Restaking’s Next Act: EigenCloud Bets on Verifiable AI

Restaking earned its reputation as Ethereum’s yield machine: lock staked ETH a second time, rent that security out to oracles, bridges and rollups, collect a second yield on top of the base staking reward. That pitch built a multibillion-dollar sector, and it has also, by mid-2026, produced roughly twenty explainer articles across this site alone walking through the mechanics. This is not another one of those.

The largest restaking protocol quietly stopped describing itself as a restaking protocol. EigenLayer became EigenCloud in June 2025, and its own product pages now talk about verifiable compute, AI agents and “crypto-grade” trust rather than actively validated services and slashing conditions. That is a real pivot, backed by a $70 million token purchase from a16z crypto, and it raises a question worth asking on its own terms: does restaking’s core mechanism, restaked ETH as a source of economic security, actually have anything useful to say about whether an AI model told the truth?

What follows looks at what EigenCompute and EigenAI actually do, who is using them today, and whether the numbers, EIGEN’s price, DefiLlama’s revenue tracker, the length of the customer list, back up the size of the claim.

From EigenLayer to EigenCloud

EigenLayer launched its mainnet restaking product in 2023 on a narrower pitch: let staked ETH double as collateral for other blockchain services, in exchange for a second yield. Vitalik Buterin warned early and publicly that overloading Ethereum’s validators with too many extra jobs could fracture the network’s social consensus if one of those jobs ever failed badly enough to need a bailout. Eigen Labs built the protocol anyway, turned on real slashing on mainnet in April 2025, and by the time EigenCloud’s AI ambitions were unveiled two months later, the underlying restaking base already secured a meaningful share of Ethereum’s staked supply.

In June 2025, EigenLayer changed its name. EigenCloud kept the same restaking mechanism underneath, but wrapped it inside a much bigger pitch: a “verifiable cloud” that developers could use the way they use AWS or Google Cloud, except every computation would carry a cryptographic or economic guarantee that it ran correctly. The rebrand landed alongside a $70 million direct token purchase from a16z crypto, on top of the $100 million the same firm had led into Eigen Labs’ Series B back in February 2024, according to CoinDesk’s reporting on the deal.

Sreeram Kannan, the CEO of Eigen Labs, framed the move as closing a gap rather than abandoning the original product: “EigenCloud will enable the next generation of disruptive, mass-market crypto apps, closing the gap between what developers want to build onchain and what blockchains allow them to build,” he said at the time. Ali Yahya, the a16z crypto general partner who led the investment, put it in similarly expansive terms, telling reporters: “EigenLayer is surmounting the technical bottlenecks of blockchains to enable a new category of applications built on verifiable solutions.”

That framing matters because it changes what EigenCloud is actually competing against. A restaking protocol competes with Symbiotic, Babylon and SSV Network for stake and AVS demand, a fight this site has covered in detail, including how Symbiotic and the rebranded Karak, now OpenGDP, have tried to win that fight without ever launching a token. A verifiable cloud platform competes with a different set of companies entirely: TEE specialists like Phala Network, zero-knowledge proof marketplaces like Boundless and Succinct, and eventually the hyperscalers themselves. EigenCloud is trying to be both at once. That two-front war is expensive to fight, and it is worth asking whether trying to be the default trust layer for both restaking-native AVS demand and general AI verifiability spreads EigenCloud’s attention, and its restaked collateral, more thinly than either fight alone would require.

Restaking in One Paragraph, for Readers Just Arriving

The short version, for anyone who has not read one of this site’s other restaking explainers: staking ETH means locking it to help secure Ethereum’s consensus in exchange for a base yield, currently in the low single digits. Restaking lets a staker point that same staked ETH, or a liquid staking token representing it, at a second job, backing an Actively Validated Service such as an oracle network, a bridge or a rollup sequencer, in exchange for extra yield and extra risk. The extra risk is slashing: if the staker’s node misbehaves on the second job, part of the restaked ETH can be destroyed.

Two things are worth knowing before the rest of this piece. First, exiting a restaking position is not instant, a friction the EigenLayer community is still actively redesigning through a proposal called ELIP-018, which this site has covered in a dedicated piece on restaking’s exit problem. Second, having a token has not obviously helped the protocols that have one: EIGEN, SSV and BABY are all down more than 90 percent from their all-time highs, a dynamic explored at length in this site’s look at the restaking protocols that chose not to launch a token at all.

The Three Pillars: EigenDA, EigenVerify and EigenCompute

EigenCloud groups its product line into three pieces, and the pitch only holds together if they work as a stack rather than three separate products.

• EigenDA: a data availability layer that rollups pay to use so they can publish transaction data cheaply while still letting anyone verify it is there.
• EigenVerify: the dispute resolution layer, deciding when two parties disagree about whether off-chain work was done honestly, backed by the threat of slashing restaked ETH from whichever operator lied.
• EigenCompute: the newest piece and the one this article is mostly about, a way to actually run code, including AI inference, inside infrastructure that produces a verifiable record of what happened.

EigenCompute runs the workload; EigenVerify is the backstop if someone disputes the result; EigenDA is the record layer everything else can point back to. In principle, that lets a developer skip writing custom cryptography for each new use case and instead reach for pre-built verifiable infrastructure, the same pitch every trust layer in crypto has made since the first oracle network, just aimed at AI workloads instead of price feeds. That stack approach is also EigenCloud’s answer to a common criticism of earlier verifiable-compute pitches, that stitching together separate point solutions for data, disputes and execution left too many seams for an attacker or a bug to hide in.

Inside EigenCompute: Confidential Computing Meets Restaked ETH

EigenCompute does not use zero-knowledge proofs. It uses trusted execution environments, TEEs, sealed sections of a processor that run code in a way even the machine’s own operating system cannot inspect, producing a hardware-signed attestation of what they ran. According to EigenCloud’s own product page, workloads run in confidential virtual machines hosted in part on Google Cloud using Intel TDX hardware, with the resulting execution “anchored onchain and backed by real economic security,” a reference to the restaked ETH sitting behind EigenVerify. The same page lists Ethereum, Arbitrum, Base, Solana, Polygon and Bitcoin as supported chains for onchain anchoring.

That combination, hardware attestation for speed plus a restaked-ETH economic backstop for disputes, is a deliberate hybrid, genuinely different from a pure TEE product like Phala Network or a pure zero-knowledge approach like Boundless. The tradeoff is that EigenCompute inherits both sets of weaknesses at once: a user still has to trust Intel and Google not to have compromised the hardware, the classic side-channel risk that comes with any TEE, and still has to trust that EigenLayer’s restaked collateral is large enough and liquid enough to actually punish a dishonest operator when it matters. TEEs are not a new idea and not flawless; security researchers have periodically found and disclosed side-channel vulnerabilities in both Intel and AMD’s confidential-computing hardware over the past decade, each one a reminder that hardware attestation moves the trust question rather than eliminating it.

EigenAI and the Bit-Exact Problem

EigenCompute handles arbitrary code. EigenAI is the narrower product built specifically for AI inference, reaching mainnet alpha in the autumn of 2025 alongside EigenCompute. It is designed as an OpenAI-compatible API, so a developer can point existing code at it with minimal changes, and its headline claim is what EigenCloud calls bit-exact deterministic execution of large language model inference on GPUs, launched initially with the open-weight gpt-oss-120b model.

That claim is harder than it sounds. GPU inference is normally not deterministic: floating-point rounding and parallel execution order can produce slightly different outputs for the exact same prompt run twice, which is fine for a chatbot but a real problem if the entire point is proving a specific output came from a specific input. Making inference reproducible enough to attest to, without the accuracy cost of forcing everything through slower, verifiable-but-simplified math, is the actual engineering problem EigenAI is trying to solve. It is also a large part of why EigenCompute leans on TEEs rather than zero-knowledge proofs for now: as Vitalik Buterin noted in an early essay on crypto and AI applications, the heaviest part of an AI computation, matrix multiplication, is one of the more ZK-friendly operations that exists, but running that math at the scale of a modern frontier model is still far more expensive to prove than to simply compute. The choice of an open-weight starting model, rather than a closed frontier model from a lab like OpenAI or Anthropic, is not incidental either: EigenAI’s attestation only proves that a specific model produced a specific output, it says nothing about whether that model is good, so starting with a model the community can independently inspect makes the verifiability claim easier to test against ground truth.

Why Verify an AI at All

The case for any of this starts with a problem that has nothing to do with restaking. Closed AI models are, as Buterin put it in the same essay, “no better than a centralized application” from a trust standpoint, because a user has no way to check that the model that answered them is the model the operator claims it is, or that its output was not quietly altered. That is an abstract concern for a chatbot. It stops being abstract the moment an AI agent holds a wallet.

On-chain AI agents, software that plans, calls tools and signs its own transactions, have grown into a real, if still developing, sector, and they have already produced real losses. Prompt injection attacks have drained agent-controlled wallets by hiding malicious instructions inside ordinary-looking inputs, including one widely reported case where an attacker used a gifted NFT and a Morse-code message hidden in a social media reply to trick a trading agent into moving funds. Payment rails built specifically for agents, such as Coinbase’s x402 protocol and Google’s Agent Payments Protocol, are already live: x402 has processed a meaningful and fast-growing volume of stablecoin payments initiated by agents rather than humans since its rollout, and Google’s protocol has signed on dozens of partners spanning payments networks and wallets. Verifiable compute is the infrastructure bet that, before agents are trusted with meaningfully more money, someone downstream needs to be able to check what an agent actually did, not just what it says it did. Framework quality has not always tracked token performance in this sector either: some of the best-known agent launchpads and frameworks have kept real developer activity even as their own tokens collapsed, a reminder that the agent economy and the tokens attached to it do not always move together, and a caution against reading EIGEN’s price alone as a verdict on whether agents actually need what EigenCompute is selling.

Four Ways to Verify a Machine

EigenCompute is one answer to a question several teams are racing to solve from different angles. It is worth laying the four broad approaches side by side, because they trade off differently, and none of them wins outright yet.

Zero-knowledge machine learning, or zkML, is the approach with the strongest theoretical guarantee, a mathematical proof rather than a promise backed by hardware or money, and this site has covered how that tradeoff plays out for privacy-sensitive use cases in more depth elsewhere. The catch is cost: proving a computation in zero knowledge is still far slower and more expensive than simply running it, which is why zkML today mostly shows up in smaller models or narrow sub-computations rather than full frontier-scale inference. Optimistic approaches sidestep that cost by assuming honesty and only proving fraud when someone challenges a result, which works well for use cases that can tolerate a delay before finality but poorly for anything needing an instant answer. Crypto-economic security, EigenCloud’s own foundation, skips cryptographic proof entirely and instead makes cheating expensive; it is fast and flexible, but the guarantee is only as good as the value of the collateral actually at risk.

Who Is Actually Building on It

The honest test of any new infrastructure layer is who is actually shipping on it, not who is talking about it. By EigenCloud’s own account, the current EigenCompute customer list includes OpenFront, a gaming platform using it to run provably fair tournament logic where players can check that match outcomes were not manipulated after the fact; Cap, an institutional credit market that also draws security from Symbiotic, using it to handle parts of its onchain credit operations; and an integration tied to Google’s Agent-to-Agent protocol, aimed at letting AI agents transact with verifiable execution sitting behind the interaction rather than a black box.

That is a real, if short, list, three named production users rather than a testnet demo, and it is genuinely further along than most competing verifiable-compute approaches, which mostly remain in developer preview or research-paper stages. It is also a modest footprint for a platform that drew a $100 million Series B round in February 2024 plus the $70 million EIGEN purchase described above. Whether that list is three logos or the start of a much longer one is the actual open question for anyone deciding whether to take the AI pivot seriously.

The Restaking Sector, Seen Through a Compute Lens

Zoom out to the wider restaking sector and the AI pivot looks less like an industry trend and more like one large player’s specific bet. None of EigenCloud’s direct restaking competitors have followed it toward verifiable compute.

Price and supply data above are from CoinGecko’s EigenCloud page, SSV Network page and Babylon page, fetched August 1, 2026; all three tokens are volatile and figures should be treated as a snapshot rather than a fixed reference point. Symbiotic has spent 2026 building Liquid Lane, a shared-collateral product aimed at real-world-asset credit markets, a pivot this site has covered in its look at the restaking protocols still operating without a token. Babylon has stayed focused on its original niche, letting Bitcoin secure other proof-of-stake chains without wrapping or bridging the underlying BTC. SSV Network has stayed closest to restaking’s original security-sharing pitch, expanding its distributed validator technology toward what it calls Based Applications. Karak’s rebrand to OpenGDP went furthest of all, dropping restaking language from its site entirely in favor of positioning itself as infrastructure for tokenizing real-world economic activity, a bet that the restaking category itself, AI-enabled or not, was not big enough to be worth staying in. EigenCloud is, for now, the only one of the five betting a meaningful part of its roadmap on AI infrastructure rather than AVS security alone.

The Money Question

A pivot is only as good as the revenue it eventually produces, and this is where EigenCloud’s numbers get harder to defend. Data aggregated by DefiLlama has repeatedly shown protocol-level revenue for EigenCloud sitting close to zero even while its total value locked, once above $19 billion at its 2025 peak, has settled somewhere in the mid-single-digit billions through 2026 depending on which tracker’s methodology is used. That gap, a large collateral base against minimal fee income, is not unique to EigenCloud among restaking protocols, but it is a specific problem for a company now asking to be valued partly as a cloud computing business, an industry where revenue per customer is usually the entire point. Some of that gap is structural: EigenCloud’s own Incentives Committee proposal, adopted in December 2025, tried to redirect emissions toward AVSs that generate real fees rather than idle deposits, precisely because the team recognized the emissions-to-revenue gap as a problem worth fixing at the protocol level.

It is worth being precise about what revenue even means here, since EigenCloud does not charge in a conventional subscription sense. EigenCompute and EigenAI are metered more like cloud infrastructure, usage-based fees for compute and attestation, which should in theory show up as protocol fee revenue on trackers like DefiLlama as adoption grows. The fact that this line item has stayed close to flat through more than half a year of the product being live on mainnet alpha is precisely why the revenue gap reads as a leading indicator of adoption, or the lack of it, rather than an accounting quirk.

EIGEN’s own price has not moved as though the market believes the pivot is working yet. The token traded at roughly $0.18 as of August 1, 2026, down about 97 percent from its December 2024 all-time high of $5.65, according to CoinGecko, with the token’s own scheduled unlocks, including one landing the same day this article was written, adding fresh sell pressure regardless of product news. A protocol can be building real infrastructure and still see its token discounted for reasons that have nothing to do with that infrastructure, emissions, macro conditions, general restaking-sector fatigue, but it does mean EigenCloud cannot point to price action as evidence the AI bet is landing.

The Skeptics’ Case

The strongest version of the skeptical argument is not that EigenCompute does not work; the product page and the named integrations suggest it does, at some level. It is that restaked ETH may not be the necessary ingredient the marketing implies. A TEE either produces a valid hardware attestation or it does not, and that check does not obviously require billions of dollars of Ethereum sitting behind it; Phala Network offers TEE-based verifiable compute today without a restaking layer underneath it at all, and its throughput on OpenRouter has at times exceeded a billion tokens processed in a single day. If the hardware attestation is doing most of the actual trust work, restaked ETH’s role starts to look more like a business model, a way to route fee revenue back to EIGEN holders, than a technical necessity.

There is a second, more structural version of the same worry. EigenLayer’s original AVS thesis always carried the risk academics call correlated slashing: a single bug or bad operator affecting every service that reuses the same underlying stake. Academic work modeling these dynamics, including a paper by Columbia researchers Naveen Durvasula and Tim Roughgarden, has shown that with the right overcollateralization buffers a small stake loss can be contained rather than cascading, but that only holds if operators and AVSs are configured with those buffers in the first place, a governance and design choice rather than an automatic property of the system. Layering AI inference on top of that stake does not remove the risk, it adds a new category of thing that can go wrong, a compromised confidential VM, a subtly incorrect attestation, to a security budget that already has to cover oracles, bridges and rollups. Whether restaked ETH is deep enough to credibly back all of that at once remains the same open question restaking has faced since EigenLayer’s mainnet launch, just applied to a new workload.

The token side of the story adds another layer of skepticism. EIGEN’s circulating supply keeps growing on a fixed emissions schedule regardless of how EigenCompute performs commercially, and August 1, 2026, the day this article was written, brought another scheduled cliff unlock of roughly 36.82 million EIGEN, the same size as the tranche released exactly a month earlier, adding fresh supply into a market that has shown little sign of absorbing it well. Emissions-funded ecosystems can outrun genuine demand for a long time before the mismatch shows up clearly in price, and skeptics argue EIGEN’s chart already shows that mismatch playing out in real time.

Restaking’s Own Track Record on Trust

EigenCloud is asking users to trust new infrastructure at the exact moment restaking’s track record on new infrastructure is mixed at best. The clearest example remains the Kelp DAO incident in April 2026, when an attacker exploited a cross-chain bridge, not EigenLayer itself, to mint roughly 116,500 rsETH out of thin air and use it as collateral to borrow against on Aave, an episode this site covered as part of a wider look at 2026’s wave of bridge exploits. The bridge, not the restaking or lending logic sitting on either side of it, was the point of failure. Aave founder Stani Kulechov spent weeks helping coordinate a recovery effort, writing at the time, per CoinDesk: “Aave is my life’s work and we’re working nonstop to find the best possible outcome for users.”

The lesson generalizes uncomfortably well to EigenCompute. New rails, whether a cross-chain messaging bridge or a confidential-VM execution environment, tend to be judged as secure right up until the specific way they fail becomes public. That is not a reason to dismiss EigenCompute, every piece of blockchain infrastructure carries that same start-up risk, but it is a reason to treat “backed by restaked ETH” as a statement about incentives rather than a guarantee that nothing will go wrong.

The Regulatory Gray Zone

Verifiable compute for AI agents sits in a part of the regulatory map that United States regulators have not really mapped yet. The SEC’s Division of Corporation Finance issued staff statements in May and August of 2025 concluding that certain protocol staking and liquid staking activities are not, by themselves, securities transactions, a framing this site has traced in detail in its ongoing look at the SEC’s own crypto enforcement record. Both statements went out of their way to exclude restaking and liquid-staking “variations” by name from that comfort, and staff statements carry no formal legal force in any case; they describe the current SEC staff’s thinking, not binding law. The broader legislative picture is not much clearer: the CLARITY Act, the closest thing to a comprehensive US market-structure bill for digital assets, has moved through Congress in fits and starts through 2026 without a final vote, and none of its public drafts have specifically addressed verifiable-compute products like EigenCompute, which sit somewhere between infrastructure, security service and financial collateral management.

EigenCompute complicates the picture further because it is not obviously a staking product at all. It is closer to a cloud computing service that happens to use restaked ETH as a security deposit, a category regulators have not addressed directly, and one that could plausibly draw scrutiny from a completely different direction if AI outputs verified through the system end up feeding into anything resembling investment advice or automated trading decisions, an area where the SEC has been clear that AI does not suspend existing rules around advisor conduct.

What Would Actually Prove This Works

A handful of concrete signals would tell readers whether EigenCloud’s AI pivot is working rather than just narrated. The first is the customer list actually growing past its current handful of named integrations into a broader set of production applications, not pilot programs. The second is DefiLlama’s revenue figures for the protocol moving meaningfully off zero, since a verifiable-compute business that never converts its usage into fees is, whatever else it is, not yet a real business. The third is whether EigenAI’s bit-exact execution claim holds up against outside scrutiny once more developers than EigenCloud’s own team have had a chance to try to break it.

The fourth, and probably the fairest test, is whether any of this changes how EIGEN trades. As of this writing it has not: the token has kept falling through every stage of the AI rollout described above, alpha launch, named integrations, the a16z investment, right through to today’s price near $0.18. Markets are not always right in the short term, and restaking-sector tokens have been depressed broadly enough that EIGEN’s slide may say more about sentiment toward the whole category than about EigenCompute specifically. But a pivot this ambitious eventually has to show up somewhere other than a blog post, and so far it has not shown up in the one number EigenCloud does not control.

There is a simpler, more human test worth watching too: whether other AVS teams and infrastructure builders start describing their own roadmaps in EigenCloud’s language, verifiable compute, agent accountability, crypto-grade trust, the way a genuinely successful pivot usually gets echoed by competitors trying to catch up. So far, Symbiotic, Babylon and SSV Network have each kept talking about their own, narrower priorities instead, which may be the most honest signal yet about how seriously the rest of the restaking sector is taking EigenCloud’s biggest bet.

Frequently Asked Questions

What is EigenCloud and how is it different from EigenLayer?

EigenCloud is the rebranded and expanded version of EigenLayer, the protocol that pioneered restaking on Ethereum. EigenLayer’s original pitch let staked ETH secure other services, called Actively Validated Services, such as oracles and bridges, in exchange for extra yield. EigenCloud keeps that restaking base but adds a second business on top: renting out the same cryptoeconomic security, plus new trusted-hardware infrastructure, to verify computation generally, including AI model outputs. The rename followed a $70 million token investment from a16z crypto in June 2025 that helped fund the buildout of EigenCompute and EigenAI, on top of the $100 million Series B the same firm had led into Eigen Labs in February 2024.

What is EigenCompute?

EigenCompute is EigenCloud’s product for running application code, including AI agents, inside confidential virtual machines, hardware-isolated environments known as trusted execution environments, hosted partly on Google Cloud using Intel TDX chips. The execution result is anchored onchain and backed economically by restaked ETH through EigenVerify, so that if an operator is later shown to have produced a false result, part of their restaked collateral can be slashed. In practice, this lets a developer deploy an application or an autonomous agent without writing custom cryptography, while still producing a verifiable record that specific code ran on specific inputs and produced a specific, attestable output.

How does restaking secure AI agents?

Restaking does not secure an AI model’s weights, training data or reasoning. What restaking-based verifiable compute can do is attest that a specific piece of code ran on specific inputs inside tamper-resistant hardware, backed by a financial penalty if the operator running that hardware is caught cheating. For an autonomous AI agent that holds a wallet and executes transactions on a user’s behalf, that gives a counterparty something concrete to check after the fact, an attested record of what actually ran, rather than a bare claim from the agent’s operator. It is a check on execution integrity, not a guarantee that the underlying AI model’s decisions were good ones.

Is EigenCloud’s verifiable compute live on mainnet?

EigenAI and EigenCompute reached mainnet alpha around late September 2025 and, as of mid-2026, list a small number of named production integrations, including a gaming platform, an institutional credit market, and infrastructure tied to Google’s agent-to-agent payments work. That puts EigenCloud further along than most competing verifiable-compute approaches, which mostly remain in developer preview. At the same time, independent trackers like DefiLlama have shown the protocol’s on-chain revenue staying close to zero even as billions of dollars in restaked collateral sit behind it, so mainnet availability and meaningful commercial adoption are not yet the same thing.

What is the difference between zkML and TEE-based verifiable compute?

zkML uses zero-knowledge cryptography to generate a mathematical proof that a computation was performed correctly; the proof itself is cheap and fast to check, even though generating it is computationally expensive, which today makes zkML impractical for very large AI models. TEE-based approaches, including EigenCompute, instead run the computation inside sealed, attestation-producing hardware; verification is fast and cheap from the start, but the guarantee ultimately rests on trusting the chip manufacturer and cloud host rather than pure mathematics. Neither approach is strictly better; they sit at different points on a tradeoff between cryptographic certainty and practical performance.

Priya Nathan covers DeFi infrastructure for HOGE Wire.