Advisory research · Compute · Cross-venue basis · 3 August 2026
Kalshi and Polymarket both list GPU rental price contracts. Both settle them against Ornn. Not against two Ornn products, not against an Ornn index and an Ornn-derived index — against the same dashboard URL, named in both contract texts, which we verified directly rather than inferred. That makes the Kalshi–Polymarket spread the first live compute spread with no index basis inside it, and on the face of it the one clean cross-venue observation the asset class has produced. It is still unreadable. Matching the contract structure collapses the headline gap from $2.17 to $0.10, and what remains is smaller than the bid-ask on one venue alone. The obstacle turned out not to be the index.
Extends the practice's compute corpus: the implied forward curve work, the index methodologies catalogue, the partition rule, and the lender hedge program design. All venue observations 3 August 2026 unless dated otherwise.
dashboard.ornnai.com — verified in contract text, not inferred from a press releaseThe single question this piece was commissioned to answer, and the one place the public record is unambiguous.
Kalshi's compute series carry a machine-readable settlement source in their series metadata.
For KXB200MON — the monthly terminal-value B200 ladder — it reads, verbatim:
Polymarket's resolution text for GPU rental prices (B200) end of 2026 reads:
Two venues in two regulatory regimes, with disjoint participant sets, settle their B200
contracts against the same URL. ornnai.com 302-redirects to
ornn.com — same administrator, confirmed at the DNS layer rather than assumed. There
is no index basis in this spread. Our forward-curve
piece identified Kalshi-vs-ICE as the one cross-venue compute spread with no index basis in
it; that pair is still pending on the ICE side. Kalshi–Polymarket is the same condition, live
today. That claim needs updating, and §07 does it.
Two qualifications, because absence of published detail is a claim about documentation and not about practice — house convention, and it cuts both ways here. First, Kalshi's contract text never writes "OCPI". It names "Ornn" and a URL. Ornn separately brands its published series OCPI-H100 SXM, OCPI-H200, OCPI-B200, OCPI-B300 and OCPI-RTX5090. The mapping from a Kalshi B200 contract to OCPI-B200 is overwhelmingly likely and is not written down anywhere we could fetch. Polymarket, which does write "Ornn B200 Index", is in this narrow respect the better documented of the two.
Second, and worse: Kalshi's published contract-terms documents for these series
(GPUA.pdf, GPUMON.pdf) are unfilled boilerplate. The operative sentence in
the live PDF is, literally, "The Underlying for this Contract is the value of <index>
<in/at> <time period>" — placeholders, unreplaced, on a designated contract
market's own asset server. Every enforceable term for the compute complex lives in per-market API
strings rather than in the contract-terms document those strings point at.
We found no licensing agreement, no benchmark-administration statement, and no continuity obligation from either venue to Ornn, or from Ornn to either venue. Ornn's own site names no exchange partners at all. Both venues reference a free public dashboard operated by a 2025-vintage startup by URL, with unilateral fallback discretion written in — Polymarket resolves on "the latest data available" if the print is not finalised within seven calendar days; Kalshi disregards revisions made after expiration. Zero index basis and zero documented entitlement to the index are, it turns out, entirely compatible.
This is where the spread actually lives. The two venues list payoffs that are not the same instrument, and in one case are not even the same kind of instrument.
Both venues run ladders on the same index, and both label them by chip and by date. Underneath, four distinct payoff types are in play — and the two that carry the year-end headline are not comparable to each other at all.
| Venue / family | Payoff | Operative language | Reads as |
|---|---|---|---|
Kalshi MONe.g. KXB200MON-26AUG31 |
Terminal value, single date | "…is above $5.71 on Aug 31, 2026" | Terminal CDF |
Kalshi WSweekly |
Terminal value, single timestamp | "…above $9 on Sep 11, 2026 at 4 PM ET" | Terminal CDF |
Kalshi MSmonthly average |
Asian — arithmetic mean of hourly prints | "…the average value… in July 2027", calculated as "the arithmetic mean of hourly values reported by Ornn" | Average, ≠ terminal |
Kalshi MAX / Qyearly / quarterly |
American touch barrier. Settles the moment the barrier prints | "…is above $7.07 by Dec 31, 2026"; can_close_early: true |
Running max — NOT a forward |
| Polymarket "end of <month>" | Terminal value, mutually exclusive brackets | "…according to the finalized Ornn B200 Index price for August 31, 2026" | Terminal PDF |
| Polymarket "hit ___" | Two-sided American touch (↑ high and ↓ low strikes in one event) | "…displays a finalized price equal to or beyond the listed price for any day between market creation and December 31, 2026" | Running max / min |
KXB200MAX-26DEC31-7.070 — nominally a December 2026 contract — carries
status: finalized, result: yes,
occurrence_datetime: 2026-07-19T16:43:11Z, expiration_value: 7.28.
It resolved on 19 July 2026, five and a half months before the date in its own ticker,
because B200 printed $7.28 and the barrier knocked in. KXH100MAX-26DEC31-3.080 did
the same on 22 June at $3.17. Anyone reading the Kalshi year-end ladder as a distribution over
the 31 December price is reading a running-maximum distribution whose lower half is already
worth par. B200 spot sits near $5.5–5.7 while every year-end strike up to $7.07 is settled at
100. That is not a bullish forward. It is a different contract.
Both venues list both structures. So the like-for-like comparison exists — it is simply not the one the headline numbers are built from.
Polymarket lists a B200 terminal bracket for 31 August 2026 and a B200 touch ladder through 31
December 2026. Kalshi lists KXB200MON-26AUG31 — a terminal threshold ladder on the
same date, same chip, same index — and KXB200MAX-26DEC31, a touch ladder over the same
window. Two clean pairs. Below, each pair reconstructed to a single implied level.
Kalshi's ladders are overlapping "above $X" thresholds, so a quote set is a discretised survival function and E[X] = x0 + ∫ P(X>x)dx. Polymarket's brackets are a partition, so the mean is the probability-weighted bracket midpoint. Neither raw quote set is monotone or sums to one (§05), so both are cleaned by pool-adjacent-violators isotonic regression before integration, with linear tail decay. These are our reconstructions from stated assumptions, not venue-published forwards — illustrative structure, not forecasts. The point of the exercise is precisely how much the answer moves with the cleaning rule.
| Pair | Kalshi bid | Kalshi mid | Kalshi ask | Polymarket | Spread (mid) |
|---|---|---|---|---|---|
| Terminal — B200 value on 31 Aug 2026 | $6.00 | $6.22 | $6.37 | $6.56 | +$0.34 |
| Touch — B200 running max through 31 Dec 2026 | $7.57 | $7.72 | $7.88 | $7.83 | +$0.10 |
| Mismatched — Kalshi touch vs Polymarket terminal the pairing the headline comparisons use | — | $7.72 | — | $5.55 | −$2.17 |
The third row is the one in circulation. Both venues label a ladder "B200 … 2026"; one is a running maximum and the other is a terminal value, and pairing them manufactures a $2.17 gap out of a barrier premium. A widely shared comparison put the year-end Kalshi–Polymarket B200 gap near $0.79. We could not verify that figure against a primary source and do not assert it — but we note that the pairing choice alone spans $0.10 to $2.17 on today's tape, which brackets it comfortably. Any number in that range is producible without either venue's participants disagreeing about anything.
Spread decomposition — pick the pairing, pick the side of the book
Live reconstruction from the 3 August 2026 quote sets embedded below. The bar is the Kalshi bid-to-ask band; the marker is Polymarket. When the marker sits inside the bar, the cross-venue spread is smaller than the measurement error on one leg — which is the finding.
A cross-venue gap on one underlying decomposes into four terms. Here the first is verifiably zero, the second is unquantifiable, the third is not a price at all, and the fourth is smaller than the noise.
| Term | Status here | Evidence |
|---|---|---|
| (i) Index / settlement basis | Zero — verified | Identical settlement URL in both contract texts. The only compute spread in the corpus where this term is measured rather than assumed away. Residual: differing finalisation and revision conventions (7-day fallback vs post-expiry revisions disregarded), which bite only on a disputed print |
| (i·b) Contract-structure basis the term the four-way decomposition omits |
Dominant | $2.17 of a $2.17 naive gap. Touch vs terminal vs arithmetic-average are three different payoffs on one index. Not a basis in any economic sense — a category error priced as one |
| (ii) Liquidity & fee differential | Not separable | Kalshi's 31 Aug B200 ladder averages a 15.7-cent bid-ask across 20 strikes; 11 of them have never traded. Polymarket's entire live compute complex is roughly $140k lifetime volume across ~16 events. Fees are asymmetric and, on the Polymarket side, not established for this category at all (§06) |
| (iii) Regulatory / access differential | Segmented, not sealed | The GPU markets sit on the offshore, Polygon-settled polymarket.com, unreachable from US IP addresses; Kalshi is a CFTC-designated contract market. For retail the participant sets are disjoint. For eligible contract participants they are not — a registered swap dealer can face a US institution bilaterally and carry the offshore leg on its own book. That channel is live as of mid-2026 (§06a) and still does not compress this spread, for reasons that have nothing to do with access |
| (iv) Genuine forecast disagreement | Below the noise floor | Structure-matched residual is $0.10 (touch) and $0.34 (terminal). The Kalshi leg's own bid-ask band is $0.31 and $0.37. The residual is 0.33× and 0.92× the measurement error on one venue alone |
Our methodologies catalogue concluded that undisclosed index construction makes compute spreads uninterpretable. This piece reaches the same destination from the opposite direction and it is the stronger result: we removed the index-basis term entirely — verified to zero, not assumed — and the spread is still uninterpretable. Eliminating the disclosure problem does not deliver a readable basis, because the binding constraints are contract-structure heterogeneity, quote sets that violate their own internal arbitrage conditions, and a participant-set partition that prevents the spread from being traded into line. Index disclosure is necessary. It is visibly not sufficient.
Before any cross-venue number means anything, each ladder has to satisfy the conditions that make it a distribution. Neither does.
The 31 August B200 event is a six-bracket, mutually exclusive, collectively exhaustive partition
of the index outcome — <$5.50, $5.50–6.00, $6.00–6.50,
$6.50–7.00, $7.00–7.50, $7.50+. Our
partition work gives the condition:
the Yes prices of an exhaustive partition must sum to one, or a riskless box exists. Displayed Yes
quotes on 3 August were 20.9¢, 35¢, 33¢, 22¢, 35¢ and 34¢.
P(X>x) must be non-increasing in x. On the 20-strike 31 August B200 ladder, mid quotes violate this at four adjacent pairs:
| Lower strike | P(>lower) | Higher strike | P(>higher) | Implied bracket prob. |
|---|---|---|---|---|
| $5.41 | 0.715 | $5.51 | 0.720 | −0.005 |
| $5.71 | 0.570 | $5.81 | 0.600 | −0.030 |
| $6.41 | 0.335 | $6.51 | 0.350 | −0.015 |
| $6.61 | 0.295 | $6.71 | 0.335 | −0.040 |
Breeden–Litzenberger extraction — the technique our forward-curve piece applies to these ladders — requires a monotone survival function to return a non-negative density. Applied raw to this ladder it returns negative probabilities at four points. The cleaning step is therefore not cosmetic: it is the step that determines the answer. Running the same integration on the bid side, the mid, and the ask side of the identical book produces $6.00, $6.22 and $6.37. The choice of which side of a stale book to read moves the implied forward by 37 cents; the entire cross-venue spread is 34.
Kalshi launched Compute Forward Curves on 14 July 2026 for B200, H200 and A100 — H100 excluded, despite H100 markets existing and H100 being Ornn's flagship series. The curves are described as "derived from Kalshi's own market data on weekly and monthly chip prices." No construction methodology is published. Given that the underlying ladders violate monotonicity at four of nineteen adjacent pairs and carry 15.7-cent average spreads with 11 of 20 strikes untraded, the undisclosed cleaning rule is not an implementation detail — on today's tape it is a larger input to the published number than anything the market is saying. This is the same finding as the index catalogue's, one level up the stack: the curve derived from the ladder inherits an additional undisclosed methodology on top of the undisclosed index methodology beneath it.
Three practical findings that determine whether any of the above is tradeable, and one that questions whether Polymarket's compute history is a single series at all.
| Kalshi | Polymarket (GPU markets) | |
|---|---|---|
| Legal status | CFTC-designated contract market | Offshore polymarket.com, Polygon-settled. Not Polymarket US / QCX LLC, the group's DCM |
| US retail access | Yes | No — the international venue is not available from US IP addresses |
| Compute contracts self-certified with the CFTC? | Yes — product certifications published for GPUA/GPUMON, though the linked contract-terms documents are unfilled templates | Not established. Sampled QCX product certifications cover athletic-outcome and reality-television contracts; none we retrieved mention GPU, compute, or any price index. Our negative finding is sampled, not exhaustive — the CFTC filing index did not return a complete machine-readable list |
| Trading fees | Taker 0.07·C·P·(1−P); maker 0.0175·C·P·(1−P); no settlement fee. Compute series carry fee_type: quadratic, fee_multiplier: 1 — no discount | Not established for this category. Geopolitical/world events are fee-free; sports carry an explicit formula; Finance and Tech are listed as planned. Polymarket US publishes a separate schedule (taker 0.06, maker −0.0125 rebate) that does not govern these markets |
| Live compute liquidity | Deepest observed single market: 84,769 contracts volume / 33,148 OI (KXB200Q-26JUN30-6.270). Front-month terminal ladder far thinner — 11 of 20 strikes never traded | ~$140k lifetime volume across ~16 live events. Largest live event $42.9k. Open interest not displayed |
| Institutional footprint | Compute Forward Curves product, 14 Jul 2026 | First institutional block trade on a GPU instrument, 2 Jun 2026 — six figures, FalconX × AneraLabs, settled against OCPI on Polygon |
The access wall above is a retail wall. It is not a wall for eligible contract participants, and the bridge was built during the eight weeks before this piece was written. A CFTC-registered swap dealer can face a US institution bilaterally under an ISDA and carry whatever offshore leg it likes on its own book; the client never touches the venue. Three transactions establish that this is operating rather than theoretical:
| Date | Dealer | Trade | Reference |
|---|---|---|---|
| 27 May 2026 | FalconX FalconX Bravo, Inc. — CFTC-registered swap dealer, NFA member | "First OTC compute forward" — an over-the-counter swap referencing the forward price of compute, with Robert Leshner (Superstate) as counterparty | OCPI-H100 directly |
| May 2026 | Galaxy Galaxy Derivatives — swap-dealer unit | $10m event swap with Arca, under existing ISDA documentation | Event outcome |
| 2 Jun 2026 | FalconX | Six-figure block trade on a GPU instrument, with AneraLabs — hedging Anera Exchange forward inference capacity. FalconX stated an intention to act as dedicated market maker for future Polymarket block trades | OCPI, settled on Polygon |
| Jun 2026 | Galaxy | Prediction-market OTC derivatives desk launched — event swaps for hedge funds and family offices, explicitly so institutions can "trade event risk without directly accessing prediction-market exchanges" | Events listed on Kalshi and Polymarket |
So the honest version of term (iii) is segmentation, not prohibition. The line the CFTC polices is not offshore exposure — it is the form of the intermediation. The 2024 Falcon Labs order, $1.77m in disgorgement and penalty, charged unregistered FCM activity for an "Edge" product that gave US institutions cross-margined direct access to offshore venues via sub-accounts and API keys, without identifying those customers to the exchange. Routing a client's order to the offshore book is FCM activity. Facing that client as principal in a bilateral swap and hedging your own book is swap-dealer activity, and both firms above are registered for it. The client's exposure in the permitted structure is to the dealer, not to Polymarket.
Arbitrage requires the channel and an executable edge, and three of the four obstacles in this piece survive intermediation untouched. Structure mismatch is unaffected — no dealer can arb a touch barrier against a terminal bracket either, so the $2.17 is not a trade for anyone. Depth forecloses size: Polymarket's entire live compute complex is roughly $140k lifetime volume against a $10m single event swap on the OTC side, so the dealer cannot hedge institutional size on Polymarket — meaning the OTC quote does not transmit Polymarket's price into US hands, it replaces it with the dealer's own mark. And the edge is inside the spread: $0.10–$0.34 of structure-matched difference against a $0.31–$0.37 bid-ask band on the Kalshi leg alone, before Kalshi's taker fee, the dealer's undisclosed markup, and collateral on both sides.
Note what the channel does to the measurement problem, which is the reverse of what one would expect. The FalconX compute swap references OCPI directly — not the Polymarket price. An institution reaching this risk through a dealer is not getting the offshore venue's price at all; it is getting a bilateral quote on the same underlying, with a markup no one publishes, under documentation no one files. Intermediation converts a venue basis into a dealer basis. The count of unmeasurable layers goes up.
The largest compute market either venue has ever run is Polymarket's GPU rental prices (H100) hit ___ in February, at $337,175 volume — 2.4× the entire live compute complex today. Its resolution text does not mention Ornn. It reads: "…if, for any day between February 2 and February 28, 2026, the Silicon Data H100 Index (SDH100RT) has a price equal to or above the listed price." Polymarket's compute contracts referenced Silicon Data early in 2026 and reference Ornn now. We could not establish any corporate relationship between the two administrators, and found no announcement of the switch. Two consequences. First, Polymarket compute prices are not a continuous time series — anyone backtesting across the change is splicing two benchmarks. Second, it is worked precedent that the settlement reference of a live compute complex can be changed by the venue, unilaterally, with no published notice — against a benchmark it holds no documented licence to. That is the risk our lender piece asks credit agreements to paper as a methodology-change fallback, and it has already happened once.
Two published claims need qualifying, in opposite directions.
| Published claim | Status | Revision |
|---|---|---|
| Compute index design: Kalshi is the only live implied compute forward curve; the Kalshi-vs-ICE spread is the one cross-venue compute spread with no index basis inside it | Needs qualifying | Polymarket lists comparable terminal-bracket ladders on B200, H100, H200, A100 and RTX 5090, from which a forward is extractable by the same construction. And the zero-index-basis condition is already live on the Kalshi–Polymarket pair while ICE's OCPI future remains pending. Both claims should be narrowed to "the only regulated-venue live curve" and "the first zero-index-basis pair between two CFTC-regulated venues" |
| Index methodologies: four administrators catalogued; undisclosed methodology makes cross-venue comparison uninterpretable | Reinforced, and extended | The catalogue's completeness claim needs a fifth settlement-reference row — Polymarket, referencing Ornn by dashboard URL, with a documented prior reference to Silicon Data. And the conclusion strengthens: disclosure is necessary but insufficient, because structure and microstructure bind even at zero index basis. Addendum drafted separately |
The one-paragraph version. Kalshi and Polymarket settle GPU rental contracts on the same Ornn dashboard URL, verified in both contract texts — so the index basis between them is zero, the first time this corpus has been able to measure that term rather than assume it away. The spread is still unreadable. Pairing the two venues' year-end B200 ladders produces a $2.17 gap that is entirely an artefact of comparing an American touch barrier against a terminal bracket; matching the structure collapses it to $0.10 and $0.34, both smaller than the bid-ask band on the Kalshi leg alone. Underneath, each venue's ladder violates its own internal arbitrage conditions — Polymarket's exhaustive partition sums to 180%, Kalshi's survival function is non-monotone at four adjacent strikes — by more than the cross-venue spread they are being used to measure. A channel to trade it does exist for eligible contract participants — registered swap dealers began writing OTC compute and event swaps against these references in May 2026 — but it compresses nothing, because no dealer can arb a barrier against a terminal value either, because $140k of offshore depth cannot absorb an institutional hedge, and because the resulting bilateral quote replaces the offshore price rather than transmitting it. The conclusion is the methodologies catalogue's, arrived at from the opposite end: index disclosure is necessary and it is not sufficient.