In the first 90 days of 2026, Akash crossed an all-time high of $5 million in compute spend. A major tokenomics upgrade went from testnet to mainnet, a new product category came online with Homenode, the agent meta arrived on decentralized compute, and the community kept building applications that met the demand of our users.
Here’s everything you missed.
BME: The Largest Upgrade to AKT Economics
The problem BME solves is one every DePIN network eventually hits. You can offer stable pricing and watch your token become irrelevant to actual network activity, or you can require native token payments and watch enterprises walk away because no one deploys a month-long AI training job when compute costs can swing 20% on a liquidation event.
Akash’s previous solution, AEP-23, allowed tenants to pay in USDC while providers received stable settlements. It worked well for driving adoption. But the tradeoff became harder to ignore: when tenants pay in USDC, demand for AKT weakens. Usage grew while AKT’s role as the network’s core economic unit eroded.
BME resolves that tension without reintroducing volatility for users or providers.
For anyone who missed the testnet: BME is the most significant change to AKT tokenomics since the network launched. The mechanism creates a direct link between compute demand and AKT. When a user buys compute, their payment automatically buys AKT on the open market, burns it to produce ACT, a non-transferable stablecoin, and uses ACT to settle the transaction. Providers receive ACT and can mint it back to AKT at the current market price.
The result: buyers and providers both get price stability, and every compute transaction on the network drives demand for AKT.
The testnet was designed with one mandate: find the holes before mainnet. The community showed up, over 10 test categories were run, up to 250 participants put the system through its paces, and more than $10,000 in incentives were distributed.
The BME upgrade officially went live on March 23rd, 2026 following the passage of Proposal 318.
Akash Homenode: Earn From Your GPU
Akash Homenode Beta opened sign-ups in Q1, introducing a new supply-side category that did not exist on Akash before: consumer and prosumer GPU compute contributed directly by individual hardware owners.
The product is purpose-built for AI inference. GPU owners register at homenode.akash.network, connect their hardware to the network, and earn from compute demand without managing enterprise infrastructure. The program launched accepting RTX 4090s, RTX 5090s, and Quadro RTX 6000 Ada GPUs. Owners of other hardware are encouraged to sign up regardless, as the program will expand GPU support based on what the community brings.
The case for this model goes beyond cost and accessibility. The conflict in the Middle East brought into focus a risk the centralized cloud industry rarely discusses: what happens when a data center gets hit. Infrastructure concentrated in a single facility, region, or jurisdiction is a single point of failure, and recent attacks on data centers in the Arabian Gulf made that concrete. Homenode’s architecture eliminates that vulnerability by design. When compute is distributed across nodes in homes across multiple countries, a grid collapse or physical attack in one region is immediately absorbed by nodes elsewhere. If a data center in the Middle East takes a strike, nodes in Seoul and Frankfurt keep running. Resilience is not a feature that gets added on top of this model. It is what the structure produces naturally.
Homenode extends Akash’s supply beyond traditional data centers and into a geographically distributed mesh of hardware that is otherwise sitting idle. That is a different kind of compute availability, and it directly supports the inference workloads AkashML was built to serve.
The core team and Akash Insiders are running rigorous testing before the alpha release, stress-testing provider onboarding, hardware compatibility, and inference performance under real conditions.
Early access is live at homenode.akash.network.
Greg Osuri Discusses the AI Infra Crisis at NEARCON SF
Akash founder Greg Osuri took the stage at NEARCON SF this quarter to make a case the industry tends to avoid: AI has a physical ceiling, and we are closer to it than most roadmaps acknowledge.
The core argument is that AI’s energy demands are scaling faster than the infrastructure to support them ever can. Transformers take up to four years to procure. Nuclear permits take 10 to 15 years. Siemens Energy is sold out of turbines for seven years. “AI moves in months, energy moves in years.” Osuri famously quotes.
The demand side is accelerating the problem. Agents consume 10 to 40 times more tokens than a human interacting with AI directly, and Gartner projects 40% enterprise agent penetration by the end of 2026. Northern Virginia, the US data center hub, is already projected to fall out of electricity reliability standards by June 2027 because of AI load.
Osuri’s answer is the model Homenode is built on: local inference, local energy, peer-to-peer compute trading across a mesh of homes rather than a hub-and-spoke cloud. And as agents get more intimate, knowing your sleep patterns, your conversations, your daily routines, the privacy case for keeping that data local gets stronger alongside the energy one.
Lowering the Barrier to Agentic AI
The release of Akash Agents this quarter was about one thing: making agentic AI simpler to deploy on decentralized compute.
It started on Akash Console. Before a dedicated platform existed, OpenClaw launched as a deployment template, letting developers run their own private AI assistant on decentralized infrastructure without dedicated hardware and without giving up data control. It was an early signal that demand was there. Developers were already choosing Akash to host one of the fastest-growing open-source AI projects in recent memory, an autonomous agent that hit 145,000 GitHub stars in weeks and connects to WhatsApp, Telegram, Slack, and Discord to execute tasks, automate workflows, and maintain memory across conversations. It is still deployable directly from the Console here.
Akash Agents was iterated and launched weeks later, built by community member Sandeep Narahari to push that momentum further. Powered by AkashML inference and hosted on decentralized compute out of the box, it brings one-click agent deployments to anyone, no infrastructure management, no config overhead.
At launch, the platform supports two agents: OpenClaw, the viral open-source AI assistant, and Hermes by Nous Research. Getting started takes minutes: plug in your API keys from AkashML and Akash Console, pick your agent, and you are up and running on decentralized compute.
AkashML on OpenRouter: 1.7B Tokens a Day and Climbing
AkashML is now officially listed as a provider on OpenRouter, putting decentralized compute on the same menu as every major AI infrastructure player in the space.
The early numbers speak for themselves. AkashML is already processing 1.7 billion tokens per day on OpenRouter and outpacing Cloudflare in daily token usage.
OpenRouter is where developers go when they want model flexibility without vendor lock-in. Being listed there means AkashML is now a live option for any developer routing inference traffic across providers, alongside the centralized incumbents.
Hackathons On and Off Campus
Something shifted this quarter in how developers are reaching for Akash. Not as an experiment, but as the infrastructure of first choice when the stakes are real and the clock is running.
In February, Akash co-hosted the Open Agents Hackathon in San Francisco alongside Venice. The challenge was specific: design, deploy, and ship AI agents on permissionless compute using open APIs, with architectures that prioritize transparency, composability, and resilience over centralized cloud and model providers. What came back from the Akash track was not what most hackathons produce.
These were the winning projects from the Open Agents Hackathon.
1st Place: HealthGuard, a 24/7 autonomous health agent that monitors vitals, analyzes wound photos, transcribes voice notes, detects dangerous patterns, and alerts doctors automatically. The privacy architecture is the point: the agent processes everything it sees and forgets it the moment the task is done. Same vigilance, zero retention. One breach can expose a million patients. The team built against that reality.
2nd Place: BioVault Agent monitors clinical documents every 30 seconds using AkashML with MiniMax M2.5, standardizing and enriching data before validating it against international healthcare standards. It catches dose variances, unknown drug names, and missing fields before errors reach patients. A single mistake on a chemotherapy chart can cost a life.
3rd Place: AkashGuard is self-healing infrastructure. The agent checks the health of all running services every 30 seconds, diagnoses failures using Llama 3.3 70B on AkashML, and if confidence is above 70%, automatically closes the affected deployment, spins up a new one, accepts a provider bid, and waits until the replacement is live. Venice AI’s vision model takes a screenshot to verify everything is working. No human required, including at 2am.
Hackathons serve a second purpose beyond exposure. They generate direct feedback from developers who push on the product in ways the core team does not always anticipate, and that feedback has historically moved improvements into the console and deployment tooling faster than internal roadmap cycles alone.
The campus side of this quarter told the same story, just louder.
At the Penn Blockchain Conference, Akash Student Ambassadors from Indiana University took first place in the research competition. They built a full-stack on-chain analytics pipeline for Cronos Chain, from raw RPC ingestion to investment-grade research output, powered by Akash.
That same weekend, two Akash Student Ambassadors from UMass Amherst won at YHack 2026 at Yale, competing against 600+ hackers nationwide. They built HandFlow in 24 hours: a computer vision system that turns a $5 glasses-mounted camera into a full input device with real-time gesture control, macro pad functionality, and a virtual interface. Custom models were trained and iterated live during the sprint, using Akash for parallel GPU training when iteration speed was everything.
Looking Ahead
Akash entered 2026 on the back of its strongest quarter of native network activity on record, and Q1 did not slow down from there.
BME rewired the relationship between compute demand and AKT at the protocol level, making every workload on the network a deflationary event for the first time in Akash’s history. Homenode introduced a category of distributed GPU infrastructure that no decentralized cloud network had built before, extending supply beyond data centers and into a global mesh of individual hardware owners. The agent platform gave developers a dedicated home for the fastest-growing workload type in AI, with one-click deployments on decentralized compute out of the box. And AkashML crossed 1.7 billion tokens processed per day on OpenRouter, outpacing Cloudflare in daily token usage.
The community matched the pace. Student ambassadors won first place at two separate hackathons in the same weekend. Builders at the Open Agents Hackathon shipped production-grade healthcare AI and self-healing infrastructure on Akash under time pressure. Contributors built and launched platforms that became core product surfaces.
Q1 2026 demonstrated that while hyperscalers are rationing capacity and building waiting lists, Akash is already running.
