A buyer of compute can hire a broker, post an order, join an auction, call a trading desk, route a request, read five benchmarks and, offshore, trade a future on the price. What it cannot hire is anyone to decide where its work should run. The buyer still makes that decision itself, one vendor call at a time.
This piece does two things. First, it inventories the firms that stand between a workload and a chip: 64 of them, in six seats, from brokers that have worked the phones for years to an inference auction that opened in October. The Venues and the Tranches mapped eleven venues from the side of an owner with a cluster to fill. This one widens the lens to every seat, on the chip side and the token side, and looks from the side of the buyer with work to place.
Second, it asks what the inventory lacks. Each firm sells one piece of a deal: a match, a quote, a route, a price, a contract, a check. Nobody sells the whole decision. Power markets gave that job a name long ago. The dispatcher takes the load, reads every generator’s offer and its limits, and decides which units run. Compute has loads and it has offers. It has no dispatcher.
The conclusion, stated up front: the dispatcher will be software, it already exists for tokens, and on the chip side it stops at about a month. What stops it is not a lack of intelligence. A term deal for compute has no common description, cannot be handed to someone else, and has no shared proof that it was delivered. Fix those three and the seat fills itself. The firms that gain most are the ones with no sales desk and no procurement desk: small operators, owners of private clusters, and buyers too small to hold a broker’s attention.
Six seats, more than sixty firms. Each one sells a piece of the deal.
The seats are sorted by the job a firm does in a transaction, because the job decides what a dispatcher could hand to it. A firm that does two jobs is listed under the one it is known for, and the other is noted.
| Firm | What it does | Takes the other side | Status |
|---|---|---|---|
| Brokers and arrangersfind the other side, for a fee | |||
| Epilogue | Matches renters with vetted supply for large private deals. Describes a standard specification sheet, a standard contract and a shared data room for each deal. | No | Per an indexed copy of its site |
| American Compute | The buyer owns the servers. American Compute sources them, sets them up and runs them for flat fees, and finds a renter for 2.5% of the contract if it signs. It also introduces lenders and arranges cover on resale value and on a renter that stops paying. | No | Live |
| Compute Exchange | Requests and auctions across more than 100 providers on terms of 1 to 36 months, a used-hardware market since July 2026, and contracts that fix a token price for up to six months. Sister company to Silicon Data. Takes no positions. | No | Live since 2025 |
| Andromeda | Standard contracts with service levels across more than 100 providers, a certification engine that benchmarks each cluster, and one invoice. | As reseller | Live |
| Shadeform | One interface and one bill across more than 30 clouds. Takes reservation requests, returns provider quotes and arranges dedicated cluster builds. | No | Live |
| NVIDIA DGX Cloud Lepton | NVIDIA’s marketplace across 20 listed partners, from AWS and CoreWeave to SF Compute, behind one interface. | No | Live since 2025 |
| Architect’s broker arm | A registered broker for block trades in capacity and for swaps on compute and memory. | No | Live |
| BGC Group | A broker between dealers that says it is building a private market for compute. Data partner of Liquid Compute. | No | Building |
| Stoa and Nodestream | Two August 2026 launches for hardware, not hours: Stoa takes quote requests for GPUs and servers, and Nodestream runs a verified marketplace for equipment. | No | Launched August 2026 |
| Principals and deskstake the other side | |||
| Jump Trading, with the Itô team | Pools capacity and resells it in smaller blocks on more flexible terms. Itô called itself “an agent-run OTC desk for sourcing, selling, hedging, and contracting GPU compute” at one fixed rate. Its public code describes contracts of 30 to 365 days. | Yes | Desk since October 2026 |
| DRW | Three teams on compute. Its founder co-founded Compute Exchange, invested in Silicon Data and calls the firm’s physical trading so far “dabbling”. | Yes | Limited so far |
| FalconX | Has completed a private swap tied to future compute prices. | Yes | One trade reported |
| Hyperbolic | Gathers supply from neoclouds, idle enterprise capacity and customer resale. Has published a model trade: buy a 36-month block, sell it in short blocks and hedge the last year with a put. | Yes | Live |
| Vessl | Korean firm that buys bare metal on long terms and resells it with a management layer. | Yes | Live |
| Order books, auctions and marketplacespost a price and match the trade | |||
| SF Compute | An order book for H100 node-hours. Fills from one node for one minute. Reserved time resells under the same rules. It operates the clusters it sells. | No | Live |
| Computable | GPU hours by the calendar week at a public price, with unused hours sold back. A Y Combinator company from the summer 2026 batch. | No | Live |
| Liquid Compute | Building an order book for physical capacity, with published pricing data. Has applied to run a futures exchange and a clearing house. Raised $15 million in seed funding. | No | Launched September 2026; approvals pending |
| Prime Intellect | Compute, training and inference under one interface. Its “liquid compute” resells unused capacity. It added an inference service in October. | No | Live |
| Vast.ai and io.net | Open marketplaces that now sell to agents directly: per-second billing at Vast.ai, and a tool interface for agents at io.net since March 2026. | No | Live |
| Akash, Spheron, RunPod, Salad, Lium | Open rental marketplaces. Often the lowest prices, with reliability that varies by host, on one third-party comparison. | No | Live |
| OneChronos, with Auctionomics | An auction that clears bundles of compute, power and dates as one problem. Its lead product has been described as a swap on dollars spent on leading models. | No | Announced July 2025; not launched |
| National Compute | A new company launched by Anjney Midha, a former general partner at Andreessen Horowitz. Its planned “Grid” would pool capacity that members rent but do not use, across providers and chip types. The business model is not final. | No | Announced October 2026 |
| Exascale | Presents itself as an exchange for GPU compute, with spot, forward and futures markets, beside its own cloud and model interface. | No | Per its site |
| Routers and token venuessend each request to a model or a host | |||
| OpenRouter | One interface in front of more than 80 providers and more than 400 models. Picks the host for a named model, with fallback. Charges 5.5% on credit purchases. Stripe agreed to buy it in August 2026; press reports put the price at $7 billion or more. | No | Live |
| Architect Liquid Inference | An auction for every request. The lowest offer that meets the buyer’s rules on cost, speed, region and data handling wins. The price cap locks before the first token, and each job gets a receipt. | No | Live since October 7, 2026 |
| Requesty, Portkey, LiteLLM, Vercel AI Gateway | Gateways that route by rule and handle fallback, budgets and logs. Requesty adds 5% to model cost. Vercel adds nothing. | No | Live |
| Martian and Not Diamond | Choose the model itself, prompt by prompt, on predicted answer quality. | No | Live |
| DigitalOcean and You.com | Routers run by hosts. You.com sends each request to its own capacity first. DigitalOcean also sells its capacity through OpenRouter. | Own capacity | Live |
| TrustedRouter | Routes to open models on encrypted, verifiable hosting. Closed a $6 million seed round in October. Plans a “New York Token Exchange” where many hosts compete for each request. | No | Router live; exchange planned |
| Antseed | A peer-to-peer market for model access, bought from independent providers. Raised $2.4 million in October, led by Spark Capital. | No | Live since May 2026 |
| Flop Labs and MARA | Announced, not live: a token meant as a claim on delivered compute (Flop Labs, August 2026) and a marketplace that would price inference by energy used per token (MARA, at the hiring stage). | n/a | Announced |
| Prices and contractssay what it is worth, and let it be hedged | |||
| Silicon Data | Rental benchmarks for the H100 and B200. CME’s planned futures settle on them. Backed by DRW and Jump. | n/a | Live |
| Ornn | Transaction-based indices for chips, memory and tokens. Its index is behind ICE’s and Kalshi’s contracts, and Architect named it for its offshore perpetuals in January 2026. Building a capacity aggregator. Raised $33 million in seed funding. | n/a | Live |
| Compute Desk, with General Index | Daily benchmarks administered by a regulated third party. Behind Nodal’s announced futures, Architect’s planned United States contracts and, by Architect’s own filing, its offshore futures. | n/a | Live |
| NativX and Inferra | Two further index publishers: NativX’s energy-normalised index, licensed to ICE, with an open-source forward model; and Inferra’s on-chain index by chip model. | n/a | Live |
| Setara Financial and 252 Capital | Two proposals for a common measure. Setara is building what it calls a governed reference layer for compute markets, with values marked as pre-production. 252 Capital has published a draft “Effective Compute Hour”, a unit adjusted for performance. | n/a | Pre-production |
| CME, ICE, Nodal, Architect, Kalshi, Polymarket | The listed venues. Architect’s offshore perpetuals and Kalshi’s event contracts trade, and Polymarket’s United States venue certified contracts on the same Ornn index in August. CME’s futures are under review until January 8, 2027. ICE has published contract months on Ornn’s index; no listing is confirmed. Nodal’s futures are announced. | n/a | Mixed |
| Touchmark | A Y Combinator startup building a forward market for prepaid inference tokens. | n/a | Early |
| Plumbingcheck it, move it, pay for it, insure it | |||
| Compute Clear and ComputeConnect | Compute Desk’s delivery layer and its network with Architect, meant to turn a futures position into delivered capacity using published adjustment tables by chip, memory and location. | n/a | Announced July 2026 |
| SemiAnalysis ClusterMAX | Grades providers. Third edition published September 2026. | n/a | Live |
| SkyPilot | Runs one job description on more than 20 clouds and picks by cost and availability. Raised $20 million in July 2026. | n/a | Live |
| x402, Machine Payments Protocol, Agent Payments Protocol | Payment standards built for software that pays. Live, and sized for small payments. | n/a | Live |
| USD.AI, ForwardCompute, Princeps, NVIDIA | Credit and cover. USD.AI lends against chips on chain, with a warranty on resale value. ForwardCompute plans outage cover, resale-value guarantees and credit protection through the Lloyd’s market; its site is marked as coming soon. Princeps is a Y Combinator-backed insurer for compute, founded this year. NVIDIA has rented back idle capacity from its partners at a fixed rate since July 2026. | n/a | Mixed |
Three things stand out when the table is read as one market.
Almost nobody holds risk. Five firms take, or say they will take, the other side of a compute deal. Jump’s desk and Vessl are the two that clearly do so in physical capacity today. The other seats match, route, measure or check. That is the pattern of a young market: plenty of introductions and very little warehousing. The Contract Has a Date. It Needs a Dealer. made the same point about the futures. It holds for the physical market as well.
The two sides barely touch. Seats A to C sell chip-hours. Seat D sells tokens. A handful of firms appear on both sides: Architect runs a venue for each, Compute Exchange auctions chips and fixes token prices, Ornn measures both, and Hyperbolic and Prime Intellect rent chips and sell tokens. None of them offers a buyer one decision across the two.
Standards are private. Epilogue describes a standard contract. Andromeda has one. SF Compute has standard units. Compute Desk plans standard adjustment tables. Each works inside one firm’s network, and none is shared. The market has several standards, which is another way of saying it has none.
Software closes the short end of the market. People close the long end.
Read the map from the top left. A model request is bought and sold by software in seconds: a router picks the host, or an auction does. An hour of chip time is the same. SF Compute’s book fills from one node for one minute, and Vast.ai and io.net now describe themselves as built for agents to buy from directly. Computable stretches the same idea to a calendar week.
Now read the bottom right. A contract for one to five years is found by a broker, negotiated by lawyers and signed by people. Epilogue’s pitch is to make that process less painful: one specification sheet, one contract form, one data room. American Compute will find a renter for a fee. The neoclouds sell direct. This is where most of the money is.
The middle row is the interesting one. Compute Exchange returns quotes from verified providers on terms of 1 to 36 months, in under two hours by a competitor’s account, and then people sign. Andromeda matches a request to a certified cluster on a standard contract. Itô, before Jump hired its founders, ran what it called an agent-run desk: software quoted a fixed rate for 30 to 365 days and the desk took the other side. That is the nearest anyone has come to the empty corner: a term deal that software can close.
The corner is empty for a reason, and the reason is not that nobody thought of it. Section 05 lists what stands in the way.
Since July, the trading firms, the exchange builders and a payments company have moved in.
Many of the seats in Table 1 have been filled for a year or more. What changed in the last four months is who is sitting down.
| Date, 2026 | Who | What |
|---|---|---|
| November 4 | CME and the Commission | CME consents to a longer review of its H100 and B200 futures. The end date moves from November 9 to January 8, 2027, or sooner if a comparable contract lists elsewhere. |
| October 22 | You.com | Launches an inference router that sends each request to its own capacity first and promises 99.9% uptime. |
| October 16 | TrustedRouter | Closes a $6 million seed round led by Sierra Ventures and says it wants to build a token exchange. |
| October 12–13 | Jump Trading | Hires the two founders of Itô Markets to lead a new compute desk. |
| October 7–8 | National Compute | Announces the Grid, a plan to pool spare capacity across providers. |
| October 7 | Architect | Launches Liquid Inference, an auction for every model request. |
| October 6 | Antseed | Raises $2.4 million, led by Spark Capital, for its peer-to-peer inference market. |
| September 15 | Liquid Compute | Leaves stealth with $15 million and applications pending to run an exchange and a clearing house. |
| September 3 | Compute Desk and Nodal | Announce GPU futures on Nodal Exchange, cleared by Nodal Clear, subject to regulatory compliance. |
| August 21 | Commission staff | Stay a Kalshi contract amendment until November 19, citing novel or complex issues. |
| August 19 | Stripe | Agrees to acquire OpenRouter, the largest router. |
| July 21 | SkyPilot | Raises a $20 million seed round led by Lux Capital. |
| July 17 | Compute Exchange | Opens a market for used and refurbished hardware. |
| July 8 | Architect and Compute Desk | Announce ComputeConnect, a network to turn futures positions into delivered capacity. |
| July | Itô Markets | Executes its first fixed-rate contracts, by its own account. |
Three kinds of firm arrived. The first is the trading firm. Jump did not buy a marketplace. It brought in the two founders of a startup that had built an agent to quote compute and hold the risk, and gave them a balance sheet. DRW has three teams on the subject and backs a benchmark and a marketplace. These firms bring the one thing Table 1 is short of, which is someone to take the other side.
The second is the exchange builder. Architect now has a venue on each side of the market and a broker between them. Liquid Compute has applied to run both an exchange and a clearing house over a physical order book. The third is a payments company. Stripe agreed to buy the largest router, and it co-wrote one of the payment standards that lets software pay for a service without a person approving each charge. A router joined to a payments network is most of a dispatcher for tokens.
The dispatcher takes a workload and its limits, and decides where it runs and under what deal.
The two sides trade different units on different clocks. The token side sells by the request and settles in seconds. The chip side sells by the job or by the term and settles over hours, months or years. A dispatcher runs the same four steps on each. Table 3 sets those steps against the firms in the inventory, to show what already exists and what does not.
| Step | Token side today | Chip side today | Still missing |
|---|---|---|---|
| 1. Describe the need | Solved in practice. A request names a model. Liquid Inference lets the buyer add rules on cost, speed, region and data handling. | Private answers only: Andromeda’s certification, Epilogue’s specification sheet, SF Compute’s requirement lists, ClusterMAX grades, Compute Desk’s planned adjustment tables. | One public set of fields that every seller fills in. |
| 2. Collect firm quotes | Live. Liquid Inference shows an order book. OpenRouter lists each host’s price. | Live by the minute on SF Compute. By request, in hours, on Compute Exchange. By phone from brokers. Silicon Data, Ornn and Compute Desk publish a level, not a quote. | A firm price at one month to three years that software can ask for and get. |
| 3. Choose and structure | Routers pick the host. Martian and Not Diamond pick the model. Compute Exchange will fix a token price for six months. | Itô at Jump fixes a rate. Hyperbolic has modelled buying long and selling short. Architect brokers swaps. American Compute arranges cover. | The choice between tokens and chips. Capital behind the choice. |
| 4. Pay, verify, move | Pay per request, with a receipt. Payment standards for agents are live. | Deposits and schedules. SF Compute refunds failed machines. Compute Clear is announced. SkyPilot moves a job between clouds. | Proof of delivery that two venues both accept. A right to hand the contract on. |
Why nobody chooses between tokens and chips
A buyer of an open model has two ways to get the same output. It can buy tokens from a host, or it can rent chips and run the model. The cheaper way changes with token prices, chip rents and how full the chips will be. No firm in Table 1 makes that choice for a buyer. There are three reasons, and this is our reading, not a finding from any source.
- The term range has a hole in the middle. Capacity bought on demand is usually firm while it runs, but it carries no promise of supply or price tomorrow. So the buyer sees two ends: short and cheap but open to interruption, or firm but years long with a large deposit and no easy exit. Tokens need no commitment at all. The choice is not one price against another. It is a purchase with no commitment against a multi-year bet on the buyer’s own usage. A dispatcher needs firm blocks of days, weeks and months that can be resold. Jump’s desk describes that product, and it is still thin.
- The spread already belongs to someone. The firms that The Refiner and the Merchant called toll refiners, such as Together, Fireworks and DeepInfra, rent chips and sell tokens. They pool many buyers, so their chips run fuller than one buyer’s would. A buyer that makes its own tokens competes with that pooling, and usually loses unless its volume is large and steady.
- The choice exists only for open models. A closed model cannot be self-hosted, and most enterprise spending goes to closed models.
So the first customer for this part of the dispatcher’s job is probably not a typical enterprise. It is a refiner deciding where to rent, or a heavy and steady user deciding whether to become one.
Three designs already point at the seat
Nobody has built the dispatcher. Three firms have built something next to it, and they disagree in a useful way about where the hard part is.
| OneChronos, with Auctionomics | Itô Markets, now at Jump | Architect Liquid Inference | |
|---|---|---|---|
| What it is | An auction that clears packages | An agent that acts as a dealer | An auction for every request |
| How it handles supply that is not standard | Each bidder states its own package and limits. An optimiser clears all the bids together. | The agent adapts each contract to the customer and takes the other side. | It does not need to. A token from a named model is already close to a standard unit. |
| What trades | Bundles of compute, power and dates. The lead product has been described as a swap on dollars spent on leading models. | A fixed rate per GPU-hour for 30 to 365 days | One model request |
| Who holds the price risk | Buyers and sellers | The desk, hedged across providers | The buyer, one request at a time |
| What it needs to work | Many bidders, and a basket of models that everyone trusts to settle against | A balance sheet | Hosts willing to quote |
| Status | Announced July 2025. No launch on the public record. | First contracts July 2026. Team joined Jump in October. | Live since October 7, 2026 |
Each design carries a lesson for whoever fills the seat.
- Detailed bids can stand in for a standard contract. OneChronos does not ask every seller to offer the same thing. It asks every side to state its limits in the same form and lets a matching engine do the rest. People cannot write package bids by hand. Software can. A dispatcher is the natural bidder in that kind of auction, which means the auction and the dispatcher need each other.
- Flexible terms need capital. Owners want long contracts from strong credits. Renters want short terms and often have no record. Itô closed that gap by standing in the middle, and joined Jump about three months after its first contracts. A dispatcher without capital can match two parties. It cannot change the term between them.
- The unit decides the speed. Liquid Inference could launch as a working auction because its unit was already standard. Everything slow about the chip side traces back to a unit that is not.
Eight frictions. The first three decide whether software can close a term deal at all.
The ranking is our judgement. Each lane ends with whether a dispatcher could remove the friction by itself.
A market with few frictions needs five things. Four of them are what the venues piece asked for.
An efficient market for compute would not make every chip alike. It would make every deal describable, priced, transferable, checked and backed. That is a shorter list than it sounds.
The Venues and the Tranches ended with a four-part combination for the owner of a cluster: a public block standard from the certifier, an order book that accepts any graded block, a holder that stands between the book and the lender, and a cleared contract at the middle tenor. Come at the market from the buyer’s side and the same four appear, plus one.
- A public term sheet. This is the block standard, widened. One set of fields that every seller fills in, for chips and for tokens: what the product is, where and when, the service promised, the price and how it is paid, the rights to resell or extend, and how data is handled. The dispatcher does not need one standard contract. It needs every contract written in the same fields, with published adjustments between grades. Compute Desk’s planned tables by chip, memory and location are the nearest thing on record, and Setara and 252 Capital are each working on the measure itself.
- A firm price at every tenor. This is the open order book and the middle-tenor contract. Software can already get a firm price for a minute and for a week. It cannot get one for six months without a person.
- A contract that can be handed on. This is new to the list, and the hardest. A right to resell or reassign, written into the term deal, is what turns a commitment into a position. It is also what a lender will resist first, because the loan was sized on the customer’s name.
- Proof of delivery that travels. This is the grade, measured after the sale as well as before it. A test that two venues both accept does for delivery what a shared grade does for listing.
- Credit that travels with the trade. This is the holder and the clearing house. Until one exists for physical deals, the substitutes are a dealer’s balance sheet and an insurer’s cover on a renter that stops paying.
None of the five asks for chips to become interchangeable. Some frictions are permanent. Compute cannot be stored. A training run needs one cluster. Every chip loses value on the day the next generation ships. A market can work around all three, as power markets work around the fact that electricity cannot be stored either.
What a dispatcher does to the tranches
The venues piece found that lenders already sort chip loans by three things: who signed the contract behind the chip, how long that contract runs, and what the chip is and where it sits. Counterparty, tenor and basis. On the public deals, a chip with an investment-grade contract behind it borrows at about 2.25 points over the benchmark rate. The same chip with no contract pays 9 to 10 percent, and smaller operators pay 10 to 15.
A working dispatcher touches all three, and this is an argument, not an observation.
- Basis becomes measurable. The venues piece called basis the attribute nobody prices. A public term sheet is a description of basis. Once every block is written in the same fields, the difference between two blocks is a number.
- Tenor stops being a cliff. A loan that outlasts its contract is priced for the gap. If the block can be re-let through a dispatcher at a visible price, the gap is a re-letting risk with a history, not a blank.
- Counterparty widens from a name to a pool. The cheapest loans rest on one customer’s signature. A block that many buyers can reach, and that has been re-let before, rests on demand in general. That will never price like a contract with a large technology company. It could price well inside a loan on chips alone.
If that is right, the loans that reprice most are the ones at the bottom of the ladder: the uncontracted and the small. The dispatcher does not make the top tranche better. It makes the bottom tranche something a lender can read.
The gains go to whoever has no sales desk and no procurement desk.
A dispatcher replaces search and negotiation. The firms that are worst at those two things gain the most. The firms that are best at them give something up.
| Party | What it gains | What it gives up |
|---|---|---|
| Small buyers, and enterprises on open models | Many quotes for the cost of one request. Shorter commitments. An exit from a deal that no longer fits. Compute Exchange’s former chief executive described a startup that paid 30 to 40 percent over the going rate through a supplier tied to its own adviser. That premium is a search cost. | Little. They hold no pricing power today. |
| Owners of private and sovereign clusters | A way to sell the surplus that When the Building Outlives the Chip said they would have and could not place. These owners are not sellers by trade and have no desk to do it. | Some control. Listing a block means describing it to a public standard and letting it be tested. |
| Small and mid-size operators | Demand without a sales force, and fuller chips. If the bottom of the loan ladder reprices, cheaper money as well. | The margin that came from buyers not knowing the price. |
| Large neoclouds | A deeper market for capacity that comes off contract, and a second use for older chips. | Their model is the long contract with a strong customer, financed against that customer’s name. Shorter terms and rights to reassign cut against how their loans are sized. They have the least to gain soon and the most to renegotiate later. |
| Hyperscalers | A cleaner market in which to rent other operators’ fleets, which the largest already do. | They sell compute inside a bundle with a platform and discounts for committed spending. Routers already list their gateways as hosts beside independents, priced side by side. For open models that turns the gateway into one quote among many. Closed models and enterprise credit stay with them. |
| Frontier labs | As buyers they are the largest term customers, and a market with exits lets them resell what they over-commit. As they come to own more of their compute, they become owners with surplus too. | Their models cannot be self-hosted, so the choice between tokens and chips never reaches them. Model routing does. Routers that pick the model prompt by prompt move easy work to cheaper models, which presses on the middle of a price list more than the top. |
| Trading firms and dealers | The spread for turning long contracts into short ones, and a dispatcher as distribution. Itô into Jump is the template. | The spread narrows as the market fills. That is the trade in every market they have made. |
| Lenders and insurers | Data. A record of deliveries tested and blocks re-let is what an underwriter lacks today. | The comfort of lending on a name alone. |
| Brokers | Volume, for the ones who adopt shared fields early and become the dispatcher’s counterparties. | The information advantage. A broker whose value is knowing who has chips is selling a search cost. |
One pattern runs through the table. The dispatcher moves value from firms that profit from the market being hard to read to firms that suffer from it. That is the same shift that price reporting brought to oil and that open access brought to power, and in both cases the incumbents with the best assets did well afterwards anyway. Opacity was never what made their assets good.
Six signals, in the order they would arrive.
Pages opened on November 10, 2026, unless a date is given.
Brokers, desks and marketplaces
- Managed GPU Deployments and American Compute, program pages; Compute Offtake Agreements, American Compute, April 8, 2026 (the review of private contracts). American Compute’s insurance programme is placed by an affiliated surplus-lines broker, so the publisher has a position in the market it describes.
- Epilogue, as indexed by search. The address redirected to a different site when fetched on November 10, so the description rests on the indexed copy.
- Jump Trading Hires Startup Founders to Lead Compute Desk, Bloomberg, October 13, 2026; We’ve joined Jump Trading, Affaan Mustafa, October 12, 2026 (contract terms and deposits); Jump Trading on What It Takes to Trade Compute, AI Street, November 10, 2026 (free section); Itô launch page and Itô Markets on GitHub (30 to 365 days, first contracts)
- Jump and DRW Gear Up to Trade AI Compute Like a Commodity, Briefs (DRW, FalconX); Jump Trading buys Ito Markets to trade compute capacity, The Primary, October 23, 2026 (the clearing remark)
- Liquid Compute Launches with $15M, Business Wire via Yahoo Finance, September 15, 2026 (Liquid Compute, BGC)
- Auctionomics and OneChronos Partner on First Tradable Financial Market for GPU Compute, July 29, 2025; OneChronos Wants to Build a Financial Market For Frontier Compute, Dave Friedman, May 5, 2026
- National Compute and the Grid, Runtime Wire, secondary coverage of the October 2026 announcement
- The Programmable Data Center, Latent Space, October 20, 2026, and a summary of the interview (the 40% example); a competitor’s comparison of Compute Exchange’s quoting, treated as secondary
- SF Compute, Computable, Hyperbolic, Compute Exchange and Andromeda terms are carried from The Venues and the Tranches, opened November 4, 2026. Company pages for Shadeform, DGX Cloud Lepton, Prime Intellect, Vast.ai and Exascale; ClusterMAX 3.0, SemiAnalysis, September 23, 2026 (Vessl); Andromeda terms of service (it aggregates and manages capacity from third-party providers under its own terms)
Routers and token venues
- Architect Financial Technologies Launches Liquid Inference, October 7, 2026; Architect’s comment to the Commission (the index licence)
- Stripe agrees to acquire OpenRouter, August 19, 2026. Neither company disclosed a price.
- OpenRouter challenger closes seed round, Semafor, October 16, 2026 (TrustedRouter); Antseed; You.com Inference Router, MarkTechPost, October 22, 2026; DigitalOcean vs OpenRouter
- Best LLM routing platforms compared, Requesty (a vendor comparing itself with rivals); gateway fee table; Martian alternatives
Prices, contracts and plumbing
- NYMEX letter extending the review of Submission 26-370, November 4, 2026; staff letter staying the Kalshi amendment, August 21, 2026; ICE product page for the Ornn B200 future, whose October 5 date one trade report describes as a release of reference data, not a listing
- Silicon Data, Compute Desk, General Index, NativX and Inferra are carried from Five Indices, One Price and the Inferra note. Backing Ornn, Galaxy; Architect and Compute Desk on ComputeConnect, July 8, 2026; Setara Financial; 252 Capital
- SkyPilot nabs $20M, SiliconANGLE, July 21, 2026; Agentic payment protocols compared, Fireblocks, and a count of x402 volume (the 20-cent average)
- ForwardCompute; Princeps, Y Combinator profile. USD.AI and the NVIDIA rent-back are carried from The Customer Is the Collateral and The Venues and the Tranches.
Assumptions and limits
- The six seats, the map in Figure 1 and the placement of each firm are Kinetic Alpha’s reading of published terms.
- The count of 64 names each firm or standard in Table 1 once, however many seats it holds. It is an inventory, not a census: providers that only sell their own capacity, such as the neoclouds and the inference hosts, are left out because they are the supply, not the seats between.
- The dispatcher, its four steps, the three reasons nobody chooses between tokens and chips, the ranking of frictions and the five conditions in section 06 are a framework, not a description of any product announced.
- The effect on chip-loan tranches in section 06 is an argument from the loan ladder in The Venues and the Tranches. No lender has said it would price a re-lettable block this way.
- Table 5 is judgement. The firms named in it have not been asked whether they agree.