY dominate short-term risk — a live liquidation-cascade and oracle-risk case study, with an interactive cascade lab and a full dashboard blueprint.
The power market goes hourly — Nodal's 168 hourly futures, ElectronX's bounded hours, and three years of the 24-hour curve
On August 31, Nodal Exchange lists a futures contract for every hour of the day at seven hubs — ERCOT North/South/West/Houston and PJM Western/AEP-Dayton/N Illinois — while ElectronX already trades the same hours as bounded, fully-collateralized futures and DA-strike binaries, and ICE's TB4 future settles the battery spread outright. Underneath them: 374,550 hourly settlement prints, a summer evening hour that averages 10× the overnight hour at ERCOT North, TB4 distributions with a $3,133 three-sigma day, and a day-ahead premium that makes the flat-priced binary wrong at every hour. Spread monitor with 99.7% bands, block economics, and the load/temperature spike maps — six-tab dashboard inline.
Four curves, one commodity — Kalshi's implied compute curve, and the race to price the GPU term structure
Kalshi launched compute forward curves on July 14, 2026 — binary event ladders settling on Ornn prints, the first liquid, executable forward pricing in the compute complex, live before either announced futures contract has listed. That makes four venues on three curve technologies and two settlement philosophies: event-implied (Kalshi), perp-funding + EFP futures (Architect), quote-assessment term curves (CME × Silicon Data), and transaction-VWAP futures (ICE × Ornn) — and three of the four settle on Ornn-family indices. Launch-day implied forwards from the real strike ladders (H100 ≈ $2.52, B200 ≥ $7.00, a stale monthly ladder pricing an implausible 24% two-week collapse), an eight-entry arbitrage monitor from the desk-plan taxonomy, and the compute spark spread with both legs finally quoting.
Single-stock futures vs. the swap desk — a balanced threat assessment for the $34.5B prime & financing franchise
CME lists the first US security futures since OneChicago on July 27, 2026 — into a year when equity perps went live on a US exchange, single-stock perps hit $62B/month offshore, and the SEC-CFTC opened the portfolio-margining question. Anchored on the leveraged single-stock ETF swap tape (T-Rex/Tuttle and Defiance MSTR funds paying OBFR +13-17% to Cantor, Marex, and Clear Street), the PB netting math a listed contract can't replicate, and the index-TRF precedent that already ran to completion. Plus the no-arbitrage rebuttal to "lower margin, no debit rate on shorts" — worked to the dollar. Threat map, moats ranked by durability, interactive 4-tool dashboard, two PDFs.
Offshore perpetual futures — margin, liquidation, and the October 10 stress test
$19B liquidated in hours, 1.62M accounts, 87% longs — the largest crypto liquidation in history exposed how each major offshore perp venue's design choices actually perform under stress. Side-by-side comparison of Binance, Bybit, OKX, Hyperliquid, dYdX, and the regulated Coinbase / Deribit alternative across 18 dimensions: margin methodology, liquidation routing, insurance funds, ADL ranking, mark price, funding mechanics. Interactive dashboard + flash-crash simulator.
Litigation outcomes as predictive contracts — a phased listing plan and a perpetual on the index
$79B in top-10 US class action settlements in 2025, a $19.4B litigation finance market, and $300B+ in pharma patent cliff exposure through 2030 — all currently absorbed by D&O insurers and shareholders rather than a liquid hedge. A four-phase listing plan (binary → multi-state → timing → perp-on-index), 16 candidate cases with $1T+ aggregate exposure, full perpetual design with cash-carry funding, and why the Kalshi precedents clear most of the regulatory path. Interactive case explorer inline.
A Gold/Silver Ratio Perpetual — turning a structural macro trade into a single-tick product
A proposed CFTC-regulated perpetual referencing COMEX GC/SI front-month VWAP, with cash-carry-anchored funding. The arbitrage triangle against two-leg cleared and the ETF pair, full bid-ask and margin economics, the index methodology, and the funding-rate formula — with an interactive dashboard for cost-benefit, funding decomposition, and venue ranking.
The Compute Complex — congealed electricity, the index dispersion, and the pre-listing trade
CME × Silicon Data and ICE × Ornn filed compute futures in May 2026 on structurally different indices — quote-based assessment vs Asian-averaged transaction VWAP. A six-level hierarchy (benchmark / grade / region / firmness / tenor / venue+credit), the four legs of the SD-vs-OCPI dispersion (the first listed-market trade), and how the compute supply curve is sitting in public interconnection queues right now, pricing PJM and ERCOT basis 12-36 months ahead of compute. Interactive dispersion dashboard inline.
Predictive market ETFs: what's filed, how the swaps work, and two concepts the market hasn't priced
Three sponsors filed 24 prediction-market ETFs (Roundhill BLUP/REDP/BLUS/REDS/BLUH/REDH; Bitwise; GraniteShares); SEC paused them May 5. A walk through the TRS plumbing that makes a 1940 Act fund possible on a CFTC event contract, the binary return profile, plus two unbuilt concepts — predictive signal ETFs and overlay products — with an interactive sizing dashboard.
Perps come onshore: what the CFTC's May 29 approvals change about contract design
The CFTC approved Kalshi's BTCPERP as the first US-regulated perpetual, issued a policy statement on listing perps, and cleared a Coinbase pathway to Deribit — all in the same 48-hour window. A walk through what a perp actually is, what knobs designers turn, and how the offshore (Hyperliquid HIP-3) and onshore (Kalshi DCM) paradigms compare.
Analytical tools
Built alongside the research.
Seven in total; these three are comprehensive enough to use as analysis tools in their own right. The engines behind them are the same ones we build against on client work.
Energy complex decomposition
430+ risk factors across 470+ contracts on ICE, NYMEX-CME and Nodal, decomposed into dated factor legs with open interest and cross-exchange offset detection.
Compute × Power workbench
Compute supply curve through to PJM, ERCOT, WECC and CAISO basis — spark spread, take-or-pay optimisation, index decomposition and trade synthesis across 14 modules.
Perps × predictive margin
BTC and SPX perpetuals against event-contract strips on one underlying, with a 5,000-path Monte-Carlo of an eight-cluster portfolio margin framework.
Working together
The same capability, on commission.
Everything above is the working evidence — published openly because the argument for the method is the method. The same capability is available to exchanges, venues, index providers, trading firms and infrastructure builders.
Index & benchmark design
Methodology design, IOSCO-principles review, and denominator analysis for new asset classes.
Contract & product design
Full instrument specification — settlement, funding, delivery, and pre-listing gap analysis.
Risk, margin & collateral
Margin frameworks, clearing-house risk architecture, collateral mobility and liquidation design.
Market structure research
Commissioned research, venue and regulatory-pathway diligence, and internal analytical tooling.
Open to partnerships, commissioned research, and consulting engagements.
If you are designing an index, listing a contract, building a margin framework, or working out the structure of a market that does not have one yet — that is the work.