If your agents run on OPAQUE's confidential AI platform, nobody has told you who supports it next quarter, so plan as if the answer is no one until you hear otherwise in writing. Anthropic has licensed OPAQUE's technology and hired part of its team, terms undisclosed, per Constellation Research on October 13. OPAQUE's own announcement calls it "a new agreement with Anthropic that carries OPAQUE's work forward" and says nothing about the product, customer contracts, support or the open-source projects, according to the company's post.
For a Claude buyer, the deal points to where Anthropic wants to go with confidential, sovereign and on-premises deployments. It doesn't give you anything you can buy this quarter.
What Did Anthropic Actually Get From OPAQUE?
Anthropic got a licence to OPAQUE's technology and some of its people, and no primary source says it got the company. Dealroom's summary calls it an acquisition of OPAQUE Systems, but the Constellation report it cites describes a licence-and-hire. Constellation's Larry Dignan describes the transaction as an acquihire in which Anthropic licenses the technology and brings on part of the team, without saying how many people move or what happens to the products. Korea's Digital Today reported the same structure a day later and added no further terms. Anthropic has not published a statement on the deal.
An acquihire is a deal in which a buyer pays mainly for the team, often through a licence plus hiring rather than by buying the company's shares. The legal entity, its contracts and its customers can stay behind with whatever staff remain.
What OPAQUE built is worth knowing in detail, because it tells you what Anthropic now has rights to:
- OPAQUE spun out of UC Berkeley's RISELab and was founded in 2021, per SiliconANGLE. Its platform runs AI on sensitive data inside hardware-isolated enclaves and produces cryptographic evidence of what ran and where.
- In May, OPAQUE acquired multiparty computation, fully homomorphic encryption and post-quantum protections from the UAE's Technology Innovation Institute, SiliconANGLE reported, extending the platform from inference into training.
- At the Confidential Computing Summit on June 23, OPAQUE launched Agent Manifest (cryptographic identity for agents) and Confidential MCP (governance controls on Model Context Protocol calls), with 3.0 support for AMD SEV, Intel and NVIDIA confidential computing and general availability scheduled for July, per Converge Digest. We covered that launch in our OPAQUE 3.0 analysis.
The money was modest by frontier-lab standards. OPAQUE raised a $24 million Series B in February led by Walden Catalyst at about a $300 million post-money valuation, bringing total funding to $55.5 million, Crowdfund Insider reported. Abu Dhabi's Advanced Technology Research Council joined that round as a new investor.
One detail changes how to read the whole thing: Anthropic was already on OPAQUE's customer list. The February funding coverage names ServiceNow, Anthropic and Accenture, and SiliconANGLE's May piece lists ServiceNow, Anthropic, Accenture and Encore Capital Group as customers. Anthropic licensed a vendor it already used.
What Happens to OPAQUE's Platform Customers?
Nobody has said, and every OPAQUE customer has to plan around that this week. Read the announcement against the questions a customer would ask and it answers none of them: it doesn't say the platform will keep being sold or supported, doesn't mention contracts, data or migration, and doesn't say whether the open-source projects stay maintained, per the post. It doesn't call the deal an acquisition either.
Compare the last big licence-and-hire in AI. When Google paid a reported $2.4 billion to license Windsurf's technology and hire its CEO and top researchers in July 2025, the coverage carried an explicit continuity line: most of Windsurf's roughly 250 employees stayed, an interim CEO took over, and the company "will continue offering its AI coding tools for enterprise customers," TechCrunch reported. The licence was nonexclusive and Windsurf kept the right to license to others. OPAQUE's announcement has no equivalent sentence. Its tone is a retrospective: it credits a decade of work from Berkeley and thanks "every OPAQUEling" along with customers and investors.
The strongest case for calm goes like this. Anthropic was a paying customer, so it has its own reasons to keep the technology running. OPAQUE may well be telling customers privately what it isn't saying publicly. And a licence does not by itself end OPAQUE Systems, Inc. as a company. All three may be true. None of them is a support commitment you can put in front of your auditor.
The exposure is sharpest where OPAQUE's selling point was regulatory. Constellation notes OPAQUE was adopted in Europe and the UAE and had released open standards for sovereign AI and on-prem deployments. A bank or ministry that chose OPAQUE so its attestation evidence would satisfy a regulator now has a control whose vendor may not exist in a year. We have seen this pattern before: Klaviyo's Agency acquihire ended the product within weeks, and Harvey took the Guardrails AI team after the hub went dark.
Who Owns the Open-Source Projects Now?
The answer depends on which project you use, and for some of them nobody has said. OPAQUE's post lists AgenTrust, Agent Governance Toolkit, TRACE, Confidential MCP and Weight Custody Manifest as open efforts on agent and model trust, without naming a licence for any of them.
They fall into two groups:
- Agent Governance Toolkit is Microsoft's. Its repository sits under Microsoft's GitHub organisation, and Microsoft shipped .NET MCP extensions for it on May 21, per Microsoft's developer blog. OPAQUE's Agent Manifest extends AGT, so if you depend on AGT itself, this deal does not change who maintains it.
- Confidential MCP and Agent Manifest are OPAQUE's. If their maintainers are among the people moving to Anthropic, the projects may slow or stop. The code stays available when a maintainer leaves, but security fixes tend to stop with them.
We traced the same risk when Cognition hired the Dioxus team and the Blitz maintainers went with it. The fix is the same: pin a version, mirror the repository, and decide who on your side would patch a CVE.
What the Deal Tells a Claude Buyer
The deal signals that Anthropic intends to sell confidential and sovereign deployments; it doesn't change what you can buy from Anthropic today. Constellation reads it as fitting Anthropic's push for enterprises weighing on-premises deployments, open models and sovereign AI architectures. That is the analyst's inference. Anthropic hasn't said it.
Anthropic's own public work on this is research, not product. In June 2025 it published a design for confidential inference using trusted virtual machines, in which a small attested model loader runs in an isolated VM and a keyserver releases decryption keys only to environments that prove what code they run, per Anthropic's research post. The same post calls the work "just a sketch," says "it is too soon to forecast" what features follow, and admits that sharing encrypted host memory with accelerators is not yet well established, so Anthropic relies on its compute providers for hypervisor security in the meantime.
OPAQUE's team built that missing piece: attested execution across AMD, Intel and NVIDIA confidential hardware, with evidence a third party can verify. The deal gives Anthropic the people who shipped it, and no ship date came with it.
So if a sovereign or on-prem requirement is the reason Claude is losing your internal bake-off, nothing has changed this quarter. The alternatives still look like the options in our sovereignty clause scorecard and the self-hosted route IBM took when it dropped Claude for air-gapped Bob. What has changed is the question you can put to your Anthropic account team: when, and on whose hardware.
What to Do About It
This Week:
- If you run OPAQUE, email your account contact and ask four questions in writing: is the platform still sold, who provides support and security patches after the deal, what is the end-of-support date, and how do you export your policies, attestation records and audit logs. Put the reply in your vendor-risk file.
- Find every place OPAQUE sits in a control. Pull the list of agents, MCP servers and data pipelines that depend on its attestation, and note which ones a regulator or customer audit relies on.
- Pin and mirror Confidential MCP and Agent Manifest if you use them. Record the exact versions in production and copy the repositories somewhere you control.
This Month:
- Price the replacement. Your realistic options are the hyperscalers' own confidential computing services on AMD SEV, Intel or NVIDIA hardware, or another attestation vendor. Get a migration estimate before you need one.
- Read your OPAQUE contract for assignment and change-of-control language. A licence-and-hire may not trigger a change-of-control clause at all, which is exactly why these deals rarely give customers an exit right. We covered how this played out at Nvidia and Poolside.
Before Your Next Claude Renewal:
- Ask Anthropic for a dated confidential computing roadmap, naming which clouds and which hardware, and whether attestation evidence will be available to customers. If the answer is "we can't comment," note it and keep your alternatives warm.
The Bigger Picture
Frontier labs are now buying the teams behind the controls regulated buyers ask for, and the products those teams sold are the part left unaccounted for. Google's Windsurf deal at least came with a public line about continuing to serve enterprise customers. OPAQUE's came with a thank-you note. The pattern has repeated across this year's acquihires, from Nvidia's pursuit of Reflection AI to the vendors that joined OpenAI months before their customers heard: the buyer gets what it paid for, and the customer has to go and ask.
If you are an OPAQUE customer, send the four questions today and log the date you sent them.
Continue Reading
- 77% Wrote AI Agent Policies. Only 26% Can Enforce Them.
- Nvidia Hired 109 Poolside Engineers. No Clause Fired.
- Guardrails' Hub Died Aug 25. Harvey Bought the Team.
- Klaviyo Bought Agency. The Product Dies August 31.
- Samsung Led Mistral's Round. Score the Clauses, Not the Flag.
- IBM's Air-Gapped Bob Drops Claude for Nemotron and Laguna
