Akash Network in 2026: The GPU Market That Left Cosmos

In the second quarter of 2026, CoreWeave, the Nasdaq-listed AI cloud company, booked $2.58 billion in revenue and sat on a contracted backlog of roughly $104 billion, figures it disclosed on 11 August 2026 as demand for Nvidia clusters kept outrunning supply (CNBC). In the same period, Akash Network, the best-known decentralized answer to that build-out, recorded about $253,000 in on-chain GPU lease revenue for the entire first quarter, according to Messari’s State of Akash Q1 2026. One company turns over more in an afternoon than the open marketplace did in three months.

That gap frames everything Akash is trying to do this year, and it explains the strangest move in the project’s history: it is preparing to abandon the blockchain it spent five years building. The founder, Greg Osuri, has put Akash’s sovereign Cosmos chain up for deprecation and launched a public search for a new home, with Solana named as a strong contender. It is a rare thing to watch a live network shop for a new base layer while a real product runs on top of the old one.

This piece looks at what Akash actually is, why it is leaving Cosmos, what the numbers say about real demand, how cheap its GPUs really are against Amazon and Microsoft, how the AKT token and its new burn mechanism work, and where the whole decentralized-compute trade sits against a centralized rival that grows by a CoreWeave every quarter. The short version: the product is maturing, the discount is real, and the two open questions, honest usage and honest computation, are still unanswered.

What Akash Is: A Reverse Auction for GPU Time

Akash Network is a permissionless marketplace for cloud compute. Overclock Labs, founded in 2018 by Osuri and Adam Bozanich, launched the mainnet in September 2020 for CPU, storage and memory, then bolted on the piece that mattered for the AI cycle: a GPU marketplace, shipped through the Mainnet 6 upgrade nicknamed Supercloud on 31 August 2023. That release let providers list Nvidia hardware, from H100 and A100 data-center cards down to consumer RTX GPUs, and rent it to anyone who could write a deployment file.

The mechanic that makes Akash distinctive is a reverse auction. A tenant writes a manifest, a short YAML file called an SDL that describes the CPU, GPU, memory and storage a workload needs. Providers see the request and bid, competing to offer the lowest price. The tenant is not forced to take the cheapest bid; it can weight a provider’s reputation, region or uptime, then open a lease and settle in AKT, the network’s token. Tools such as the Akash Console handle the deployment so a user does not have to touch the chain directly. In practice it behaves like a spot market for idle silicon, and the pitch is simple: the world is full of underused GPUs, and a market can price them far below what a hyperscaler charges for the same card.

Where centralized clouds sell reliability, service agreements and a support desk, Akash sells price and neutrality. There is no know-your-customer gate to rent a GPU, no account manager, and no single company that can revoke access. That openness is the product’s appeal and, as later sections show, the source of its hardest problems.

The 2026 Headline: Akash Is Leaving Its Own Blockchain

In October 2025, Osuri announced that Akash would deprecate its sovereign Cosmos-SDK chain and migrate to a shared-security network, an evaluation The Block reported at the time as a search across candidates inside and outside the Cosmos ecosystem (The Block). The formal vehicle is a request-for-proposals called AEP-79, Akash on Shared Security, published on the project’s roadmap and framed as a comparison of the security, community, liquidity and growth on offer from established Layer 1s (akash.network).

Akash has described the exercise as one of the most comprehensive evaluations any DePIN project has run, spanning roughly fifteen ecosystems whose foundations reached out after the announcement. Solana is the most-cited option, and Osuri has called it a strong contender while promising to be, in his words, extra prudent about the destination. Whatever chain wins, Akash says it will keep Inter-Blockchain Communication support so it stays reachable from the Cosmos apps it grew up with. The roadmap slots the transition into the fourth quarter of 2026, and as of late August no chain has been chosen and no binding governance vote has been held.

For a network that markets decentralization, walking away from your own validator set is a loaded decision. It concedes that running a standalone chain is a cost, not a moat, and it hands a chunk of Akash’s security and token narrative to whichever L1 it lands on. The upside Akash is chasing is focus: stop paying to secure a blockchain and spend that energy on the compute product instead.

Why Leave Cosmos: A License Fight and the Cost of Staking

Two forces pushed Akash toward the exit. The first is economic. Securing a sovereign proof-of-stake chain means locking large amounts of AKT in staking to deter attacks, capital that sits idle rather than doing productive work. Osuri’s argument is that this is capital-inefficient by design, and that a pay-per-use security model, buying protection from an established L1 only when needed, would free that value and cut operational overhead. The idea rhymes with the pooled-economic-security thesis behind Ethereum restaking, where one pool of stake secures many services; readers who want that mechanism in depth can see our explainer on restaking in 2026 and the great unwinding. Akash is essentially asking whether it can rent security the way its own users rent GPUs.

The second force is a licensing dispute that turned a slow strategic question into an urgent one. In 2026 the Cosmos SDK’s Enterprise module changed its license from the permissive Apache 2.0 to a Source Available Evaluation License that permits testing and study but requires a separate commercial deal for production use. Osuri said the change would stop Akash from deploying the component in production, using it commercially, or offering it as a service to third parties, and reported outlets covered the shift and his response (DePIN Scan). His public statement was blunt about the stakes: he wrote that Cosmos leadership had changed the license of a critical component in a way that would prevent Akash from deploying it in production, using it commercially in any way, or offering it as a service (Greg Osuri on X).

Shared security is not a free lunch, and the tradeoff is the heart of the debate. Under the model Akash is chasing, a larger Layer 1 provides the economic weight that makes attacking the chain expensive, and Akash pays for that protection rather than bootstrapping it from its own token. The upside is obvious: a small network inherits the security budget of a large one instead of trying to match it. The cost is dependence. Akash would tie its liveness, its upgrade cadence and part of its fate to a host chain it does not control, and a serious problem on that base layer becomes Akash’s problem too. For a project whose pitch is neutrality, swapping self-reliance for a landlord is a real philosophical concession, however sound the economics look on a spreadsheet.

That leaves staking in limbo. Osuri has said that staking, rather than vanishing, is evolving into something superior, hinting at designs beyond pure consensus security, but no technical specification exists yet. For AKT holders who stake for yield, that is the uncomfortable part of the migration: the reward that pays them today is tied to a consensus job the network intends to outsource.

The Numbers: What Akash Actually Rented in Q1 2026

Marketing aside, the cleanest read on real demand comes from Messari’s quarterly report. The first quarter of 2026 shows a marketplace that is being used more often but paid less, and served by a shrinking base of providers.

The tension is right there in the table. New leases rose more than 27% to 43,540, the third straight quarter of growth, yet lease revenue fell about 45% to roughly $253,000, according to Messari. More deployments paying less money is what you get when the price of renting a GPU falls faster than volume rises, and when a lot of new activity is small, cheap or short-lived. Provider count told the same story: an average of 58 active providers, the lowest Akash has recorded, with only 84 of 334 available GPUs actually rented at any time. For the full year 2025, on-chain lease revenue came to about $3.15 million, up 128% year over year, real growth off a tiny base.

None of this is fatal, but it punctures the tidy narrative that AI demand is flooding into decentralized compute. Demand is arriving in leases, not in dollars, and the supply side is consolidating rather than expanding.

The Transparency Gap: $5 Million or $253,000?

Here is where an investor has to read carefully. Akash’s own first-quarter report claimed roughly $5 million in compute spend and pointed to inference traffic and managed-service momentum as evidence of demand (Akash Network). Messari’s independently tracked on-chain lease revenue for the same quarter was about $253,000. That is close to a twentyfold difference, and it deserves an explanation rather than a headline.

The likely reconciliation is scope. Messari counts settled on-chain leases, the money that actually moves through the protocol. Akash’s larger figure appears to bundle managed and off-chain products, including its hosted inference service, alongside gross activity that never settles as a raw on-chain GPU lease. That is a normal distinction between protocol revenue and total platform spend, and it is not, by itself, evidence of bad faith. The problem is that Akash has not published a public reconciliation that maps one number to the other, so readers are left to trust the framing. Independent commentators have pressed the same point; a mid-2026 piece from Crypto Daily asked directly whether Akash can prove real utilization after the AI rally cooled (Crypto Daily).

The practical rule for anyone valuing AKT: use the audited on-chain number as the floor and treat the platform figure as a ceiling that includes products the token does not directly capture. When a project quotes a revenue number, ask which pocket it lands in.

How Cheap Is Cheap? Akash Against the Hyperscalers

The reason anyone tolerates the rough edges of a decentralized market is price. On the flagship card, the Nvidia H100, the spread between hyperscalers and specialist or decentralized providers is large and consistent across trackers.

The pattern is clear: decentralized and specialist clouds routinely price the H100 well below the major providers, often 40% to 85% cheaper on-demand, per a 2026 pricing comparison from Spheron (Spheron). Two caveats keep this honest. First, Akash’s own pricing pages are rendered client-side and do not expose live figures to simple scrapers, so the Akash numbers here come from third-party trackers and vary by GPU model, region and provider; treat them as indicative rather than a quote. Second, price is not the whole cost. Hyperscaler rates buy service agreements, instant scaling, dedicated networking and a support relationship; a marketplace bid buys a card from a provider you may not know, with reliability that varies. For a training run that must not fail, that difference matters. For a batch inference job, a research experiment or an agent that can retry, the discount usually wins.

Raw hourly price also flatters the marketplace in a way worth spelling out. Large training runs need many GPUs wired together with high-bandwidth interconnect, the kind of tightly coupled InfiniBand fabric that hyperscalers and specialists such as CoreWeave build on purpose. A pool of scattered consumer and small-provider cards cannot easily replicate that, which is why decentralized markets win inference and lose frontier training. Add the softer costs, the engineering time to handle a provider dropping offline, the retries, the lack of a support contract, and the effective discount narrows from the sticker gap. It is still a discount, and for the right workload a large one, but the honest comparison is total cost of ownership, not the price of a single card-hour.

AKT Tokenomics: Inflation, Staking, and Thin Real Yield

AKT is the payment and staking token. As of 24 August 2026 it traded near $0.5523 with a market capitalization around $164 million, ranking about 190th, on roughly 297 million tokens circulating against a max-supply parameter near 388.5 million, according to CoinGecko. That price is about 93% below the April 2021 all-time high of $8.07, a reminder that the token round-tripped an entire hype cycle before the product matured.

The monetary policy is a two-lever system. Issuance rewards stakers and providers, while a community-pool tax skims a share of that issuance into a treasury the network controls. Governance has repeatedly tightened both levers: Proposal 283 in March 2025 cut the inflation ceiling from 13% to 8% and set a 4% floor, and a companion proposal lifted the community-pool tax to 50% (Akash Network). The catch for stakers is arithmetic. When roughly half of new issuance is routed to the community pool rather than to stakers, a nominal staking yield in the 7% range does not keep pace with the dilution the same issuance creates, so real yield sits close to flat or slightly negative. Staking AKT is therefore closer to a governance and alignment decision than a reliable income stream, a distinction worth keeping straight for anyone comparing it with, say, the staking economics covered in our look at where the big exchanges stand on custody and yield in the Coinbase, Binance, Kraken and OKX bank test.

The migration adds a further wrinkle. If Akash outsources consensus to another chain, the staking reward that exists to pay validators has to be rethought from first principles. Osuri insists the design will improve rather than disappear, but until there is a specification, the honest position is that AKT’s yield model is under construction.

Burn-Mint Equilibrium: A Burn That Does Not Burn

The most-discussed 2026 change to AKT is Burn-Mint Equilibrium, or BME, activated on 23 March 2026 through the Mainnet 17 upgrade. The design is meant to smooth the token’s role in settlement. When a tenant pays, AKT is burned in exchange for a USD-pegged settlement credit, and the provider is later paid AKT valued at settlement time. In theory, tying payment to a stable unit shields both sides from AKT’s volatility during a lease, an idea that echoes the stable-settlement logic now spreading across payments; our coverage of the GENIUS stablecoin rulebook reaching offshore traces where that logic is heading in regulated finance.

The nuance that marketing tends to skip is what the burn actually does. In the first days after activation, on the order of 53,520 AKT moved into the BME system (Messari), but this is not a permanent, deflationary destruction of supply in the way a fixed token burn would be. The tokens back the settlement credit; net removal from supply only happens if AKT appreciates between the moment a credit is minted and the moment it settles. If the price is flat or falls, the mechanism recycles rather than shrinks the float. BME is better understood as a settlement and backing layer than as a guaranteed scarcity engine. It is a sensible piece of plumbing, but investors who read the word burn as an automatic supply cut are misreading it.

Homenode: Turning Gaming PCs Into Cloud Providers

To grow supply beyond its shrinking pool of professional providers, Akash opened early access to Homenode on 25 February 2026, a program that lets owners of consumer cards contribute compute without standing up a Kubernetes cluster (PR Newswire). The first phase targets laptops, desktops and edge machines with RTX 4090 or RTX 5090 GPUs, exactly the hardware sitting in millions of gaming rigs and workstations. The vision, which Osuri has pushed since 2025, is to turn the long tail of idle consumer silicon into cloud supply, the way peer-to-peer networks once turned home bandwidth into a service.

The strategic question is whether more supply is the constraint that binds. With utilization near 34%, Akash already has more GPUs listed than tenants renting. Adding hobbyist capacity risks widening that gap unless demand grows to meet it. Homenode makes more sense as a bet on future workloads that suit scattered consumer cards, batch inference, rendering, agent tasks and fine-tuning of small models, than as a fix for today’s revenue problem. It also raises the reliability bar: a data-center provider and a gaming PC under a desk are not equivalent counterparties, and the marketplace’s reputation and matching systems will have to price that difference.

AkashML and the Pivot to Inference

The clearest sign of product-market fit is not raw GPU rental but Akash’s managed inference layer, AkashML, which serves open models such as Llama, DeepSeek and Qwen behind an interface that hides the underlying marketplace. Throughput has climbed fast: from roughly 1.7 billion tokens a day measured on OpenRouter in the first quarter (Akash Network), to about 5 billion a day in May 2026, to more than 10 billion a day by early July (Own Your Mind), with named users including Venice.ai, ElizaOS, Morpheus and Gensyn. Akash also launched an Agents platform on 26 March 2026 for one-click deployment of AI agents, an attempt to convert the surge of interest in autonomous agents into steady compute consumption.

Inference is the right target. Serving a model is bursty, forgiving of a failed request and far more price-sensitive than a mission-critical training run, which is precisely the workload a marketplace can win. But there is a strategic irony worth naming. The more Akash leans on managed products such as AkashML, the more it looks like a thin cloud provider that happens to source hardware from a marketplace, and the further real revenue drifts from the on-chain leases that AKT captures. That is the same wedge behind the transparency gap: growth in tokens served does not automatically become growth in protocol revenue.

The aggregation layer matters as much as the raw capacity. Much of AkashML’s measured traffic reaches it through OpenRouter, a router that sends a developer’s request to whichever backend is cheapest or fastest, which means Akash competes for that flow on price and latency against both centralized inference providers and other crypto networks. Winning there is a genuine signal of competitiveness, since a router does not care about a project’s token or its decentralization story, only its numbers. It also means the demand is only as loyal as the next price quote. That is a healthier kind of usage than incentivized activity, but it is mercenary by design, and it will follow a cheaper venue the moment one appears.

The Field: Render, io.net, Nosana, and the CoreWeave Yardstick

Akash is one of several tokens competing to sell decentralized GPU time. The prices below, pulled on 24 August 2026, show a sector that has kept building products while its tokens sit far below their 2024 peaks.

Render, the largest of the group by market value at about $765 million, comes from GPU rendering and has leaned into AI workloads (CoinGecko). io.net aggregates GPUs into clusters and markets a large registered device count, though its active-device figures are far smaller than its headline; its token trades near $0.1413 for a market cap around $54 million (CoinGecko). Nosana, focused on Solana-native inference, is smaller still at roughly $27 million (CoinGecko). Notably, Akash’s migration would put it on the same base layer, most likely Solana, as three of its rivals, turning a cross-chain contest into a same-neighborhood one.

Then there is the yardstick. Add the four leading decentralized-GPU tokens together and you get a combined market capitalization of roughly $1 billion. CoreWeave books more than twice that in revenue in a single quarter, guided full-year 2026 revenue of $12.4 billion to $13.2 billion, and signed a $21 billion capacity agreement with Meta running through 2032 (Fortune). CoreWeave’s chief executive, Michael Intrator, has framed the moment bluntly on the company’s earnings calls: the constraint in AI, he argues, is no longer whether enterprises and AI labs want to deploy, but how quickly high-performance, reliable AI cloud capacity can be delivered. Decentralized markets are not competing for that top-end, contract-and-guarantee demand. They are competing for the price-sensitive tail underneath it, which is a real market, just a far smaller one than the token valuations once implied.

Rented Compute Is Not Verified Compute

There is a technical limit worth stating plainly, because it shapes what Akash can safely be used for. Akash rents you a GPU; it does not prove what that GPU did. When a provider returns the output of an inference job, nothing in the base protocol cryptographically guarantees that the provider ran the model you asked for, at the precision you specified, rather than a cheaper substitute that produces plausible-looking results. A tenant can benchmark and spot-check, but that is trust and reputation, not proof.

This is the line between a compute marketplace and verifiable compute, and it is not a detail Akash has solved. Vitalik Buterin, in his widely cited essay on the intersection of crypto and AI, argued that the most robust use of crypto in AI is verifiability, and flagged the awkward tradeoff between cheap unverified computation and expensive verified computation (vitalik.eth.limo). Akash sits firmly on the cheap, unverified side of that tradeoff, which is fine for workloads where a wrong answer is cheap to catch and retry, and a real risk for anything where correctness must be guaranteed. Projects trying to close that gap, from optimistic and zero-knowledge machine-learning schemes to trusted-execution approaches, are a separate design space; our look at whether you can tokenize an AI model inside opML’s real test covers what it takes to make a computation provable rather than merely cheap. Until Akash adopts something like it, buyers should treat the network as low-cost infrastructure, not a source of trustless answers.

A concrete example makes the risk tangible. Suppose you pay to run a 70-billion-parameter model at full precision, and a provider quietly serves a heavily quantized version that is far cheaper to run. The outputs will look reasonable on most prompts, the provider pockets the difference, and you may never notice unless you built your own checks. That is not a hypothetical unique to Akash; it is the default failure mode of any market that pays for compute without proving it. Reputation systems, staking-and-slashing and sampling audits all help at the margin, but none turn an unverified answer into a verified one. For agents that transact or code that ships, that distinction is the difference between a cost saving and a silent liability.

Where US Regulators Stand

For a US reader, the regulatory question is whether AKT looks like a security. The Securities and Exchange Commission still applies the investment-contract test from the Howey framework it summarizes on its own site (sec.gov), which turns on whether buyers expect profit primarily from the efforts of others. AKT complicates that story because it is, first and foremost, a payment token: it is the unit you burn to rent a GPU, a real metered service with a real cost. Tokens that function as consumptive fuel for a working network sit further from the classic investment-contract cases than tokens sold mainly as bets on a team’s future work.

Through 2026, US market-structure policy has moved toward drawing a clearer security-versus-commodity line, and infrastructure tokens tied to a delivered service have generally been treated as closer to commodities than securities, though nothing is settled and enforcement posture can shift. The migration adds an interesting angle rather than a legal one: moving AKT’s settlement to another chain, such as Solana, does not change what the token does, so it should not change the analysis. The staking overhaul is the part regulators may watch, since a redesigned reward that pays holders for less active participation could be read differently from fees earned for validating a chain. For now, AKT reads as a utility and payment asset, not an obvious security, but holders should treat that as the current weather, not a permanent climate.

What to Watch Through the Rest of 2026

Four questions will decide whether Akash’s year is remembered as a pivot or a plateau. First, the chain choice: which Layer 1 wins the AEP-79 process, and whether the community ratifies it without a fight. A clean, well-argued selection would be a credibility win; a contested one would stall everything downstream. Second, whether the staking redesign arrives with a real specification rather than reassurance, because a lot of AKT is staked on the assumption that the reward survives the migration.

Third, the numbers. The next Messari report should show whether the first quarter’s pattern, more leases but less revenue, was a blip or a trend, and whether provider count stops falling. If demand for AkashML inference keeps compounding and some of it lands as on-chain revenue rather than off-chain platform spend, the transparency gap narrows on its own. Fourth, verification: if Akash or its ecosystem ships a credible way to prove computation, it graduates from a cheap-compute story to a trust story, which is a much larger market. The physical economics are the quiet backdrop to all of it. As Osuri put it in the Q1 report, AI moves in months while energy moves in years, the same hard constraint that governs Bitcoin’s power-hungry build-out, which we track in our coverage of the hashrate climb toward two zettahashes. Cheap idle GPUs are only cheap while someone else is paying to keep them powered and cooled.

The honest summary is that Akash in 2026 is a better product on a shakier foundation than a year ago. The reverse auction works, the discount against hyperscalers is real, inference volume is climbing, and the team is willing to make hard structural changes. Against that, real on-chain revenue is small, the token has been diluted through an entire cycle, the flagship burn does not reliably burn, and the network is mid-flight between blockchains. It is a credible bet on a real slice of the AI build-out, priced and reported with enough ambiguity that it demands a skeptical reader.

Frequently Asked Questions

What is Akash Network and how does it work?

Akash is a permissionless marketplace for cloud compute, best known for renting Nvidia GPUs. A tenant describes the hardware they need in a deployment file, providers bid to supply it in a reverse auction, and the tenant picks a bid and settles the lease in the AKT token. It launched its mainnet in 2020 and added GPU rentals in August 2023.

Why is Akash Network leaving Cosmos?

Akash argues that running its own Cosmos-based chain is capital-inefficient, since large amounts of AKT must be locked in staking to secure it, and it wants a pay-per-use shared-security model instead. A 2026 change to the Cosmos SDK Enterprise module’s license, which restricted commercial use, added urgency. Through the AEP-79 process, Akash is evaluating roughly fifteen Layer 1s, with Solana a leading candidate, and targets the fourth quarter of 2026.

Is Akash cheaper than AWS for GPUs?

Yes, usually by a wide margin on raw hourly price. On the Nvidia H100, decentralized providers including Akash often price in the range of $1 to $3.50 per hour against roughly $7 on AWS and $11 to $12 on Azure and Google Cloud. The tradeoff is that hyperscaler rates include service agreements, instant scaling and support, while a marketplace bid buys a card from a provider whose reliability can vary.

What is Burn-Mint Equilibrium on Akash?

BME, activated in March 2026, is a settlement mechanism: a tenant burns AKT for a USD-pegged credit, and the provider is paid AKT valued at settlement. It stabilizes payment against AKT’s volatility, but it is not a guaranteed deflationary burn. Tokens back the credit rather than being permanently destroyed, and net supply only shrinks if AKT appreciates between minting and settlement.

How much is Akash Network actually being used?

By the independently tracked on-chain measure, modestly. Messari recorded about $253,000 in GPU lease revenue in the first quarter of 2026 on 43,540 leases, with GPU utilization near 34%. Akash’s own report cited roughly $5 million in compute spend for the quarter, a much larger figure that appears to include off-chain and managed products. The audited on-chain number is the safer floor for valuing the AKT token.

By Marcus Okafor, senior markets writer at HOGE Wire, covering crypto infrastructure and the AI-compute economy.