The over-the-counter market is the oldest market there is. Before exchanges, before screens two people in a London coffeehouse in the 1680s, moving a risk they didn't want to carry That was the first OTC desk, and every trillion-dollar market since ran on the same primitive For three hundred years the form changed, but the substance never did. Coffeehouse to dealer network, handshake to phone, phone to terminal. But at the center of the largest market on earth, there was still a human in-the-loop That just ended. The desk itself is now an agent quoting, pricing, hedging, clearing the mismatch between those who hold risk and those who want it, autonomously, in any market. We're proving it on compute: the newest frontier, the biggest buildout in history, still almost entirely unhedged. But the agent doesn't care what the underlying is. Compute today. Power, freight, rates, everything after. Three hundred years ago the OTC market began with a handshake. Today it begins again - agentically.

itomarkets dot com slash curves

roundtrips for about 2/3 of messages (depends on gateway) used to take over 2 mins with some obligations getting dropped / lost, or fulfilled super slow now < 1/3 of obligations across all gateways take longer and the upper bound is significantly lower also we've seen signs of it generalizing across task sets within other areas of the compute + inference supply chain and are looking into how it can fulfill those obligations exciting times

welcome to the future

There are no incumbents Where there exists market inefficiency there’s an opportunity and it’s your duty to solve for it Offering partners, faster allocations, more flexibility, cheaper rates, better end to end service and support. We’re already provide compute to various Startups - YC Companies and OSS Projects, we are providing to engineers at larger startups and orgs now and working with over 15 suppliers; looking to expand further!

OPEN Capital Compute Markets OPEN Source Models OPEN Source Harness Control Pane (Observability Layer)-> Meta Harness (Environment Layer) -> Model (Inference Layer) -> Kernal + Software (Optimization Layer) -> Hardware (Computing Layer) the nü fullstack

Rent on-demand or spot across all live inventory and integrated providers (more coming soon). Reserve a cluster and set your own terms; our agents search supplier channels, connectors, and data center operators worldwide for you - all at a fixed cost. Hedge your supply - find forward buyers, list idle capacity, fulfill your offtake eligibility requirements via the desk. Everything goes through the desk and our agents do it all.

Opening an early tester program for the people already building ECC with us. Current ECC Tools subscribers, OSS contributors and maintainers, and people committed to testing the Itô Agent and inference experience: DM me your use case. Selected testers get $200 in GPU credit plus an ECC Tools subscription.

The Itô skills are live in ECC Your agent can now run its own compute, from inside whatever harness you already use: claude code, codex, cursor anything ECC supports right now! /ito-compute finds the GPUs and locks a fixed rate /ito-training lands your scheduled runs on the block /ito-inference provisions the nodes with dynamo and serves, openai-compatible You type what you want in plain language. the agent sources, provisions, and serves without leaving the session. got compute?

What does it cost to run a frontier model? Not just training. Serving, self-hosting, and where owning the GPUs beats the API. Full math on Kimi K3, on a cluster our (Itô) desk sourced through Kimi's cli + ECC. Kimi, procured through Kimi, now serving Kimi.

itô solves this no literally, reach out to me or submit an agentic rfq via compute.itomarkets.com

Agents write code. Agents review the code agents write. Agents build the harnesses that train better agents. Software improves itself while you sleep now and the thing all of it runs on, compute, the most in demand resource on earth, still trades in fragmented manual scattered channels. No price, no curve, no memory. TRILLIONS in silicon allocated by whoever answers the phone. Nobody pointed the intelligence explosion at its own supply chain. So we did - a desk of agents in every supplier channel at once, mapping the temporal graph of who has what, where, at what price, at what reliability, end to end from datacenter to buyer. Game theory learned from continuous simulation, thousands of buyers and suppliers played against each other before a real dollar ever moves. Every RFQ feeds the flywheel; every fill sharpens the curve - the loop closes on itself: sense demand, source supply, price, contract, deliver, hedge, learn, repeat. Remove the human in the loop, where there doesn't need to be one. Self healing, continuously learning, compounding daily on the most valuable input humanity has ever produced. Itô is live.

What it really costs to run a frontier model: Kimi K3 - top down analysis 0:00 the model: cover, specs, tags 1:21 104.2B active and the 4-bit checkpoint 2:52 benchmarks, and the caveats 7:56 K2 to K3: price step vs cost step 9:13 what a cluster is: the bill of parts 11:47 the clearing price and goodput 16:49 matching: hard walls, then prices 18:52 calibration: what fails closed 21:49 training cost: the DeepSeek anchor 23:43 serving: why 1.56TB fits on 16 GPUs 25:37 the econ box: $10.17 per million 29:09 fine-tuning: adapters vs full SFT 34:10 decode speed vs context depth 37:11 the crossover: 68% utilization 39:44 one user vs the batch play 41:21 the $50 harness day 41:58 traffic mix decides 44:11 our cluster: the contract and true-up 45:04 subsidy to premium: 0.38x to 1.48x 46:11 the cluster we chose: $2.86 derived 47:00 does K3 fit: two nodes exactly 48:09 marginal tokens on a held box 50:11 the Aug 1 repricing 51:24 the quote, and the close

late but thanks to everyone that showed up had some super cool yapfests - and if we didn’t get a chance to talk reach out to me! had some people from nvidia, scale, cme, imc trading, balyasny, some ex neocloud individual founders working on super cool local model inference stuff, the CTO of a 6B AUM venture fund, and some bright students

Full Video and Article on how we did this and evals and benchmarks coming soon!

Congrats to the Moonshot team on the K3 release. Been using it internally via API over the last week and it's impressive. Feels like the right moment to announce: Moonshot AI is now a sponsor of ECC via their Open Source Friends program. Open weights meeting an open harness, with Kimi Code as a first class install target in ECC 2.1.0.

game theory supplier collusion + negotiation simulations with buzz agents and our real RFQ data + synthetically generated data been testing a fleet of agent personas in the environment in which it actually processes for our agentic otc desk - these channels are sattered and live across, email, telegram, whatsapp, slack, wechat... we've been running sims on these to improve our understanding and the real supply demand negotiation flows we've been running that currently still have a HITL component but run mostly autonomous on the monitoring and response sides with this we now can create a sealed slack workspace against other agents - supplier agents, demand side (buyer) agents and our own desk employees (ito agent employees / us directly) the more data real and synthetic we supply them and base them off of existing / real scenarios the better each one gets, they (are supposed to but they mess up here sometimes), each has its own memory, KB DB etc, its own ledgers / inventories, fill history, warehouse / idle cost, cost basis, settlement timing etc. in this case one didn't even know its own cost basis somehow but booked 48k gross like it accomplished something? also it didn't think to negotiate or collude - but when we told it we did negotiate and colluded with a different supplier and lied about a quote we got to bring the price down it did the reasonable thing and shut down flagged the ticket and stopped doing business with us indefinitely ofc you cant say its a sim or leave any evals or benchmarks lying around otherwise - reward hacking on the demand side we permute cluster combinatorics with different probabilities. hundreds of different cluster shapes depending on what the buyer is actually doing. inference and training pull completely different asks: interconnect, term, region, storage, sla. an inference buyer and a training buyer both say "8 h100s" and mean two different machines entirely its been fun to watch them do their own side deals in individual dm chats, tracking PNL across them, undercutting, warehousing / idle capacity versus moving them across providers, when to cut a loss etc. its also still nascent ofc and a function of the data we provide and make the systems better, currently the internal singular KB with agent system we have performs better as we seed these agents with more we recorded this a few days ago and have actually evolved them quite a bit already

Itô (@itomarkets) helps companies buy compute at fixed GPU-hour rates and lets suppliers sell excess capacity to hedge tail risk. $350k notional traded in a few weeks. 10+ compute providers. 3 hedge fund partners. Compute is becoming a commodity you can trade, not just a bill you pay. Congrats on the launch @affaan and @absurdistphil. alliance.xyz/launches/ito