Sovereign AI at the application layer: no sovereignty without frontier performance

Event recap · 07 October 2026

From Sovereign AI at the Application Layer, an AI Alliance meetup with Neo4j and IBM, October 7 at IBM One Madison Avenue, New York.

The claim on the table

Anthony J. Annunziata, co-founder and chair of the AI Alliance, opened by putting the day's premise up for argument.

Directed towards the comfortable version of sovereignty: open weights, self-hosted, inside your own perimeter. That version provides progress, but if the sovereign option costs you capability, almost no one will choose it. “It's a trade-off too big for most to make.”

Then he turned the argument. The world's most valuable material sits outside the training corpus: “There's know-how that's implicit in operations, in people, in workflows. It's not available by and large to large-scale foundation model training efforts.”

If the data that would make a model useful in a regulated industry never reaches frontier training runs, sovereignty stops looking like a performance tax. It becomes a performance source nobody has tapped. Project Tapestry exists to test that thesis. It's also why the AI Alliance, now past 200 organizations across 29 countries, put a technical roadmap on screen instead of a manifesto.

IBM made its own case as host: a sovereignty timeline from System/360 in 1964 through on-premises software, Linux, hybrid cloud, watsonx and IBM Sovereign Core, backed by more than $50 billion in open-source company acquisitions over seven years.

We constructed the room to put the claim to test: IBM, Neo4j and OpenMined alongside NYU, Columbia and Wharton, plus startups, banks and consultancies. The sharpest moments came from three questions the speakers couldn't fully answer.

1

Context is the product

Nyah Macklin took the first layer: making a firm's own data legible to a model. She framed it as liability. Agents have left the chat window and started making decisions on loan approvals and Medicare and Medicaid eligibility. “We can't just have rogue agents making decisions, and then we have no idea why they made the decisions that they made.” A transaction log tells you what happened. It can't tell you why.

She brought numbers from an ablation study by Jiang et al. on a 3GPP question-answering task.

Source: Jiang et al., IEEE Communications Magazine, March 2026.

She also noted her tool's limits, which made the 91% easier to trust: “graph is not the end-all, be-all solution for all problems that we have.”

Her context graph pairs a knowledge graph with three memory tiers: session state; a long-term entity model built on the POLE fraud-detection ontology; and reasoning memory, which records which policy an agent applied and why. That third tier turns a decision history into something an agent or an auditor can query.

She tied it back to the thesis with Neo4j composite databases. Using local and remote aliases queried in Cypher, one bank can read another bank's approved fraud-pattern summary while every customer record stays put. The live demo crashed. She switched to pre-recorded visuals and kept the room, and with this audience the recovery probably earned her more credibility than a clean run.

2

Scale the network, not the node

Daniel J. Beutel joined remotely and led with three deployments: the first vision transformer trained in orbit, with StarCloud; a 70-billion-parameter model pre-trained across GPU clusters at three national labs, with Sandia; and the NHS connecting UK biobanks for cancer research. Three very different environments, he said, share one constraint:

In regulated industries, pooling the world's data in one place is impractical and undesirable. “Every organization should retain full control over the data that they have.” So Flower builds a network. A server sends out the model; each organization trains locally; only weights come back, and the server aggregates and repeats. The server never sees training data.

He backed it with production customers: Banking Circle, a B2B bank that processes 10% of European e-commerce traffic, trains anti-money-laundering models across regional datasets. Flower also runs in more than 500 hospitals and health institutes, covering over two million patient records.

3

Bring the question to the data

Ronnie Falcon covered the layer that makes the other two permissible. She sized the gap.

About two million times less. And the missing data is the data that matters: “the data that would make AI really useful and smarter is data that you're not allowed to move outside of the organization.”

She argued against pooling on two counts. If you copy the data, you lose control and the audit trail. Blend it into a training set, you can't delete it or show a regulator where an answer came from.

OpenMined shifts the contribution from training time to inference time. The network broadcasts a query. Each institution checks it against its own usage policy, runs it on local expert models, redacts, and returns an answer. An open-source aggregator synthesizes the results. Call it federated retrieval, with governance at each data store.

In August, OpenMined's technology enabled the first double-blind evaluation of a frontier model: Singapore's AI Safety Institute studied a DeepMind model's internals, weights, and user logs without either party handing over what it held. OpenMined is an official Tapestry partner and co-leads the Data Requirements Working Group with Common Crawl.

The $0 audit

Elaine Chan, an independent Tapestry contributor with no institution behind her, drew the loudest reaction of the day. She re-ran the M0 evaluation on an Apple M4 MacBook Air with 24 GB of memory: a 14.6 GB model, under 2 GB of headroom, fully offline after staging. No API, no cloud, no judge model. Budget: $0.00.

She found a gap between the recorded training loss of about 2.28 and the loss of 5 published with the artifact, raised it, and reported the team's explanation: restart spikes early in training, while state fluctuates.

That makes open weights a governance mechanism. Partners should also hear her closing point: you can contribute without a PhD or a cluster.

What the room pushed on

The speakers named three open questions on stage.

We're reporting these because the pushback showed the room cared.

The economics

Kaushik Bhatta closed by moving from architecture to money, where every participation decision lands. He defines sovereignty as agency, and the fear he hears is proprietary data leaking into a closed lab: “You talk to a CTO or CIO of a major ratings agency, they'll tell you privately they're not letting that data touch a closed lab. That's the proprietary advantage that they have to maintain their edge.”

Then the structural case. English accounts for 5% of native speakers and 50% of the web corpus, so frontier models under-represent most languages, cultures and regulatory regimes from the start.

The web answered that question with advertising. Without a new business model, he expects worse this time: “We're selling our civilizational output back to be rented back to us.”

His precedent was the Linux kernel. “Open has won before. When you create an ecosystem that itself is open, it tends to be a heck of a lot bigger than the closed ecosystem.” Then the concession: “We're not doing that right now with intelligence. Frankly, open's losing.”

What happens next

Milestone 1 will be the first Tapestry model anyone can download and use, targeted for the end of this year, likely in health. The spec: three or more compute nodes, two or more cultural contexts, an open reasoning base, and evaluation against task ground truth and clinician review on top of benchmarks. The research question: does collaboration beat matched local-only training?

Sovereign nodes

In India the work runs through BharatGen and IIT Bombay, with public backing from Professor Ajay Sood, Principal Scientific Adviser to the Government of India, who has aligned Tapestry with India's research agenda.

The team makes architecture decisions in the open on GitHub and runs contributor workshops twice a week. The parameter count is still under public debate.

Frequently asked questions

What is sovereign AI?

Sovereign AI is AI built so that a country or organization keeps control of the data, the model weights and the terms of use, rather than depending on a system designed and governed elsewhere. At this event, Kaushik Bhatta of the AI Alliance defined sovereignty more simply as agency. The event's central claim was that sovereignty only counts if it comes with frontier-level performance, because a sovereign option that costs you capability is a trade-off most organizations will refuse.

What is Project Tapestry?

Project Tapestry is the AI Alliance's initiative to train open frontier models across sovereign nodes. Contributors supply data, compute and expertise to a shared base model while keeping ownership of their raw data, then derive their own private models from the common base. Partners include Flower Labs, OpenMined, Common Crawl, Monash University, BharatGen with IIT Bombay, and NII.

What did Project Tapestry's Milestone 0 prove?

Milestone 0 proved that consortium training works across geographies. The team trained OLMo 3 7B across AWS Sydney and AWS Virginia, merged the result in Ohio, using eight H100s per site and two rounds of three billion tokens per site, with zero training data moved between sites. The final merged model recorded a loss of 2.2865 and perplexity of 9.78, and advanced to a new checkpoint, showing the distributed run improved the model. The checkpoint was deliberately not published, because Milestone 0 was a proof of concept rather than a release.

What is federated learning, and why does finance care?

Federated learning trains one model across many organizations without moving their data. A central server distributes the model, each participant trains on its own data locally, and only the updated weights are returned and aggregated. The server never sees the training data. Regulated industries care because data residency rules, client confidentiality and competitive sensitivity all prevent pooling, and regulators are increasingly encouraging the approach for use cases such as anti-money-laundering.

What is network-sourced AI?

Network-sourced AI, a term coined by OpenMined, moves the contribution of private data from training time to inference time. Instead of pooling data into a training set, a query is broadcast across a network of data holders. Each one checks the request against its own usage policy, runs it against local models, redacts, and returns an answer, and an open-source aggregator synthesizes a combined response with sources and an audit log. Data owners keep control and their rules, and AI users get answers drawn from data they could never otherwise reach.

What is a context graph?

A context graph pairs a knowledge graph with three tiers of agent memory: short-term session state, a long-term entity model, and reasoning memory that records which policy an agent applied and why. The reasoning tier is what turns a decision history into something an agent or an auditor can query, which matters when agents make decisions on loans or benefits eligibility.

When is Project Tapestry's Milestone 1?

Milestone 1 is targeted for the end of 2026 and will be the first Tapestry model anyone can download and use, most likely health-oriented. The spec calls for three or more compute nodes, two or more cultural contexts, an open reasoning base, and evaluation against task ground truth and clinician review on top of standard benchmarks. Architecture decisions are being made openly on GitHub.