Decentralized compute networks explained: GPU marketplaces for AI, from Akash to Aethir

Who this is for: Founders and ML engineers deciding where to rent GPUs, investors assessing DePIN compute tokens, and analysts separating real usage from marketing; after reading you will be able to price a workload across a DePIN, a neocloud and a hyperscaler and judge whether a network’s revenue is real.

Every AI product runs on rented silicon, and for most of 2023 and 2024 that silicon was scarce. Nvidia’s H100 sold for $25,000 to $40,000 a card, and cloud rates for a single H100 reached $7 to $10 an hour at the peak, according to Aethir’s September 2026 review of the rental market. Scarcity created an opening for a crypto idea that had been waiting for a use case: pool the GPUs sitting idle in data centres, render farms, mining sheds and gaming PCs, coordinate them with a blockchain, and sell the capacity to anyone with a card or a wallet.

Three years on, H100 rates have fallen 75 to 80 percent from their 2023 peak and the question has changed. It is no longer “can you find a GPU?” but “can a permissionless marketplace beat CoreWeave, Lambda and the hyperscalers on price, reliability and trust for the workloads that matter?” Some networks now publish revenue; others publish token prices. This guide separates the two, from how the networks work and what they cost to verification, token models, the bitcoin miner pivot and the state of decentralized training. For the wider category, see our guide to what DePIN is.

Decentralized compute by the numbers

What a decentralized compute network is

A decentralized compute network, or GPU DePIN, is a two-sided marketplace with a blockchain in the middle. On one side are providers: data centre operators with spare racks, render studios, former crypto miners and individuals with gaming cards. On the other are tenants: AI startups renting GPUs for inference, researchers fine-tuning open models, studios rendering frames, and increasingly AI agents buying compute programmatically.

The blockchain does three jobs: it records who offered what capacity at what price, it escrows payment so a provider is paid per unit of work delivered, and it issues a token that can reward supply before demand exists, the bootstrapping trick every DePIN relies on. The compute itself happens off-chain, inside containers the network’s software schedules and monitors.

In practice a provider installs node software, passes hardware checks and advertises capacity, often posting a stake that can be slashed; a tenant describes a workload and accepts a posted price or runs a reverse auction; the protocol escrows payment in stablecoins, fiat or the native token and releases it as the lease runs; and token mechanics convert that spend into a burn, a fee to stakers or emissions. Networks differ mainly in how permissionless they are. Akash runs an open auction where any provider can bid. Aethir and io.net lean on enterprise data centre partners vetted by the company. Render accepts work only from curated creative and AI applications. Bittensor turns the whole thing into a competition in which validators score miners and the chain pays the winners.

How the main networks work

The projects below are the ones a buyer or investor is most likely to meet. They are not interchangeable; several barely compete with each other.

Akash Network (AKT): the open marketplace

Akash is the oldest of the group, live since 2020 on its own Cosmos chain. Tenants post a deployment, providers bid, and the lowest acceptable bid wins. Its Q1 2026 report said all-time compute spend crossed $5 million; Q3 2025 lease income was about $851,700 per Messari data cited by BlockEden in March 2026. Small numbers next to Aethir, but the most transparent in the sector because every lease settles on-chain. Akash added AkashML, a managed inference service launched November 2025 that served about 1.7 billion tokens a day through OpenRouter by Q1 2026, a HomeNode beta for RTX 4090 and 5090 owners, and confidential compute in July 2026. Founder Greg Osuri set out the original vision in a 2021 Crypto Coin Show interview.

Render Network (RENDER): rendering first, AI second

Render began as OTOY’s distributed rendering service for 3D artists and moved its token to Solana in November 2023. AI inference and robotics work made up an estimated 35 to 40 percent of volume in 2025, per BlockEden, and monthly burn rose from 20,452 RENDER in January 2025 to 120,928 in September 2025. A frame is a self-contained job that can run on one card anywhere and be verified by looking at it, the purest fit for distributed hardware.

io.net (IO): aggregation at scale

io.net, on Solana since 2024, aggregates GPUs from independent data centres, miners and other DePINs and clusters them for customers who want tens or hundreds of cards in one job. Its site listed more than 30,000 available GPUs across 130 or more countries in October 2026; BlockEden put its annualised on-chain revenue at about $20 million in October 2025 with 56 enterprise clients. A 2024 episode in which fake GPU metadata inflated its supply count is a reminder that “available GPUs” is not the same as rented GPUs.

Aethir (ATH): the enterprise GPU cloud

Aethir is the revenue leader. The Block reported $36 million of first-year revenue in July 2024; for 2025 the company reported $127.8 million, with quarterly revenue rising from $28.5 million in Q1 to $39.9 million in Q3, per figures compiled by BlockEden in March 2026. In September 2026 Aethir said it operates more than 430,000 GPU containers across 94 countries. It is a GPU-as-a-service business with a token layer: enterprise data centre partners supply capacity, Checker nodes verify containers are live, and AI and gaming customers pay largely in fiat. In September 2026 its affiliate Axe Compute, in which the Aethir Foundation is the largest shareholder, took full ownership of a 2,304-GPU Nvidia B300 cluster in Georgia on a take-or-pay contract expected to generate $364.6 million through 2031.

Nosana (NOS) and Hyperbolic: the long tail and verified inference

Nosana runs single-GPU inference jobs on Solana: in October 2026 it listed an RTX 4090 at $0.40 an hour, an A100 80GB at $0.90 and an H100 at $1.50 across 269 GPU hosts, a small-team marketplace rather than a place to rent a cluster. Hyperbolic sells raw GPUs and a serverless inference API; in October 2026 it listed H100 SXM and H200 rentals at $2.40 an hour, repriced weekly, and Llama 3.1 8B inference at $0.10 per million tokens. Its Proof of Sampling scheme, developed with researchers at UC Berkeley and Columbia, spot-checks a random fraction of outputs.

Gensyn, Prime Intellect and Pluralis: the training protocols

These are research labs making training work across untrusted machines rather than rental marketplaces. Gensyn released a Reproducible Execution Environment in March 2026 and shipped open-1b, which it calls the first model with an auditable training record, in September 2026. Prime Intellect built the INTELLECT series; Pluralis works on protocol models in which no participant ever holds the full weights. Their results are in the training section below.

Bittensor (TAO) subnets: compute as a competition

Bittensor is a network of subnets, each producing a commodity such as inference, storage, prediction or raw compute. Miners produce the output, validators score it, and TAO emissions flow to both by score. Several subnets sell GPU time or inference, but the chain holds no leases or escrow in the Akash sense: the buyer pays the subnet’s app and the token rewards the best-scoring miners, so subnet revenue and TAO emissions are only loosely linked.

0G Labs (0G): an AI operating system with a compute layer

0G describes itself as a decentralized AI operating system with chain, storage and compute. Its compute network offers two inference modes: TeeTLS, where a broker inside a trusted execution environment attests the routing path, and TeeML, where the model itself runs inside a TEE on Intel TDX with TEE-enabled GPUs, so prompts arrive encrypted and responses are signed in the enclave. GLM-5.3 moved to TeeML on September 3, 2026 at $1.40 per million input tokens and $4.40 per million output. On September 28, 2026 0G launched Infinite AI: stake 0G, mint a derivative called iAI and receive 1.271 compute credits a day per iAI, turning staking into a standing right to inference. 0G has also published DiLoCoX, a method it says trained a 107 billion parameter model across decentralized nodes.

What these networks are good for, and what they are not

The physics of AI workloads decide where decentralized compute can win. Training a large model means thousands of GPUs exchanging gradients many times a second, so the connections matter as much as the chips. Inference is one GPU or a small group answering a request; rendering is one GPU drawing a frame.

Pluralis quantified the gap in January 2026: consumer links at 80 Mbps are 100 to 1,000 times slower than the 100 Gb/s InfiniBand and roughly 1 TB/s NVLink inside a data centre. DiLoCo, DisTrO and subspace compression cut the data exchanged by orders of magnitude but still trail dense data centre training on wall-clock time. The practical conclusion: rent decentralized GPUs for serving and tuning, rent a neocloud cluster for serious training.

Verification: how a network knows the work was done

The hardest problem in decentralized compute is not finding GPUs but proving that a stranger’s GPU did what it was paid for. A provider could run a smaller model than promised, return cached answers or fake hardware specs. Networks use five approaches, often combined.

1. Hardware attestation. Node software reads device identifiers and benchmarks, and a staked bond is slashed if claims prove false. This is the baseline on Akash, io.net and Aethir, and what io.net tightened after its 2024 spoofing episode.
2. Liveness checking. Aethir’s Checker nodes continuously probe provider containers and pay rewards only for verified uptime.
3. Sampling. Hyperbolic’s Proof of Sampling recomputes a random fraction of inference calls on a second node, so a cheating provider risks detection on every request.
4. Trusted execution environments. 0G’s TeeML runs the model inside an Intel TDX enclave with TEE-enabled GPUs and publishes attestations; Akash’s July 2026 confidential compute does the same for general workloads. TEEs protect the tenant’s data as well as proving the model ran, at the cost of trusting the chip vendor.
5. Reproducible execution. Gensyn’s REE (March 2026) aims to make training steps deterministic so a verifier can replay a slice of the work; its open-1b (September 2026) ships with an auditable training record. The only approach aimed at training, and the least mature.

Pricing: DePINs vs AWS vs CoreWeave, with real numbers

The table uses prices visible on provider pages and trackers in the first week of October 2026. Marketplace rates move weekly; hyperscaler list prices are heavily discounted for committed spend.

Two things stand out. First, the gap between a DePIN and a well-run neocloud is now small: Hyperbolic at $2.40 and Lambda at $2.49 are within pennies, and the 70 percent savings io.net and Akash advertise are real only against hyperscaler list prices that few serious buyers pay. Second, prices are falling everywhere. Aethir’s September 2026 review noted that H100s launched at $7 to $10 an hour and now fetch $2 to $4, and that used H100 cards trade at $6,000 to $22,000, 40 to 60 percent below list. An asset that depreciates that fast rewards operators with the lowest power cost and the shortest payback, which is the lens for the miner pivot below.

Revenue, utilisation and what counts as real

Token market caps in this sector have often exceeded a billion dollars; revenue is smaller and unequal.

Cautions: Aethir’s revenue is company-reported and largely off-chain, so it cannot be audited from the blockchain the way Akash’s can; Akash’s 80 percent utilisation applies to a small fleet, and a network that is always full may simply be short of supply; io.net’s 30,000 GPUs is availability, not rentals. The question for any network is what was paid, by whom, for how many GPU-hours, and whether it is visible on-chain or in a filing.

Token models: burn-and-mint, staking and subsidies

Early DePINs paid providers in freshly minted tokens and hoped demand would arrive, which produced large supply and little revenue. The models now in use try to tie token value to spend.

Burn-and-mint equilibrium (BME)

Render popularised the design: customers pay in dollars or RENDER, the protocol burns tokens equal to the dollar value of the work, and a fixed schedule mints new tokens to node operators, so supply shrinks when burns exceed emissions. Akash adopted its own version on March 23, 2026 (Mainnet 17, Proposal 318). A tenant’s dollar payment buys AKT on the open market and burns it to mint ACT, a non-transferable, dollar-pegged compute credit; providers are paid in ACT and redeem it for AKT minted at the oracle price at settlement. In Akash’s example, $1,000 of credits bought at $1.14 burns about 877 AKT; if AKT is $1.50 at settlement only 667 are minted and 210 leave supply for good, while a falling price means net minting, with circuit breakers throttling new ACT at a 0.90 collateral ratio. BME removes token volatility from provider revenue and makes every dollar of usage a buyer of the token, but it only helps holders if usage is large relative to supply.

Staking, bonding and emissions

Providers on Nosana, Aethir and Bittensor stake tokens to participate. Staking creates a cost of misbehaviour and locks supply, but ties a provider’s economics to the token price, exactly the exposure BME removes. 0G’s Infinite AI inverts the idea: stake 0G, mint iAI and receive a daily inference allowance that expires if unused, a prepaid compute subscription. At the other end, Bittensor’s TAO emissions reward miners and validators by score whether or not a customer paid, which is why TAO’s value and subnet revenue can diverge. Akash returned 3,370,484 AKT of unused provider incentives to its community pool across 2025, keeping subsidies conditional. For any network, estimate what share of provider income is customer payment versus emissions; one whose providers would leave if emissions stopped is a subsidy, not a business.

The bitcoin miner pivot: the competitor to watch

The companies best placed to undercut everyone on GPU hosting are former bitcoin miners. They own what AI needs most and DePINs lack: grid interconnects measured in hundreds of megawatts, cooled buildings and balance sheets. From 2024 they began converting mining sites into high-performance computing campuses leased to AI labs and neoclouds.

• IREN announced a Microsoft GPU-cloud contract in November 2025, then signed $2.8 billion of new contracts with AI developers in July 2026 and raised its 2026 annualised revenue target from $3.7 billion to over $4 billion. In June 2026 it closed a $3.65 billion GPU financing facility.
• Cipher Digital (formerly Cipher Mining) extended its Barber Lake, Texas lease, with Fluidstack and Anthropic as the counterparties, to 20 years in September 2026, lifting expected revenue above $9 billion.
• Core Scientific, whose CoreWeave hosting contracts made it the template for the pivot, said in August 2026 that its Muskogee, Oklahoma site would expand to 1.5 GW after acquiring Polaris DS.

This matters for DePINs twice over. Miners that might once have plugged idle GPUs into Akash or io.net now sign multi-year take-or-pay leases with labs instead; the Axe Compute deal in Aethir’s orbit (2,304 B300s on a 36-month take-or-pay contract, September 2026) shows the DePIN world adopting the same structure. And the H100 capacity released as miners and neoclouds upgrade to Blackwell is what pushes marketplace rates toward $1.50 to $2.50: good for tenants, hard for anyone who bought H100s in 2023 expecting $4 an hour forever. For the broader story of bitcoin companies repurposing balance sheets, see our guide to bitcoin treasury companies.

Decentralized training: what has been shown

The claim that open networks can train frontier models deserves scrutiny, because it is both the most exciting idea in the sector and the one furthest from proof.

• Prime Intellect, INTELLECT-1 (November 29, 2024): a 10 billion parameter model trained across five countries with OpenDiLoCo, the first globally distributed run at that size.
• Prime Intellect, INTELLECT-2 (May 11, 2025): a 32 billion parameter model whose reinforcement learning rollouts ran on permissionless GPUs, showing that the loosely coupled part of RL post-training suits distributed hardware.
• Nous Research, Consilience 40B (reported September 29, 2025): described by Nous as the largest distributed pretraining run to date, coordinated by Psyche smart contracts on Solana and using DisTrO to compress communication. Nous has not published node counts or final benchmarks, and it was a testnet exercise: a capability demonstration, not a production model.
• Prime Intellect, INTELLECT-3 (November 26, 2025): a 106 billion parameter mixture-of-experts model post-trained from GLM-4.5-Air on 512 H200 GPUs connected by InfiniBand in one cluster over about two months. A strong open model, but a centralized run, and the company has not claimed otherwise.
• Pluralis, Node0 (January 2026): a 7.5 billion parameter model trained on 36 billion tokens over three weeks by 303 active participants on roughly 1,700 consumer GPUs with links as slow as 80 Mbps, reaching a loss of 2.75. In July 2026 Pluralis post-trained an 8B mixture-of-experts model with GRPO on 14 consumer Macs and one B200, lifting an agentic search score from 29 to 63 percent.
• 0G Labs, DiLoCoX: a method 0G says trained a 107 billion parameter model across decentralized nodes; without published token counts or benchmarks it is not yet comparable to the runs above.

The summary as of October 2026: distributed pretraining works at 7 to 40 billion parameters on tens of billions of tokens, far below the trillions frontier labs use; distributed RL post-training is practical today; and the one 100B-class model from this ecosystem was trained centrally. The trajectory is real, the gap is still large.

How we got here: a timeline

Worked example: renting 8 H100s for a month

How to evaluate a compute network: a checklist

• Is revenue visible and attributable? On-chain settlements (Akash) or company figures with a quarterly cadence (Aethir) are good; “annualised” numbers from one strong month are not. Ask what share is customer payment versus emissions.
• What is rented versus listed capacity? A network listing 30,000 GPUs with 3 percent rented is a directory, not a cloud. Look for GPU-hours sold per quarter.
• How is work verified? Attestation alone is weak. Sampling (Hyperbolic), checker nodes (Aethir), TEEs (0G, Akash) and reproducible execution (Gensyn) each close a specific hole; know which matters for your workload.
• Does the token capture usage? Burn-and-mint ties spend to supply; emissions-only models do not. Model the burn at current spend against annual emissions before assuming “deflationary”.
• Who are the providers? Three enterprise data centres are reliable but not decentralized; 269 hobbyist hosts are decentralized but not reliable. Match the mix to your SLA needs.
• Can it cluster? If you need more than one node, confirm the nodes in a lease share a fabric (InfiniBand or at least 400 Gb/s Ethernet) at a single site. Many marketplaces cannot promise this.
• Is there a Blackwell story? With H100 rates down 75 to 80 percent from the 2023 peak and take-or-pay leases absorbing the newest GPUs, is the network a secondary market for ageing H100s or does it have a path to B200 and B300 supply?

Risks and open questions

Price deflation cuts both ways. Falling GPU rates attract buyers but squeeze provider margins and the token burn that depends on dollar spend. If H100 rates settle near $1.50, a network must sell far more hours to burn the same value of tokens.

Verification remains incomplete. Inference verification relies on sampling or on TEEs that trust Intel and Nvidia; training verification is a research problem. Enterprises with regulated data will keep paying a premium for a named counterparty with contractual liability until that changes.

Token value and business value can diverge. Aethir’s revenue is company-reported and largely off-chain; Bittensor’s emissions are unlinked to customer payments; io.net’s listed supply far exceeds its rented supply. A token can trade on narrative long after fundamentals move, in either direction.

Regulatory status is unsettled. Compute credits like Akash’s ACT and 0G’s iAI are designed as prepaid services, not securities, but no regulator has ruled on them, and networks spanning 90-plus countries face export-control questions on advanced GPUs. Our guide to SEC versus CFTC jurisdiction covers the framework these tokens would fall under.

What to watch next

Glossary

GPU-hour

One graphics processor running for one hour; the standard billing unit for AI compute.

Hyperscaler

One of the largest cloud providers (AWS, Microsoft Azure, Google Cloud) that sells GPUs alongside a full suite of services.

Neocloud

A cloud provider specialised in GPUs, such as CoreWeave, Lambda, Nebius or RunPod, usually cheaper than hyperscalers for raw compute.

Interconnect

The network linking GPUs within and between servers; NVLink and InfiniBand are the high-bandwidth fabrics large training runs depend on.

Inference

Running a trained model to answer a request, as opposed to training it; the workload best suited to distributed GPUs.

Burn-and-mint equilibrium (BME)

A token model in which customer payments burn tokens equal to the dollar value of work while a schedule mints new tokens to providers, so net supply tracks usage.

Trusted execution environment (TEE)

A hardware-isolated enclave, such as Intel TDX paired with TEE-enabled GPUs, that runs code where the machine owner cannot read it and produces attestations proving what ran.

Proof of Sampling

A verification scheme that recomputes a random fraction of outputs to deter cheating without doubling the work.

Take-or-pay

A contract in which the customer pays for reserved capacity whether or not it is used, common in miner-to-AI-lab leases.

Subnet

In Bittensor, an independent market with its own incentive rules where miners produce an output and validators score it for TAO rewards.

Why it matters

Decentralized compute is the first DePIN category to generate revenue a traditional analyst would recognise, and that is both its achievement and its test. Aethir’s $127.8 million and IREN’s multi-billion-dollar contracts come from the same demand, and the networks that survive will offer what neoclouds and miners cannot: permissionless access, verifiable execution, consumer and edge supply the big players will never aggregate, and compute an AI agent can buy with a wallet rather than a procurement process.

For buyers, serving and tuning open models on a marketplace at $1.50 to $2.50 an hour is a sound default in late 2026 if the work tolerates interruption. For investors, price the business, not the narrative: count GPU-hours sold, check the burn against emissions, and watch where Blackwell supply lands first. Track related coverage in the Crypto Coin Show interview archive and the research hub.

Sources

1. Akash Network: Q1 2026 Report, April 1, 2026
2. Akash Network: What Burn-Mint Equilibrium Means for Akash, March 18, 2026
3. Aethir: Aethir on GPU Resale Value and Rental Pricing, September 29, 2026
4. Aethir: Axe Compute Takes Full Ownership of Georgia AI Cluster, October 6, 2026
5. CoreWeave: GPU Cloud Pricing, accessed October 6, 2026
6. AWS: EC2 Capacity Blocks for ML Pricing, accessed October 6, 2026
7. Hyperbolic: Pricing, accessed October 6, 2026
8. Nosana: GPU Marketplace, accessed October 6, 2026
9. io.net: Decentralized GPU Cloud, accessed October 6, 2026
10. 0G Labs: Introducing Infinite AI, September 28, 2026
11. 0G Labs: GLM-5.3 now runs fully sealed in TeeML on 0G, September 3, 2026
12. Prime Intellect: INTELLECT-3, November 26, 2025
13. Nous Research: The Next Phase of Psyche, September 29, 2025
14. Pluralis Research: Multi-party Training Stack (Node0), January 2026
15. Gensyn: REE, AXL and open-1b, accessed October 6, 2026
16. Cipher Digital: Expands Barber Lake Lease Term to 20 Years, Increasing Revenue to Over $9 Billion, September 25, 2026
17. Core Scientific: Press Releases (Polaris DS, Muskogee 1.5 GW), August 14, 2026
18. IREN: Signs $2.8bn in New Customer Contracts, Raises 2026 ARR Target to over $4bn, July 20, 2026
19. The Block: Aethir brings in $36 million in first-year revenue, July 16, 2024
20. BlockEden: DePIN Compute Revenue Pivot (Akash, io.net, Aethir, Render data), March 12, 2026
21. Spheron: H100 Price Per Hour 2026, updated October 3, 2026
22. Spheron: Nvidia B200 Cloud Pricing 2026, updated October 3, 2026

Disclosure: Ashton Addison, founder of Crypto Coin Show, holds a token warrant in 0G Labs. They had no input into this guide. This guide is for education only and is not investment, legal or tax advice.