Akash in 2026: Real AI Demand, a Shrinking GPU Supply

Michael Intrator, the chief executive of CoreWeave, keeps returning to one line about the artificial-intelligence build-out: the constraint is no longer whether enterprises and AI labs want to deploy, it is how quickly reliable capacity can be delivered. Decentralized-compute projects have been quoting a version of it ever since. For Akash Network, the permissionless GPU marketplace that has spent two years selling itself as the cheaper, decentralized answer to exactly that capacity crunch, the line cuts in two directions at once. By its own numbers, demand for Akash compute has never been higher. By an independent analyst’s numbers, the supply of machines willing to serve that demand has never been lower.

The demand side reads like a growth story. Akash’s own first-quarter report for 2026 announced that cumulative compute spend across its products had crossed $5 million for the first time, that its managed inference layer, AkashML, was processing about 1.7 billion tokens a day on OpenRouter (a figure the company said outpaced Cloudflare, and one that has climbed since), and that new leases on the network rose 27.1% quarter over quarter. Founder Greg Osuri has toured conferences with the message that AI has finally found a use for spare GPUs, and the token even briefly rallied on it.

The supply side reads like a warning. Messari’s State of Akash for the first quarter put the number of active providers at 58, the lowest on record, down from 63 the previous quarter and 69 a year earlier. GPUs advertised on the network fell more than 57% quarter over quarter, to about 334. Average utilization sat near 33.7%, meaning roughly two-thirds of the advertised cards were idle at any given moment. And despite all those extra leases, lease revenue actually fell 45% quarter over quarter, to about $253,250. More customers, fewer and cheaper machines, less money changing hands.

AKT, the token that is supposed to capture all of this, tells the same story as the supply side rather than the demand side. It changed hands at about $0.4957 on 16 September 2026, down close to 15% on the week and roughly 94% below its 2021 peak, for a market value near $148 million, according to CoinGecko (crypto prices move fast, so treat any single figure as a snapshot). The first three questions the market asked about Akash (is it actually cheap, is the demand actually real, and should it leave its own Cosmos chain) have all been argued at length over the past year. The 2026 question is the supply-side one, and it is harder: if renting out a GPU on Akash barely pays, who is going to do it, and what is Akash doing to change that? Answering it means starting with the auction that sets a provider’s paycheck.

A quarter when the leases rose and the machines left

The clearest way to see Akash’s problem is to lay the first-quarter figures side by side, because they point in opposite directions.

Read down the column and the tension is obvious. Tenants signed up for more deployments than in any prior quarter, yet they did it on a shrinking pool of hardware, and the network collected far less rent than it had three months earlier. The usual explanation for falling revenue is falling demand. Here demand rose. What fell was the number of operators willing to keep expensive accelerators plugged in and advertised, and the price those operators could command. Messari’s own read was blunt: providers cut capacity faster than usage dropped, which is why utilization looked stable even as the absolute size of the network shrank. A marketplace can survive slow demand. It cannot easily survive its sellers leaving, and Akash’s sellers have been leaving for more than a year.

How the reverse auction actually pays a provider

To understand why operators are walking away, you have to understand how they get paid, and Akash’s payment mechanism is a reverse auction. A tenant who wants compute writes a short manifest in a format Akash calls the Stack Definition Language, describing the CPU, memory, storage, and GPU model the workload needs. That order is broadcast to the network’s providers, who respond with bids. Crucially, the price moves downward: providers compete to offer the lowest acceptable rate, and the tenant then picks a winner. The tenant is not obliged to take the cheapest bid, since reputation, location, and hardware attributes matter, but in practice price does most of the work, and the mechanism is designed to drive it toward the marginal cost of the cheapest willing seller.

Once a lease is struck, the provider runs the tenant’s containers inside a Kubernetes cluster and is paid per block into an on-chain escrow account, from which the tenant’s balance is drawn down over the life of the deployment. Since March 2026 that settlement runs through a mechanism called Burn-Mint Equilibrium, which prices the lease in US dollars and handles the AKT accounting underneath (more on that, and its limits, below). The short version is that a provider’s revenue is set by an auction built to push prices toward the floor, denominated in a volatile token, on a marketplace where the buyer can walk to Amazon or CoreWeave the moment the experience disappoints. That is a hard place to earn back the cost of an accelerator.

The unit economics that do not close

Put real numbers on it. An NVIDIA H100, the workhorse accelerator of the current AI cycle, rents on Akash for something like $1.33 an hour, according to the network’s own pricing write-ups. Run it flat out for a month and it grosses just under $1,000. But Akash’s fleet did not run flat out; it ran at about a third of capacity in the first quarter. Apply that 33.7% utilization and the same card grosses roughly $300 to $350 a month in rental income.

Now weigh that against the cost of the card. A data-center H100 has traded in the $25,000 to $30,000 range for most of its life. At $325 a month, the card alone takes on the order of seven years to pay back, and that is before electricity, cooling, hosting, bandwidth, hardware failure, or the operator’s own time. The exact figure swings with the card price, the power contract, and how often the machine actually rents, so treat this as an illustration rather than a forecast. But the shape of the problem is not subtle: at Akash’s current prices and utilization, a merchant provider buying hardware specifically to rent on the network is underwater for years. It is a harsher version of the arithmetic that has made pure-play Ethereum staking a marginal business once you count the capital tied up against the yield on offer, a squeeze we walked through in our look at validator economics and the cost of locked-up ETH. The difference is that a provider’s capital is not a liquid token it can unstake, but a depreciating physical asset whose resale value drops every time NVIDIA ships a new generation.

The counterweight is supposed to be that Akash is dramatically cheaper for the tenant, and it is.

Those savings are real, and for a startup fine-tuning a model or an agent developer running inference, paying a third or a quarter of the hyperscaler rate is a genuine draw. But the tenant’s saving is the provider’s lost revenue. The reverse auction hands the benefit of decentralization to the buyer and leaves the seller to make the capital math work on thin, volatile margins. That is the engine of the supply problem.

The Kubernetes wall

Price is only half of why supply is scarce. The other half is friction. Becoming an Akash provider has traditionally meant standing up and maintaining a Kubernetes cluster, exposing it to the internet safely, keeping it patched and online, and handling the operational reality that a failed lease damages your standing on a network with no formal service-level agreement to fall back on. Akash leans on provider attributes and an audited-provider system in place of hard SLAs, so a provider’s reputation is effectively its product. For a professional data center that already runs Kubernetes at scale, this is routine. For everyone else, it is a wall.

That wall interacts badly with the economics. The operators most likely to accept thin margins (hobbyists, small shops, owners of idle gaming GPUs) are exactly the ones least equipped to run production Kubernetes. The operators who can run it comfortably (enterprise data centers) have better-paying customers than a marketplace whose average card sits idle two-thirds of the time. Akash’s two big supply-side initiatives of 2026, StarCluster and Homenode, are best understood as an attempt to break that bind from both ends at once: recruit the enterprises with a purpose-built vehicle, and remove the Kubernetes barrier for everyone else.

StarCluster, the answer from the top

StarCluster is Akash’s attempt to manufacture supply rather than wait for it to appear. Described in Messari’s third-quarter 2025 report and Akash’s own materials, it is a protocol-owned GPU network rolled out in phases, and its first phase is financed by an instrument the project calls Starbonds. A Starbond sells for $1,000, pools community money, and the pooled capital funds the purchase of roughly 7,200 NVIDIA Blackwell-generation GPUs. Those machines are run not by anonymous bidders but by vetted operators Akash calls Nodekeepers: enterprise-grade facilities, often telecom data centers with long-term power agreements, positioned near population centers for low-latency inference. Rental revenue from the fleet is split between the Nodekeepers who run it and the Starbond holders who funded it.

This is a very different animal from Akash’s permissionless marketplace. Instead of many independent providers each earning only when their own cards rent, StarCluster pools revenue across a managed fleet to smooth the yield, and it puts professional operators with cheap, stable power behind the hardware. In effect it is a decentralized financing wrapper around a conventional, well-run GPU data center. It solves the two problems the merchant-provider model cannot: it guarantees the machines are competently operated, and it spreads the utilization risk that would otherwise sink a single owner whose card sits idle. The catch, which later sections return to, is that funding GPUs by selling revenue-sharing bonds to the public is a securities activity, and that pulls a decentralized network back toward a very centralized, very regulated core.

Homenode, the answer from the bottom

If StarCluster recruits the professionals, Homenode is aimed at everyone else. Launched in early access on 25 February 2026, Homenode lets owners of consumer GPUs (initially the NVIDIA RTX 4090 and RTX 5090, plus some workstation cards) contribute compute without ever touching Kubernetes. It ships as a dedicated operating system that a user can dual-boot or run on a partition, turning a gaming rig or workstation into an Akash provider node with a setup closer to installing a game than administering a cluster.

The ambition behind it is not new. Back in early 2025, Osuri laid out a plan to scale Akash’s supply toward 11,000 professional data centers and as many as 7 million edge machines, dangling free high-end GPU clusters for operators who committed capacity. Homenode is the consumer end of that vision made real: instead of buying accelerators to rent, people who already own them for gaming or content creation flip on a switch and earn from the idle hours. The limits are inherent to the hardware. Consumer cards lack the memory, error correction, and fast interconnects that serious training needs, so Homenode supply is best suited to inference and smaller jobs, precisely the workloads AkashML and the agent platform are generating. Whether millions of gamers actually turn their GPUs into always-on servers, with the electricity bills and wear that implies, is the open question. But as a way to add supply that is indifferent to the reverse auction’s thin margins, because the hardware is already bought and the marginal cost is mostly power, it is the more creative of the two bets.

Paying providers to show up, and the dilution question

There is a third lever, and it is the bluntest: pay operators directly, in AKT, to bring capacity online. That is the substance of AEP-53, Akash’s on-chain provider-incentives proposal, which would fund pools by infrastructure type and pay providers in tokens after verifying that the promised resources were actually delivered. It is a rational response to a market that does not clear on its own: if the auction cannot pay providers enough, the protocol tops them up.

The problem is where the tokens come from. Subsidizing supply with issuance is dilution, and Akash has spent the past two years trying to tighten its monetary policy, not loosen it. In March 2025 the community passed Proposal 283, which cut maximum annual inflation from 13% to 8% and the floor from 8% to 4%, while a companion measure raised the share of issuance routed to the community pool from 40% to 50%. That larger community pool is exactly what could bankroll provider incentives, so the two policies sit in tension: a network trying to look less inflationary to holders may end up printing tokens to rent hardware it cannot otherwise attract. Every AKT paid to a provider to compensate for an uneconomic lease is an AKT that dilutes everyone else, and it only pencils out if the resulting usage lifts the token enough to cover the giveaway. So far, as the price chart shows, it has not.

The CoreWeave contrast, and the megawatt war

It helps to zoom out to the scale of the industry Akash is bidding into. CoreWeave, the centralized AI cloud that went public in early 2025, reported second-quarter 2026 revenue of $2.58 billion, up 112% year over year, with a revenue backlog of $104 billion (up 246%), 51 active data centers, about 1.5 gigawatts of power in operation, and guidance to spend somewhere between $35 billion and $39 billion in capital this year. Akash’s entire market capitalization, near $148 million, is a rounding error against a single quarter of CoreWeave’s capital budget.

That gap frames the real constraint, which is not software but power and capital. As Intrator has told investors, “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.” Osuri has his own version of the tension. “AI moves in months, energy moves in years,” he wrote in Akash’s first-quarter report, a line that quietly concedes the bottleneck is the physical build-out Akash is trying to route around. It is also why Bitcoin miners have been converting their sites into AI data centers as fast as they can sign leases, chasing the same scarce megawatts, a pivot and power crunch we traced in our piece on Bitcoin’s hashrate stall a year after the zettahash. Akash’s thesis is that it can aggregate capacity it does not own and undercut the hyperscalers on price. StarCluster is a tacit admission that aggregating other people’s idle machines is not enough, and that at some point you have to put competently run gigawatts behind the marketplace, which is precisely the capital-heavy game CoreWeave is winning.

Does AKT capture any of this?

Suppose the supply problem gets solved and Akash’s machines fill up. Does the token benefit? This is where late-2026 Akash gets uncomfortable, because the mechanism meant to link usage to token value does less than it appears. Burn-Mint Equilibrium, activated on 23 March 2026, is often described as a burn: tenants burn AKT to pay for compute. In practice, a tenant mints a US-dollar-pegged credit and the AKT that backed it is released to the provider at settlement. Tokens are recycled through the system rather than permanently destroyed. Net supply only shrinks if AKT appreciates between the moment a lease is priced and the moment it settles; if the price is flat or falling, the burn is close to a wash.

That would be a footnote if Akash published the net figure. It does not. As the analyst Zoha Imdad Ali argued in a pointed TECHi piece built around the idea that Akash runs real AI compute but AKT cannot prove it captures the value, the project reports gross tokens flowing into the BME vault (about 53,520 AKT in the mechanism’s first nine days) but not the net amount actually removed from supply over time. Without that number, holders cannot tell whether real compute demand is translating into real token scarcity. The price offers a circumstantial answer: usage metrics keep setting records while AKT trades 94% below its all-time high and below $0.50. Something in the chain from compute demand to token value is leaking, and the undisclosed net burn is where a skeptic would look first.

The trust gap Akash has not closed

There is a deeper reason the addressable demand, and therefore the revenue that would make providers whole, stays capped: Akash rents you a GPU, but it does not prove what happens on it. A tenant has to trust that the provider ran the requested workload honestly, did not snoop on the data, and did not tamper with the result. That is fine for rendering a video or running an open model, and it is a real problem for anything sensitive or high-value. Akash is not verifiable compute; renting the hardware and proving the computation are different guarantees, and the projects trying to prove computation live in a different part of the stack, as we explored in our look at Ritual and whether a public blockchain can keep an AI secret.

Akash’s answer is confidential computing, specified in AEP-65 and built on hardware trusted-execution environments such as AMD SEV-SNP and Intel TDX, which would let a provider attest that a workload ran inside a sealed enclave the operator cannot inspect. It was targeted to ship by the end of July 2026 and, as of mid-September, had not. Until it does, a category of enterprise and privacy-sensitive demand stays off the network, which means the revenue that could lift provider margins, and through BME the token, stays off the network too. The supply problem, the value-capture problem, and the trust gap are the same problem viewed from three angles.

The token field: AKT, Render, io.net, Nosana

Akash is not alone in this bind. The whole decentralized-GPU category shares the same gap between usage narratives and token prices, and the same tiny scale relative to centralized clouds. Combined, the sector’s networks were annualizing on the order of $180 million to $220 million in revenue earlier in 2026, less than a tenth of what CoreWeave books in a single quarter.

The models differ. Render grew out of 3D rendering and layers AI inference on top; io.net aggregates GPUs from data centers and independent operators into on-demand clusters; Nosana focuses on inference on Solana. But the pattern rhymes: each token is down roughly 90% or more from a 2024 peak, each network points to rising usage, and none has convincingly shown that usage flowing back into token value. Akash’s specific distinction is that it is furthest along on the managed-inference pivot (AkashML and the agent platform) and the only one pairing that with a protocol-owned fleet and a consumer-provider program. Whether that adds up to a durable edge or just more moving parts is the bet.

The SEC question hanging over Starbonds

The supply fix carries a regulatory tail that the marketplace never had to worry about. AKT itself reads as an infrastructural utility token: you spend it to buy compute, and it is peripheral to the parts of US securities law that bite hardest. Starbonds are different. A revenue-sharing instrument sold to the public to fund an asset, with returns paid from that asset’s cash flows, is close to the textbook definition of a security, and Akash has structured the offering as a regulated one in the United States rather than pretending otherwise.

That places a chunk of Akash’s supply strategy squarely under the Securities and Exchange Commission, inside the same commodity-or-security debate that has governed everything from token listings to fund approvals; we mapped that gate in our coverage of crypto ETF approvals and the commodity-or-security line. The irony is worth sitting with: to expand a permissionless, decentralized compute network, Akash is issuing a centralized, regulated, revenue-sharing security and handing the machines to vetted operators. It may be the pragmatic way to get gigawatts online, but it is a long way from the trustless ideal, and it gives US regulators a clear handle on a piece of the network.

What to watch into the fourth quarter

Several threads come due before year-end, and together they will decide whether the supply story turns.

• Chain selection. Akash’s plan to abandon its sovereign Cosmos chain for a shared-security model, formalized as AEP-79, still has no chosen host as of mid-September; Solana has been called a strong contender, and the roadmap targets a decision by the end of 2026. Whatever wins reshapes staking, security costs, and where AKT lives.
• The next Messari report. The second-quarter State of Akash had not published by mid-September. It is the cleanest read on whether the provider count stabilized or kept falling, and on whether the lease-revenue slide reversed.
• Confidential computing. AEP-65 slipped its July target. Shipping it would unlock the enterprise demand that could finally make provider margins work.
• StarCluster and Homenode traction. The proof is not the announcements but the machine count: does the fleet come online, and does the provider number climb back off its record low?
• Token2049 Singapore. Akash is among the projects at the September conference, the kind of venue where a chain-selection hint or a StarCluster milestone would most likely surface.

One caution runs through all of it. Akash’s roadmap has a habit of slipping (confidential computing is already late, the chain decision keeps receding into the next quarter), and crypto infrastructure has a broader pattern of endgames that never quite arrive, which we have written about in the context of account abstraction and the endgame that keeps slipping. The demand thesis for Akash is, at this point, proven: people want cheap decentralized inference and they are using it. The supply-and-value-capture thesis is not. Until providers can make money without a subsidy, until the net burn is disclosed and positive, and until the network can prove what runs on its machines, the record-low provider count is the number that matters most, and the one worth watching into 2027.

Frequently Asked Questions

Why is Akash’s active GPU provider count falling while AI demand rises?

Because supplying a GPU on Akash often does not pay. Its reverse-auction model pushes rental prices down, average utilization sat near 33.7% in the first quarter of 2026, and revenue per card is low, so many operators cut capacity faster than demand fell. Active providers dropped to 58, a record low, even as new leases rose more than 27% quarter over quarter.

How much can you earn renting out a GPU on Akash?

Less than most operators expect. At roughly $1.33 an hour for an H100 and about a third average utilization, a single card grosses only around $300 to $350 a month before power and hosting, so a $25,000 to $30,000 card can take years to pay back on rental income alone. Actual returns depend on the card, the power contract, and how often it rents.

What are Starbonds and StarCluster?

StarCluster is Akash’s protocol-owned GPU network, and Starbonds are the way it is funded. Starbonds are a regulated US securities offering, sold at $1,000 each, that pool community money to buy roughly 7,200 NVIDIA Blackwell GPUs operated by vetted datacenters called Nodekeepers. Rental revenue is shared between the Nodekeepers and the bondholders.

Does using Akash burn AKT?

Not in the way a permanent burn does. Since the Burn-Mint Equilibrium went live in March 2026, tenants pay by minting a US-dollar-pegged credit and the AKT backing it is released to providers at settlement, so tokens are recycled rather than destroyed. Net supply only shrinks if AKT rises in price between mint and settlement, and Akash has not published a running net-burn figure.

Is Akash cheaper than AWS or CoreWeave?

Usually yes on the raw hourly rate. An H100 on Akash runs around $1.33 an hour against roughly $4 to $7 or more on major clouds, though prices vary by up to two times across sources. The trade-offs are that Akash offers no formal service-level agreement, no built-in confidential-compute guarantee yet, and a much smaller, more variable pool of machines.

Marcus Okafor covers AI and crypto infrastructure for HOGE Wire.