AKT: The Compute Demand Test
Akash has built a real open compute market and connected every new workload to AKT. The next question is harder: can demand become large, steady, and valuable enough to outweigh idle capacity and token issuance?
- Coverage
- Akash Network / AKT
- Evidence cutoff
- July 29 · 00:25 UTC
Chapters
The mechanism works. Scale is still the missing proof.
- Akash has real demand, and it improved after a weak first quarter. Lifetime marketplace spend reached $5.72 million. Q2 2026 spend rose 92.5% from Q1 to $488,090, and the first 28 days of July added another $207,902.
- The workload base is useful, but it is not yet compounding smoothly. Average active leases rose from 583 in Q1 to 699 in Q2, then moved back to 646 in July. Spend can grow while the number of running leases falls if the remaining workloads use more expensive resources.
- BME finally makes compute usage matter to AKT. Since March 23, compute payments have bought and burned about 1.30 million AKT to mint stable ACT credits. That is a direct demand path that did not exist under stablecoin-only settlement.
- Gross burn overstates the permanent supply effect. Providers can redeem ACT and remint AKT. About 1.01 million AKT has returned to supply, leaving roughly 288,800 AKT net burned. Net burn, not gross burn, is the metric investors should track.
- GPU demand is the strongest resource signal, but capacity is still underused. Akash listed 370 GPUs with 125 active, equal to 33.8% utilization. The top five providers controlled 54% of capacity, and utilization varied sharply across operators.
- The product is moving toward harder, more valuable workloads. Confidential Compute launched on July 28, Console split managed and self-custody users, and Akash is developing Homenode, virtual machines, agents, and managed inference. Confidential Compute is still experimental, and remote attestation tooling remains incomplete.
- Current burn is too small to make AKT deflationary at the network level. With 4% inflation, current supply implies about 11.85 million AKT of annual issuance. The observed net-burn pace annualizes to roughly 0.83 million AKT, offsetting about 7% of issuance. AKT also remains down about 66% over one year.
Akash is a marketplace for cloud capacity that no single company owns.
A customer describes the compute it needs. Independent providers compete to host it. The winning provider runs the workload and earns payment while the agreement remains active.
The marketplace works through reverse auctions. Instead of one cloud company posting a fixed price, providers bid for a deployment. The customer chooses a bid based on price, hardware, location, reputation, or other requirements. The lease and settlement state live on the Akash chain, while the actual container runs on the provider's infrastructure.
Akash is not one uniform datacenter. It combines different providers, processor generations, GPU models, storage systems, and network locations. That makes the market flexible and permissionless. It also makes quality, performance, and reliability harder to summarize with one capacity number.
Akash is trying to make cloud capacity behave like an open commodity market, then make AKT the economic bridge between customers who want stable prices and providers who want stable payment.
Demand recovered in Q2, but the network is still small and uneven.
Akash's public series shows $253,533 of marketplace spend in Q1 2026, $488,090 in Q2, and $207,902 from July 1 through July 28. Q2 nearly doubled from the weak first quarter. July's average daily spend was also about 38% above the Q2 average.
That is the strongest current operating signal. It should still be kept in scale. Lifetime spend across the network is $5.72 million. That proves real usage, but it is tiny beside centralized cloud revenue or the capital value placed on AI infrastructure.
Lease count and spend answer different questions. An active lease is one running agreement. A small CPU service and an eight-GPU training job each count as one lease, but their economics are very different. Spend can therefore rise without a proportional increase in active leases.
The mix appears to be moving toward more expensive workloads. In 2025, Akash reported that deployments grew 466% and dollar spend grew 128%, while year-end active deployments fell 69%. The network described this as a shift toward shorter, higher-frequency AI inference. The data supports a changing workload mix. It does not identify unique customers, customer concentration, or retention cohorts.
Akash has recurring marketplace activity and a visible Q2 recovery. It cannot yet prove broad customer adoption because public metrics do not separate unique tenants, repeat customers, or revenue concentration.
BME fixed the connection between product usage and AKT.
Before BME, stablecoin settlement made the product easier to use but weakened the token's role. Customers could buy compute without creating meaningful AKT demand. Mainnet 17 changed that on March 23.
A customer funds a deployment in stable terms. The system buys AKT, burns it, and mints ACT, a non-transferable dollar-denominated compute credit. The provider receives ACT as the workload runs. When the provider redeems ACT, the system can remint AKT at the current market price.
This last step matters. Gross burn tells us how much AKT entered the mechanism. Net burn tells us how much remained out of supply after provider redemption. At the cutoff, BME had burned 1.30 million AKT and later reminted 1.01 million AKT, leaving 288,806 AKT net burned.
The BME dashboard also showed $825,602 of ACT minted, $621,558 of ACT redeemed, $204,043 of ACT outstanding, and a 1.278 collateral ratio. Those figures demonstrate that the system is operating beyond a test deployment.
BME does not guarantee permanent deflation on every workload. The direction and size of net burn depend on AKT price movements between customer funding and provider redemption. The mechanism creates demand. The market and settlement path determine how much of the initial burn persists.
Akash has useful GPU inventory, but demand is not yet pulling the whole network tight.
At the cutoff, Akash listed 370 GPUs, with 125 active and seven pending. That equals 33.8% point-in-time utilization. GPU was the most utilized major resource. CPU utilization was 23.8%, memory was 16.1%, and storage was 5.3%.
GPU counts hide large differences. An H200, H100, A100, RTX 4090, and older datacenter card are not interchangeable. Performance, memory, networking, price, and workload support vary. A simple utilization rate therefore shows capacity use, not compute quality.
Supply is also concentrated. Thirty-five online providers listed GPU capacity. The largest five controlled 54% of total GPUs, and the largest ten controlled 77%. Some large providers were nearly full, while others had most of their capacity idle.
Provider count was 56, close to the Q1 average of 58 reported by Messari and below prior quarters. Fewer providers can mean better filtering of uneconomic supply. It can also mean less geographic and operational diversity. The useful signal is whether spend, utilization, and provider depth improve together.
Akash is building toward workloads that need more than cheap containers.
The next stage is about trust, easier onboarding, and support for workloads that enterprises cannot move into a basic container marketplace.
Confidential Compute
Akash launched network-wide Confidential Compute support on July 28. A tenant can request CPU or CPU-plus-GPU trusted execution in its deployment file. Supported providers use AMD SEV-SNP or Intel TDX, Kata Containers, and NVIDIA confidential GPU modes to isolate workload memory from the host operator.
The release is important because privacy is a real barrier for enterprise inference, proprietary models, and sensitive data. It is also early. Akash's documentation labels the feature experimental. The launch note states that requesting a trusted environment does not by itself prove the environment is genuine and uncompromised. Remote attestation tooling is still under development, provider capacity is limited, and some GPU generations have performance constraints.
Homenode, Agents, and managed inference
Homenode is designed to let owners of consumer and prosumer GPUs join without operating a full datacenter provider. Early support includes RTX 30, 40, and 50 series cards and RTX Pro 6000 hardware. This can widen supply, but the public material still describes pilot testing and staged rollout rather than a mature contributor network.
Akash Agents lowers the deployment barrier for self-hosted software agents. AkashML provides managed inference and is listed on OpenRouter. Akash has cited 1.7 billion daily tokens through AkashML, but that is a first-party product claim and is not the same as onchain lease spend or corporate revenue.
Console Air and the enterprise roadmap
Akash split its user experience in May. Managed-wallet and credit-card users remain in Console, while self-custody users moved to the open-source Console Air interface. This gives each group a simpler path without changing existing onchain deployments.
The 2026 roadmap targets virtual machines, managed services, provider incentives, private networking, and a move to shared security. These could make Akash easier to adopt and cheaper to operate. They remain roadmap items until live usage appears.
AKT now has a usage link, but issuance still dominates the supply equation.
AKT secures the chain, pays gas, participates in governance, and supports the BME vault. The chain reported 296.24 million AKT of total supply, 84.45 million bonded, 4% inflation, and a 4.21% staking APR. The bonded share was about 28.5%.
At the current supply and inflation rate, estimated annual issuance is about 11.85 million AKT. Net BME burn from March 23 through the cutoff was 288,806 AKT. Annualized, that pace is about 0.83 million AKT, equal to roughly 7% of estimated issuance.
This does not make BME irrelevant. It makes compute demand the variable that matters. If spend grows and the net-burn relationship remains favorable, the offset can rise. If usage stalls or provider redemption absorbs most gross burn, supply growth remains dominant.
At the cutoff, AKT had a market value of about $132 million and $4.4 million of 24-hour volume on CoinGecko. The token remains far below its 2021 all-time high. That gives the demand thesis leverage if network economics improve, but it also shows that product releases alone have not reversed the market trend.
The upside comes from one loop getting stronger.
The AKT case no longer requires users to hold a volatile token before buying compute. Customers can think in dollars, providers can price in dollars, and the protocol can still make each workload touch AKT. That is a cleaner product design than forcing token exposure on both sides.
The strongest path is straightforward. Better onboarding brings more workloads. More workloads raise provider utilization and improve provider economics. Higher spend drives more AKT purchases through BME. More net burn offsets more issuance. Better economics attract deeper, higher-quality supply, which makes the product more useful.
Quarterly spend grows without depending on one customer, promotion, or workload burst.
GPU utilization rises while provider count and high-end supply remain healthy.
Permanent burn grows faster than issuance and becomes material to circulating supply.
The product roadmap can accelerate this loop. Confidential workloads, virtual machines, private networking, managed services, home GPUs, and simpler billing each remove a reason not to use Akash. The risk is that the roadmap widens faster than demand.
AKT wins if Akash becomes a useful open compute market first, and a token story second. BME gives product success a path into the token. It cannot substitute for product success.
The thesis breaks if better infrastructure does not create durable demand.
- Demand scale. $5.72 million of lifetime spend proves usage, but it remains small. A few customers or temporary inference programs can move the series.
- Customer visibility. Public metrics do not show unique tenants, retention, customer concentration, churn, or gross margin.
- Idle capacity. Two thirds of listed GPUs were not active. Supply can leave if provider returns remain weak.
- Provider concentration. The ten largest GPU providers control most capacity. Related operators can make apparent diversity look larger than operational diversity.
- Gross-burn framing. Marketing can emphasize 1.30 million AKT burned while ignoring the 1.01 million AKT later reminted.
- Inflation. Current net burn offsets only a small part of estimated issuance. AKT can remain inflationary even while BME works as designed.
- Experimental privacy. Confidential Compute is early, remote attestation tooling is incomplete, and provider support can be scarce.
- Roadmap execution. Virtual machines, managed services, provider incentives, and shared security are targets, not current operating proof.
- Cloud competition. Hyperscalers, specialized GPU clouds, decentralized rivals, and direct provider markets all compete on price, reliability, tooling, and trust.
- Token security and governance. BME relies on oracle inputs, governance-controlled parameters, market liquidity, and correct vault operation.
Watch the economic loop, not the announcement count.
| Metric | What confirms the case | What weakens it | Where to verify |
|---|---|---|---|
| Quarterly compute spend | Repeated growth with less daily volatility | Q2 recovery reverses or depends on one workload | Akash Stats and Console API |
| Active leases and tenants | More running workloads plus public retention data | Spend rises while the workload base keeps shrinking | Akash Stats and future customer reporting |
| GPU utilization | Rises above 50% without supply leaving | Idle capacity stays high or high-end GPUs exit | Console provider inventory |
| Provider depth | More reliable providers across regions and GPU tiers | Further consolidation or repeated downtime | Console provider inventory |
| Net AKT burn | Permanent burn grows faster than gross burn and issuance | Reminting absorbs nearly all gross burn | Akash BME dashboard |
| Confidential Compute | Remote attestation, more capacity, and paid workloads | Feature remains experimental with few bids | Akash docs, providers, and lease data |
| Roadmap delivery | Virtual machines and private networking create usage | Dates slip or releases fail to change spend | Akash roadmap and releases |
The report uses live protocol data where the data exists.
Methodology and limitations
Network spend is the value paid through the Akash marketplace. It is not Overclock Labs revenue, protocol profit, or value accruing directly to tokenholders.
Active leases are running agreements, not unique customers. A lease can represent very different amounts and qualities of compute.
Resource utilization is a point-in-time ratio of active to total listed capacity. GPU counts do not normalize H200s, H100s, A100s, consumer cards, memory size, networking, or uptime.
Provider concentration uses current online providers and their listed capacity. Separate provider addresses can have related operators, ownership, or infrastructure.
Gross AKT burned for ACT is not permanent supply reduction. The report subtracts AKT reminted during provider redemption and uses net burn for supply analysis.
Annual issuance multiplies current total supply by the current 4% inflation rate. Annualized net burn extends the observed March 23 through July 29 pace across one year. It is an illustration, not a forecast.
Product announcements and throughput claims from Akash are first-party evidence. The report labels experimental and planned features separately and does not treat roadmap dates as completed work.
Market figures are fixed to the evidence cutoff. The mentioned-asset card updates live and can differ from the article figures.
This report is for information and research. It is not investment advice or an offer to buy or sell an asset. Digital assets are volatile, and protocol parameters, usage, and market data can change quickly.