Tracker

Decentralized compute, tracked and sourced.

Twelve decentralized GPU networks in one table — price, verification, uptime, payment, and the watch-outs most comparison pages leave out. Every figure carries a link and the date we checked it.

DCF Editorial Last checked Oct 4, 2026 Refreshed monthly

How to read this table

Basis: one NVIDIA H100 80GB-class GPU-hour, on-demand, US or EU region, taxes excluded. Where a network only sells consumer cards, we say so and quote that tier instead — we never mix an H100 with an RTX 4090 and call the result a price.
Every row links to the page we read and shows when we read it. Marketplace prices move continuously, so a number here is a dated snapshot, not a quote. Confirm it at the source before you commit.
Uptime figures are self-reported network averages, not audited SLAs. No decentralized network below publishes a contractual uptime guarantee, so treat every figure as a claim rather than a promise.
Watch-outs are the part other comparison sites skip. We only list ones we can document. If we can't source a claim, it doesn't go in the table — including claims that would flatter these networks.

Network	Best for	H100-class rate	Verification	Reported uptime	Payment	Watch-outs
Akash NetworkAKT provider checked 09-12	InferenceFine-tuningKubernetes	$2.04–$3.16live weighted average	On-chain attestation; audited provider tier	99.4%	AKT · USDC	Popular regions sell outQuality varies by provider tier
io.netIO provider checked 10-04	InferenceFine-tuningMulti-node	$1.49–$2.20H100 80GB, mixed modes	Proof-of-time-difference; hardware scans	98.6%	IO · USDC · card	No contractual SLAFat tail on residential nodes
Render NetworkRENDER tracker checked 09	RenderingInference	$1.85network figure	Proof-of-Render; node tiers	99.8%	RENDER	Rendering-first job flow, not a general shellEnterprise AI support is newer
AethirATH tracker checked 09-12	Cloud gamingInference	no public buyer ratehost estimate $1.45–$3.50, paid in ATH	Enterprise agreements; network-level checks	not published	ATH	Host payouts denominated in tokenDisclosed host economics assume 95%+ utilization
NosanaNOS tracker checked 10-04	InferenceConsumer GPU	from $0.048no H100 — consumer cards only	Bonded reputation; not published in detail	not published	NOS	No datacenter-grade H100Earnings and uptime undisclosed
Hyperbolic— tracker checked 10-04	Inference APIFine-tuning	$3.19H100 80GB	Not published	not published	crypto · fiat	H100 priced above most decentralized peersVerification undocumented
Spheron NetworkSPHN tracker checked 10-04	Multi-nodeSpot GPU	$2.91H100 spot · H200 $5.92	Partner facilities: ISO 27001 / SOC 2 / HIPAA	not published	crypto · fiat	Supply is aggregated from a fixed partner set, not permissionlessNewest cards are spot-only
Theta EdgeCloudTHETA tracker checked 09-17	Edge inferenceRendering	$3.69H200 · RTX 4090 $0.38	Not published	not published	card · crypto	Beta-grade tooling — no K8s, no persistent storage in some tiersNo SOC 2, no reserved instances, no SLA
Clore.aiCLORE tracker checked 10-04	Consumer GPURental market	from $0.027RTX 5090 · H100 available	Not published	not published	CLORE	Consumer-grade supply dominatesLowest independent reliability score in this set
FluenceFLT tracker checked 10-04	General computeInference	from $2.80H100 80GB	Verification-first design; GPU support younger	not published	FLT	Compute lineup is CPU-centric; GPUs are newerGPU pricing above peers
Prime Intellectno token tracker checked 09-24	Multi-nodeHosted training	$2.43H100 · spot $0.94 · H200 from $0.47	None published — no hardware attestation	not published	USD (fiat)	No formal SLANot crypto-native — no token, fiat rails, KYC not documented
Gensyn— tracker checked 09	Multi-nodeVerifiable training	~$0.40 claimtraining; not a posted H100 rate	Proof-of-compute; optimistic verification + arbitration	not published	token (early)	Still early — small market, nascent toolingUploading your own model/data is a real exposure

Legend: "provider" = the network's own pricing page · "tracker" = an independent third-party index we read and cross-checked where possible. "not published" means the network does not disclose it, not that the value is zero.

Reading this table honestly

Cheap is not the same as usable. The $0.03/hr consumer card and the $2.04 H100-hour in the same table are not competing products. One fits a 13B model on a good day; the other trains. Comparing the two columns is how a comparison site turns a real 40% saving into a fake 98% one — that is the error the "save up to 90% vs AWS" headline is built on.

Verification is the only moat that survives. Price converges fast — every network reprices toward the cheapest credible supply within weeks. What does not converge is proof that the GPU you paid for ran the job you asked for. Networks with real verification (on-chain attestation, proof-of-render, proof-of-compute) can eventually charge a premium for it; the ones that run on "trust us" cannot, and their discount is permanent because it is paying for the risk you carry.

Token payment is a cost paid twice. Paying in a network token is not the same as paying in dollars. Between funding a wallet and a job settling, the token can move — and the accounting, invoicing and tax trail a normal purchase produces has to be rebuilt by hand. Networks that added USDC and card rails (io.net, Akash) are pricing that pain correctly; networks that did not are passing it to the customer.

The gap that is not closing is interconnect. Single-GPU and embarrassingly-parallel jobs migrate to decentralized supply easily. Synchronized multi-node training does not, because distributed nodes have no NVLink and no InfiniBand — the cluster is only as fast as its slowest residential uplink. That is why the decentralized tier is winning inference and losing pre-training, and why the "multi-node" tag above means "possible", not "comparable to a hyperscaler".

Excluded on purpose

"Decentralized" is doing a lot of work in this sector. These are widely compared alongside the networks above, and we left them out — with reasons:

Vast.ai — the original "Airbnb for GPUs", but a company-run marketplace: no token, no permissionless supply. Useful, just not decentralized.
RunPod, Lambda, Together, CoreWeave — centralized neo-clouds. They are the benchmark, not the category.
Salad — consumer supply, centralized control, no token. Notable for something rare: it publishes measured per-GPU earnings.
Bittensor (TAO) — an incentive network for machine learning, not a place where you rent a GPU-hour. Different product, different buyer.
Grass — bandwidth, not compute. Frequently listed next to GPU networks; it does not sell a GPU-hour.

What we don't know

A tracker is only as honest as the gaps it admits. For every network above, these are still open:

Paid utilization. Networks publish registered GPUs. Almost none publish paid hours. A network with 30,000 nodes and 4% utilization is a smaller business than its supply headline suggests.
Dropout during long runs. Nobody publishes node-failure rates mid-training — the number that decides whether cheap hours are actually cheap.
Host earnings net of power. Reward estimates are quoted gross. Only Salad publishes measured earnings; nobody publishes them net of electricity.
Whether "verified" tiers are audited. "Verified provider" is often a self-declared label. We cannot confirm third-party audits for most of the rows above.
Enterprise reality. Invoicing, tax residency and liability are undocumented for most of these networks, which is why enterprise procurement stalls on them.

Methodology and corrections

Figures were collected from each network's own pricing page or marketplace order book where one exists, and from independent third-party indexes where it does not; the source cell in each row says which. Ranges reflect that marketplace pricing is continuous — where a network publishes a single live rate, we quote it as a point and note that it moves.

Rows are normalized to a single H100 80GB-class GPU-hour for comparison, except where a network does not sell one; those rows are labelled with the tier they actually sell. We do not adjust for performance, reliability or egress, because a table that quietly folds those into one number is the thing this page exists to replace.

Corrections are welcome and get published: if a figure here is wrong or out of date, mail editor@omc.center with a link to the current source. Disclosed conflict: DCF is operated by the Omniverse Compute (OMC) team, which is building a competing network. We include no OMC row and quote no OMC price here, precisely so this comparison cannot be read as self-promotion.

Update log

2026-10-04 — first edition: 12 networks, 5 workload filters, watch-outs and open questions added.

About this tracker

DCF is written and operated by the Omniverse Compute (OMC) team — a decentralized GPU network project on BNB Chain, currently in public testnet. That is a conflict of interest for a page like this one, so we handle it the blunt way: no OMC row, no OMC price, and every competitor figure sourced to its own page. Judge the table on whether it is checkable, not on who runs the site.

About OMC →

Decentralized compute, tracked and sourced.

Reading this table honestly

Methodology and corrections

Update log

Comments