Start with a price. Rent an H100 — the workhorse chip of the AI boom — from Amazon Web Services and the meter runs roughly $4.50 to $5.50 an hour. Rent comparable capacity on Akash Network, a marketplace with no sales team and no headquarters, and the price drops to $1.20 to $1.80. That gap is not a rounding error or a promotional week. It is the entire argument for decentralized AI inference, stated in two numbers, and in 2026 the networks making it stopped being whitepapers.
The default arrangement has been simple for years. Nearly every AI product in use runs on somebody else's computer, almost certainly one of three companies: Amazon Web Services, Microsoft Azure or Google Cloud. Together they control the overwhelming majority of cloud AI. Rent from them, or don't ship. That has been the deal.
Alongside the hyperscalers, a parallel infrastructure has been quietly building: decentralized networks gather idle GPU capacity, spread inference work across scattered nodes, and pay providers in tokens instead of a monthly invoice. Bittensor's subnets process real inference volume. Akash set records for compute spend. io.net says its pool spans hundreds of thousands of GPUs in more than a hundred countries. The questions now are what these systems deliver, what they charge, and where they fall short.
The price gap is the argument
The economic case starts with waste: GPUs cost a lot, and renting them from the cloud costs more than any other route. For a startup running inference at scale, that hourly meter never stops. Meanwhile, enormous quantities of GPU capacity sit idle — in crypto mining farms repurposed after proof-of-work declined, in enterprise data centers between batch jobs, in gaming PCs overnight. The supply exists. The demand exists. What the networks sell is the meeting point, without the hyperscaler margin: a token layer handles discovery, payment and, in some designs, verification.
Money isn't the only argument, though. There is a sovereignty case: centralized providers answer to the laws of the countries where they are based, and the U.S. CLOUD Act is the example everyone cites. For certain users and governments, running inference on infrastructure that no single company controls counts as a feature in its own right, not merely a cheaper price. The discount gets the headlines. What holds governments' and regulated buyers' attention is that jurisdiction question, which stays put even when token prices swing.
There is also resilience. When thousands of small providers form a mesh, there is no single point's failure brings the whole thing down — a property that matters more with every year that AI workloads become critical infrastructure. Price brings people to the door; jurisdiction and resilience are why some of them stay.
Bittensor pays for answers, not hours
Bittensor is the most ambitious of these networks, and the most unusual: it doesn't rent out raw GPUs at all. Instead it runs specialized subnets — 128 of them active — where miners compete at AI tasks: text completion, image generation, prediction, retrieval. Validators judge how useful the work was, and that judgment sets payment in the network's TAO token. You are not buying machine time. You are buying judged output.
The figures point to a mechanism gaining ground. Bittensor posted approximately $43 million in Q1 2026 revenue, the highest-earning project in the DePIN — decentralized physical infrastructure — sector, with market capitalization around $2.7 billion as of May 2026. Its Chutes subnet, focused on serverless LLM inference, was the first to pass $100 million in cumulative inference volume. Names such as Nvidia and Polychain Capital appear among its clients and validators.
The caveats are real, though. On busy subnets, a hundred or more miners chase the same fixed reward pool, and once electricity costs are paid most are left with near-zero. In practice, a hobbyist running a single RTX 4090 is generally outmatched, and because token prices move sharply, the economics of taking part swing along with crypto markets. And Bittensor answers "which outputs are useful?" through validator consensus — a different trust proposition than a contractual SLA with a cloud provider. That distinction runs through the whole sector, and it is the gap the verification projects below are trying to close.
That consensus mechanism deserves a plain statement. A cloud contract gives you a counterparty: a company that can be sued, audited and held to a service level. A validator network gives you a market judgment: output is worth paying for because independent participants, staking their own positions, scored it useful. Trust by law versus trust by incentive — they fail differently, and a buyer who cannot say how each fails does not yet understand the purchase.
Auctions, aggregators and ghost GPUs
Akash Network and io.net are closer to what most people picture: marketplaces for raw compute. Akash runs a permissionless reverse-auction model: users post workloads, providers bid, the lowest price wins. It processed a record $5 million in compute spend in Q1 2026, at that $1.20 to $1.80 per hour for H100 capacity against AWS's $4.50 to $5.50 — a 60 to 75 percent discount. Akash launched its Burn-Mint Equilibrium upgrade in March 2026: every dollar of compute spend destroys an equivalent amount of AKT tokens, so token demand is tied directly to real usage — scarcity tracks real demand rather than speculation alone. On OpenRouter, its AkashML service now processes 1.7 billion tokens per day, reportedly ahead of Cloudflare's volume.
io.net takes the aggregation approach, pooling idle GPUs worldwide, including repurposed crypto-mining hardware, into clusters for AI training and inference. It reports 370,000+ GPUs across 138 countries and more than $20 million in cumulative on-chain revenue. The honest footnote, raised by independent analysts, is that actual utilization rates and the proportion of "ghost GPUs" — listed but unavailable capacity — remain disputed. The network exists and transacts, but from outside it is harder to verify how much headline capacity can genuinely be used at any given moment. Announcing capacity is easy, auditing it is hard.
The auction model has a second effect: the price discovered is a market price, not a list price — moving with real supply and demand in a way a hyperscaler's published rate card never does. The fine print is that a market price cuts both ways; it can rise as fast as it fell, and it settles in tokens whose own price is a market too.
Render Network, originally built for GPU rendering, has pivoted substantially toward AI compute, with AI tasks now 35 to 40 percent of its job volume. Aethir has taken the enterprise route, reporting $166 million in annualized revenue for Q3 2025 and 1.4 billion compute hours delivered, at prices roughly 70 percent below AWS. Aethir's model, though, depends on a few large clients instead of truly distributed nodes. Decentralized in branding, less so in architecture.
Trust is the unsolved line item
Gensyn attacks a different problem: not cheaper GPUs, but verifiable computation. Its protocol uses cryptographic verification to prove that a claimed AI computation actually happened correctly, the fundamental trust question of decentralized inference: can you trust a GPU you don't own? Its token and mainnet went live in 2026, but its official RL Swarm runs were paused as of September 2026, leaving Delphi as the live surface. Treat it as a watchlist project: interesting verification technology, but still not a drop-in training cloud.
Adjacent efforts point the same way: Ritual focuses on on-chain verifiable inference, and Prime Intellect released INTELLECT-2, its model with TOPLOC verification for decentralized training. The industry is converging on the view that trustless AI compute needs cryptographic proof, not economic incentives alone — but the production-grade implementations are still early.
Then there are the limits. Latency also varies: sending inference across a global mesh of heterogeneous nodes cannot match the sub-50ms consistency of a hyperscale region, and that matters for real-time applications. Reliability is uneven; individual providers can go offline, and redundancy is not the five-nines engineering of AWS. Compliance is the enterprise blocker. Regulated industries need audit trails, data-residency guarantees and support contracts — the same obligations now being codified into formal AI governance frameworks — which token networks do not provide. Capacity claims deserve a careful reading as well: "GPUs listed" does not mean "GPUs available," and token incentives can keep alive usage that would not survive unsubsidized. Token-price volatility feeds straight into provider economics, and provider economics feed into reliability — a chain the hyperscalers' monthly invoices don't have.
None of this makes the movement hollow. Bittensor's subnet revenues, Akash's record compute spend and io.net's transaction volume are real economic activity, not projections. While multi-agent setups leave pilots behind for production and enterprises serve open-weight models on infrastructure they control themselves, most serious workload still runs on hyperscalers for compliance and support reasons — the same reasons open-source models overtaking proprietary ones in benchmarks hasn't by itself moved enterprise contracts. The most accurate framing: decentralized inference has become a genuine secondary market for AI compute — cheaper, more open and more censorship-resistant than the hyperscalers, but not yet a replacement. It is already a serious option for cost-sensitive batch inference, open-source model serving and sovereignty-conscious deployments. The cloud giants remain the default where workloads are latency-critical and compliance-heavy.
Break the verdict down by workload and it sharpens. Cost-sensitive batch inference — work that can queue, retry and tolerate variable latency — is where the discount is worth most and the limits bite least. Open-source model serving fits the same networks naturally: the models are public and the serving stack is portable. Sovereignty-conscious deployments are the strategic case: buyers for whom the question was never only price. Against those three, set the workloads the mesh still cannot honestly take: the latency-critical application needing sub-50ms consistency, and the compliance-heavy contract needing an audit trail, data-residency guarantee and someone to call. The line between the two lists moves with verification, not with token prices.
Decentralised inference is an argument about who owns the hardware layer. The centralised end of that market is being rebuilt around NVIDIA's Blackwell B200 superchips and the liquid-cooling engineering they demand. Sovereignty pushes from the other side too: Huawei's Mate 90 and its Kirin 9050 Pro are pitched as an AI stack answering to no foreign cloud.
Sources
- Startupik — "Bittensor Alternatives 2026: io.net, Akash, Render, Ritual" (network status, Gensyn RL Swarm paused, Sept 2026): startupik.com
- HOGE Wire — "Decentralized Inference in 2026: Can You Trust a GPU You Don't Own?" (Chutes $100M volume, io.net 100K+ GPUs, Akash burn-mint): hoge.gg
- LinkedIn / Yang Shanhe — "Decentralized Compute Projects: A Survey from DePIN Experiments to AI Infrastructure" (Bittensor $43M Q1 revenue, Akash $5M record, pricing data): linkedin.com
- Coinroop — "10 Best Gensyn Rivals for Decentralized AI Training" (Akash 594K leases, Render 60,000 GPUs via Salad): coinroop.com
- gaboesquivel.com — "2025-11: Web3 and AI" (Bittensor subnets, Fetch.ai agents, marketplaces overview): github.com
