The competition.
Every player in the GPU-rental picture and where we sit among them, written plainly: the marketplaces we read, the brokers we compete with, the comparison sheets we out-live, and the giants that set the ceiling. Strengths and weaknesses ours included — a map drawn honestly is the only kind worth navigating by.
compiled 2026-08-24 from public sources · …
The matrix.
Capability by player class — the landscape below names every player. An honest tick for a rival is worth more than a flattering blank.
| Capability | us | Shadeform | Prime Intellect | SkyPilot | GPUsMarket | Sheets | Source markets | Hyperscalers |
|---|---|---|---|---|---|---|---|---|
| Live cross-provider book, normalized $/GPU/hr | ✓ | ✓ | ✓ | ✓ | — | ~ | — | — |
| Price history you can query | ✓ | — | — | ✓ | — | — | — | ~ |
| Price alerts, no account | ✓ | — | — | — | — | — | — | — |
| Standing limit orders (trigger → signed fill ticket → execution) | ✓ | — | — | — | — | — | — | — |
| Standing orders survive preemption (auto-refill from the book) | ✓ | — | — | ✓ | — | — | — | ~ |
| Renter protection: spend watchdog + kill-cap | ✓ | — | — | ✓ | — | — | — | ~ |
| Will-it-fit: model → VRAM math → cheapest holding machine | ✓ | — | — | — | — | ~ | — | — |
| As-delivered trust tape (advertised vs measured, public) | ✓ | — | — | — | — | — | ~ | — |
| Workload-aware pricing (training / inference lenses) | ✓ | — | — | — | — | ~ | — | — |
| MCP / agent-native | ✓ | — | — | ~ | ✓ | — | — | — |
| Rents machines through it | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ |
| Provider price published unmarked-up | ✓ | — | — | ✓ | — | — | — | — |
| Keys/custody never leave you | ✓ | — | — | ✓ | — | — | — | — |
| One invoice across providers | — | ✓ | ✓ | — | — | — | — | — |
| Hyperscaler prices in the book | — | ~ | — | ✓ | — | ✓ | — | n/a |
| Owns hardware / gives SLAs | — | — | — | — | — | — | ~ | ✓ |
✓ has it · ~ partly · — doesn't · hover a cell for the caveat
The throne table.
Same rows as the matrix, grouped by who the buyer is, with two more columns: who is strongest at each thing today, on public evidence, and what we do to take that seat — or, where the seat is ours, to make it harder to take. Each section says where we stand, so the table can drive the choice between pressing an advantage and closing a gap. Only the rows we actually compete in; owning hardware and selling SLAs is not our game and is not listed. Sources: our research file of 2026-08-28 (sourced, dated); "us" claims are the live product.
| Capability | Strongest today | To take the throne — or keep it |
|---|---|---|
| Agentic — the machine as the buyer — where we stand: ahead. Execution, orders, custody and the MCP surface are ours; the rail that lets an agent pay has moved real money once, house-to-house (0.41 USDC on Base, reconciled on-chain by an outside reviewer); its first outside payer is still open. | ||
| MCP / agent-native | RunPod — official MCP server (npm) over an OpenAPI-first v2 API; GPUsMarket in the registry | stay ahead: keep the widest tool surface (24, execution included), serve x402 v1 and v2, and keep the 402 challenge disclosing its fee in-band (live since 28 Aug 2026) |
| Price alerts, no account | us — we have not found another accountless alert (checked 28 Aug 2026) | stay ahead: alerts that act — a tripwire is already a standing order; next, alerts an agent can arm from the MCP server with a spend cap attached |
| Standing limit orders (trigger → signed fill ticket → execution) | us — no competitor has order books on rentals | stay ahead: keyless fills over x402 so an order can settle without a provider key at all; publish fill receipts on the tape from fill one |
| Standing orders survive preemption (auto-refill from the book) | SkyPilot — cross-cloud auto-failover for whole workloads | take: match its failover semantics at the order level (re-place on the next-cheapest holding machine, state carried), and print the refill on the receipt |
| Renter protection: spend watchdog + kill-cap | SkyPilot autostop; Vast.ai's agent guide ships instance control with no spend cap at all | stay ahead: move the kill-cap into the keyholder so it survives a station restart (the external reviewer's open finding), and disclose its own failure mode in every receipt |
| Keys/custody never leave you | us (keyholder sidecar); SkyPilot keeps credentials local; Shadeform requires your keys on disk | stay ahead: remove the key from the equation — x402 keyless rental is the end state; the keyholder is the bridge |
| Rents machines through it, keylessly | Shadeform / Prime Intellect — one API across 20–50+ providers with real settlement volume (with your keys) | take: the keyless rail. Pay with a 402, no key, prepaid blocks, fee disclosed, receipt on a public tape — Akash first, others as written consent allows |
| Renters and developers — the pains people actually name — where we stand: competitive, with one real gap. The book and the fit tools hold up; host quality — the loudest complaint on the record — is the seat we do not yet contest. | ||
| Live cross-provider book, normalized $/GPU/hr | GetDeploying — 105 providers, 4,357+ live prices, daily; Shadeform for a rentable book across 20+ clouds | take: backfill supply (a dozen feeds → the long tail, hyperscaler spot as labelled reference rows) and keep printing which feeds refresh in 5 min and which in 30 — breadth they have, freshness-with-receipts they do not |
| Provider price published unmarked-up | Vast.ai by policy (zero markup, hosting fee 0%); us | stay ahead: keep the display lever at 0 forever, publish both fee levers, and let the sealed tape prove the price was never touched |
| Will-it-fit: model → VRAM math → cheapest holding machine | Hugging Face's model-memory estimators for the arithmetic; nobody joins it to a live book | stay ahead: tie fit to cost — tokens per dollar for the workload on the machine it fits, with the assumptions listed, not hidden |
| Workload-aware pricing (training / inference lenses) | us — spreadsheets approximate it by hand | stay ahead: benchmark-backed lenses (SemiAnalysis measured 10–20% MFU variance on identical hardware across clouds) so the price column reflects delivered throughput, not spec sheets |
| As-delivered trust tape (advertised vs measured, public) | ClusterMAX (SemiAnalysis) — 209 providers tracked, 84 evaluated, medallions; the renter's loudest pain is silent host downclocks and dead pods | take: a per-host quality signal from our own probes and renter reports, published beside the price — the pain renters actually name, and the one gap in this book |
| Enterprise — teams and procurement — where we stand: behind, by choice so far. No invoice, no hyperscaler quotes, no SLA. The rail changes the first; reference rows change the second; the third is not our game. | ||
| One invoice across providers | Shadeform / Prime Intellect | take: prepaid x402 blocks are one invoice by construction — every fill lands on one receipts tape under one account, without us holding a balance |
| Hyperscaler prices in the book | GetDeploying; Vantage's instances.vantage.sh | take: labelled reference rows for AWS/GCP/Azure/OCI spot so the enterprise shopper stops bouncing — reference, never blended into the index |
| Institutional and academic — desks, exchanges, researchers — where we stand: ahead on one property — as of 28 Aug 2026, across every compute price index and tracker we have checked, we found none that seals, hashes and third-party-timestamps each published day so a past number cannot be silently changed. Reproducibility alone is shared (ccir.io, gpumarkets.dev, SGPI); the un-restatability is ours. The demand here is institutional, not grassroots; the 5 October listing is the first dated audience. | ||
| Price history you can query | SemiAnalysis — hourly since 2023, on Bloomberg as SAH100SC; GetDeploying for per-model charts | take: time, and proof. Every day sealed, Merkle-anchored and reproducible from public files; a dated permalink per chip-day; from 6 Oct, our number beside the settled index (the basis). Theirs is longer; ours is the only one a stranger can recompute |
| Sealed, reproducible price tape + index (not in the matrix above; added 2026-08-28) | nobody — Silicon Data, Ornn, SemiAnalysis and GetDeploying are all vendor-attested; a vendor-attested index can restate its own history and nobody outside it can prove it did not | stay ahead: more days, more feeds, a second independent verifier implementation, and the basis series — the history is the moat; the format is free to copy |
The landscape.
Four kinds of player. A supplier can also be a rival; every feed we aggregate would rather you came to them directly.
| Player | What it is | Relationship |
|---|---|---|
| Feeds we aggregate today — 12 | ||
| Vast.aithe deep book | Peer GPU marketplace; auction-priced, the cheapest interruptible capacity anywhere and the widest consumer-card variety. ~80% of our book — our single largest concentration, and the platform risk we carry. | source |
| RunPod | GPU cloud with community and secure tiers; serverless endpoints. Both our rental rails run here and on Vast. | source |
| Verda (ex-DataCrunch) | EU dedicated GPU cloud (Finland, green energy; rebranded Aug 2026); clean per-model catalog. | source |
| Akash | Decentralized compute marketplace, crypto-settled; we read the network's minimum ask per model. | source |
| Cudo Compute | GPU cloud; closed to new customers since March 2026 — availability now "by request", which caps how useful its quotes are. | source |
| io.net | Decentralized GPU network (Solana); undocumented explorer market feed, read per-product. | source |
| Hyperstack | NexGen Cloud's GPU platform; flavors × pricebook × live stocks behind a free key. Live in the book since 2026-08-24. | source |
| Novita | GPU cloud with a products catalog and per-region offers; keyed. Live in the book since 2026-08-24. | source |
| Prime Intellectalso a rival | Aggregator we also read — its feed carries Lambda's book among others. Full row under direct competitors. | source competes |
| Shadeformalso a rival | Aggregator we also read (19 clouds via its feed). Full row under direct competitors. | source competes |
| Scaleway | European sovereign cloud; keyless public catalog with real per-SKU stock (Paris, Warsaw — H100, L40S, L4, B300). EUR, converted at the ECB daily rate. Live in the book since 2026-08-24. | source |
| OVHcloud | European cloud; keyless public order catalog (A10 → H200, published flavor sheet). EUR at the ECB daily rate; regions not stated per flavor, so ours stay honest-unknown. Live in the book since 2026-08-24. | source |
| Supply we can add, key-gated — 3 | ||
| TensorDock | GPU marketplace; inventory API answers only for funded accounts ($5 minimum). Adapter written, waiting on the deposit. | pending |
| Lambda | AI cloud with on-demand instances and clusters; free API key exposes prices + live regional capacity. | pending |
| Salad | Distributed consumer-GPU network; customer prices need a key — its public feed is supplier earnings, the wrong side of the market to quote. | pending |
| Direct competitors — aggregation & brokerage | ||
| Shadeformthe funded analog | YC-backed GPU marketplace ($20M raised): one API and console over 30+ clouds, unified billing, routes to the cheapest instance. A reseller — you buy through them. Since 2026-08-24 also a feed in our book. | competes source |
| Prime Intellect | "Compute exchange" aggregating ~a dozen clouds, paired with decentralized training; rentals go through their platform. Since 2026-08-24 also a feed in our book. | competes source |
| SkyPilotGPU Compass | Open-source multi-cloud launcher; its GPU Compass dashboard compares 20+ clouds and keeps spot-price history. You bring your own cloud accounts — the closest thing to our BYO-key stance. | competes |
| GPUsMarket | Agent-first GPU rental with a CLI + MCP server — proof we are not the only ones who noticed agents buy compute; sells its own market rather than aggregating. | competes |
| Compute Exchange | Compute auction/brokerage platform positioning around the 2026 capacity crunch. | competes |
| Comparison sheets | ||
| GetDeploying · ComputePrices · cloud-gpus.com · GPU Finder… | Static or periodically-refreshed price tables. Useful bookmarks; no live book, no history you can query, no alerts, no execution. This is the graveyard tier — and what we would be if we stopped at a table. | competes |
| The backdrop | ||
| AWS · GCP · Azure · OCI | The price ceiling and the quota wall. H100s at multiples of marketplace rates; enterprises pay it for the SLA. We don't quote them (yet) — deliberate, our buyer is price-driven. | context |
| CoreWeave · Nebius · Crusoe… | Neoclouds that own the metal. They are who the brokers resell; several are also feeds inside Shadeform and Compass. | context |
Where that leaves us.
Strengths
- Agents are the first customer, not a feature. Native MCP over the same core as the UI, listed in the official registry. Only GPUsMarket shares the thesis, and they sell one market — we aggregate.
- Prices we do not touch. One process on commodity hosting, so the book carries the provider's own number with no markup in it — checkable against their page. The platform fee sits on execution alone, capped per lease per day. A $20M burn rate cannot copy this cost base.
- BYO-key brokering. We place rentals with your provider key and never store it. Every funded competitor wants to be your wallet instead.
- The tape. Price history per model and class recorded every half hour since first boot — a past no late entrant can backfill. Only SkyPilot keeps anything comparable.
- The tripwire → limit orders → orders that survive. Under-the-market alerts with no account — and since 2026-08-24 they execute: standing limit orders via signed fill tickets and the open-source keyholder sidecar, custody-free. Same day, they learned to outlive the machine: a standing order whose rental is preempted refills from the current best offer on the book, cross-provider. Nobody else has the pipe, and only the layer that sees every provider can do the last step.
- Idle capacity as the editorial spine — interruptible-first ranking, the idle index, the discount instrument. The cheap half of the market is our home turf, not an afterthought.
Weaknesses
- A dozen feeds vs Shadeform's thirty. Coverage is our thinnest wall; it grows a signup at a time.
- The tape is young. Days of history where Compass shows quarters. Time fixes this and only time.
- No unified billing. A team that wants one invoice across clouds goes to Shadeform or Prime Intellect today. That is Stage 3, unbuilt.
- No hardware, no SLA, no enterprise story. We are an instrument over other people's machines and say so.
- No brand, no track record. Unnamed product, zero testimonials, a book that has never had a paying renter through it.
- No hyperscaler quotes — deliberate, but it means the enterprise comparison shopper bounces.
Points of difference
- Instrument, not reseller. Everyone funded wants to sit between you and the provider and clip the flow. We read the market, point, and step aside — the renter pays the provider's own price.
- Machine-first distribution. Registry listing, llms.txt, keyless REST — discovered by agents, not ad spend.
- Honesty as interface. Stale flags shown, unrated models named, "not a forecast" printed on the timing instrument, fee stated in-band — and now a trust tape that measures the market's oldest lie, advertised-vs-delivered, from probes renters run themselves. In a market full of "from $0.99*" tables this reads different because it is.
- A product with a world. The pond, the frog, the field-journal voice — nobody mistakes this for another dashboard template, and nobody forgets it either.
- The five-minute money book. The feeds people actually rent from refresh every five minutes; the sleepy catalogs every thirty — and we print which is which. Sheets refresh weekly and hide it.
- Every price carries its own tape. Each row in the market table draws that model's last 24 hours inline, off history we recorded ourselves. A table a competitor can copy in a day; the tape behind it they cannot.
- Accountless to the bone. The book, the API, the MCP endpoint and the price alerts all work with no signup, no email, no cookie wall. Metering is an optional bearer token. Most others gate at the inbox.
- The workload lens. The price column itself answers "for MY work": flip the market table to training and it re-prices every card in FP16 TFLOPS-hours per dollar; flip it to inference and it prices the VRAM that actually gates serving; flip it to bandwidth and it prices the ceiling for memory-bound work. Spec-sheet ceilings, never benchmarks, unrated cards named and sunk — the honesty is what makes the comparison usable, and nobody else re-prices a live book this way. Its sibling /fit answers the question before that one: will the model fit at all, and what does the cheapest machine that holds it cost right now.
- Safety rails built for unattended buyers. Rentals dry-run by default, price caps absolute, an optional budget guard that destroys the rental at the cap — its own failure mode disclosed in the receipt — idempotent fills that cannot double-rent on a retry, and a zombie-spend watchdog that asks "still yours?" when a machine outlives your spend or age line. Exchange-grade paranoia, documented.