Akash Network Explained: Crypto’s GPU Marketplace Bet

What Is Akash Network?

Akash Network is a decentralized marketplace for cloud computing, built on its own blockchain rather than running on top of Amazon Web Services, Google Cloud or Microsoft Azure. Anyone with spare server capacity, whether a data center operator, a crypto mining outfit repurposing hardware, or someone with a single gaming PC, can list it for rent, and anyone who needs compute can bid for it in an open, reverse-auction market. The project describes itself as a ‘supercloud,’ a layer that lets buyers and sellers of computing power transact directly, with price set by competition among providers rather than a single company’s rate card.

The network’s native token, AKT, is used to pay for leases, to stake and help secure the underlying Cosmos SDK blockchain, and to vote on protocol upgrades. As of mid-July 2026, AKT traded around $0.57, giving the token a market capitalization of roughly $167 million, according to CoinGecko, a small fraction of its April 2021 all-time high of $8.07.

Akash launched as a general-purpose compute marketplace in 2020, years before generative AI turned GPU scarcity into a mainstream story. Its pivot toward GPUs in 2023 made it one of the more closely watched names in the decentralized physical infrastructure network category, alongside projects like Render Network and io.net. Three years on from that pivot, the network has enough usage data to move the conversation past narrative and into an honest look at what is actually being rented, by whom, and at what price, which is what the rest of this piece works through.

From Idle Servers to a GPU Marketplace: A Brief History

Akash was founded in March 2018 by Overclock Labs, a software company started by Greg Osuri and Adam Bozanich that had previously built multi-cloud deployment tools. The team raised roughly $1.8 million in seed funding that year to build a blockchain-based alternative to renting compute from a handful of large providers. After several testnets, the Akash mainnet went live in September 2020, initially supporting only CPU, memory and storage workloads, essentially competing with commodity cloud hosting rather than AI infrastructure.

That changed with the Mainnet 6 upgrade, which The Block reported went live on August 31, 2023, adding support for Nvidia H100 and A100 GPUs alongside consumer-grade cards, following a testnet phase that involved roughly 1,300 participants. The timing was not an accident. It landed less than a year after ChatGPT’s public launch triggered a scramble for GPU capacity that left many smaller AI developers unable to get Nvidia chips at any price from the major cloud providers. Akash rebranded the effort ‘Supercloud’ and began marketing directly to AI developers priced out of AWS, Azure and Google Cloud waitlists.

The GPU marketplace has gone through several more upgrades since, most recently Mainnet 18 in June 2026, each aimed either at expanding the pool of usable hardware or at making pricing and availability data more reliable, both covered in more detail further down.

How the Reverse Auction Marketplace Works

Renting compute on Akash starts with a tenant writing a deployment file, in Akash’s own Stack Definition Language, describing what it needs: how many GPUs, of what type, how much memory and storage, and for how long. That request is broadcast to the network, and any provider running the Akash Provider software that can meet the specification may submit a bid. Because bidding runs as a reverse auction, providers compete by offering the lowest price rather than the tenant negotiating off a rate card, and the tenant can accept any bid, not necessarily the cheapest one, weighing factors like a provider’s uptime history or physical region.

Once a bid is accepted, the two sides enter a lease and the tenant’s workload deploys onto the provider’s hardware, typically a Kubernetes cluster the provider has configured with the Akash Provider stack. Most tenants manage this through the Akash Console or the community-built Praetor dashboard rather than the command line. Payment flows for the duration of the lease and, since March 2026, routes through Akash’s Burn Mint Equilibrium mechanism rather than a simple wallet-to-wallet AKT transfer, covered in the tokenomics section below.

The model’s appeal is straightforward. A provider sitting on underused GPUs, whether a data center with excess enterprise capacity or an individual with a gaming rig, can turn that hardware into income without signing a contract with a single buyer, and a tenant can shop a live, competitive market instead of a fixed price list. Akash has claimed its GPU auctions settle 80 to 90 percent below equivalent on-demand AWS pricing, according to research published by Yellow. The tradeoff is that supply is fragmented across many independent operators of varying reliability, rather than the uniform, heavily redundant infrastructure a hyperscaler operates.

Inside the Numbers: Providers, Leases and GPU Utilization in 2026

How much of that marketplace is actually being used is where the picture gets more complicated. Independent tracking from crypto research firm Messari, aggregated by Own Your Mind’s review of the network, shows Akash’s GPU footprint shrinking through late 2025 and into 2026 even as AI compute demand elsewhere kept climbing. Average active providers fell from 69 in the first quarter of 2025 to 63 in the fourth quarter, and to 58, the lowest on record, in the first quarter of 2026.

Akash and Messari have not published the exact methodology behind that 33.7% utilization figure, so it should not be read as simply active GPUs divided by available GPUs; the two rows in the table above come from the same source but do not resolve to the percentage through basic division, a reminder that these metrics are not fully standardized across the sector yet.

Lease activity told a similarly mixed story. Messari counted lease volume up 27.1% quarter over quarter in Q1 2026, but total tracked lease revenue fell 45% over the same period to $253,250, which annualizes to roughly $1 million, a sign the network processed more, smaller or cheaper deployments rather than more valuable ones. A live snapshot of the network in mid-July 2026 showed the marketplace listing around 121 GPU providers and 253 active GPU leases, with roughly 75.5TB of GPU-attached storage available, per Akash’s own network statistics dashboard, though that dashboard uses a different, real-time methodology that is not directly comparable to Messari’s quarterly figures.

AkashML and Akash Chat: Turning Spare GPUs into Inference Endpoints

Rather than leaving every GPU buyer to configure a deployment from scratch, Akash has increasingly built a managed layer on top of the raw marketplace. AkashML, the network’s managed inference product, offers API access to open-weight models including Llama 3.3-70B, DeepSeek V3 and Qwen3-30B-A3B, hosted across a claimed 65-plus data centers, without requiring the buyer to pick a provider or write a deployment file. A companion product, AkashChat, offers a free, login-free chat interface running on the same GPU fleet, storing conversation history in the browser rather than on a server.

Akash’s own Q1 2026 report put AkashML’s OpenRouter-routed traffic at 1.7 billion tokens processed per day, which the company said outpaced Cloudflare’s daily token volume over the same stretch. Growth continued from there. Network-wide AkashML throughput climbed from roughly 5 billion tokens a day in May 2026 to more than 10 billion by early July, hitting repeated highs, according to Own Your Mind’s tracking. Named users include Venice, ElizaOS, Morpheus and Gensyn, all themselves AI-crypto projects running production inference on Akash’s GPUs instead of a hyperscaler.

For a network whose raw GPU lease revenue looks thin next to a company like CoreWeave, discussed below, the inference business is what Akash’s team points to most often as evidence the marketplace has found real usage, since it captures demand that never shows up directly in the lease-auction numbers.

The AKT Token: Staking, Governance and the New Burn Mint Equilibrium

AKT does three jobs. It is the gas token for transactions on the Akash blockchain; it is the asset validators and delegators stake to secure that chain under a delegated proof-of-stake model, in a structure similar to how validators get paid on other Cosmos and proof-of-stake networks; and it is the governance token that decides protocol parameters, including its own inflation rate. Akash has run more than 300 on-chain governance proposals over its history, with seven approved in the first quarter of 2026 alone, requiring a 33.4% quorum to pass, and the network currently counts between roughly 85 and 99 active validators out of 270 registered, per Own Your Mind’s tracking.

That inflation rate is capped at 8% annually under Proposal 283, passed in March 2025, though realized annualized inflation ran slightly hotter than the cap in early 2026: 8.94% in the first quarter versus 8.12% in the last quarter of 2025, per Messari data. Nominal staking rewards sit around 7.3% APY, and because roughly half of new issuance is diverted to the community pool rather than paid out to stakers, the real, inflation-adjusted return for someone staking AKT has recently run negative, an important caveat against any headline staking-yield marketing. Unlike Ethereum, where liquid staking tokens such as stETH let holders stake without giving up liquidity, AKT staking remains mostly the plain, locked variety.

The network’s answer to that dilution problem is Burn Mint Equilibrium, or BME, live since March 23, 2026 after Proposal 318 passed with roughly 99.7% of the vote. Under BME, a tenant paying for compute burns AKT to mint a dollar-pegged settlement credit called ACT, and the provider is paid out in AKT at the prevailing settlement price, tying token burn directly to real marketplace usage instead of a fixed schedule. In its first nine days live, the mechanism burned 53,520 AKT, an average of roughly 5,950 AKT a day, a start that would need to scale considerably to offset an inflation schedule minting far more than that daily, but one that at least links the burn side of the ledger to actual compute spend rather than speculation.

Mainnet 18: Homenode, Resource Reclamation and Lowering the Barrier to Supply

Akash’s June 2026 upgrade, Mainnet 18, activated at block 27,230,465 and introduced what the team calls Oracle v2, a Resource Reclamation feature under proposal AEP-82, and fixes to how the chain handles market-order-close events. Resource Reclamation lets providers automatically recover capacity a tenant leased but stopped using, which matters for a marketplace whose core pitch is accurate, live pricing: idle but technically leased GPUs that never get reclaimed understate real availability and distort the auction.

The bigger supply-side bet is Homenode, a beta program that lets owners of consumer Nvidia cards, including RTX 4090s, RTX 5090s and Quadro RTX 6000 Ada workstation cards, list them on the marketplace without running a full Kubernetes cluster, the technical barrier that had until now limited providers mostly to data-center operators and hobbyists comfortable with container orchestration. It is a direct attempt to widen who can supply the network beyond enterprise hardware.

The strategy traces back to a plan Osuri laid out in February 2025, when Akash was reporting utilization near 70% and roughly $4.3 million in annualized revenue, well ahead of where the Q1 2026 numbers above landed. At the time, Osuri told The Defiant the network was targeting supply from 11,000 professional data centers and as many as 7 million edge locations worldwide, offering incentives such as free H200 GPU edge clusters to providers willing to commit capacity, alongside product additions like agent launchpads and vector database hosting. Homenode and the Mainnet 18 changes are, in effect, the consumer-facing piece of that plan finally shipping, a year and a half later and against a weaker utilization backdrop than Osuri was describing at the time.

Akash vs the Competition

Akash is not the only project trying to build a marketplace for GPUs outside the big three clouds. Render Network started as a marketplace for 3D rendering power and has since expanded into AI inference; io.net aggregates GPU clusters from data centers and individual machines into pools aimed at machine-learning training jobs; and Nosana runs a Solana-based marketplace focused on lower-cost AI inference jobs. Bittensor takes a different approach again, rewarding machine-learning contributions across specialized subnets rather than simply renting hardware by the hour, so it competes with Akash for AI-crypto investor attention more than for the same GPU-rental use case. All of the direct rental competitors price themselves well below list-price hyperscaler rates, and all of their tokens have fallen sharply from their highs.

io.net’s own documentation claims access to more than 100,000 GPU devices across its network, and the company says its monthly active provider addresses grew from around 8,000 in the first quarter of 2025 to more than 45,000 a year later, alongside A100 cluster pricing around $12 to $28 an hour against roughly $32.77 an hour for a comparable AWS instance, according to Yellow’s research cited above. Render’s market cap, tracked on CoinGecko, has been considerably more volatile than Akash’s, sitting above $1.5 billion in early May 2026 before falling to under $800 million by mid-July, a decline broadly in line with the pullback across AI-linked tokens over the same stretch, AKT included. Nosana, smaller still by market cap, said it processed 985,000 jobs in 2024 with $92.4 million worth of NOS staked by node operators.

None of these decentralized marketplaces are close to competing with the centralized ‘neocloud’ tier on raw scale. CoreWeave, the Nvidia-backed cloud provider that listed on Nasdaq in March 2025 at $40 a share, reported first-quarter 2026 revenue of $2.078 billion, up 112% year over year, and has guided toward $12 billion to $13 billion for the full year, commanding an estimated 18% of the dedicated AI training and HPC GPU market. That single company’s quarterly revenue dwarfs the entire decentralized GPU sector’s estimated combined annualized revenue of $180 million to $220 million. The gap says less about whether the decentralized model works and more about how early it still is: the economics of financing and depreciating GPU hardware look a lot like the economics Bitcoin miners have wrestled with for years, where scale, cheap capital and utilization matter as much as the underlying pitch.

The AI Infrastructure Gap Decentralized Compute Is Betting On

The bet underlying Akash and its peers is that AI compute demand will keep outstripping supply for years, and that a meaningful share of that demand will go to whoever offers the cheapest available capacity rather than staying loyal to a single hyperscaler. ‘AI moves in months, energy moves in years,’ Akash founder Greg Osuri said in the network’s Q1 2026 report, pointing to the mismatch between how quickly AI compute demand scales and how slowly the power and data-center capacity needed to serve it can be built. McKinsey estimates cited by Yellow put the global AI infrastructure market on track to exceed $700 billion a year by 2030, and Amazon, Microsoft and Google together are estimated to control around 65% of existing data center GPU capacity, leaving a sizable pool of developers who cannot get allocation from any of the three, or do not want to commit to their pricing and lock-in.

Crypto-focused investors have made a version of this case for a while. Andreessen Horowitz’s crypto arm named DePIN as one of the sectors with the strongest product-market fit signals in its State of Crypto research, arguing that projects like Akash aggregate existing physical assets rather than needing fresh capital formation to grow supply, which partly insulates them from crypto’s boom-bust funding cycles. The Homenode push described above is a direct application of that logic: instead of Akash or its providers buying new GPUs, existing consumer hardware becomes investable supply overnight.

Where that thesis meets resistance is demand quality. A growing share of AI workloads, particularly autonomous agents that plan, transact and call tools on their own, need infrastructure they can trust as much as infrastructure that is merely cheap. Projects like Ritual are building specifically around verifiable AI execution rather than undercutting cloud pricing, a reminder that decentralized compute and trustworthy compute are related but distinct pitches, and Akash has so far competed almost entirely on the first one.

The Transparency Problem: Why Akash’s Own Numbers Don’t Always Match

One recurring criticism of decentralized compute networks generally, and Akash specifically, is that self-reported metrics and independently tracked ones do not always line up. Akash’s official Q1 2026 report described an all-time high of $5 million in quarterly compute spending, while Messari’s independent tracking put lease revenue for the same quarter at $253,250, annualizing to roughly $1 million, a gap of roughly five times between the two figures.

Some of that gap is plausibly a scope difference rather than a discrepancy. Akash’s compute-spend figure likely includes AkashML’s managed inference revenue and other product lines alongside raw GPU and CPU leases, while Messari’s number appears to track lease-marketplace activity specifically. But the network has not published a clear reconciliation showing exactly how the two figures relate, which leaves outside observers to guess. That matters more for a project whose pitch rests partly on being a transparent, auditable alternative to opaque centralized billing; if the marketplace layer itself is hard to measure consistently, it weakens part of the argument for using it over a hyperscaler.

The context that should not get lost is that Akash’s full-year 2025 revenue, on Messari’s own accounting, grew 128% year over year to roughly $3.15 million, so the criticism is about measurement clarity and the shape of the growth curve, a strong 2025 followed by a soft first quarter of 2026, rather than a case of a network manufacturing activity from nothing.

Security and Trust: What a GPU Marketplace Does and Doesn’t Verify

It is worth being precise about what Akash’s marketplace actually guarantees. A lease confirms that a provider has committed specific hardware for a specific period at an agreed price, enforced through escrow-style payment and a reputation system that penalizes providers for downtime or non-delivery. What it does not do is cryptographically prove that a given job, an AI inference request for example, ran on the hardware promised, used the model weights the tenant expected, or returned an untampered result. That is a trust assumption, not a mathematical guarantee, similar to the trust assumption DeFi protocols make about their price oracles rather than verifying every feed cryptographically.

The technical term for closing that gap is verifiable compute, and it is a genuinely different engineering problem from renting hardware cheaply. Approaches include zero-knowledge proofs of correct execution, trusted execution environments that rely on hardware attestation, and crypto-economic staking-and-slashing schemes that make dishonesty expensive rather than impossible. None of these are mature enough yet to verify a large language model’s output cheaply at scale, which is part of why marketplaces like Akash and verification-focused networks have, so far, stayed on separate tracks rather than merging into one product.

For a tenant running a sensitive workload, that distinction is practical rather than academic. Akash is a reasonable place to shop for cheap, roughly commodity compute where the cost of an occasional bad or dishonest run is low, and a harder sell for workloads where provable correctness matters as much as price.

Regulatory Status: Where AKT Stands with the SEC in 2026

Infrastructure tokens like AKT, ones that pay for a real, live service rather than representing a claim on a company’s profits, have generally had an easier regulatory path in the US than tokens sold explicitly as investments. That distinction got more formal backing in March 2026, when the SEC and CFTC issued a joint interpretive release establishing a five-category taxonomy for digital assets, covering digital commodities, collectibles, tools, stablecoins and securities, explicitly stating that not every token is a security.

The SEC added three crypto-specific items to its 2026 rulemaking agenda on July 7, covering a pathway for crypto asset offerings, broker-dealer amendments for crypto custody and trading, and market-structure rules for digital-asset trading venues, all still at the proposal stage, according to Crypto Briefing’s coverage of the agenda. SEC Chair Paul Atkins framed the push as an effort to make the US the ‘crypto capital of the world’ with ‘clear rules of the road for capital raising with crypto assets.’

None of this amounts to a specific ruling on AKT. But the general framework matters for how a token like it gets evaluated. Under the Howey test the SEC has long used to identify securities, a token that functions primarily as a payment and access mechanism for a live, decentralized network sits further from an investment contract than a token whose only real use is speculation. That does not make AKT immune from scrutiny, particularly given how much of its value has historically traded on AI-narrative momentum rather than marketplace fundamentals, but it does mean the 2026 taxonomy shift is a meaningfully friendlier backdrop than the enforcement-heavy years earlier in the decade.

Risks Facing Akash Network

Akash’s pitch, cheap, permissionless GPU access, is genuinely differentiated, but the 2026 data points to several concrete risks worth weighing before treating the network as a settled AI-infrastructure winner.

• Utilization and demand risk: Q1 2026 GPU utilization of 33.7% and a 45% quarter-over-quarter drop in Messari-tracked lease revenue suggest the marketplace still has far more supply than paying demand most of the time.
• Competitive risk from both directions: centralized players like CoreWeave are scaling revenue in the billions per quarter with access to far cheaper capital, while larger-fleet decentralized rivals like io.net and better-capitalized ones like Render compete for the same AI-developer audience.
• Dilution risk: at 8.94% realized annualized inflation against a BME burn rate that started at under 6,000 AKT a day, token supply growth still outpaces burn most days, meaning holders and even stakers can face negative real yield.
• Measurement risk: the gap between Akash’s self-reported compute spend and independently tracked lease revenue makes it hard for outsiders to verify growth claims precisely, a problem for a project whose value proposition includes transparency.
• Verification risk: the marketplace confirms hardware and uptime, not that a given AI job executed correctly, which limits its fit for workloads where provable correctness matters as much as price.
• Regulatory risk: the 2026 SEC-CFTC taxonomy is a friendlier backdrop than prior years, but it is new, largely untested through enforcement, and could still evolve in ways that affect how infrastructure tokens are treated.

Frequently Asked Questions

What is Akash Network?

Akash Network is a decentralized, open-source marketplace for cloud computing built on its own blockchain. Instead of renting server capacity from a single company such as AWS or Google Cloud, tenants request CPU, storage or GPU resources and independent providers bid to fulfil that request in a reverse auction, with payment settled in Akash’s native AKT token. Since 2023 the network has focused heavily on renting out Nvidia GPUs for AI training and inference, positioning itself as a cheaper, decentralized alternative to hyperscaler cloud compute.

How does Akash Network’s GPU marketplace work?

A tenant describes the compute it needs, including GPU type, quantity and duration, in a deployment file, and independent providers running Akash’s provider software bid to fill that lease, usually with the lowest qualifying bid winning. Providers historically needed a Kubernetes cluster to participate, but Akash’s Homenode program, launched in 2026, lets owners of consumer GPUs such as RTX 4090s and RTX 5090s join the marketplace without that technical overhead.

What is the AKT token used for?

AKT pays transaction fees on the Akash blockchain, is staked by validators and delegators to secure the network under a delegated proof-of-stake model, and is used to vote on governance proposals, including the network’s own inflation rate. Since March 2026, AKT is also burned through a Burn Mint Equilibrium mechanism whenever a tenant pays for compute, linking token burn directly to marketplace usage.

Is Akash Network cheaper than AWS or Google Cloud?

Akash and similar decentralized GPU marketplaces have reported cost savings in the range of 60 to 90 percent versus on-demand hyperscaler pricing for comparable hardware, according to third-party research. Those figures reflect competitive auction pricing rather than a guarantee of equivalent uptime, support or service-level agreements, and 2026 usage data shows actual GPU utilization on Akash’s marketplace remains well below hyperscaler-level capacity usage.

Is Akash Network’s compute verifiable, or could a provider cheat?

Akash’s marketplace guarantees that a provider has committed specific hardware and enforces uptime through payment and reputation mechanisms, but it does not cryptographically prove that a given job executed correctly or used an untampered model. Verifying execution itself is a separate problem, called verifiable compute, that other projects address using techniques such as zero-knowledge proofs and trusted execution environments.

Written by the HOGE Wire research team.