Akash vs io.net vs Render: The DePIN GPU Race in 2026

The decentralized AI-compute trade is loud again. Akash Network’s AKT changed hands near $0.78 on 5 October, up roughly 8% on the day and about 17% over the week, according to CoinGecko. It was not alone. Nosana’s NOS had jumped more than 70% in seven days, io.net’s IO was up high single digits, and Render had pushed its market cap back above $1 billion. The whole basket of tokens that promise to rent out graphics cards re-rated together, in a week when Bitcoin pushed into Uptober near $85,000 and the macro backdrop flipped bullish.

Moves like this usually say more about risk appetite and sector rotation than about any one network’s fundamentals. The same low-float dynamic that drove Gensyn’s $AI through an AI rotation it could barely catch is at work across the whole group: thin circulating supply, a strong narrative, and a wall of buyers arriving at once. So this is a good moment to step back from the candles and ask the more useful question. If you are buying exposure to decentralized GPUs, what are you actually buying? This piece puts the main contenders side by side, Akash, io.net, Render, Nosana and Phala, against the centralized benchmark they all measure themselves against, CoreWeave, and scores them on price, real usage, tokenomics and trust.

What a decentralized GPU marketplace really sells

The pitch is the same across the sector. The world is full of underused graphics cards, sitting in crypto-mining sheds, independent data centers, research labs and gaming rigs, while demand for AI compute has outrun the big clouds’ ability to pour concrete. A permissionless marketplace, the argument goes, can match that idle supply to that demand and clear at a price well below Amazon, Microsoft or Google. You package your workload into a container, the network finds a machine, and you pay by the hour, or in some cases by the block.

The differences live in the details: how each network matches supply and demand, what hardware it attracts, how its token actually fits into the money flow, and whether a serious enterprise can trust an anonymous stranger’s machine with its data and model weights. It is also worth drawing a boundary. This is the GPU-rental group. Bittensor (TAO) is often lumped in with it, but it runs a different model entirely, rewarding machine-learning contributions across competing subnets rather than renting you an H100 by the hour, so it sits outside this comparison. The five networks below, plus CoreWeave, are the ones you can actually hand a container to and get a GPU back.

Akash: the reverse-auction incumbent

Akash is the oldest of the group and the one with the most distinctive matching mechanism. Built by Overclock Labs and live on mainnet since 2020, it added GPUs with its Supercloud upgrade in August 2023. The defining feature is a reverse auction. A tenant writes a short manifest in Akash’s Stack Definition Language (SDL) describing the hardware and image they want; providers bid to host it; and, crucially, the tenant is not forced to take the lowest bid. You can weigh a provider’s price against its reputation and audited status before you sign a lease. Payment runs through on-chain escrow, drawn down block by block, and tenants can settle in AKT, in USDC or in IBC-bridged tokens, which means a buyer can rent a GPU without ever holding the native token.

On top of the raw marketplace, Akash has built an inference layer. AkashML and the consumer-facing Akash Chat serve open models such as Llama, DeepSeek and Qwen, and the project reported throughput climbing from roughly 1.7 billion tokens a day early in 2026 to more than 10 billion a day by midsummer. The marketplace also crossed a cumulative milestone of around $5 million in all-time compute spend, per Akash’s own Q1 2026 report. Founder Greg Osuri has framed the company’s urgency around the gap between software and infrastructure timelines, writing that “AI moves in months, energy moves in years.” Hold onto that $5 million figure; it collides with a very different number later in this piece.

io.net: orchestration, and the GPU-count question

io.net approaches the problem from the software side. Built on Solana, it leans on Ray, the open-source framework for distributed Python, to stitch scattered GPUs into clusters that behave like a single machine for machine-learning training and inference. Where Akash sells you one provider’s box, io.net’s pitch is that it can assemble a cluster on demand across many providers, with an orchestration layer and, more recently, an IO Intelligence product and developer tooling layered on top. In its own head-to-head comparison, io.net argues it is the better fit for GPU-heavy ML work while conceding Akash’s strength in general container workloads. Treat that as a vendor’s framing, not a neutral verdict.

The recurring question around io.net is how much of its advertised supply is real. The network has marketed figures in the tens of thousands of connected devices, even six figures in its documentation, but independent tallies of GPUs actually visible on its explorer have at times come in an order of magnitude lower, and the project spent part of 2025 purging duplicate and fake hardware from its counts. IO’s circulating supply has also kept climbing, to roughly 410 million of an 800 million cap by early October, which means token holders face steady emission even when demand is flat. The token traded near $0.17 with a market cap around $71 million, well off its $6.43 high.

Render: vertical integration, from pixels to AI

Render is the odd one out, and the largest by market cap. It did not start as an AI-compute play at all; it grew out of OTOY, the company behind the Octane renderer, as a way to farm out 3D rendering jobs to idle GPUs. That heritage matters, because it gives Render something the others lack: a built-in, paying customer base of artists and studios, with AI and general compute bolted on as an expansion rather than the whole thesis. Governance runs through Render Network Proposals, and the token migrated from Ethereum to Solana. Render also pioneered the Burn-Mint-Equilibrium model that Akash later adopted, so the two networks share a tokenomic lineage even as they chase different workloads.

Render’s market cap pushed back above $1 billion in the October move, trading around $2.01, helped through the year by broader retail access after a Robinhood listing. That scale cuts both ways. It makes RENDER the blue chip of the group, but it also means the token already prices in a great deal of optimism relative to the sector’s modest on-chain revenue, a tension that runs through every name here.

Nosana and Phala: inference-first and TEE-native

Nosana is the specialist. Also on Solana, it began life aimed at continuous-integration pipelines and pivoted to a GPU grid optimized for AI inference, the cheaper, bursty, latency-sensitive workloads that do not need a full training cluster. Its token design is unusually clean for the sector: a fixed 100 million supply, fully in circulation, so there is no emission overhang diluting holders. That clean float is part of why NOS was the sharpest mover in the October rotation, up more than 70% on the week to around $0.72, a reminder that a small, fully-diluted token can swing hardest in both directions.

Phala competes on a different axis: trust. It built its entire thesis on trusted execution environments (TEEs), the on-chip secure enclaves that let a workload run on hardware its owner cannot inspect. For confidential AI, where the data or the model weights are the sensitive asset, that is the whole ballgame, and it makes Phala the most direct rival to Akash’s own confidential-compute push rather than a straight price competitor. PHA is a small-cap, trading in the region of $0.06 to $0.07 for a market cap somewhere around $50 million to $65 million depending on the source, against roughly 857 million tokens circulating of a 1 billion cap, per CoinMarketCap. We will come back to why confidentiality and verifiability are not the same thing.

CoreWeave: the centralized giant in the room

No honest comparison of decentralized GPU networks can skip CoreWeave, because it is the yardstick that exposes just how small the token-based sector still is. CoreWeave is a centralized AI cloud, listed on Nasdaq, that does the same job the DePIN networks aspire to, renting Nvidia GPUs to AI labs, except at a scale that makes the rest look like a rounding error. In its second-quarter 2026 results, CoreWeave reported revenue of about $2.58 billion, up 112% year on year, a backlog near $104 billion, and a footprint of 51 data centers running some 1.5 gigawatts of active power. It guided to $12.4 billion to $13.2 billion of revenue for the full year and planned capital spending of $35 billion to $39 billion.

Put bluntly, CoreWeave books more revenue in a single quarter than the entire decentralized-GPU sector earns on-chain in years. Chief executive Michael Intrator has argued that the binding constraint in AI has shifted from demand to delivery: “The constraint in AI is no longer whether enterprises and AI labs want to deploy. It is how quickly high-performance, reliable AI cloud capacity can be delivered.” That framing is the bull case for every network on this page, and the bear case too. The demand is real and enormous. The question is whether a permissionless marketplace of strangers’ machines can capture a meaningful slice of it, or whether the capital-heavy incumbents simply out-build everyone.

The price of a GPU-hour, side by side

Price is the sector’s headline selling point, so start there. The decentralized networks genuinely do undercut the hyperscalers on a sticker-price basis, though the gap is narrower and noisier than the marketing implies, and comparisons are muddied by availability, cluster size, contract length and reliability. The numbers below are indicative ranges for a single Nvidia H100 by the hour, drawn from the networks’ own quotes and third-party trackers; sources disagree by almost two to one, so treat them as a map, not a quote sheet.

Two things stand out. First, the decentralized discount against on-demand hyperscaler pricing is real, often roughly half to a third of the price for a comparable card. Second, that discount is eroding from both ends. The big clouds have introduced committed and reserved options that close much of the gap for customers willing to sign longer deals, while the DePIN networks are now competing on price with each other, a race to the bottom that squeezes the very providers they need to attract. Cheapness, in other words, is not a durable moat. It is the opening bid.

How much is actually being used?

Price means nothing without utilization, and this is where the sector’s story gets uncomfortable. The most rigorous public data on Akash comes from Messari’s quarterly state-of-network reports, and the first-quarter 2026 edition described a network whose demand metrics were rising while its supply quietly collapsed. New leases grew 27.1% quarter on quarter, to 43,540, which looks healthy. But the count of active providers fell to 58, a record low, GPUs available on the marketplace dropped more than 57% to 334, and utilization sat around 34%. Lease revenue actually fell 45% on the quarter to about $253,000, which annualizes to roughly $1 million. Full-year 2025 revenue was $3.15 million, up 128% from the prior year, impressive growth off a tiny base.

Here is the collision promised earlier. Akash’s own report trumpeted a cumulative $5 million in all-time compute spend, while Messari tracked about $253,000 in a quarter. Those two numbers are not lying to each other, but the gap needs an explanation the project has never fully published. The $5 million is a cumulative, all-time figure spanning every product, including AkashML and managed offerings, while Messari’s number is strictly on-chain lease revenue for a single quarter. It is a scope-and-timeframe difference, not evidence of dishonesty, but the absence of a clean reconciliation is itself the problem. When a network’s headline marketing number and its audited on-chain number differ by more than an order of magnitude and nobody bridges them, investors are left guessing which one to believe.

The pattern repeats across the sector. io.net’s annualized revenue has been estimated in the low tens of millions at most, and credible independent figures for Nosana and Phala are smaller still. Taken together, third-party research has pegged the entire decentralized-GPU sector’s annualized on-chain revenue in the region of $180 million to $220 million, a number CoreWeave eclipses in roughly a week. The demand for AI compute is real. The share these token networks capture on-chain, and can prove, is still tiny.

Does the token capture the value?

Even if you believe the demand is coming, a separate question decides whether the tokens are worth owning: does usage actually flow back to holders? This is where the sector’s designs diverge most, and where Akash in particular invites scrutiny. In March 2026 Akash activated Burn-Mint-Equilibrium through a near-unanimous governance vote. In theory, tenants burn AKT to buy a USD-pegged settlement credit, providers are paid out at settlement, and the token sits in the middle as the medium of exchange. It sounds deflationary. In practice, the permanent on-chain burn is close to zero, because the tokens are not destroyed so much as cycled through a vault that backs the settlement credit; net destruction only happens if AKT appreciates between the moment it is minted and the moment it is spent.

That nuance is easy to miss and important to grasp. Akash has never published the net AKT actually burned by BME, which is the one figure that would tell holders whether the mechanism removes supply or merely recirculates it. The analyst Zoha Imdad Ali put the critique sharply in a TECHi piece titled, in effect, that Akash runs real AI compute but AKT cannot prove it captures the value, with the central complaint being exactly that the net burn goes unreported. Layer on the token’s monetary policy, inflation capped at 8% a year with half of issuance routed to a community pool rather than to stakers, and the nominal staking yield of around 7% translates into a slightly negative real yield. Holders who stake are, in effect, running to stand still, an echo of how restaking’s premium melted once a risk-free rate was on offer.

The through-line is that none of these tokens has yet demonstrated a tight, provable link between network usage and token value. Nosana’s fixed supply is the cleanest design on paper; Render’s is the most battle-tested; Akash’s is the most ambitious and the least transparent. That matters because in a sector-wide rally like October’s, all of them move together on narrative, as the market snapshot showed, which tells you the buying is driven by sector beta, not by any single network proving it monetizes compute better than the rest.

Trust is the real moat: confidential versus verifiable

Price gets a startup in the door; trust is what lets a bank or a hospital actually deploy. When you rent a GPU from an anonymous provider, you are handing a stranger your data and, often, your model weights, and asking them not to copy, leak or tamper with either. For a weekend hobby project that is fine. For regulated or commercially sensitive work, it is a dealbreaker, and it is a close cousin of the broader question of who you are really trusting once the keys, or the hardware, are out of your hands.

Akash’s answer landed in July 2026 with confidential compute, built on trusted execution environments. In a launch post, Osuri described the user experience as almost trivial: “One line of SDL, and your workload runs where the machine’s owner can’t see it.” There are no new SDKs or rebuilt images; a tenant adds a single parameter and the workload runs inside an encrypted enclave. As Osuri put it, “memory encryption is managed directly by the CPU, rendering the data inaccessible to the host operating system, the hypervisor, and anyone with physical access to the infrastructure.” Under the hood it uses AMD or Intel secure enclaves on the CPU paired with Nvidia’s confidential-computing mode on the GPU, with an attestation step so the tenant can verify the enclave is genuine before sending any secrets.

This is a genuine advance, and it is where Phala, which built its whole network on TEEs, meets Akash head on. But it is critical to understand what confidential compute does not do. It protects confidentiality and attests that a specific image is running; it does not prove the output is correct. A provider using a TEE cannot read your data, yet nothing in the enclave stops them from running a cheaper or different model and returning plausible garbage. Proving correctness is the domain of verifiable compute, zkML and optimistic schemes, which remain far heavier, often hundreds to thousands of times the cost of running the model normally. TEEs also root their trust in Intel, AMD and Nvidia, and have a documented history of side-channel attacks. Confidential is not the same as verifiable, and buyers should not conflate them.

The chain underneath: Akash leaves Cosmos

The biggest structural story at Akash has nothing to do with GPUs directly. The network is leaving the sovereign Cosmos SDK chain it has run on since launch. The formal proposal, AEP-79, calls for migrating to a shared-security model, where another, larger Layer 1 provides the security Akash currently pays for by locking up staked AKT. The roadmap entry frames this as a capital-efficiency fix: instead of immobilizing hundreds of millions of dollars of AKT to secure the chain, Akash would buy security on a pay-per-use basis and redirect its energy to the product. The evaluation has run across roughly fifteen ecosystems, with Solana repeatedly cited as a strong contender, and IBC interoperability is a stated requirement. As of early October no chain has been chosen, and the target for completion is the end of 2026.

The move accelerated after a licensing dispute. In April 2026, the Cosmos SDK’s enterprise module shifted from the permissive Apache 2.0 license to a source-available license that barred commercial use without a separate deal. Osuri called the change “hostile” on X, saying it would prevent Akash from deploying the component in production or offering it as a service. Whatever chain Akash lands on, the episode underlines a risk that is easy to overlook when you buy a compute token: you are also betting on the long-term viability of the base layer beneath it, and that base layer can change the rules.

Who will actually supply the GPUs?

The record-low provider count is the quiet crisis under every Akash bull case. Demand can rise all it likes, but if the supply of machines keeps shrinking there is nothing to rent. Part of the cause is friction: running an Akash provider has historically meant operating Kubernetes, a real barrier for the small operators and gaming-rig owners the network theoretically wants. And the reverse auction that gives tenants such good prices also pushes providers toward a race to the bottom, where thin margins make it hard to justify buying new cards.

Akash is attacking the problem from both ends. Top-down, there is StarCluster, a protocol-owned GPU mesh whose first phase is being financed through Starbonds, a regulated US investment instrument sold in $1,000 units with an offering cap reported up to $75 million, to fund on the order of 7,200 Nvidia GB200 Blackwell GPUs operated by vetted enterprise operators it calls Nodekeepers. Bottom-up, there is Homenode, which reached early access in February 2026 and lets owners of consumer cards such as the RTX 4090 and 5090 become providers without touching Kubernetes. A separate proposal, AEP-53, would pay on-chain incentives in AKT to providers who bring specific hardware online, which helps supply but adds dilution. The strategic bet is a pincer: enterprise-grade capacity from the top, long-tail capacity from the bottom.

This is also where the centralized comparison bites hardest. CoreWeave is pouring tens of billions into data centers and power contracts precisely because the binding constraint in AI is increasingly megawatts, not chips, and the same miners who once chased Bitcoin hashrate are now repurposing their power deals for AI. A marketplace of volunteers is a clever way to aggregate idle capacity, but it does not, by itself, build gigawatts. Whether Akash’s pincer can out-pace a shrinking provider base is the single most important operational question facing the network into 2027.

What the SEC sees

For a US investor, the regulatory picture around this sector is less dramatic than the headlines suggest, but it is not empty. AKT itself reads as an infrastructural utility token: its primary job is to pay for and meter compute, not to represent a claim on a company’s profits, which keeps it well away from the center of the Securities and Exchange Commission’s attention. The SEC has not brought an action against Akash, and the asset looks more like a commodity-style network token than an investment contract, the same commodity-versus-security line the agency has been drawing around the crypto ETFs it has let through, with all their custody and tax fine print.

The more interesting detail is that Akash itself seems to understand the line. When it needed to raise real capital for StarCluster, it did not issue more tokens and hope; it structured Starbonds as a regulated US investment instrument, the compliant path for something that genuinely is a security. That split, a utility token for the network and a regulated security for the capital raise, is a sign of a project trying to stay on the right side of the rules rather than paper over them. Confidential compute adds a second regulatory angle: by keeping data invisible even to the machine’s owner, it opens the door to workloads bound by privacy and data-handling obligations that would otherwise never touch a decentralized network. The compliance questions there are about data protection, not securities law, but they are exactly the questions that gate enterprise budgets.

The scorecard, and what to watch into year-end

Put it all together and there is no single winner, because the networks are not really competing for the same job. Each owns a different corner of the same market. The scorecard below is a way to hold the whole field in your head at once.

The October rally does not settle any of this. When Nosana, AKT, IO and Render all jump together in a week, the signal is sector beta, money rotating into an AI-compute theme, not a verdict that any one network cracked the code on monetizing GPUs. The durable advantages are narrower and less exciting than a green candle: Akash’s reverse auction and new confidentiality layer, io.net’s clustering, Render’s paying creative base, Nosana’s clean supply, Phala’s head start on TEEs. The cheap-GPU story that launched the sector is real but fading as both the hyperscalers and the networks themselves compete the discount away.

Into year-end, a handful of concrete events will matter more than price. Akash is due to name the chain it migrates to under AEP-79, with completion targeted for December. Token2049 in Singapore opens on 7 October, the kind of stage where these projects tend to make announcements. Messari’s second- and third-quarter reports, still unpublished as of early October, should finally show whether the provider base recovered or kept shrinking. And a hard number nobody has disclosed, how many providers actually offer confidential compute, would tell you whether Akash’s enterprise pitch is real or aspirational. For a token buyer, the single thing to track is the gap between usage and value-capture; until a network closes it, rallies like this one will keep being about rotation, not fundamentals.

Frequently Asked Questions

Is Akash cheaper than AWS for renting GPUs?

Usually yes on sticker price. Indicative rates for a single Nvidia H100 on Akash run roughly $1.30 to $2.00 an hour, against about $3.90 to $7.00 for AWS on-demand, so the headline discount is real. But the gap narrows once you factor in reserved and committed cloud deals, reliability, and the effort of managing a decentralized provider, so the saving is smaller in practice than the raw numbers suggest.

Akash vs io.net, which is better for AI workloads?

It depends on the job. Akash is the more flexible general-purpose marketplace and now offers confidential compute, while io.net is built to stitch many GPUs into large clusters for training and heavier machine-learning work. io.net markets a larger GPU supply, though independent counts have often come in well below its advertised figures, so verify availability for your specific model before committing.

Does burning AKT make it deflationary?

Not in the way it sounds. Akash’s Burn-Mint-Equilibrium cycles AKT through a vault that backs a USD-pegged settlement credit rather than permanently destroying it, so the net burn is close to zero unless the token appreciates between mint and settlement. Akash has not published the net AKT actually burned, which is the figure that would settle the question, so treat any deflation claim with caution.

Is Akash really leaving the Cosmos blockchain?

Yes, that is the plan. Under proposal AEP-79, Akash intends to deprecate its own Cosmos SDK chain and move to a shared-security model on a larger Layer 1, with Solana among the contenders and completion targeted for the end of 2026. No destination chain had been chosen as of early October 2026, and the project says it will keep IBC interoperability.

Can you run private or confidential AI workloads on Akash?

Yes, since July 2026. Akash’s confidential compute uses trusted execution environments so a workload runs where the provider cannot see the data, enabled by adding a single line to the deployment file. Just remember that confidential is not the same as verifiable: the enclave hides your data but does not by itself prove the provider ran the correct model.

By Marcus Okafor, HOGE Wire. Market data as of 5 October 2026; nothing here is investment advice.