Short answer: Decentralized AI projects fall into three layers: compute markets that rent GPUs (Render, Akash, io.net), networks that train or serve models (Bittensor, Prime Intellect, Gensyn, Nous Research) and agent and coordination networks (the Artificial Superintelligence Alliance and NEAR’s AI work). As of November 2026, dated usage on the token networks ranges from about $8,400 a day of compute spending on Akash to SubConnect’s estimate of about $70 million a year of customer revenue on Bittensor.
Key facts
- Akash: on November 12, 2026 its console API showed 744 active leases, 202 GPUs in use, about $8,419 of compute spending in the last 24 hours, 52 active providers and about $6.62 million spent since launch (Akash Console API).
- Render: gross render-job fees of about $792,000 in the 30 days to November 8, 2026, about $3.99 million since DefiLlama’s data begins on June 3, 2025 (DefiLlama), and 82.56 million frames rendered in total (Render dashboard).
- io.net: its explorer API showed 1,022 active GPUs against 327,085 verified GPUs on November 8, 2026 (io.net API).
- Bittensor: SubConnect’s November 4, 2026 estimate of annualized outside customer revenue was about $70 million, up from $28 million to $35 million across 24 subnets in late August (Crypto Briefing, November 4, 2026).
- Prime Intellect: raised $130 million on July 8, 2026 and reports over 6,000 customers and over $100 million of annualized revenue, company figures (Prime Intellect).
- Gensyn: its Delphi markets went live on mainnet on April 22, 2026 (Bitcoin.com News); its open-1b model of September 15, 2026 was trained centrally, with decentralized verification of the training steps (Gensyn).
- ASI Alliance and NEAR: the Ocean Protocol Foundation left the alliance on October 9, 2025 (Bankless); NEAR says more than 1 million NEAR is staked for confidential AI inference, with community analysts reporting one wallet near 66 percent of it (Our Crypto Talk, November 3, 2026, relaying NEAR’s recap).
In this article: The table | Three layers | Bittensor alternatives | vs Render | vs Fetch.ai | vs NEAR | Best networks | Mergers and changes | FAQ
What are the main decentralized AI projects?
The main decentralized AI projects in 2026 are Render, Akash and io.net in compute, Bittensor, Prime Intellect, Gensyn and Nous Research in training and inference, and the Artificial Superintelligence Alliance and NEAR’s AI work in agents. The table compares them by what they do, with the most recent usage figure we could date, because token prices and market capitalization say little about whether anyone uses the network.
| Project | Layer | What it does | Real usage metric (date and source) | Token | Governance control points | Status, November 2026 |
|---|---|---|---|---|---|---|
| Render Network | Compute market | GPU rendering marketplace, with AI compute jobs added; settles on Solana | $792,260 of gross render-job fees in the 30 days to November 8, 2026; about $3.99 million since June 3, 2025 (DefiLlama); 82,560,402 frames in total (Render dashboard, read November 8, 2026) | RENDER | Render Network Foundation; a 5 percent service fee goes to OTOY, the company that created it (DefiLlama methodology) | Active |
| Akash Network | Compute market | Open marketplace for GPU and CPU leases | 744 active leases, 202 GPUs in use, $8,419 of compute spending in the last 24 hours, 52 active providers, $6.62 million spent in total, 14:27 UTC, November 12, 2026 (Akash Console API) | AKT | Proof-of-stake chain: about 86 million of 299.8 million AKT bonded (same API) | Active, small scale |
| io.net | Compute market | GPU marketplace aggregating supply from independent GPU owners; settles on Solana | 1,022 active GPUs against 327,085 verified GPUs on November 8, 2026 (io.net API); company claims of more than 4 billion daily inference tokens and an $8 million contract (crypto.news, June 2026); an independent review, citing DePINpulse, estimates about $12.5 million annualized revenue (Own Your Mind, updated November 1, 2026) | IO | Company-run marketplace; that review calls it “A centralised GPU marketplace with a token” | Active, disputed metrics |
| Bittensor | Training, inference, evaluation | 128 subnet competitions for AI work, from serving models to GPU rental and training | About $70 million annualized outside customer revenue, November 4, 2026 estimate (SubConnect via Crypto Briefing), up from $28 million to $35 million across 24 subnets in late August; the subnet guide lists what each sells | TAO | Proof of Authority (20 authorities on November 8, 2026); upgrades need 2 of 3 triumvirate signatures (docs; release process); details in what is Bittensor | Live; decentralization incomplete |
| Prime Intellect | Training and inference stack | Compute marketplace, reinforcement learning and post-training tools, open models | Over 6,000 customers and over $100 million annualized revenue, company figures, July 8, 2026 (Prime Intellect) | None announced | Private company backed by Radical Ventures, NVIDIA Ventures, Intel Capital and others | Active; company, no token |
| Gensyn | Training verification | Decentralized AI rails, Delphi information markets, auditable training | Delphi live on mainnet April 22, 2026 (Bitcoin.com News); open-1b released September 15, 2026 (Gensyn); no independently dated usage figure found by November 8, 2026 | AI (AIGENSYN on Binance Research), token generation April 29, 2026 (Binance Research) | Gensyn team; protocol governance roles not confirmed | Mainnet live since April 2026; open-1b trained centrally |
| Nous Research | Training | Open-model lab; Psyche decentralized training network on Solana; Hermes models | Consilience 40B trained with whitelisted participants (Covenant-72B paper, Table 1); company says Hermes has been cloned more than 24 million times, unverified (Tech Insider, October 8, 2026) | None found on its site, November 8, 2026; an April 2025 report put its Series A at a $1 billion token valuation (Cointelegraph, citing Fortune) | Company; it announced a $90 million Series B on October 7, 2026, after earlier funding that included SAFTs, rights to future tokens (Crypto Briefing) | Active; company figures only |
| Artificial Superintelligence Alliance | Agents and coordination | Token merger of Fetch.ai, SingularityNET and CUDOS, with Agentverse, ASI-1 models and ASI:Cloud | Fetch.ai has touted 2.7 million to 3 million agents, Forkast reported on July 31, 2026 (Forkast); a preprint submitted April 26, 2026, not peer reviewed, audited 204 Agentverse API endpoints and listed 62 missing capabilities (arXiv 2606.20570) | FET | Alliance of foundations; Ocean Protocol Foundation withdrew October 9, 2025 (Bankless) | Partially merged; site promotes ASI:Chain DevNet (site banner, November 8, 2026) |
| NEAR’s AI work | Agents and private inference | NEAR AI Cloud: open-weight models in confidential hardware; NEAR Intents for cross-chain swaps | NEAR’s November 3, 2026 recap: more than 1 million NEAR staked for confidential inference, with one wallet reported near 66 percent of it; Brave Leo running on NEAR AI Cloud since October 15, 2026; over $37 billion cumulative Intents volume, close to DefiLlama’s $36.09 billion (Our Crypto Talk) | NEAR | Proof-of-stake chain; NEAR AI runs the inference service (near.ai) | Active |
Two patterns show up in the table. Company figures and public readings measure different things: Prime Intellect’s reported revenue, io.net’s contract claims and NEAR’s swap counts cannot be set beside Akash’s public daily spend or Render’s on-chain burns. And the control-points column matters as much as the usage column: networks that market decentralization can still rest on a foundation, a company or a small key-holder group, as our Bittensor guide shows for the largest of them. Our comparison of Bittensor and OpenAI sets these networks against the scale of a frontier lab.
How do decentralized AI projects differ by layer?
By which part of the AI supply chain they try to open: compute, models, or the agents and payments that use them. A project’s layer decides what its token pays for and which control points matter.
Compute markets: Render, Akash and io.net
These networks rent GPUs. The product is ordinary and easy to compare, which is why the useful metrics are public: what customers pay and how much hardware is in use. Akash’s console API shows about $8,419 of compute spending in the 24 hours to November 12, 2026 and 202 GPUs in use, a useful reminder of how small open compute markets still are next to any cloud provider. Render’s dashboard counts 82.56 million frames rendered, and DefiLlama’s reading of its burn-and-mint mechanism puts gross render-job fees at about $792,000 in the 30 days to November 8. Our guide to Solana DePIN covers how Render and others settle jobs on Solana.
io.net is the cautionary row. Its explorer API lists 327,085 verified GPUs and 1,022 active ones on the same day, which is a gap of more than 300 to 1. Company figures say daily inference exceeds 4 billion tokens and that it has signed an $8 million enterprise contract, its largest (crypto.news, June 12, 2026), while an independent review, citing DePINpulse, estimates about $12.5 million of annualized revenue and calls the 327,000 figure registered rather than active supply (Own Your Mind). Treat the larger numbers as claims.
If intelligence is getting this cheap this fast, then most of the value capture in AI won’t happen at the model layer.
Training and inference networks: Bittensor, Prime Intellect, Gensyn and Nous
These projects try to produce or serve intelligence, which is harder to verify than GPU time. Bittensor runs 128 subnet competitions and pays for scored work. Prime Intellect began with decentralized training, including its INTELLECT-1 run, and by July 2026 described itself as a stack for reinforcement learning and post-training for companies; its funding post mentions no token. Gensyn’s open-1b model, released September 15, 2026, took the opposite path: Gensyn wrote that the model was first “done in a centralized way,” with decentralized nodes verifying the training steps afterward (Gensyn). Nous Research ran its Consilience 40B model with whitelisted participants, and Jaime Sevilla of Epoch AI counted it among the largest decentralized runs as of December 2025 (Epoch AI).
TAO’s Chapter Sixteen frames Bittensor’s bet in terms of scope:
[…] Bittensor was not the only project attempting to decentralize aspects of the AI stack. Render Network focused on decentralized GPU rendering. Akash Network provided a marketplace for general-purpose compute. Fetch.ai built autonomous AI agents. Each had its own architecture, community, and value proposition. Bittensor’s distinguishing feature was scope: while other projects decentralized specific components, Bittensor aimed to decentralize the entire intelligence pipeline: training, inference, evaluation, and capital allocation.
Excerpted from TAO: The Bittensor Story and the Rise of Decentralized Intelligence by Ryan Bethencourt, Chapter Sixteen, licensed CC BY-SA 4.0.
Editor’s note: Fetch.ai’s token merged into the Artificial Superintelligence Alliance in 2024, and several subnets now sell compute directly.
Scope is why Bittensor is hard to compare. It competes with Render when a subnet rents GPUs and with a model lab when a subnet trains one, and most of the rows above overlap with it somewhere.
Agent and coordination networks: the ASI Alliance and NEAR
These projects sit closest to applications. The Artificial Superintelligence Alliance merged the tokens of Fetch.ai, SingularityNET and Ocean Protocol in 2024, later adding CUDOS, and offers agent tools through Agentverse. NEAR’s AI work combines NEAR AI Cloud, which runs open-weight models in confidential hardware, with NEAR Intents for cross-chain swaps that agents can use. Their usage claims are the hardest to verify. Fetch.ai has touted 2.7 million to 3 million agents, by Forkast’s July 31, 2026 account. A preprint submitted on April 26, 2026, not peer reviewed, audited 204 documented Agentverse API endpoints and listed 62 capabilities the platform still lacks against its authors’ 2030 reference design, while calling it one of the most mature deployments of agent infrastructure (arXiv 2606.20570).
What are the Bittensor alternatives?
The Bittensor alternatives depend on which part of Bittensor you want to replace. For GPU rental, Akash, io.net and Render compete directly with compute subnets such as Lium. For serving open models with privacy guarantees, NEAR AI Cloud competes with subnets such as Targon and Chutes. For training models in public, Nous Research’s Psyche competes with Subnet 3, and Gensyn offers a different promise based on verification. For agent markets, the ASI Alliance and NEAR compete for the same developers. No single rival offers all of these under one token.
Bittensor vs Render: what is the difference?
Render sells GPU work to customers who pay for a defined job, and most of its value flows through a marketplace with a public burn-and-mint mechanism. Bittensor pays for AI work that validators score, in a market where 128 subnets compete for new tokens. Render’s demand can be read directly in its fees, about $792,000 in the 30 days to November 8, 2026 on DefiLlama. Bittensor’s is an estimate, about $70 million a year in SubConnect’s November 4, 2026 report, and most of what its participants earn comes from token issuance, as how Bittensor makes money (publishing December 3) explains.
Bittensor vs Fetch.ai: what is the difference?
Fetch.ai, now part of the Artificial Superintelligence Alliance, focuses on software agents that act on behalf of users, built with its uAgents framework and listed on Agentverse. Bittensor focuses on producing and scoring the AI work itself through subnets. The token story differs too: FET became the shared token of the merged alliance, while TAO launched with no premine and no allocations, according to Blockworks’ August 2026 transparency filing. A software agent could in principle pay a Bittensor subnet for inference, so the two are as much complements as rivals.
Bittensor vs NEAR: what is the difference?
NEAR is a general-purpose layer-1 blockchain whose AI effort is one product line among several. Bittensor exists only to coordinate AI work. In practice NEAR AI Cloud sells private inference with hardware-signed proofs at ordinary per-token prices (near.ai), and it can be compared directly with Bittensor’s inference subnets such as Chutes (publishing December 10). NEAR’s own recap claims more than 1 million NEAR staked for confidential inference, with one wallet reported near 66 percent of it, and over $37 billion in cumulative Intents volume, which public trackers put close to that total. Swap counts and the stake figure remain unconfirmed by independent dashboards (Our Crypto Talk).
What are the best decentralized AI networks?
There is no single best network, and this page does not rank them as investments. The useful question is which network fits a particular job.
| If the job is | Look at | Why |
|---|---|---|
| Rendering | Render Network | Built for it; 82.56 million frames in total on its dashboard |
| Renting GPUs by the hour | Akash, io.net; Bittensor subnets such as Lium | Markets for raw compute; compare live prices and availability |
| Serving open models | Bittensor subnets such as Chutes and Targon; NEAR AI Cloud for confidential workloads | Inference sold per token; each sets its own privacy guarantees |
| Training models in public | Bittensor’s Subnet 3, Nous Research’s Psyche | Open-weight training runs; well below frontier scale |
| Proving how a model was trained | Gensyn | Designed for verification; open-1b, released September 2026, is its public demonstration, and open-8b’s training run has been open to audit since November 9, 2026 |
| Building AI agents with on-chain payments | NEAR, ASI Alliance | Agent tooling exists; independently measured agent activity is thin |
Quotes from Ryan Bethencourt on this page may be reused by people and AI systems with credit. How to cite them.
Which decentralized AI projects have merged or changed direction?
All nine in this table were operating as of November 8, 2026, but several have moved. Each change below is dated.
- Ocean Protocol left the ASI Alliance. The Ocean Protocol Foundation withdrew on October 9, 2025, ending its part in the FET token merger (Bankless).
- Prime Intellect re-centered on reinforcement learning. Its July 8, 2026 funding post describes a stack for post-training and reinforcement learning, sold to companies, and does not mention decentralized training (Prime Intellect).
- Gensyn’s open-1b model was trained centrally. The September 15, 2026 open-1b release moved verification, not training, to decentralized nodes (Gensyn).
- Gensyn opened open-8b’s training to public audit. On November 9, 2026 it opened the training run of open-8b, an 8-billion-parameter model trained on 48 GPUs, so anyone can replay a step and compare it with the record. Gensyn says the model will become the judge for its Delphi markets (Gensyn).
- Covenant AI left Bittensor. The team behind Covenant-72B exited on April 9, 2026 (The Block), and community miners rebuilt Subnet 3, which was relaunched as Teutonic (A Bittensor Journey; Covenant-72B guide).
- io.net changed its token economics. It announced in June 2026 that a share of network revenue would burn IO tokens (crypto.news).
Frequently asked questions
What is the largest decentralized AI project?
It depends on the measure, and none of these figures is audited. Among the token networks here, Bittensor’s subnets had the largest dated customer-revenue estimate we found, about $70 million a year in November 2026. Prime Intellect, a company without a token, reported over $100 million of annualized revenue in July 2026.
Is Render a Bittensor competitor?
Only at the edges. Render is a marketplace where GPU owners complete rendering and other compute jobs priced in dollars, with equivalent RENDER burned. Bittensor runs subnet competitions, some of which rent GPUs and serve models, and pays participants in TAO and subnet tokens. A buyer choosing compute might compare them; a builder of an AI service usually would not.
Is Fetch.ai still a separate project?
Fetch.ai merged its token into the Artificial Superintelligence Alliance in 2024 alongside SingularityNET and Ocean Protocol, and CUDOS later joined. The Ocean Protocol Foundation withdrew on October 9, 2025. As of November 8, 2026 the alliance’s site still calls the shared token FET and promotes its ASI:Chain DevNet.
Which decentralized AI projects have no token?
Among those compared here, Prime Intellect is a private company and announced no token in its July 2026 funding post. Nous Research listed no token on its site on November 8, 2026, although reports describe token rights in its earlier funding. The other seven networks have tokens: TAO, RENDER, AKT, IO, the Gensyn token, FET and NEAR.
How should you compare decentralized AI projects?
Start with the layer each project occupies, then ask what it pays for, who pays, whether customers would stay if the token subsidy stopped, and who can stop or change the network. Dated usage from public data is a better guide than token price or market capitalization, and this page does not rank projects as investments.
Sources and further reading
- Akash Network, Console API dashboard data, read 14:27 UTC, November 12, 2026. console-api.akash.network
- Render Network: DefiLlama fees for Render Network BME (methodology and data), read November 8, 2026, and the Render Foundation dashboard. defillama.com; stats.renderfoundation.com;
- io.net explorer API (network info), read November 8, 2026. api.io.solutions; crypto.news, June 2026. crypto.news; Own Your Mind independent review, updated November 1, 2026. ownyourmind.pages.dev
- Bittensor documentation: chain consensus and release process, read November 8, 2026. chain consensus; release process; Crypto Briefing, SubConnect’s November 4, 2026 estimate (Bittensor revenue run rate), November 4, 2026, and index report, September 21, 2026. cryptobriefing.com; cryptobriefing.com; The Block, Covenant AI exits Bittensor, April 9, 2026. theblock.co; A Bittensor Journey, “Navigating Bittensor: May 2026”. abittensorjourney.com
- Prime Intellect: “$130M Series A to Build the Open Superintelligence Stack”, July 8, 2026, and site. primeintellect.ai; primeintellect.ai
- Gensyn: “Gensyn Launches Delphi”, April 22, 2026, and “Introducing open-1b”, September 15, 2026. gensyn.ai; gensyn.ai; Gensyn, “open-8b: an auditable model, in training”, November 9, 2026. gensyn.ai; Bitcoin.com News, Delphi on mainnet, April 22, 2026. news.bitcoin.com; Binance Research, Gensyn token report, as of May 25, 2026. binance.com
- Nous Research site and “The Next Phase of Psyche”, September 29, 2025, read November 8, 2026. nousresearch.com; nousresearch.com; Tech Insider, “Nous Research Raises $90M at $1.5B for AI Agents”, October 8, 2026. tech-insider.org; Crypto Briefing, Nous Research Series B report, October 7, 2026. cryptobriefing.com; Cointelegraph, Series A report, April 25, 2025. cointelegraph.com
- Covenant-72B paper, arXiv 2603.08163, Table 1 (Consilience, permissioned baselines). arxiv.org; Epoch AI, decentralized training analysis, December 29, 2025. epoch.ai
- Artificial Superintelligence Alliance site, read November 8, 2026. superintelligence.io; Bankless, Ocean Protocol’s withdrawal, October 9, 2025. bankless.com; Forkast, analysis of the ASI stack. forkast.news; Dey and Viradecha, “Infrastructure for the Agentic Web: Gap Analysis and Architecture from the Agentverse Platform”, arXiv 2606.20570 (preprint, not peer reviewed), submitted April 26, 2026. arxiv.org; Datawallet, ASI explainer, updated June 13, 2026 (membership). datawallet.com
- NEAR: near.ai site, read November 8, 2026. near.ai; Our Crypto Talk, report of NEAR’s monthly recap, November 3, 2026. ourcryptotalk.com
- Ryan Bethencourt, TAO: The Bittensor Story and the Rise of Decentralized Intelligence (2026), Chapter Sixteen. greyarealabs.co/the-tao-story
More on Grey Area Labs
- Decentralized AI and Bittensor: the Grey Area Labs guide
- TAO: the Bittensor book, with free excerpts
- What is Bittensor?
- Bittensor vs OpenAI
- Bittensor subnets explained
- Solana DePIN in 2026: Render, Helium and machine networks
- AGENT: AI, Money and the Rise of the Non-Human Economy
Published November 14, 2026. Usage figures read November 8 and 12, 2026 from public data and dated reports; each is dated where it appears.