NEWS
Google Cloud Splits Bank Crypto Keys After Willow Advance
Google Cloud’s Multi-Party Compute is in bank preview, but it still signs with today’s curves and lets signers override a policy denial.
Google Cloud opened a private preview of Multi-Party Compute on October 28, a key-splitting service for banks that still signs with the elliptic-curve math Bitcoin and Ethereum already use. Commercial sale is aimed at early 2027. Alphabet has started using the same stack for its own assets on Ethereum and Solana, without naming the tokens or the amounts.
Thirteen days earlier, Google Quantum AI published a Willow result that cut magic-state errors 40-fold. The custody pitch arrived in the same month as fresh warnings about AI-accelerated math, and the product Google is actually showing banks is built to stop stolen credentials and rogue transfers, not to replace secp256k1.
Google Multi-Party Compute, Now in Private Preview
Google Cloud showed the service at its first digital-assets symposium in New York on October 28. Nalin, who works on digital assets at Google, called it the group’s biggest launch in that business so far and said the point was security that does not trade away speed.
This is our biggest launch in the digital assets space to date 🚀 If you need to custody crypto at scale, Google MPC provides security without trading performance. Reach out to learn more.
— nalin 🇺🇸 (@nalin) October 28, 2026
Public documentation last updated October 27 describes Multi-Party Compute as a Google Cloud custody platform for digital asset signing keys. It bundles threshold signing without a full private key, transaction policy, and governance. The protocol is a hardened extension of DKLS23. Independent parties each hold an isolated share and produce a signature without assembling the complete key in one place.
A KeyPolicy sets who holds shares, how many must take part, and where those shares live: Google Cloud, the customer’s own premises, or user mobile devices. Google’s examples include 2-of-3 and 5-of-7. A bank can keep one share on its own hardware, one in Google Cloud, and one with another cloud, so no single provider, Google included, holds the full key. Google says it cannot read customer shares in plaintext, even when a share sits on its servers. Signers, policy authors, and admins also use an authorization key whose private half never leaves the customer’s custody.
Supported math in the public docs is ECDSA on secp256k1 for Bitcoin, Ethereum, and other EVM chains, plus EdDSA on Ed25519 for Solana, marked coming soon. Google has said the service can later move to post-quantum signing as chains adopt those schemes. It has not given a date.
Rich Widmann, director of digital assets strategy at Google Cloud, tied the launch to banks putting client assets on chains.
Digital asset adoption and tokenization are at an inflection point. As regulated financial institutions begin adopting blockchains and tokenization, they need scalable, proven infrastructure to secure their clients’ assets.
Rich Widmann, director of digital assets strategy, Google Cloud
That is a credential-theft problem, and it is a large one. Chainalysis put private keys as 43.8% of 2024 thefts by value, the biggest single slice of a $2.2 billion year in stolen funds.
THE 2024 KEY-THEFT TALLY
- Stolen value: Chainalysis counted $2.2 billion taken from crypto platforms in 2024.
- Private keys: Compromised keys accounted for 43.8% of that stolen value, the largest share.
- Incident count: The firm logged 303 hacks in 2024, up from 282 in 2023.
Those figures measure last year’s theft mix, not Google’s preview. They do explain why a bank that cannot unwind a confirmed transfer wants more than one machine to hold a complete key.
Signers Can Still Push Through a Denied Transfer
Before any share is used, the instance decodes the unsigned payload and runs every enforced Cedar policy against it. The default is deny. A transfer is allowed only if at least one permit rule matches, and any matching forbid rule wins. With no enforced policies, every transfer is denied.
That sounds strict until the next paragraph of the same manual. The verdict is a guardrail on automatic signing, not a lock on the people who hold shares.
The verdict guides Signers but doesn’t enforce the final decision: a transaction that the automated transaction policies deny can still be executed if enough Signers approve it.
Google Cloud Multi-party Compute overview, updated October 27, 2026
Deny by Default, Then Let Signers Overrule
On a permit, the instance will sign at once if the request already carries enough authorizations to meet the KeyRing threshold. Otherwise it waits. On a deny, it will not sign on its own. The transfer sits in PENDING_AUTHORIZATION. Signers can still authorize it, and it is signed if enough of them do. The original evaluation record stays attached for audit.
Google has described that check as running in Confidential Space, the hardware-isolated environment it has sold for this job since 2023. Isolation limits what Google’s own operators can see. It does not stop a quorum of authorized signers from agreeing to move funds the rules just rejected.
A Dollar Figure the System Does Not Check
Policies can look at on-chain amounts, destination lists, chain identifiers, and fees. They can also read metadata the initiator attaches, including a fiat amount that must be in whole US dollars. The docs are blunt about that field. The instance does not check metadata against the payload. The dollar figure is what the initiator says the transfer is worth. Other signers see it before they approve. A sample Cedar rule in the manual caps that self-reported amount at 10,000, which only holds if people treat the field as honest.
KeyLists of trusted addresses can change after a policy is approved, and every rule that points at the list uses the new entries on the next check, with no fresh policy-author vote. Keys cannot be moved between KeyRings or deleted. They can only be disabled, which stops further signing.
A 40-Fold Error Cut Landed 13 Days Before the Pitch
On October 15, Emma Rosenfeld and Craig Gidney of Google Quantum AI wrote that magic-state cultivation had worked on the Willow processor, with a 40-fold cut in magic-state errors against uncultivated states and T-states at 99.99% logical fidelity. The result ran in Science. Their own post said some estimates still put up to 95% of a future machine’s capacity into making those states, which is why a cheaper recipe matters.
Willow is not a machine that can spend bitcoin. Google was careful to say larger-scale cultivation is still required. The timing is still awkward for a custody launch that leans on AI and quantum fear. The same company that just made a scarce quantum resource cheaper is asking banks to store key shares for assets whose live signatures are ECDSA and, soon, Ed25519, the families a cryptographically useful quantum computer would target.
THE OCTOBER CALENDAR INSIDE GOOGLE
- October 1, 2026: Release notes add distributed Schnorr signatures on Curve25519 with standard Ed25519 hashing, and extend Bitcoin policy checks to amounts, destination types, and change outputs.
- October 7, 2026: Vitalik Buterin warns that AI-accelerated math has a good chance of hitting lattice cryptography over the next two years, and that ECDSA may fall faster than people expect.
- October 15, 2026: Google Quantum AI reports the Willow cultivation result in Science, with a 40-fold error cut and 99.99% T-state fidelity.
- October 27, 2026: Multi-party Compute public docs are updated, still listing Solana Ed25519 as coming soon and still describing policy verdicts as a guide for signers.
- October 28, 2026: Google Cloud presents the bank preview in New York, with a commercial target of early 2027.
On October 7, Buterin wrote that no one needed to scramble wallets that day, then drew a harder line about the math. “We should take the risks to cryptography from AI-accelerated math seriously, and minimize our exposure to not just quantum-vulnerable cryptography, but also potentially AI-vulnerable cryptography,” he said. He named ML-DSA, fully homomorphic encryption, and lattices as the new weak spot, and said hash-based schemes look better where they can be used. He also repeated the fresh-address advice, because an exposed public key is the thing a future break would eat.
MPC does not hide a public key that is already on-chain. It changes who must collude to sign the next transfer. Those are different jobs, and only one of them is in the preview.
Fireblocks, BitGo and Fordefi Already Run This Model
MPC is not new in this market, and neither is Google’s research on it. On January 24, 2023, Google Cloud published a Confidential Space post on using MPC so parties could sign without exposing key material to each other or to the platform operator. The April 2023 MPC signing architecture turned that into a reference design, with co-signers allowed on Google Cloud, other clouds, on-prem systems, or a phone. What Google had not sold, until this preview, was a packaged custody product that wrapped signing, policy, and recovery into one service. Customers who wanted that stack on Google Cloud had to assemble Cloud KMS, Cloud HSM, and their own MPC code.
Fireblocks, BitGo, and Fordefi already sell institutions tools that split signing and attach approval rules. Fireblocks documents MPC-CMP with policy inside secure enclaves and lets a client keep shares. BitGo pairs qualified custody with multi-signature and MPC. Fordefi sells MPC wallets and simulation. Google’s opening is less “we invented split keys” than “your key share can live next to the rest of your Google Cloud estate, and another share can live somewhere Google does not run.”
WHO SELLS THE SPLIT
| Firm | What it sells | Key method | Status |
|---|---|---|---|
| Google Multi-Party Compute | Signing, policy, and recovery for institutions that keep their own assets | Threshold MPC (DKLS23) | Private preview, commercial target early 2027 |
| Fireblocks | Wallet and policy platform, with an optional trust company | MPC-CMP | In production for institutions |
| BitGo | Qualified custody plus wallet tools | Multi-signature and MPC | In production for institutions |
Google did not name the banks in the preview and did not publish pricing. Speaker lists for the New York event included people from Morgan Stanley, Citi, Wells Fargo, and OKX. That is a room full of firms that already have vendors for this job, which is the point of the pitch: a cloud company they already use, rather than a crypto-native specialist they have to add.
What the Early 2027 Product Does Not Fix
The preview is a signing and policy service. It is not a patch for the math on the chains it supports, and it is not a promise that a policy denial is final.
GAPS STILL OPEN IN THE PREVIEW
- Live curves: Public docs still sign with ECDSA on secp256k1 and, for Solana, Ed25519 marked coming soon, with no date on post-quantum upgrades.
- Override path: Enough signers can authorize a transfer the Cedar rules just denied, so the policy engine is a brake on autopilot, not a veto.
- Self-reported dollars: A fiat amount in metadata is the initiator’s claim, and the instance does not match it to the on-chain payload.
- Exposed keys: Splitting a new key does not help coins whose public keys are already on-chain, which is the stock Vitalik told people to stop growing.
- Unnamed book: Testers are unnamed, Alphabet’s holdings are unnamed, and pricing is unpublished, so a board still cannot compare this preview with Fireblocks or BitGo on cost or independent signing-speed tests.
Banks that want this anyway are not buying a magic shield against Willow. They are buying a way to keep a share off Google, run approvals as code, and still settle on the chains they already use. Collusion among enough signers remains in scope. So does a bug in the policy language. So does a chain-level break of ECDSA, which no vendor’s MPC product can sign its way out of.
Alphabet Is Using It on Ethereum and Solana
Google has said its parent has begun using Multi-Party Compute to secure digital assets on Ethereum and Solana. It has not said which tokens, which amounts, or whether those wallets are treasury, testing, or both. That silence is useful for a sales meeting and useless as a holding disclosure. It tells a bank the stack has moved real shares, not that Alphabet is a hidden ether whale.
The public docs still list Ed25519 Solana signing as coming soon, a day before the symposium. Those two statements can both be true if Alphabet is on a private build, or if Solana balances are being handled on a path the public manual does not yet describe. Until Google publishes the asset list, the honest read is narrower: the parent is a design partner with live shares on those two chains, and the public Ed25519 line has not caught up.
August 20 release notes had already added Curve25519 key generation and signing. October 1 added the Schnorr and Ed25519 hashing work. The Solana row in the overview table is the piece that remains labelled as future tense for everyone else.
The Share That Lives on Google Cloud
The product Google can sell in early 2027 is a signing fabric that sits beside the rest of a bank’s cloud bill. One share can stay on the bank’s hardware. One can sit in Google Cloud. One can sit with a rival. Policy runs before a share is used, then the same people who hold shares can overrule it. Alphabet is already on the stack for unnamed Ethereum and Solana assets. Fireblocks, BitGo, and Fordefi already do the split for institutions that wanted it without waiting for a Google SKU.
Willow’s 40-fold cut and Buterin’s lattice warning are real, and they are a different project from stopping a stolen laptop from moving a client’s bitcoin. Multi-Party Compute, as documented on October 27, is the second project. The first one is still on secp256k1.
Early 2027 is the date Google has given banks. The testers, the price, and the Solana general-availability line are not on that calendar yet.
Disclaimer: This article is news reporting on a Google Cloud product and on public figures about digital-asset theft and cryptography. It is for information only. It is not investment advice, legal advice, or a security assessment of Google Multi-Party Compute, and it is not a recommendation to buy, hold, transfer, or custody any digital asset. Custody, treasury, and compliance teams should consult a qualified lawyer, a licensed financial adviser, and their own security engineers before changing key-management or custody arrangements. Figures, preview status, chain support, and commercial timing reflect the public documents and statements cited here and can change as the private preview proceeds.
-
BUSINESS5 months agoAramco Crash Kills 14 Two Days After Oil Restart
-
NEWS6 months agoSalah Captained Egypt to Their First World Cup Win
-
GAMING2 months agoZelda’s 40th Switch 2 Ships Without a Zelda Game
-
NEWS3 months agoApple Labor Day Deals Hide June’s Memory Price Hike
-
NEWS4 months agoEl Niño Puts Israel’s Needed Winter Rain on a Flood Clock
-
GAMING2 months agoPalworld 1.0.4 Adds Palpedia Search After the 1.0 Launch
-
NEWS4 months agoSaudi Visa Routes Split Nigerian Travelers Into Two Speeds
-
BUSINESS9 months agoBAE’s 60-Year Saudi Franchise Now Runs Through Taif
