INCYT
Observe Live
Collect provider rate cards, marketplaces, published indexes, perpetual futures and prediction markets on a schedule, normalize them to one unit and keep each raw response as evidence.
Compute desk
The record behind Incyt's compute brokerage: every public price for GPU compute, normalized to USD per GPU-hour and traced to its source, with venue discovery and collection operations on the same data.
Desk workflow
Discover Live
For a trade described by side, settlement, execution method and instrument, find the venues that support it, each with the evidence it was classified on.
Balance In design
Net a client's contracted, hedged and open capacity by month and size the trades that close the gap. The exposure model is specified; it is not yet connected to client positions.
Execute Planned
Release requests for quote to eligible counterparties, compare quotes, and place hedges under pre-trade limits with post-trade reconciliation. Follows order eligibility.
Coverage by market type
Collection
Needs attention
GPU classification review
Cheapest published rate per GPU model click a model for its configurations
Every live market
Median USD per GPU-hour, by hour
Observations
Model price each source's median, then the median of those ·
Price spread over time
Venue ladder who is asking what, right now
Aggregator markup the same capacity, priced twice
Term structure Azure, this GPU
Specifications the denominator
Index providers. Click to include or exclude — the figures below follow, and a hollow point on the chart is a day taken from a fallback feed.
Index levels
Where they stand
How closely they track pairwise, over shared days
Feed cross-check does a fallback reproduce the primary?
What each index measures why two of these disagree
User-defined indexes
Indexes the desk defines itself from the prices it collects: you choose which prices go in and how they combine, and the Worker prints one value per UTC day and fills in the last 120 days. They are labelled user-defined wherever they appear, with their rules beside them. A private index is yours alone; sharing one with the team needs every source it names to be licensed for benchmark construction.
Choose an index to view
Build an index
Your token can preview an index but not create one: creating needs the index_author scope, which each teammate's own token carries.
Published indexes as votes:
Projected price
Averages and revenue
Term price anchors
Monthly path
A market is a venue and a product. Pick a product above to narrow these to the venues quoting it — the same GPU on two venues is the comparison the table at the foot of this tab makes.
Where it is trading
Mark against its index
Funding
The same product, priced twice where two venues list it
Collection history
Yes/no contracts on the compute market, each event's buckets in strike order — lowest first, never alphabetical. A price is the market's probability that the bucket pays. Month-end and year-end GPU events settle on Ornn's finalized daily index, shown beside each one, and the bucket today's level falls in is highlighted.
Contract venues
Venue search
Four independent dimensions, not one list of labels. A source can reserve compute and run an order book and settle in a stablecoin, so it carries terms in every dimension that its evidence supports. Terms combine with OR inside a dimension and AND across them. Every tag shown here carries the evidence it was assigned on, its confidence and the date it was last reviewed — open why matched on any result.
high confidence — docs, contract spec, API or observed product medium — company material or consistent secondary observation low — indirect or inferred dashed outline — not an active capability
CRM access needed.
The token on this device can read market data but not client records. Each teammate has their own CRM token, which attributes their notes and changes to them. Sign out (top right) and sign in with yours.
Opportunities
Build
Choose a product.
Readiness by GPU model every required leg, read now
Market gaps required venues nobody can trade today
Internal only · OpenRouter site stats and an unofficial InferenceX pull · not for redistribution
- 1Rent GPUs$ per GPU-hour, from INcyt's live prices
- 2Make tokenstokens per second per GPU, from InferenceX
- 3Sell tokens$ per million tokens, from OpenRouter's sellers
- 4Keep the differenceprofit per day and per month
1
Rent the GPUs
Choose the chip, how many, and which of today's market prices you would pay.
Rent
2
Make tokens
Each GPU runs the model and writes tokens. How many depends on the model, the chip, and how fast each user is served: faster per user means fewer users per GPU.
InferenceX tested very long agent conversations, two to seven times heavier than typical OpenRouter traffic. The default adjusts its measured throughput to OpenRouter's mix; the conservative option uses it as measured. The truth is in between.
3
Sell tokens
Many companies already sell this model on OpenRouter, each at its own price. We assume you win customers starting from the cheapest seller and work upward, taking at most a set share of any one seller's customers. Each bar is one seller: its height is the price, its width the volume you could win from it.
Volume you could win, per seller
What you sell
Your cost per million tokens
Everything your GPUs can make
4
Keep the difference
Rent is fixed; revenue grows with how much you sell. Move along the chart to see profit at any utilization.
Revenue
Rent
Profit
Loss
No more buyers
What this says
Which model should this fleet sell? same GPUs, rent and settings for every model; click a row to switch
| Model | Market size, tokens / day | Sellers | Average price, $ / M tokens | You make / day | You sell / day | Profit / day | Profit / month | Tokens/s per GPU: needed vs available |
|---|
Sources, method and limits
- GPU prices
- Live from
/v1/model-prices: the source-balanced on-demand median and middle half (each source's median, then across sources), and the interruptible median. - Market
- Snapshot 26 Oct 2026 of OpenRouter's site stats: each seller's effective input and output price and its share of the model's tokens, with daily volume averaged over 7 complete UTC days. These endpoints are undocumented and OpenRouter's Terms (section 7) bar scraping the site; the desk chose to use this one snapshot internally. It is not collected on a schedule.
- Price per token
- Each output token arrives with the model's observed input load (about 68 input tokens per output for GLM-5.3, mostly cache hits). A seller's price is its output price plus that input at its input price; most revenue is input.
- Selling
- Your tokens sell from the cheapest seller upward, up to the chosen share of each seller's volume. Anything the market cannot absorb goes unsold.
- Throughput
- SemiAnalysis InferenceX, an unofficial manual pull of 26 Oct 2026 (no data licence; permission not yet requested). Per model and chip, the measured configurations form a frontier of tokens/s per GPU against p90 speed per user; the tab reads it at the chosen speed, linearly between points, never beyond the fastest. Agentic traces, except gpt-oss-120b (fixed 8k in / 1k out); GLM-5.3 uses GLM-5.2's results; GLM-5.3 Flash is not benchmarked. Tuned benchmarks: treat as a ceiling.
- Excluded
- Power and facility costs beyond rent, staff, model rev share, router fees, and demand outside OpenRouter.
| Model | Publisher | Weights | Sellers | Reference $/M | Best | Routed | 7-day change |
|---|
Every offer, cheapest first
| Seller | Venue | Offering | Configuration | $/M | Counts? |
|---|
Depth: taking now against posting for later
No venue publishes how many tokens a seller will fill, so the axis is seller rank, not volume. Bars are each seller's on-demand price, cheapest first; the line is the average price of splitting an order across the cheapest N. Dots are the same seller's delayed-fill price (batch, or spot that may not serve at all), and the discount is against that seller's own on-demand price, never the router's floor. Fireworks states batch at 50% off serverless without a per-model list, so it is noted here rather than drawn.
| Seller | Kind | Offering | $/M | Same seller on demand | Discount |
|---|
Venue agreement: a seller's own card against what a router reports for it
Reference price, last 30 days
Seller view
One seller's prices across every model it sells, each against that model's reference price. The summary is the median ratio over models another seller also prices, so one outlier does not make a seller look cheap or dear.
| Model | Leg | Price $/M | Via | Reference | Rank | vs reference |
|---|
Prediction markets on token prices
Open Polymarket contracts on output token prices (USD per million), each beside today's posted price for the family it asks about, as OpenRouter lists the family's newest model. Probability is the contract's displayed price. A strike today's price already meets is marked; whether the contract resolves on it depends on its own rules.
Coverage
A venue is current when its last good run is within twice its cadence; a stale venue's prices are shown but do not vote. Labels not yet approved wait in the review queue.
GPU token cost calculator
What it costs to make a million tokens on a rented GPU, against what the market pays for them. Ported from the GPU_Token_Throughput_Model sheet with its formulas unchanged. GPU prices are INcyt's live on-demand medians where we have them, the sheet's otherwise, and every cell can be edited. Throughput is the sheet's illustrative pure-decode estimate, not a measurement.
GPU prices and pure-decode throughput (tokens/s per GPU; blank = cannot serve)
Cost per million tokens
Against the market: the cheapest GPU's cost beside each model's reference price
| Model | Cheapest GPU | Cost in | Market in | Cost out | Market out | Margin on output |
|---|
Stress scenario
What to monitor the levels at which each limit trips
Balance of risk
Capital stack
Exposure by book
Stressed loss by source
Break points: loss against rental decline
Borrower coverage against rental decline average GPU loan at the horizon
Break points GPU-backed loans and the whole book, by rental decline
Trade margin break points index gap within one liquidation window
Loss waterfall base case and the chosen stress
Assumptions changed values are bold, and are kept in this browser only
Smart Order market discovery
Find the markets that could support a transaction, before worrying about whether any of them can fill it. This answers does this source generally support this kind of trade — not can it fill 500 GPUs right now, which needs inventory, credit and price and belongs to a later phase. Every stage stays on screen: change an earlier choice and everything below it recalculates.
Side
Are you seeking to acquire or dispose of the underlying product or exposure?
Settlement
What ultimately changes hands?
Execution method
How do you want to transact?
Instrument
What is being bought or sold? Selecting several matches any of them.
Scope
Markets found
Posted purchase prices for datacenter GPUs, per GPU, next to the hourly rent a published index says the same GPU earns. A baseboard is the GPUs alone; a server adds CPUs, memory and chassis; a DGX is NVIDIA's turnkey system. Payback is the per-GPU price over the index level, in months of rent at the utilisation above. Implied $/GPU-hr is the purchase price spread over the GPU's billable hours (life × 8,760 × utilisation), and its percentage is that cost as a share of the index rent: 20% means the hardware costs a fifth of what the hour rents for. Almost every posted price is a list price on an item that is out of stock or made to order, not a sale.
Price by configuration
Customer order
Order calculator a seller's posted price, a quantity, and Incyt's margin on each line
Pick a model and a seller to fill in its posted price; where a seller lists more than one unit for that model, pick the unit. Margin is gross margin per line: the customer pays cost ÷ (1 − margin), and Incyt keeps the margin's share of that price. Posted list prices: no volume discount, freight, tax, networking or support. The order is kept in this browser.
Configuration over time
Lowest posted price per GPU, by venue
Residual value
Record a quote
For a price you were given by a venue this service does not read: an RFQ answer, a broker's sheet, a call. It is stored as a quote, attributed to your token, and shown as its own venue column.