Research
Notes, papers, and interactive explorations.
The Other Side of the Crack Spread.
Every GPU-compute derivative announced this year settles on rent — CME with Silicon Data, ICE with Ornn, Architect, Kalshi. All of them reference r, the right-hand term of the producer's margin. The token price has a literature and no contract. And the term that isn't in the producer's equation at all — how many tokens it takes to get a job done — has neither. Write the buyer's side to match and the sign structure is the whole argument. The producer's spread is C = T·p − r: long the token price, long a heat rate that is physics. The buyer's is S = p·h·N: short the token price, exposed to a heat rate that is not physics but prompt design, agent loops and model generation, and no r at all unless it self-hosts. A token index would have natural counterparties on both sides. A rental index structurally cannot, because an enterprise buying through an API pays no rent. Then who is actually studying this buyer, which turns out to be a gap rather than a disagreement: the exchanges name enterprises as hedgers without naming a firm; the independent market-structure writers put their weight on fleet financing; the consultancies address governance and routing; the FinOps practice has by far the best empirical picture of the buyer and never says the word hedge; and the one academic paper treating application-layer companies as the primary buy-side is a simulation settling against an index it proposes itself. Ranked by what moves the bill: price is deflating, so what a buyer fears is not the level but the jump when a model is retired and the 50–100× dispersion that makes every routing decision a price decision. Quantity is the real risk — structurally rising as agents replace chat, correlated with the buyer's own success because a better agent spends more tokens, and correlated across buyers because a new model generation moves everyone's consumption on the same day. That last correlation is what eventually makes it an index. Two repricing events from the last fifteen months read as risk events rather than product news. And the conclusion is an odd one for a corporate hedger: with the levered producers under hedging pressure, a buyer committing to baseload volume would not be paying for certainty, it would be paid to provide it. So when the demand side shows up it won't be to lock a price. It will be to write cheap floors for producers on volume it was buying anyway, and spend the proceeds on the engineering that controls quantity. Every layer above the enterprise has found a way to be short that quantity term to the layer below. The end user is the last holder of both, with no customer to pass them to, and the only participant in the chain without a contract, a desk, or an analyst.
ReadWhere the Risk Goes.
Every risk the compute economy throws off has a natural seller, and most now have an announced instrument pointed at them. Run all of them and a specific set of exposures is still left over — and it lands, and stays, on whoever intermediates. This asks the market-maker's question rather than the product question. Eight risks in a ledger, each with four columns: who is naturally long it and wants out, who is naturally short it or paid to hold it, what instrument connects them, and what's left. The natural writer of obsolescence protection turns out to be the chipmaker itself — new-generation sales accelerate when old values fall, so it is structurally short the risk and can warehouse the long side cheapest — which is elegant and also wrong-way, since its own launch is the trigger event. Then ten structures ordered by what exists today, from listed strips and exchange-for-physical bridges through tolling, securitization and residual-value towers to the one that's missing: protection that pays on a release-date event rather than on continuous decline, the CDS-shaped instrument compute doesn't have. What no instrument absorbs is five residuals — the tenor tail past the listed strip, fleet-versus-index basis, the jump wing that delta-hedging can't touch, utilization embedded in structured deals, and correlated credit, where the counterparty most likely to fail is failing because the market gapped. A capital argument that runs against intuition: under SA-CCR the residual 'other commodity' bucket carries an 18% supervisory factor while electricity alone carries 40%, so compute would attract less than half the capital of a power derivative despite a jumpier history and a far thinner settlement layer. That's a mis-calibration, and it runs in the direction that should worry a risk committee rather than a client. The analytical core is that the inventory is unhedgeable outright but almost entirely priceable. Seven spreads do it: the inference crack marks utilization exposure and supplies the financial buyer the ledger appeared to lack; the compute toll strips the power leg into a deep existing market so only the new part is warehoused; curve-implied depreciation replaces accounting assumptions with the traded calendar slope, and arrives with the first credible settlement rather than with volume; event ladders are the only existing instrument that pays in the jump. The warehouse becomes financeable not when every exposure has a perfect hedge but when every exposure has a price. And the inventory's shape saves it — a strip of monthly deliveries that self-liquidates, with jump risk calendarized against telegraphed release dates, so a book seasons as it survives them. Compute inventory ages like a loan book, not a bond position. The bottleneck was never instruments.
ReadSilicon to Tokens.
Every economic question in AI — capex sustainability, token pricing, GPU-backed credit, compute derivatives — reduces to a physical one: what silicon is doing the work, how efficiently it's being used, and who bears the price risk on it. The primer for a markets and risk audience. Serving a model is two different jobs: reading the prompt is a math problem, writing the answer is a data-moving problem, and each phase uses only half the chip you paid for. That asymmetry is the whole economics. Every output token requires hauling the model's entire weights plus an accumulating conversation cache out of memory, so decode is bounded by data movement rather than arithmetic — and since output tokens price at roughly 4× input and reasoning models pushed output volume up an order of magnitude, revenue and cost both concentrate there. Gross margin collapses to one spread: the market price of an output token minus your cost to run decode. Prefill shows up in user experience; decode shows up in the P&L. Covers the 2026 accelerator lineup, the inference-optimization stack that lets the same fleet deliver several times the tokens without buying a chip — which is why serving-stack quality rather than hardware ownership is the real dispersion factor, and why defining 'a unit of compute' for a settlement benchmark is genuinely hard. Then where the constraint binds: $660–690bn of guided hyperscaler capex, fab and packaging limits, and the migration of the binding constraint from chips to electrons against utility capital plans above $1.1 trillion. An illustrative worked example follows one GPU through the cascade — training tenancy, redeployment to inference, decay, and the shutdown floor where spot rental meets variable cash cost, the identical logic to a power plant retiring on a negative spark spread. The reconciliation that matters: the asset generates good lifetime cash, while book value at the start of year three sits roughly double the present value of what's left. The asset works; the schedule doesn't. Closes on why no closed-form forward price exists for something that can't be stored, and the three relationships that bound it anyway — the shutdown floor, the substitution ceiling set by the next generation's realized cost per token, and the token-parity ceiling that ties the compute curve to the token market. Backwardation is the normal state, and the informative prints are the violations: a legacy curve in contango is a macro signal about the AI capex cycle available nowhere else.
ReadA Signal Is Not a Curve.
Compute has prices worth watching. It does not yet have prices you can lean on, size into, and exit from — and the difference decides who can manage risk today. A trader's read of Liquid Compute's primer 'Compute Is Already Trading', which is right about more than the physics: compute can't be stored so nothing ties forwards to spot by arbitrage, value is lost in jumps on telegraphed release dates, its single-name-credit analogy is the best framing of GPU obsolescence anyone has published, and its reframe — if you operate a fleet, buy capacity on contract, or lend against hardware, you already have a position on this curve — is the sentence this market needs said to it. Two things it doesn't earn. First, a price is tradable, as opposed to informative, only when five things are true at once: fungibility, two-way size, transferability, settlement integrity, credit intermediation. Run its exhibits through that test and the pattern is that the further into the piece you go, the more they are signals wearing the costume of trades — a posted rate is an offer, term clears are real prints but private and bespoke with no secondary bid, the index trades are a backtest it labels as such itself, and the one genuinely executable trade in the document is executable precisely because it needs racks and a sales team rather than financial infrastructure. The most tradable thing in compute today is the thing that looks least like trading. Second, the causality — which is pointed one way here when it runs both. Trading does help build a market: early flow pays for the price data, forces the standardization nobody volunteers for, and capitalises the desks that later warehouse basis, and NYMEX didn't wait for crude to be fungible, since the delivery spec is part of what made it so. What trading can't do is stand in for the preconditions. So the claim is about sequence, not denial: trading accelerates what it cannot replace, and run ahead of a settleable reference or a margin model built for jump risk it manufactures the appearance of depth and none of the substance. Bandwidth in 2000 had the desks, the curves, the published indices and the conference-stage confidence, and no two-way commercial flow underneath — though the disanalogy cuts in compute's favour, since a GPU-hour with form factor and contract basis resolved is far more standardisable than a New York–London OC-3 ever was, and the demand is booked rather than forecast. What transfers is the failure mode, not the outcome. The test itself is a gradient rather than a gate: WTI in 1983 would have failed two-way size on its first morning, every market is born failing this, and the question is which conditions are missing and how fast they're closing. The analytical core is a waterfall: the headline tenor discount, $5.15 an hour or 57% for committing three years, is not a premium you can harvest but five things priced as one number — posted-rate air that was never real, expected depreciation which is a path and not a premium, utilization equivalence earned by operators running a sales effort, flexibility and credit that options and clearing exist to carve out, and a residual clean forward risk premium on the order of $0.35, perhaps a tenth of the headline, that no instrument isolates. Each strip is also a build order. Then an inventory of who can actually hedge what today — lenders are stuck carrying almost everything, which is what double-digit all-in pricing with maintenance covenants was telling us — and the sequence that has to come first: benchmark integrity, assignability, margin built for jump rather than diffusion risk, dealers willing to warehouse basis, then event instruments and insurance. Closes on the primer's own sign-off, which is truer than its pull-quote and proves the opposite of what it's deployed to suggest: the trades are old, but they move in bundles, once, between neighbours close enough to underwrite each other. That is barter. Oil traded forward for a century before 1983; NYMEX didn't invent the crude trade, it changed who could hold the risk.
ReadOpen Source Is Not an Audit Trail
UPDATED September 1, 2026 — NATIVX responded to this piece by shipping commit 6e2459f on August 31, and section 11 re-verifies every finding against it: the reproducibility failure is closed (64 of 64 published index values now recompute exactly from their own constituents, and all four checksums verify as SHA-256 over canonical JSON), the contract-unit fix is partial (form factor and topology resolved, contract basis still pooled in 6 of 16 buckets), the decay estimator is untouched, there is still no backtest, and the gap between the forward model and the platform’s own published history has widened from 4.7x to 6.3x — which is the argument of the piece in miniature, since reproducibility did not repair the model, it made the model’s error provable. The original audit: On August 25 NATIVX published a complete compute-benchmark engine to GitHub under Apache-2.0 — normalization across six clouds, six aggregation methodologies, a source-ablation stress test, a provenance ledger, point-in-time snapshots with SHA-256 checksums, an implied forward curve, twelve dashboard views, twenty-six passing tests. Every line that produces every number is readable by anyone, which makes it the most inspectable compute price index in existence and 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, and see whether the number comes back. It does not. Seven of thirteen published indices fail to reconcile as medians, two are indices for SKUs with no observations anywhere in the repository, and the sharper finding is what the reconciling ones turn out to be — ten of the thirteen equal a single RunPod on-demand list price to the cent, an eleventh equals CoreWeave’s, and not one is a median of a constituent set. Run the same test on the three earlier snapshots and twenty-five published values match no observation at all, in the series the dashboard charts as settlement prices. All four snapshots ship empty constituent arrays beneath a column labelled SHA-256 Checksum, and the four checksum values are not hashes of anything present. The piece is careful about where the fault sits: the snapshot constructor computes a real hash and stores its constituents, but never writes the field the history chart reads, so engine output and demonstration data are two artefacts that look identical in the interface — and nothing distinguishes a number that was computed from a number that was typed. Beyond reproducibility it documents a contract-unit error that pools nine observations spanning $1.99 to $12.29 into one reported H100 market where seven deliverable instruments exist; a sync that replaces almost the whole tape silently; a forward anchored 46% above the index it forecasts, falling at 11% a year on a dashboard whose own history falls at 51%, with no backtest anywhere; and an uncertainty band pinned at a flat 122% of the forward out to five years, which is why nothing downstream of it can be margined. It is equally explicit about what is worth taking — the ablation engine, the sensitivity matrix, the provenance ledger and the snapshot design are the right instruments, permissively licensed, and better than most administrators publish. Closes on why both August moves point the same way: the defensible asset is regulated administration, contributor governance and transaction data, not normalization code. Includes a per-claim verification ledger recording where every figure was checked, which claims are inferences, and which could not be verified at all.
ReadSame Algos, Same Book. Not the Same Margin.
Three announcements landed inside thirty-six hours, and they are more interesting read together than any one is alone. CME said it would enter the wind market — financially settled Wind Power futures and options on five modeled-generation indices across Germany, the UK, Victoria and ERCOT, listed on NYMEX in the fourth quarter pending review. FalconX and Kemet said Kalshi event contracts would sit inside the execution and risk platform institutions already use for options, perps and spot: same algos, same book, same risk model, in Kemet's CEO's words. Hours later Kalshi named The Weather Company its trusted source for verifying weather-related market outcomes, for a climate and weather vertical it reports growing 500% year over year toward $1.1 billion annualized. The middle announcement is the genuine leap — workflow was the binding constraint on institutional event-contract participation and now it mostly isn't. But a risk layer is worth what the balance sheet behind it can carry, and that capacity stops at the digital-asset boundary precisely where the flow has started crossing it. Because the risk arriving in that normalized book is not digital-asset risk. It is weather, carbon, rates and inflation wearing an event-contract wrapper, and it shares a factor with none of the crypto book and all of a commodity book. The piece sets out one exposure in four wrappers — event binary, listed weather future, parametric reinsurance, OTC swap — each with its own regulator, margin regime and netting set, and shows that only the dealer wrapper currently nets across the others. Then it maps three collateral pools with no bridge: a fully collateralized venue where maximum loss is posted in cash up front, listed margin at a clearing member, and reinsurance capital in trusts. A dealer who takes down a client's event hedge and lays the residual off in the listed market funds the Kalshi leg gross and the CME leg at its FCM, with nothing anywhere recognising that the two offset. The hedged book pays for its own prudence twice. Closes on the three routes by which the span gets built — the venue, the FCM complex, or dealers internalising it — with the third the de facto answer today, the second the likely end state, and the gap between them the place franchises get built, exactly as they were in early WTI.
ReadFive Indices, One Price.
Five providers now publish a compute price — Silicon Data, Ornn, Compute Desk, NATIVX and SemiAnalysis — and the differences between them are real. But step back and the shape is uniform: every one measures the price of an on-demand GPU-hour, and the on-demand market is the residual. The capacity that matters commercially is sold two to five years forward at a fixed price. Five providers are competing to publish the flat price of the thin part of the market. Underneath sit four series that would each carry a real exposure and that nobody publishes: an attested meter with realised revenue per GPU-hour, a term and renewal curve built from disclosed prints, a grade-standardised token price with attested precision and reference throughput, and a settlement-grade residual series. The piece is careful about which absences are total and which partial, because that decides what a commercial decision can fix. Term structure exists — SemiAnalysis's ten-tenor table, Silicon Data's forward curve — just not from disclosed prints with disclosed reporters. Residual value exists in three incompatible forms at once: a model-implied curve derived off rental forwards, bands built from arm's-length secondary transactions, and indicative supplier quotes. None is daily or settlement-grade, and residual swap strikes are still negotiated bilaterally rather than struck against a reference. Utilization is the single flat absence: no series exists gated or free, and no methodology for one — which matters because the word published, in benchmark administration, means methodology disclosed, not data free. Platts and Argus give their methodology guides away and charge for the assessments; IOSCO's Principles use the word Subscribers. A paywall is not a deficiency. An absent series is. There are six working models of benchmark administration, and the four compute has adopted are precisely the four that need no contributor to agree to anything: scraping needs no permission, deriving a forward needs no permission. The two it lacks — the meter and the contributor mutual — are the two that require negotiating with a data-holder who has something to lose. Which points at the harder deficiency, which is not data but independence. Appendix C asks a designated contract market to verify its index provider minimizes the 'opportunity or incentive to manipulate,' and incentive is not addressable by policy. Every administrator here has sat inside an adjacent commercial position, all publicly disclosed, none improper. August produced the first two exceptions, and they bought different things. Compute Desk's benchmarks are now calculated, published and overseen by GX Benchmarks, an FCA-regulated subsidiary of General Index that holds no position in compute — which buys independence institutionally, but moves the test down a layer, since the transaction data underneath is still Compute Desk's. NATIVX open-sourced FSKU, which buys reproducibility technically — posted rates in, a modelled forward out, provenance ledger and source-ablation view included. Neither closes the gap that matters: not one administrator names a contributor, though naming sources is one of the three safeguards Appendix C explicitly suggests. FSKU names its sources because they are public price sheets that cost nothing to name. A reader still cannot tell, for any published compute index, whose private prints are in it. The insurance industry ran this experiment over five decades: ISO formed from a consolidation of rating bureaus in 1971, ceded control to a majority-independent board in 1995, converted to for-profit in 1997, and listed as Verisk in 2009 in a wholly secondary offering. The order matters, and so does the part usually skipped — ISO was a named antitrust defendant, and the standard-forms machinery was the alleged instrument of the conspiracy.
ReadThe Compute Crack Spread.
The compute futures that list on October 5 settle on the one price in this complex that does not float. Rent has been nearly still — seven-day changes under 1% across Silicon Data's SKUs, the hyperscaler series unchanged on 39 of 50 trading days. The price that does float, what a token sells for across roughly a hundred providers repriced in hours, has no index at all. So we built the series that should exist. A refiner doesn't hedge crude and hedge gasoline; it hedges the crack. A generator doesn't hedge power and hedge gas; it hedges the spark spread. The compute version is C = T·p − r — reference token revenue per GPU-hour minus the rent — and its quotient, the market heat rate H* = r/p, is the tokens per GPU-hour a fleet must produce to break even at a given venue's price. Built from a live OpenRouter endpoint book for Llama-3.3-70B and a deposited vLLM throughput sweep, the series reads as a merit order. Two thirds of the active book clusters at 3.5–3.6 million tokens per GPU-hour, within a few percent of one another, which a saturated H100 covers three times over for a crack of +$5.37. At 10 requests per second almost nobody clears. The sign of a host's P&L is set by utilization, not by price — the same conclusion the spark-spread work reached from the cost side, now visible from the revenue side. Dispersion across active endpoints is five-fold for one declared good on one screen, which search costs cannot explain and which any real series would have to absorb with a grade standard. The piece then reads the CFTC's August 19 request for comment as a specification rather than a threat: fungibility, standardization, liquidity, opaque bilateral price formation, dominant participants' pricing power, and manipulation by adjusting a posted rate. Every one of those is true of the rent leg and mostly false of the token leg — and where the token leg does fail, it fails for reasons that are fixable rather than structural. Closes on how physical adoption actually arrives: not by re-papering a five-year take-or-pay to float on a rent index, which no operator will do and no lender would allow, but through a token-indexed toll — a fixed capacity charge plus a conversion fee that floats with a graded token price. That is the gas-era netback contract, and it is the first physical compute contract with an indexed element that anyone has a reason to sign.
ReadThe Contract Has a Date. It Needs a Dealer.
CME lists GPU rental futures on October 5 and ICE has promised two more complexes. The question nobody has answered in public is who is on the other side — not which speculator, but which commercial balance sheet carries a floating exposure to a GPU-hour that a cash-settled monthly average can actually offset. We went looking in the filings, the earnings calls and the lender decks, and found the opposite of what a benchmark needs. Every participant with a real exposure prices it fixed: the neo-cloud sells two-to-five-year take-or-pay, with CoreWeave's revenue backlog at $104 billion as of June 30; the lender sizes term debt against that contracted revenue and explicitly excludes the merchant tail; the buyer signs the same contract from the other side. Not one named operator, lender, insurer, hyperscaler or enterprise buyer has said it will hedge, and CoreWeave's Q2 call contains no mention of futures, hedging or an index at all. The primary documents are equally thin where it matters. NYMEX submission 26-370 is a Rule 40.3(a) voluntary request for approval rather than a self-certification — CME chose to ask rather than certify and list — and the two exhibits that would let anyone outside the exchange evaluate the cash market are filed under separate cover, one of them confidential. More consequentially, the filing never states which Silicon Data tier settles the contract, even though the neo-cloud and hyperscaler readings differ by roughly a factor of three; a hedger cannot tell from the public record which tier it is hedging. History says what has to happen next, and it runs the other way round from what is happening. WTI did not deepen because producers hedged directly; it deepened because physical barrels began pricing off the screen and banks warehoused the bilateral risk, laying the residual off in futures — swap dealers still hold about 30% of WTI open interest short. Henry Hub deepened because Order 636 unbundled the cash market first. Compute is running the sequence backwards: futures before formula pricing, with neither a floating physical leg nor a dealer. What it does have, and neither oil nor gas had at this stage, is lenders who already own the merchant tail. That is where the first real hedger comes from — and the contract design helps it less than it should.
ReadThe Long End Has Two Sellers Now.
The AI build-out has become the largest private borrower of duration in history, and it arrived at the moment the US Treasury needs the same buyers most. This piece takes the debt-crisis argument — Dalio's, Dimon's, the IMF's, the BIS's — as a conditional rather than a forecast, and asks the question the macro literature skips: if it happens, how does it actually reach a data center? The answer is not the intuitive one. Three paths are held apart, because each sends the shock somewhere different: a term-premium shock in the gilt-2022 shape, where the long end gaps while the Fed holds; fiscal dominance, where the curve is capped and inflation is the price; and a disorderly dollar, the only path on which the short rate moves against borrowers. Most capital already deployed turns out to be insulated. Hyperscaler long bonds are fixed — a +150bp move takes 18–20% off the price of Meta's 40-year, but that loss belongs to the bondholder, and Meta, having locked 5.75% for four decades, is if anything a winner. Neocloud and GPU debt floats against SOFR, which follows the funds rate, so a term-premium shock that takes the 30-year from 5.2% to 6.7% while the Fed sits at 3.50–3.75% does not raise the coupon on a single existing GPU loan. What is exposed is narrower and more specific: the flow of new projects, where debt service per MW rises 13–32% and capex rises 6–13% on roughly 60% import content; the 2027–29 refinancing of short GPU debt, where spread rather than base rate is the binding variable; and tenant credit. Then the worked case. An illustrative 100 MW facility run through four regimes breaches 1.0x coverage in year four or five under every one of them — including today's curve with no macro shock at all — because a 22% annual rental decline meets a five-year amortizer. Rates move equity IRR from 16% to negative and pull the breach forward a year, but they do not cause it. The merchant project dies of deflation; the rate shock is the second derivative. The contracted variant, on a five-year take-or-pay, survives every path. Closes with a five-phase chronology, a risk architecture borrowed from merchant power, and five reasons the episode would be less dangerous than it sounds.
ReadThe Other Side of the Spark Spread.
OpenAI is hiring a Power Trading Lead — the first role at a frontier lab whose mandate reads like a merchant commercial desk rather than a procurement seat. Everyone else is still hiring originators: Anthropic's Data Center Energy Lead pays a higher cash band than OpenAI's trading seat and never mentions a hedge, a forward or a swap. So this piece walks the five functions of a merchant generator's floor — origination, term and structuring, day-ahead and cash, real-time dispatch, risk and mark-to-market — and asks of each whether the position transfers, inverts, or has no analog at all. The organizing model is a three-leg chain with two heat rates in it, and only one of them is physics. Leg one is silicon: a DGX H100 draws about 10.2 kW across eight GPUs, so at a 1.2 PUE the site burns roughly 1.5 kWh per GPU-hour, which at $45/MWh is $0.069 — about 2.7% of what a rented H100 sells for. Leg two is not fixed by anything except how hard the fleet is run: a published vLLM sweep puts throughput at 2.88 million tokens per GPU-hour when lightly loaded and 11.16 million when saturated, which swings token revenue from $2.24 to $8.68 and the inference spread from −$0.29 to +$6.15 on the same chip drawing the same power. Power is 0.8% of saturated token revenue. That number is the whole argument: the lab's trader is not hedging margin, because a doubling of power prices matters less than a ten-percent move in utilization. What the seat hedges is absolute dollars and the tail — a 10 GW portfolio at 80% load factor moves about $700 million a year per $10/MWh, more than three times Constellation's illustrative CCGT sensitivity, and pointing the other way. The tail is where the load side diverges most: summer hour-ending-20 real-time prices at ERCOT North averaged $205/MWh against $21 overnight across three years of hourly settles, and a flat inference load buys every one of those hours. Includes an interactive spread-chain calculator, a survey of published hedge ratios at Vistra, Constellation and NRG against the complete absence of any disclosed lab hedge ratio, and the observation that a Meta affiliate quietly obtained market-based rate authority in November 2025 — the first lab-side entity licensed to run a two-sided book.
ReadThe Other Side of the Block.
On August 19 Cantor Fitzgerald opened institutional block trading in event contracts on Kalshi to roughly 3,000 clients. Strip the release to its mechanics and three roles fall out: Cantor is the introducing broker and takes no position, Kalshi is the designated contract market whose block rule lets two eligible contract participants agree a price away from the book, and Susquehanna Predictions is the principal — the firm that quotes, takes the other side, and owns the risk until resolution or offset. That third role is where the economics live, because a block is not a matched order but a bilateral risk transfer, and somebody has to hold a $2 million binary on a September rate cut or a $500,000 position on a Super Bowl winner until it resolves. The piece works the offset menu in order of cost — internalize against retail flow, work it into the book, cross-list, hedge in correlated listed instruments, lay it at sportsbooks, or warehouse it — and then runs two blocks through it. An FOMC block back-to-backs almost exactly: 2,000,000 contracts at 63¢ against roughly 1,920 fed funds futures 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. A second sort follows, on how informed the counterparty is likely to be, which splits the market into four quadrants rather than two and puts elections (unhedgeable, and the best-informed counterparties are campaigns and officials) and corporate events (easily hedged in single-name options, but carrying MNPI and a live security-based-swap reclassification question) in classes of their own. Seven candidate moats are assessed and only some survive: the Cantor pipe is contestable within months, but warehousing capacity is not, because the only way to acquire it is to lose money building a model and a risk appetite — as on the June 10 night when one Knicks tip-in took $22.4 million out of the market-making community. Closes on a thirteen-entry risk register and the number that will actually settle the question: the block print series itself.
ReadThe compute risk premium got its first print.
In the window between the May futures announcements and the October 5 NYMEX listing, the academy filed its first brief. Federico Bandi and Yinan Su of Johns Hopkins posted "(Early) AI Compute Asset Pricing" to arXiv, with both index administrators at the center of the listed complex supplying data directly. It is not quite the first paper to model compute derivatives — an earlier SSRN working paper fit a reduced-form forward curve to the same non-storable commodity — but it is, as far as we can establish, the first to price the risk. The framework does three things right. Cost-of-carry is dead on arrival, because a GPU-hour not used today cannot be carried forward, which puts compute closer to electricity and freight than to oil: every intuition imported from crude desks fails and every intuition imported from power desks transfers. Reserved-rental term curves are contaminated proxies for forwards, because a physical term rental bundles price insurance with what the authors call a capacity-locking option — so term-implied forwards are upper bounds, loosest at the long end, and the gap has a name: the physical access wedge. And once listed, futures price risk rather than carry, with the sign of the premium set by hedging pressure in the Bessembinder–Lemmon tradition — the fleets press harder, so futures should clear below expected spot. Then the number: hold-to-maturity returns on synthetic futures built from term curves, annualized at 7.5% on A100, 26.2% on H100 and 11.2% on B200, with the authors' own illustration that a $7.00 expected B200 rent supports a futures print near $6.25. The paper also independently corroborates three positions already published here — and goes further than one, since its correlation network finds some cross-provider index pairs negatively correlated, which upgrades "your hedge is noisy" to "your hedge may have the wrong sign." Four corrections follow, all pointing the same way: the market the paper calls unlaunched has been trading since May on curves that carry no wedge at all; the synthetic panel sits partly on an index that was restated 35–40% (flagged as plausible, not confirmed); an exchange-for-physical facility restores the convergence discipline the paper writes off; and every object in the model inherits the utilization blind spot, so the premium is estimated from returns on quotes.
ReadThe contest can't be hedged. The financing can.
BIS Working Paper 1367 does what no bank research desk would: it writes down a model in which the AI build-out is rationally oversized by 40–50%, puts roughly even odds on a bust, and prices the damage at $378 billion — three-quarters of it a debt-financed fire sale. Crucially it does not argue anyone is behaving irrationally. Five coalitions over-invest because over-investing is each player's best response to the others: a contest externality means every dollar of capex partly captures revenue a rival would have earned, so each coalition behaves as if its investment earns 20% more than it socially does, and that wedge compounds into 1.4× the efficient level. The early-deploy premium is what causes the leverage — remove the head-start motive and the model coalition takes on no debt at all, and bust odds fall from 50% to 38%. This piece takes the paper as a risk specification and asks the practitioner's question it deliberately doesn't: can anyone hedge this, and with what. The organizing observation 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 structurally unhedgeable by choice, and empirically they behave that way: they are sellers of protection, not buyers. The demand migrates one layer out to everyone who financed them without a ticket to the prize — which is exactly where single-name CDS volume went, $4.6B in Q1 2026 against $759M a year earlier. Then the sizing: roughly $1.5T of external financing need against about $12.5B of net protection outstanding, a gap near 120:1. Each of the paper's four loss channels is mapped to the instruments that actually exist — compute futures, token indices, hourly power curves, the credit complex — with honest residuals where nothing settles. Closes on the two distinct risks travelling under the name "dark GPU," which have opposite hedges, and the electricity arithmetic that decides which one you get: roughly 220 GW of gas-turbine backlog and 2029-plus delivery for a unit ordered today, which makes power rather than silicon the binding constraint through 2028.
ReadThe speculation sleeve.
Half of Gen Z investors moved money meant for investing into sportsbooks over the past year, and 26% describe sports betting as a deliberate part of their long-term financial strategy — against 14% of Millennials, 6% of Gen X and 1% of Boomers. The damage is measured rather than hypothetical: a June 2026 Journal of Financial Economics study tracking ~184,000 households finds legal sports-betting access cuts net brokerage investment by roughly 20%, with heavy bettors cutting deposits by more than half. The industry's answer — stop betting, buy index funds — is losing to a product that pays out Sunday night, and fifteen years of responsible-gambling research says voluntary restraint tooling barely moves net losses. So the piece asks a different question: if a generation insists on holding a book of binary risk, what is the least destructive market structure for it, what guardrails actually bind, and can anyone legally manage it. The answer proposed is a capped, guardrailed sleeve of exchange-traded event contracts, because CFTC-designated prediction markets are the only venue where the house edge is a fee schedule rather than a business model — worst-case taker fees around 1.75¢ on a 50¢ contract against a blended 2025 sportsbook hold above 10% and parlay holds of 20–30%. The honest caveat is given equal weight: settled-contract data shows takers losing ~32% of stake on average, because retail crosses the spread to buy longshots that resolve YES ~2% of the time, so the gap between cheap structure and expensive behavior is precisely what the guardrails must close. Built against live Kalshi and Polymarket order books snapshotted on publication day, with a diversified event universe spanning rates, inflation, payrolls, elections, climate, AI milestones and crypto, five strategy archetypes, ten hard-stop guardrails that ration the lottery rather than banning it, and a registration analysis of who could legally operate it — the answer to which is neither a robo-adviser (no discretion, no rebalancing, wrong regulator) nor a discretionary CTA (the user keeps contract selection), but a non-discretionary rules engine attached to an RIA: closer to a 401(k) plan document for speculation, a structure adopted once that then makes the destructive versions of the saver's own behavior unavailable. Nobody has built it, and the three blockers are tax character, the absence of anything resembling rebalancing, and fiduciary optics.
ReadThe index was the wedge. The product is the risk desk.
Silicon Data entered the compute-financialization race as one index among four. Somewhere between the CME announcement in May and the October 5 listing date it became something else: an eleven-product market-data platform covering the spot price, the term curve, the listed hedge, asset underwriting, the quality assay, the resale market — and, uniquely among every provider we track, both sides of the AI production margin, the GPU-hour going in and the token coming out. The right frame is no longer "index provider" but the anatomy of a mature commodity-market data franchise — what Platts became for oil and Fastmarkets for battery metals — being assembled in advance of the market it serves. The piece audits all eleven products against that anatomy, then does the work nobody has done: maps seven recurring risk exposures (price level, term and rollover, margin, residual value, utilization, quality, site capacity) to the products that answer them, for each of five participant classes — lenders and structured financiers, venture and growth equity, inference-dependent corporates, neoclouds and operators, traders and funds — plus a sixth the suite is quietly courting, insurers. It includes an honest routing table naming where competitors are the better answer (Ornn for settlement integrity, Kalshi for short-dated views, Compute Desk for physical convergence), because a framework that never routes elsewhere is marketing. And it is direct about three places the map fails: the inputs are quotes with no published rulebook, the 2.6× counterparty-tier basis means "the price of an H100-hour" is a family of prices and the futures settle on one member, and one vendor now supplies the index, the curve, the forecast, the underwriting model and the settlement print. The blind spot is the finding: no administrator anywhere publishes an observed utilization benchmark, so rental rates can hold firm while fleets idle — the scenario that impairs operators, their lenders and their insurers simultaneously, unhedged by every instrument in the piece.
ReadThe contract got a date.
On August 11 CME Group and Silicon Data announced that the first US compute futures — Silicon Data H100 and B200 Rental Index Futures, each representing a month's worth of GPU rent — will list on NYMEX on October 5, pending regulatory review. Part two of two on the week compute's market structure arrived, and it lands the morning after NVIDIA's financing platforms made a public compute price half a trillion dollars more urgent. The piece takes inventory of what the release actually disclosed, then builds the dispersion ladder the contracts have to navigate: same-chip performance variance of roughly 1.4× normalized away inside the index model, a grade spread of about 2.0× between H100 and B200 that becomes a listed market on day one, a counterparty-tier spread of 2.6× between neo-cloud and hyperscaler H100 that is published as sub-indices but sits outside the settlement sample, a model-class spread of about 4.9× in tokens, and a cross-venue spread of up to 7× on identical open weights. The tier rung is where the design quietly takes a position: an enterprise hedging a $7.19 hyperscaler bill with a $2.73 instrument is running a correlation the sub-index history is only now long enough to test. Also assembles Carmen Li's public record — from "very TBD" on Odd Lots in June to a listing date eight weeks later, the 38% same-chip performance variance she concedes and quantifies, and the token index Bloomberg reads as the cleanest proxy for AI capex, now 20% off its May peak — plus a precise three-bin account of which index documentation is public, gated, or absent across the venue board.
ReadCompute became collateral. NVIDIA sold the floor.
On August 10, 2026 NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish financing platforms intended to mobilize over $500 billion of third-party capital for AI infrastructure — special-purpose entities issuing debt secured by GPU compute itself, with Goldman explicitly setting out to build "a market for credit backed by NVIDIA compute." The stock fell 2.9% and the semis fell harder. But the half-trillion is money NVIDIA does not supply, promised "over time," under memoranda that bind nobody until final agreements are executed — no named projects, no capital split, no advance rates, tenors or covenants. Read the release for what NVIDIA commits itself to and exactly one figure survives, and the piece argues the enforceable economics live entirely there: residual-value support on up to approximately 25% of an opportunity, case-by-case, which is a written put on the depreciation of NVIDIA's own hardware — the same instrument aircraft manufacturers have sold for decades to move metal through leasing channels. The piece walks the SPE structure step by step against its ancestors in commodity and equipment finance, then names what the structure does not yet have and every ancestor does: a settlement-grade price for the collateral, a forward curve for the revenue, and a venue on which to hedge it. It also notes that the oldest structural objection to compute derivatives — where is the natural short — now has six logos, since a lessor financing racks over five years is long residual value and long forward rents by construction. Includes an interactive collateral engine and a skeptics' ledger covering circularity and the depreciation case.
ReadThe metal keeps calendar time. The contract ticks.
On August 3, 2026 Kalshi launched 15-minute gold and silver binaries settling on Pyth Network feeds — and the interesting thing it listed is not a product but a clock. In 2002, writing in the wreckage of the dot-com bubble, Emanuel Derman argued that short-horizon speculators do not perceive risk and return in calendar time at all: they count trading opportunities — intrinsic time — and during speculative episodes capital chases an asset’s temperature, volatility scaled by the square root of its trading frequency (χ = σ√ν). A quarter-hourly settlement grid quotes the exchange rate between those two clocks in public, at a resolution no listed market has offered before: 96 times a day per metal. The piece works that claim carefully. A near-the-strike binary is to first order a bet on that window’s volatility, so the strip of 96 prices is a market-quoted intraday activity clock — a window’s implied vol relative to the day’s average is the square root of its share of intrinsic time — and for metals, whose activity is sharply structured around London, the 8:30 ET data window and the COMEX settle, that strip should look like a skyline rather than a lawn. Then it holds three distinctions apart, which is where the analysis earns its keep. Intrinsic time is endogenous and discovered; the exchange’s grid is imposed and uniform, which is precisely why it measures the terrain — a ruler only reveals unevenness because the ruler itself is flat. The product does create something new, but it is a third clock: settlement time, 96 synchronized forced resolutions a day, exchange-manufactured. And the heat is in the trader, not the metal: listing a faster-settling derivative leaves gold’s own χ untouched while raising the participant’s trading frequency by two orders of magnitude, so fees and edges that are invisible per tick become decisive per calendar year. Includes The Two Clocks — an interactive intraday activity profile against the 96-window implied-vol strip, and a tick-illusion panel showing why per-tick edges must compress toward the fee floor. Closes by re-sorting settlement designs as a hierarchy of clocks (calendar-point snapshot, breadth-defended candle close, calendar-average TWAP, volume-weighted partitions approaching intrinsic-time averaging), with the prescription that a market should be settled in the clock it actually runs on — plus a coda on CFTC Letter 26-22 and five falsifiable predictions.
ReadOne roof, four registrations, seventy-seven questions.
The CFTC’s July 30 Conflicts and Affiliations proposal (91 FR 50926) is the first comprehensive conflicts framework for the DCM/DCO/FCM stack that every crypto-native venue and nearly every prediction market spent 2020–2026 assembling. It does not break the model up — it licenses it, and prices it. Four pillars: affiliated principal trading firms barred on futures exchanges except as formal market makers whose orders are filled last at every price level regardless of time priority; boards at least 35% independent with a fully independent regulatory oversight committee; non-preference rules and independent reporting lines for affiliated clearing members; and mandatory third-party surveillance of affiliated brokers, generalizing the arrangement CME adopted voluntarily. The piece maps all four against thirteen corporate families as they actually stand — Kalshi, Rothera, Polymarket, Crypto.com, Bitnomial, Coinbase, Interactive Brokers, DraftKings, Gemini, Underdog, ProphetX, Kraken and CME — with a collision-severity matrix. The costs concentrate on the two groups whose liquidity is affiliated, barely graze the group that outsourced clearing and the group that outsourced liquidity, and are close to costless for the incumbent whose practice is being codified. Four gaps will define the comment file: the SEF asymmetry, influence below control (ICE’s stake in Polymarket), distribution contracts left entirely out of scope, and a one-commissioner Commission proposing all of it while the largest incumbent sues the agency. Comments due October 5.
ReadPerp the S&P. Sue the format.
On August 17 Coinbase Derivatives lists US500 — a perpetual-style, funding-anchored S&P 500 futures contract at up to 20x on a CFTC-regulated exchange, with a filing pending to let traders post USDC as margin. It arrives fourteen months after the perp format first went onshore, eight weeks after CME sued the CFTC to have the entire class declared swaps rather than futures, and three weeks after CME relaunched single-stock futures — the product the joint SEC–CFTC regime starved to death once already. Three fights are running at once and each feeds the others: what a future legally is, who may license the index, and whether a margin obligation that never sleeps can be met with a stablecoin during the 58 hours a week Fedwire is closed. The piece works the sequence (fourteen months of stack-building, flagship product last), the definitional matrix (who dies under each of three legal outcomes), the collateral chain end to end, and the oracle problem — the S&P 500 cash market prints 32.5 hours out of 168, so for roughly 80% of the week a near-continuous S&P perp anchors its funding to something other than the thing it tracks. Two arguments are our own: the boomerang in CME’s legal theory, where winning the case may free S&P DJI to license the S&P 500 “non-future” to everyone; and the weekend basis, where Monday’s cash open stops being the first print after the weekend. Includes a stylized carry comparator across the E-mini roll, perp funding and a single-stock futures basket, and a leverage table for the single-name field CME just walked into with a statutory 15% margin floor.
ReadThe coin can't pay yield. The fund behind it can.
In seven weeks the largest cash managers in the world launched three purpose-built stablecoin-reserve money market funds — State Street on 16 June, Fidelity on 18 June, and BlackRock's BRSRV on 3 August. Only BlackRock's is tokenized, and that difference is the whole story: the reserve asset now lives on Solana, Ethereum and Stripe's Tempo, the same rails as the coins it collateralizes. Four days before that launch, Securitize — the firm that keeps the register for BUIDL, Apollo, KKR and VanEck — added an SEC investment-adviser registration, its fifth, completing the first fully regulated vertical tokenization stack: issue, trade, register, administer, advise. The piece works through what GENIUS actually wrote (§4(a)(1)(A) enumerates the reserve and simultaneously bars the coin from paying what that reserve earns, which is why the yield migrates to the fund), why a private registrar now performs Cede & Co.'s function without Cede & Co.'s oversight regime, and how Ondo's Oasis Pro FINRA approvals show the stack is the strategy. Closes with five ways this reads as less than it looks — permissioned is not composable, a $3M minimum and whitelisted wallets make BRSRV a register replicated on three chains rather than DeFi collateral — and corrects two claims in wide circulation: BRSRV is the first tokenized GENIUS reserve fund, not the first one, and the $8.6B attached to Securitize's RIA is the vault market it will advise into, not its own assets.
ReadSoft target, hard target: the event-perp feedback loop
Can intraday crypto event contracts and the perpetual-futures complex distort each other? Yes — but not equally, and not in the direction most people assume. The two sit at opposite ends of one spectrum, how manipulation-resistant their settlement print is, and that gap decides which way distortion can flow. Kalshi settles crypto on a 60-second average of a per-second multi-exchange CF Benchmarks composite, with 20% trimming on certain markets. Polymarket's short-dated markets settle on a Binance-heavy Chainlink oracle that was a near point-in-time snapshot until 00:00 UTC on 7 August 2026, when it became a 30–60 second TWAP. Perpetual liquidation marks are engineered specifically to resist this: Binance's mark is a median of three inputs on a capped multi-exchange index; Hyperliquid's is a stake-weighted median of validator medians across eight venues. 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. This piece's own contribution is the leg that literature does not cover: because Hyperliquid publishes every wallet's liquidation price on-chain, 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. With an interactive settlement-hardness model, a cost-to-flip calculator, and a settlement-clock overlay.
ReadMarkets went continuous. Central-bank money didn’t.
The CLARITY Act is stuck on the Senate floor over a government-ethics provision that has nothing to do with market structure, and the sector is treating that as the story. It isn’t. Over twelve months the entire institutional stack — stablecoin trust charters, tokenized securities on Nasdaq and NYSE, a tokenized ETF posted as margin at CME, perpetual futures, 24/7 trading — was built by agencies acting without it, by no-action letter, exemptive order, pilot programme and self-certification. What is actually pulling stablecoins and tokenized collateral onto institutional balance sheets is not crypto adoption but the Federal Reserve’s operating calendar: US markets now trade 58 hours a week, 34.5% of every week, during which no central-bank settlement rail is open — and the Fed’s own 2028–29 expansion to six operating days closes only 22 of them. Saturday never opens, which makes a private 24/7 settlement asset permanent infrastructure rather than a bridge. With an hour-by-hour settlement gap grid, a float-versus-flow decomposition (float +11%, adjusted volume +125%, and the same divergence inside Circle’s income statement), the float-income pool the yield war is actually about, and a filterable map of every venue that takes tokenized collateral — including the Options Clearing Corporation, which considered it and declined.
ReadVolta 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, out of stealth with $300M raised and a reported $10B six-year anchor behind a 133 MW Norway build, 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. Trillium runs the pipe in the other direction, packaging a cloud 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, nine months after a 'fully collateralized' offering. The exposure ladder each structure creates, the valuation gap both share (no settlement-grade mark prices Vera Rubin capacity, and nothing at all prices a platform compute credit), the hedge program each side would need, and an aside on how one anonymously sourced Bloomberg report became four mastheads of 'confirmation' for the anchor's identity.
ReadThe number you cannot trust — AI in wealth management, Part V
Forty-five independent strategy variants are enough to manufacture a Sharpe ratio of 1.0 out of five years of pure noise. That is an afternoon's work, the result is thirty years old, and it is almost universally ignored. Part IV specified what an AI system may do; Part V asks whether it should be there at all. On 17 April 2026 the OCC, Federal Reserve and FDIC 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. Validation stopped being a compliance exercise and became a question of whether you want to know. Covers the False Strategy Theorem and the deflated Sharpe arithmetic, an even-handed reading of a genuinely contested replication literature (65% of 452 anomalies failing a t of 1.96 on one side; 82% replicating under a hierarchical Bayesian model on the other), evaluation design for a system with no track record, the judge problem — the best LLM judge held only 65% of its verdicts when the two answers were swapped — champion–challenger promotion and shadow deployment, and what drift means when a pinned model name is not a pinned model.
ReadThe token leg has a price screen. Nobody has built the index.
In March 2026 a paper designed a complete futures contract for AI inference tokens — a standard grade anchored to fixed benchmark thresholds, a volume-weighted settlement index, margin rules, circuit breakers. It 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 data feed the paper's index requires. This piece pulls that order book — 56 quotes across five model books at one morning's snapshot, 101 venues on the screen and 100 by the afternoon — then tries to build the index, with a widget that lets you be the administrator and watch the print move. It falls apart in four specific places: identical open weights quote a 7× output-price spread, so a composite is an average over a distribution that has no single price; a capability-threshold grade calibrated to a January 2024 model is stale on arrival; the declared attribute that moves the product most is unverified; and the unit of account is not the unit of value. Closes with the disclosure list a real token price index administrator would have to publish, and an appendix stating exactly how the data was retrieved and where transcription risk remains. Sequel to the inference spark spread.
ReadThe shield — AI in wealth management, Part IV
Interactive Brokers has the richest pre-trade risk control surface available to a retail investor — price collars, total-value limits, size limits, fat-finger checks — and in July 2026 it opened an MCP endpoint to every major AI client. The agent may analyse, research, monitor and draft; it may 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, order notional cap, rate limit, daily loss limit, drawdown breaker, cooling-off period 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 a constraint set an agent cannot argue with — drawing the template from the Market Access Rule regime that followed Knight Capital's $460M, 45-minute failure. With a working constraint compiler, controlled-testing evidence on what a prompt-injected agent does to a portfolio, and the observation that FINRA retired the pattern day trader rule eight days after Robinhood opened trading to agents.
ReadTwenty minutes a week — AI in wealth management, Part III
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 about twenty minutes per advisor per week — which does not become a new client, it becomes a slightly less rushed Tuesday. Efficiency in this category is almost always quoted as hours per firm per year, the one unit that makes a small number look like a department. Part III takes the advisor stack audience by audience: the jagged frontier of what actually works by category, why 64% of firms lacking a unified data layer means the data layer is a gate rather than a nice-to-have, the supervision and books-and-records obligations that attach the moment a tool is switched on (including a recording-consent exposure nobody in securities is watching), a build-versus-buy framework for solo, mid-size and enterprise firms, and a correction to the consensus on vendor risk — absorption was the wrong thing to fear; the real risk is a $100-a-seat notetaker quietly becoming your system of record. The efficiency is real, smaller than advertised, and the best firms have decided not to convert it into revenue at all.
ReadRead the filings first — AI in wealth management, Part II
Part I promised a hands-on evaluation of the retail AI advisors: identical prompts and test portfolios into PortfolioPilot, Cortex, Public's Generated Assets, Magnifi and peers. Pulling the Form ADVs, Form CRSs and advisory contracts first changed the exercise. A platform marketing 50,000 users and $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 says its AI output is "not, and should not be construed as, individualized investment advice" takes full discretion over the resulting account for 49bps. One platform's 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 on advice quality — with the full hands-on protocol published alongside, and the terms-of-service collision that makes it effectively unrunnable by anyone outside the firms.
ReadTwo venues, one index — the live compute spread and why it still cannot be read
Kalshi and Polymarket both list GPU rental price contracts, and both settle them against the same Ornn dashboard URL — named in both contract texts, verified directly rather than inferred from a press release. That makes the Kalshi–Polymarket spread the first live compute spread with no index basis inside it, and on its face the one clean cross-venue observation the asset class has produced. It is still unreadable. Matching contract structure — an American touch barrier against a terminal bracket are two different payoffs on one index, not two views — collapses the headline gap from $2.17 to $0.10, less than the bid-ask on one venue alone. And both ladders violate their own internal arbitrage conditions first: a six-bracket exhaustive partition on Polymarket sums to 180%, a Kalshi survival function is non-monotone. The result runs the opposite way from the catalogue and is stronger for it — remove the disclosure problem entirely and the basis is still uninterpretable. Index disclosure is necessary and visibly not sufficient.
ReadThe inference spark spread — what an 80% token price cut requires of the serving stack
On July 30 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 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 power less gas times the heat rate. Written that way, inference reduces to one unobservable term, tokens per GPU-hour, and inverts to a threshold: the serving efficiency a rate card requires in order to clear. The heat rate is the whole difficulty, because nobody publishes it — so this computes it, from 1,024 measured vLLM serving runs on Llama-3.1-70B across four H100s joined sample-by-sample to the NVML power traces recorded alongside them. Measured median 11.6M tokens per GPU-hour, 51 Wh per million tokens saturated. The July cards raise the required efficiency 5×. Also: the sign convention that makes labs and resellers natural counterparties on the token leg — an answer to the standing critique that compute lacks two-sided hedging demand.
ReadVerifiable is not replicable — the on-chain compute index and what it actually discloses
Our administrator catalogue closed on a stark finding: across the four providers positioned to bear settlement weight on CFTC-track 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 — the Inferra Index, a per-model GPU-hour benchmark on Solana run by a protocol that is simultaneously 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 a 20% documented cap on its own trade weight — the only quantified bound on settlement circularity among the five. And it still cannot be reproduced by anyone who settles against it. The inversion, the four-role conflicts stack, and a principle-by-principle mapping of what on-chain publication does and does not satisfy.
ReadFrom 60/40 to autonomous agents — AI in wealth management, Part I
Part I of a series. Three regimes of portfolio construction and what breaks between them; one technology with three very different buyers; the autonomy ladder — a four-rung taxonomy from copilot to autonomous agent; an interactive explorer of the current product landscape by segment and autonomy level; and the control problem: proving an AI portfolio process stays inside mandate.
ReadHedging the inference book
Inference providers run a commodity retailer's book: fixed subscriptions downstream, floating token costs upstream. Two instrument families now exist to hedge it — token forwards in $/M-token and GPU futures in $/GPU-hour — and 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 caps over fixed strips given the asymmetry.
ReadHow compute index providers actually calculate price
Four providers are positioning to become the settlement benchmark for GPU compute, and they represent four different answers to what an H100-hour costs — posted rates at scale, invoice-verified prints, market-implied forwards (H100 conspicuously excluded), and an undisclosed physical-delivery reference. Component-level methodology, data sourcing, and governance for each, the finding that no provider publishes a rulebook or uses utilization as an input, and the CFTC Part 40 filings that will force the first real disclosure. The transparency baseline the rest of our compute work has been missing.
ReadData centers are not a flat load
Measured H100 facility load profiles joined against 401,379 hourly DA settlements across eight ERCOT and PJM locations, including the DOM zone. The load-weighted price runs up to 9.4% above flat-block, the premium peaks at 40% utilization rather than saturation, and it flips sign by season. The shape basis every 7×24-hedged operator is carrying, quantified — with a live join engine over every hub, facility type, and utilization.
ReadThe parlay is a correlation market
A combo price minus the product of its legs is an implied correlation — printed thousands of times a day, quoted by nobody. The partition rule mapped across every live prediction-market category, this week's Fed ladder joined to the BTC price ladders (implied ρ ≈ 0.13), a World Cup lookback where the framework's consistency check flagged the actual final, and parlay flow read as the correlation tier that arrived before its benchmark.
ReadHedging 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 vs physical vs optional-EFP settlement (with the EFP held for the recovery state), benchmark-tier standardization with config basis charged rather than hedged, a five-check index validation program requiring no administrator cooperation, and the power leg grounded in ERCOT and Dominion settlement data with an interactive risk dashboard.
ReadImplied compute forward curves
Kalshi's compute ladders settle against Ornn's Compute Price Index, which makes each rung a digital option on an administered benchmark — and a strip of digitals is a discretised probability distribution. A forward curve has therefore existed since those contracts listed, and since July 14, 2026 Kalshi has been publishing it: the first continuously published term structure for compute, on the same index ICE's and Architect's pending contracts will settle on. Extraction, the convenience yield exposed against provider term sheets, a two-factor hedge calibrated rather than assumed, and the governance question underneath — one administrator now carries four instruments and has published no calculation.
ReadHedging corporate event risk inside a hierarchy
A prediction-market architecture where liquidity concentrates in a small number of parametric benchmark contracts and idiosyncratic exposure is bought back as basis — the ILW and CDS-index pattern applied to corporate events. Built out across six families: pre-release content compromise and mass-tort litigation in depth, then M&A deal-break, cyber, recall, and approval risk, against the CFTC event-contract regime and market-scoring-rule design. Interactive charts, tier-efficiency and tranche/correlation models, full source list.
ReadThe denominator problem — what a GPU-hour actually costs in kilowatt-hours
Every compute-to-energy model multiplies nameplate TDP by an assumed PUE — including four dashboards on this site. NLR measured it instead at 0.1-second resolution and published everything under CC-BY. Real workload duty factors run 0.795–0.880, so nameplate overstates device energy by 12–21%; facilities peak at 73–80% of rated, and peak sizing versus energy volume need different derates. What that means for the denominator of an energy-normalized compute index — and a public correction of our own published numbers.
ReadKalshi files metals perpetuals — and leaves the ratio leg on the table
Kalshi filed with the CFTC to list gold, silver, and platinum perpetuals — its first expansion beyond crypto, on a 45-day clock, inside the blast radius of CME's lawsuit. The contract it didn't file: a Gold/Silver Ratio perpetual completing a no-arbitrage triangle with the legs it just filed — funding assembled from the venue's own prints, margin on ratio vol instead of two gross legs, and the FX-cross precedent for why the third contract deepens the first two. Interactive triangle, funding decomposition, and margin comparison.
ReadThe 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 — registered plumbing anchored at the transfer agent, not an offshore wrapper. What programmable equity collateral means for prime brokerage: financing disintermediation, the wrapper-basis haircut, sec lending, netting, and the dual-book seam. With an RWA tokenization timeline and a token-vs-SSF-vs-offshore-perp comparison.
ReadPaper 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, and location. EFP mechanics are how paper stays welded to physical in every mature commodity — a deep dive through the crude, gold, gas, and metals precedents, with an EFP lifecycle explorer, basis-table simulator, and convergence lab.
ReadCompute 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 this site's 374,550 hourly ERCOT/PJM settlements: the energy content of a GPU-hour, the arb between the normalized benchmark and the real intraday shape, and the compute heat rate with every leg finally listed.
ReadThe 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: funding rates, mark price, and the oracle handoff to STAR Market × USD/CNY are the whole risk picture. Interactive cascade lab, dashboard placeholder, and a full build blueprint.
ReadThe 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 lists a futures contract for every hour of the day at seven hubs — and ElectronX already trades the same hours in bounded, fully-collateralized form, while ICE's TB4 future settles the top-4-minus-bottom-4 spread outright. Built on 374,550 hourly ERCOT and PJM settlement prints: the seasonal 24-hour shape (summer HE20 at ERCOT North averages $205 vs $21 overnight), a 24×24 hour-spread monitor with 99.7% bands, TB4 distributions whose 99.7th-percentile day is $3,133, the 41% real-time-beats-day-ahead hit rate underneath ElectronX's binaries, and the load/temperature maps behind spread spikes. Interactive six-tab dashboard inline.
ReadFour 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 — binary event ladders settling on Ornn prints, the first executable forward pricing in the complex. That makes four venues (Kalshi, Architect, CME × Silicon Data, ICE × Ornn) on three curve technologies and two settlement philosophies — and three of the four settle Ornn. Launch-day implied forwards from the actual strike ladders (H100 ≈ $2.52, B200 ≥ $7.00, with the monthly ladder internally inconsistent), an eight-entry arbitrage monitor, the decay-vs-scarcity tenor model, and the power-market link. Interactive five-view curve comparator inline.
ReadSingle-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. 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 portfolio-margin math, and the index-TRF precedent that already ran to completion. Includes the no-arbitrage rebuttal to "lower margin, no debit rate on shorts": the borrow fee is netted into the futures entry price, matched to the dollar in a step-by-step walkthrough. Threat map, moats ranked by durability, interactive 4-tool dashboard, two PDFs.
ReadOffshore perpetual futures — margin, liquidation, and the October 10 stress test
$19B liquidated in hours, 1.62M accounts, 87% longs. Side-by-side comparison of Binance, Bybit, OKX, Hyperliquid, dYdX, Coinbase Derivatives, and Deribit across 18 dimensions of margin methodology, liquidation policy, and perp design — anchored on the Oct 10, 2025 cascade that exposed how each design actually performs under stress.
ReadLitigation 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. A four-phase listing plan, 16 candidate cases with $1T+ aggregate exposure, a perpetual-on-index design with cash-carry funding, and why the Kalshi precedents clear most of the regulatory path. Interactive case explorer + downloadable PDF.
ReadA Gold/Silver Ratio Perpetual — contract design, funding mechanics, and the cost-benefit case
A proposed CFTC-regulated perpetual on COMEX GC/SI VWAP with cash-carry-anchored funding. Arbitrage triangle against two-leg cleared and ETF pairs, bid-ask and margin economics, full index methodology, and an interactive sizing dashboard.
ReadThe 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. 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 pricing PJM and ERCOT basis 12-36 months ahead. Interactive dispersion dashboard inline.
ReadPredictive market ETFs: what's filed, how the swaps work, and two concepts the market hasn't priced
Three sponsors (Roundhill, Bitwise, GraniteShares) filed 24 prediction-market ETFs; SEC paused them May 5. A walkthrough of the TRS plumbing that makes a 1940 Act fund possible on a CFTC event contract, the binary return profile, plus two unbuilt structural concepts — predictive signal ETFs (trading traditional assets on prediction-market info) and overlay products (hedge + return enhancement) — with an interactive sizing dashboard.
ReadPerps come onshore: what the CFTC's May 29 approvals change about contract design
The CFTC just approved Kalshi's BTCPERP as the first US-regulated perpetual, issued a policy statement on the listing of perps, and cleared a Coinbase pathway to Deribit. A walk through what a perp actually is, what knobs designers turn, and how the offshore (Hyperliquid HIP-3) and the onshore (Kalshi DCM) paradigms compare — with an interactive caps/floors dashboard.
Read