Akash’s Compute Is Real. Whether AKT Captures It Isn’t

Akash Network’s token has had a good autumn. AKT climbed more than 50% off its mid-September lows near $0.50 to an early-October peak close to $0.78, then eased back to around $0.68 for a market capitalization near $203 million by 14 October, according to CoinGecko. The move tracked a broad re-rating across decentralized GPU tokens, and for a few weeks it looked like the market had finally decided that renting out other people’s graphics cards was a business worth paying up for.

Underneath the price, Akash runs something real. It is a permissionless marketplace where anyone can rent GPU time, and over the past year it has pushed hard into AI inference, the workload that actually pays a recurring bill. The demand is not imaginary. What is harder to find is evidence that any of that demand flows back to the people holding the token. AKT is supposed to capture the value of the compute that changes hands on Akash. The mechanism that would prove it exists, and the network has never shown the number.

This piece is about that gap. Not whether Akash is a real product, because it is, and not whether the rally was fun, because it was, but whether AKT captures the value of Akash’s compute or is simply the asset that trades while the compute changes hands under a mechanism that may barely touch its supply. That distinction is the entire investment case, and after a year of record usage milestones it is still unresolved.

What you actually rent on Akash

Akash launched its GPU marketplace in 2023 under the name Supercloud, extending a compute network that had run CPU and storage leases since 2020. The design is a reverse auction. A tenant writes a short deployment file, called an SDL, that says what it needs: a certain number of NVIDIA H100s, so much memory, a container image to run. Providers on the network see the request and bid, each quoting a price. The tenant chooses. It does not have to take the cheapest bid, but in practice price does most of the sorting, which is the whole point. Akash exists to undercut the hyperscalers on cost.

On raw hourly rates it succeeds. Independent sector research from Yellow puts decentralized H100 rentals well below on-demand hyperscaler pricing, with Akash and its peers often quoting a fraction of what AWS or Azure charge for comparable cards. The catch, and it is the catch that drives this entire discussion, is that cheap compute is not the same as captured value. A tenant can pay for an Akash lease in USDC or other IBC-bridged tokens without ever touching AKT. The marketplace can be busy while the token sits off to one side of the transaction.

The size of that discount is worth pinning down, because it is also the thing that caps revenue. On Akash an H100 has rented for roughly $1 to $2 an hour, against something closer to $4 to $7 an hour for on-demand capacity at the major clouds, a gap that can run to two or three times depending on the comparison, according to Yellow’s sector research. That is a genuine edge for a price-sensitive developer. It is also the mechanism by which the network competes its own revenue down, because the discount that wins the tenant is the same discount that shrinks the fee the token might one day capture.

There is a second thing Akash does not do, and it is worth naming early because it bounds how much the network can charge. Akash rents you a GPU; it does not prove the provider ran your job honestly. A machine could return a cheaper, degraded result and you would have limited means to catch it. Closing that gap is the entire project of verifiable compute, and it is a different problem from confidentiality, which Akash did address in 2026. Keep the two apart. They matter later, because a marketplace that cannot prove either honesty or scarcity is a marketplace whose token has to work extra hard to justify itself.

The demand side was never the problem

Give Akash its due. The usage is real and it is growing in the place that counts. Over the past year the network leaned into AI inference through AkashML and its consumer-facing Akash Chat, hosting open models and serving them through aggregators. Akash’s own first-quarter report for 2026 put throughput at roughly 1.7 billion tokens per day on OpenRouter early in the year, a figure management said climbed into the tens of billions by midyear. New leases rose, and the company marked a cumulative compute-spend milestone of $5 million across its products. Founder Greg Osuri summed up the thesis in that report with a line worth keeping: “AI moves in months, energy moves in years.”

Read charitably, that is a company doing the hard thing. Inference is stickier than raw GPU rental, because once a developer wires an app to your endpoint they tend to stay, and it points the network at recurring revenue instead of one-off training runs. The independent data backs the direction of travel even where it quibbles with the magnitude. Messari’s quarterly report recorded new leases up 27.1% quarter over quarter, to 43,540, in its State of Akash Q1 2026. People are deploying more, not less.

It is worth being honest about scale even here. Akash’s inference business is young, and in its first months it ranked behind several better-funded rivals on third-party measures of hosted-model revenue. The pivot is the right one and the growth is real, but it is early-stage growth off a small base, not dominance. So if the question were simply whether anyone uses Akash, the answer is yes, more every quarter. The uncomfortable part arrives with the next question. Did the money follow the leases, and did any of it reach the token.

The revenue went the other way

Here is where the story turns. In the same quarter that leases climbed 27.1%, on-chain lease revenue fell. Messari recorded roughly $253,250 in lease revenue for Q1 2026, down about 45% from the prior quarter. More deployments, less money. The explanation is the reverse auction working exactly as designed. Providers compete on price, the price of a GPU-hour keeps dropping, and a tenant who deploys more while paying less per hour can lift lease counts and depress revenue at the same time. The mechanism built to win customers on cost also caps what the network can earn from them.

Pair that with the supply picture and it gets starker. Messari counted 58 active providers in Q1 2026, a record low for the network, with 334 GPUs available, down more than 57% quarter over quarter, running at 33.7% utilization. For the full year 2025, Messari put revenue at $3.15 million, up 128% year over year, which sounds healthy until you annualize the Q1 lease-revenue run rate and get a number that points the other way. The trend inside the trend is the problem. Headline usage is up, but the on-chain revenue a token could plausibly capture is thin and, in the most recent quarter with data, shrinking.

Too few providers to grow into

Thin revenue is partly a supply problem. A reverse auction only produces healthy margins when enough buyers chase a scarce resource. On Akash the opposite happened. The provider base thinned to a record-low 58 in the first quarter, available GPUs fell more than 57% from the prior quarter to 334, and a third of the cards on the network sat idle at 33.7% utilization, per Messari. A marketplace with too few providers and so much idle capacity has no pricing power, so each new tenant is won on a lower bid. Usage can rise while the revenue per unit of usage keeps sliding, which is exactly the pattern the quarter showed.

Akash’s answer is a pincer. From the top down, a protocol-owned GPU mesh called StarCluster is bringing on enterprise-grade capacity, including several thousand NVIDIA Blackwell-class cards run by vetted operators and financed through the Starbonds instrument. From the bottom up, a consumer provider program called Homenode lets owners of high-end gaming cards like the RTX 4090 and 5090 contribute without running Kubernetes, which had been a real barrier to entry and a large reason the provider base stayed small in the first place.

Both are credible fixes, and both carry a cost for the value-capture story. The planned on-chain incentive program pays new providers in AKT, which dilutes existing holders. StarCluster is funded by a regulated bond rather than by the token. The network is spending equity-like and token-like resources to rebuild the supply that would let it raise prices, which is the right move operationally and another drag on the token while it happens. A marketplace has to have enough supply to be worth bidding for before its fees can grow, and Akash is not there yet.

How BME is supposed to turn compute into AKT

Akash knows that a busy marketplace does not automatically feed its token, which is why in March 2026 it activated a mechanism built specifically to connect the two. It is called Burn Mint Equilibrium, or BME, and it went live through a network upgrade and a governance proposal that passed with near-unanimous support. The idea belongs to a small family of token designs that try to tie a token’s supply to the real work done on a network, so that scarcity is a function of usage rather than of promises.

Mechanically it works like this. When a tenant pays for a lease, their AKT is burned in exchange for a settlement credit pegged to a US dollar value. The provider is then paid in freshly minted AKT worth that dollar amount at settlement. The peg is the clever part, because it shields both sides from AKT’s volatility during the life of a lease, in the same spirit that a well-built stablecoin AMM keeps a price steady while the assets underneath it move. A tenant is quoted dollars, pays dollars of value, and the provider receives dollars of value, with AKT acting as the settlement rail in between.

On paper, this is the value-capture engine. Every dollar of compute burns AKT on the way in. Do enough compute and you are burning a lot of AKT, which, all else equal, should shrink supply and support the price. In its first days, Akash reported tens of thousands of AKT flowing into the BME system. The design is deliberate, and it is the strongest answer Akash has ever had to the question of what the token is actually for. The problem is not the ambition. It is what happens on the other side of the mint.

The burn that barely burns

A BME burn is not a bonfire. Because the provider is paid in newly minted AKT worth the same dollar amount that was burned, the two sides very nearly cancel. The tokens taken out on the tenant side come back in on the provider side. The only permanent net destruction is the small difference created if AKT’s price moves between the moment of the burn and the moment of settlement, plus whatever modest share the protocol routes elsewhere. If AKT appreciates in that window, slightly fewer tokens are minted to cover the provider’s dollar payout and the system is modestly deflationary. If AKT is flat or falling, the net effect on supply rounds to nothing.

This is not a hidden flaw. It is how the design works, and in fairness a USD-pegged settlement layer is genuinely useful to tenants and providers regardless of what it does to supply. But it means the casual mental model, that compute burns AKT and therefore AKT is deflationary, is wrong, or at least unproven. BME is a settlement mechanism with a conditional deflationary kicker, not a straightforward buy-and-burn. Whether it meaningfully shrinks supply depends on AKT’s own price action, which is the very thing it is supposed to be driving. That circularity is exactly why the next number matters so much.

The number Akash has never published

Given all of that, there is one figure that would settle the argument: the net AKT actually removed from supply by BME, reported each quarter and trending up. It is the single metric that would show compute converting into scarcity rather than into a wash. Akash has not published it.

The point was put sharply by analyst Zoha Imdad Ali in a July 2026 piece for TECHi, whose central claim is that “usage and token value are separate” questions and that Akash has not yet closed the gap between them. Ali’s observation about the first-quarter report is the one that lands: it trumpets the $5 million spend record and the billions of daily tokens, yet “it does not publish how much AKT the burn side actually removed.” A team that engineered a demand-to-burn loop this carefully, Ali argues, and then declines to report the burn, is either sitting on an unflattering number or has not yet decided that value capture is the metric it manages toward.

That is the crux of this whole piece. The machinery to answer the value-capture question now exists on Akash. The answer does not. And in crypto, when a team builds a disclosure and then does not make the disclosure, the market is usually right to assume the number is not the one the team would want to show. Silence on a metric you went out of your way to make measurable is itself a data point.

Five million that will not reconcile with $253,000

There is a related gap Akash has also never reconciled, and it feeds the same doubt. The company’s own reporting celebrates a $5 million compute-spend figure. Messari, tracking on-chain lease revenue, measured about $253,250 in a quarter. Those two numbers are roughly twenty times apart, and both come from credible places.

The likely explanation is not dishonesty, it is scope and timebase. The $5 million is a cumulative, all-time figure spanning every Akash product, including managed inference services like AkashML that may not settle as ordinary on-chain GPU leases. The $253,250 is one quarter of raw on-chain lease revenue only. Different numerator, different clock. That reconciles the arithmetic in principle. What it does not do is tell an AKT holder how much value reaches the token, because the big, impressive number lives partly in products that may bypass the on-chain lease accounting entirely, and Akash has never published the bridge between the two.

Investors trying to value AKT are left doing what Bitcoin analysts do when the models behind the price targets rest on assumptions no one can pin down. They pick which number to believe and size the position accordingly. That is a difficult place from which to underwrite a token, and it is a self-inflicted one. A single reconciliation table would remove the ambiguity. Its absence keeps the most flattering figure and the most measurable figure on separate pages.

The supply side keeps printing

While the burn side stays conditional and undisclosed, the mint side runs on a schedule. AKT is an inflationary token. A March 2025 governance change, Proposal 283, cut the maximum annual inflation rate from 13% to 8% and the floor from 8% to 4%, and a companion proposal lifted the share of issuance routed to the community pool. You can read the network’s own account of those changes in its token-economics posts. The cut was real and responsible. It did not make AKT deflationary. It made the token inflate more slowly.

The twist for anyone staking AKT to earn that inflation is that roughly half of issuance is directed to the community pool rather than to stakers, which pushes the real staking yield into negative territory once you net out dilution. You can be paid a nominal single-digit percentage to stake and still lose purchasing power in AKT terms. For a token whose pitch is value capture, paying holders to hold with freshly printed supply is the opposite motion from the burn that is supposed to define it.

And more issuance is coming. One of the items on the 2026 roadmap, on-chain provider incentives, would pay providers in AKT for bringing capacity onto the network, verified after the fact. It is a sensible way to attack the record-low provider count, but it is paid in the token, which means it dilutes holders in order to subsidize supply. Add it up and the mint side has more planned than the burn side can currently prove. The scarcity story is a promise; the dilution is on the calendar.

Shared security rewrites AKT’s job

The deepest question mark over AKT’s value capture is structural, and Akash raised it itself. In late 2025, Osuri announced that Akash would deprecate its sovereign Cosmos-SDK chain and migrate to a shared-security model on another Layer 1, a plan formalized as the proposal known as AEP-79. The network is evaluating around fifteen ecosystems, with Solana described as a strong contender, and intends to keep IBC interoperability so it can still move assets across the cross-chain plumbing it relies on. No destination had been chosen as of October 2026, with completion targeted for the end of the year, per the Akash roadmap.

The migration got an extra push in April 2026, when the license on a critical Cosmos SDK component changed in a way that barred commercial use without a separate deal. Osuri called the move “hostile” on X and framed it as proof that Akash should not build its future on infrastructure it does not control. The strategic logic was laid out when the plan was first reported. Whatever the merits of that fight, it accelerated a decision that cuts to the heart of the token.

Here is why value-capture investors should care. On a sovereign chain, AKT’s largest job is security. Validators stake it, and that locked stake is what makes the chain expensive to attack. Shared security replaces that with a pay-per-use arrangement on someone else’s Layer 1. Osuri has argued the current model is capital-inefficient, which it is. But if staked AKT is no longer what secures the network, then one of the token’s two big demand sinks, the one that locks up a large share of supply, is being retired. The whole story then leans even harder on BME, the sink Akash will not quantify. Removing a demand sink is defensible engineering. It is also a reason to look harder at the sink that remains.

A utility token, a regulated bond, and the SEC

Value capture is also a regulatory question, because what a token is allowed to do depends on what it is. AKT reads as an infrastructure utility token. You use it to pay for and settle compute, and governance aside, its design does not promise holders a share of profits. That is the profile least likely to draw the US Securities and Exchange Commission’s attention under the agency’s long-running test for what counts as an investment contract. There is an irony in it for bulls: an asset that captures value mainly by being burned for usage is also the asset least likely to look like a security, which is another way of saying its upside is tied to usage rather than to any claim on the enterprise.

The more revealing wrinkle sits right next to AKT. Akash’s top-down supply push, a protocol-owned GPU mesh called StarCluster, is financed in part through an instrument called Starbonds, structured as a regulated US investment vehicle rather than a free-floating token. That is a telling choice. When Akash needed to raise real money to buy real Blackwell-class GPUs, it reached for a regulated security, not for AKT. The thing that funds the hardware and the thing that trades on exchanges are deliberately different instruments. That should tell holders something about where the network itself expects hard value to accrue, and it is not obviously the ticker in their wallet.

The rally was sector beta, not a re-rating

Step back to the price that opened this piece. If AKT’s autumn run had been about Akash specifically proving value capture, you would expect it to have diverged from its peers. It did not. The entire decentralized-GPU basket moved together. Render, the largest of the group by market value, traded around $1.99 for a market capitalization just above $1 billion; io.net sat near $0.16 and roughly $67 million; Nosana changed hands around $0.56 for about $56 million, all per CoinGecko on 14 October. They rallied together in late September and early October, and they sold off together into mid-October. That is the signature of sector beta, a rotation into an AI-compute theme, not a verdict on any one network’s fundamentals.

The flow data under the rally supports that reading. When AKT broke out in late September, it did so on a sharp jump in trading volume and on-chain signals of large buyers rotating in, a move AMBCrypto attributed to AI-sector rotation and whale accumulation rather than to any fresh utilization print. No new Messari quarter had landed to justify it; the Q1 report was still the newest data on the network. A price that moves on rotation and volume, with no change in the underlying revenue or burn, is close to the textbook definition of beta, and beta cuts both ways, which is why the whole basket gave back ground into mid-October.

This is the same trap the post-2021 history of crypto valuation keeps setting. When a narrative is hot, price reaches for a story and leaves the fundamentals to catch up or disappoint later. The useful discipline, borrowed from the autopsy of Bitcoin’s own stock-to-flow era, is to ask which model of value actually survives contact with the data. For AKT, the model that would survive is a disclosed, rising net burn. Until that number exists, a rally in AKT is best read as exposure to the AI-compute theme, priced and timed by flows, not as the market rewarding Akash for capturing value its peers cannot.

The benchmark that does not need a token

To see how modest Akash’s economics still are, measure them against the company eating the same demand with no token at all. CoreWeave, the centralized AI cloud, reported second-quarter 2026 revenue of about $2.58 billion, up 112% year over year, with a backlog above $100 billion, according to its results as covered by CNBC. Akash’s entire 2025 revenue was $3.15 million. At CoreWeave’s run rate, it books Akash’s full-year revenue in under three hours.

CoreWeave’s chief executive, Michael Intrator, framed the demand environment in terms that explain both his company’s trajectory and the market Akash is chasing: “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 is precisely the opening Akash is built for, cheaper capacity delivered permissionlessly. But CoreWeave answers it by raising tens of billions in debt and equity and signing multi-year contracts, capturing value the ordinary way, as a company with revenue and a balance sheet. Akash is trying to route the same value through a token, and the token is the part that has not been shown to work. The contrast is not that decentralized compute is fake. It is that the centralized version has a value-capture mechanism nobody doubts: shareholders own the cash flows.

What would actually prove value capture

None of this says Akash is a fraud or that AKT is going to zero. The product is real, the inference pivot is smart, and the supply-side fixes, StarCluster from the top down and a consumer provider program from the bottom up, are serious attempts to rebuild a provider base that has thinned to a record low. A network can be genuinely useful and still have a token that does not capture its usefulness. Those are separate claims, and the argument here is that Akash has proven the first and not the second.

So here is the short list of what would change the verdict, in rough order of importance. First and above all, a published net-burn number, reported quarterly and trending up, so holders can watch compute convert into scarcity instead of taking it on faith. Second, a reconciliation between the $5 million product figure and the on-chain lease revenue, so the impressive number and the measurable number stop living on different pages. Third, clarity on what AKT secures after the shared-security migration, because if staking stops being the chain’s defense, BME has to carry the token by itself. Fourth, a reversal in the provider count and utilization, so the marketplace has enough supply to grow revenue rather than compete it to the floor.

Watch those four, in that order, through the end of 2026 and into the next Messari report. If the chain selection lands, confidential compute wins real adoption, and a burn number finally appears, the next AKT rally might be something you can underwrite on fundamentals rather than on flows. Until then, enjoy the beta, respect the product, and keep the token and the network in separate columns of the spreadsheet. The compute is real. Whether AKT captures it is still, after everything, the open question.

Frequently Asked Questions

What is Akash Network and what is AKT used for?

Akash is a permissionless cloud-compute marketplace where tenants rent GPU and CPU time from independent providers through a reverse auction. AKT is its native token, used to pay for and settle leases and to stake for network security. Tenants can also pay in USDC or other IBC tokens, so using Akash does not strictly require holding AKT.

Does Akash’s BME burn make AKT deflationary?

Not reliably. Under Burn Mint Equilibrium, a tenant’s AKT is burned for a US dollar-pegged credit while the provider is paid in freshly minted AKT of equal dollar value, so the two sides mostly cancel. Net supply only shrinks if AKT appreciates between the burn and settlement. Akash has not published the net AKT removed, so the real deflationary effect is unproven.

Why did AKT rally in late 2026?

AKT climbed more than 50% off its mid-September 2026 lows alongside the entire decentralized-GPU token basket, including Render, io.net and Nosana. The synchronized move points to a sector rotation into AI-compute tokens rather than to any Akash-specific proof of value capture, and the basket sold off together into mid-October.

Is Akash leaving the Cosmos ecosystem?

Akash plans to deprecate its sovereign Cosmos-SDK chain and move to a shared-security model on another Layer 1, a plan known as AEP-79, while keeping IBC interoperability. As of October 2026 no destination chain had been chosen, with completion targeted for the end of the year. The shift would change what AKT secures and therefore how it captures value.

How does Akash compare to CoreWeave?

They chase the same AI-compute demand at very different scales. CoreWeave, a centralized cloud with no token, reported about $2.58 billion in second-quarter 2026 revenue and a backlog above $100 billion. Akash’s full-year 2025 revenue was $3.15 million. Akash competes on price and permissionlessness, but CoreWeave captures value as a conventional company rather than through a token.

Marcus Okafor covers AI and crypto infrastructure for HOGE Wire. This article is analysis, not investment advice.