The AI Alliance has completed the first technical milestone of Project Tapestry, its open consortium for sovereign AI. Two proofs of concept trained models across four sites in India, Australia and the United States. No training data left any site; only model weights moved.
The project published every result, including the ones that went against it. That combination makes the milestone worth reading closely. It is also narrower than the sovereign AI headline suggests.
What Project Tapestry actually ran
The AI Alliance launched Project Tapestry in April 2026. It brings together more than 200 member organisations across 29 countries. Yann LeCun is chief science advisor, and Christopher Nguyen of Aitomatic is chief architect.
Milestone Zero closed on 1 September 2026. It ran two proofs of concept. The first, led jointly by BharatGen in India and Monash University in Australia, fine-tuned an OLMo 2 7B model. Each node tuned locally on separate, culturally specific partitions of a dataset called CultureInstruct, then merged compressed weight deltas. No tuning data was exchanged.
The second used the Flower federated framework. Two independently operated AWS GPU sites, in Sydney and Virginia, continued pre-training an OLMo 3 7B model. They trained on disjoint local partitions of the Dolma 3 mixture. A coordination node in Ohio combined the results.
The numbers behind the milestone
The second experiment is the more informative one. Each site ran eight H100 GPUs. It completed 7,500 local steps per round, producing about three billion tokens before aggregation. Two rounds gave each site six billion tokens and the consortium twelve billion.
The engineering details are where the difficulty sits. A single local phase took about 19.3 hours, which exceeded Flower’s default twelve-hour message lifetime. The team extended it to seven days and resumed from a saved checkpoint. The setup also absorbed a fifteen-fold latency difference between the two sites.
That is a real result. It shows that geographically separated clusters can co-train a model while their data stays home. It is also a feasibility demonstration at seven billion parameters, not a frontier model.
The catch: sovereign AI on someone else’s base
Tapestry’s architecture is what it calls “core-plus-sovereign”. A shared base model carries frontier capability, and each participant builds sovereign layers on top. The split starts near 80/20 between centralised and sovereign work.
The project is candid about the first phase. It adopts an existing open-weights model, in this case OLMo, as the base. Its own architecture record calls this “an honest, acknowledged dependency” and warns that it must not become permanent. The same document notes that training a frontier model from scratch needs more than $200m. The consortium does not yet have that maturity.
So today, sovereignty means sovereign layers on an external base. The claim to watch is whether the consortium ever trains its own.
Cultural alignment: a real result, narrower than the claim
The release says teams demonstrated cultural alignment by tuning a model for Indian and Vietnamese social and linguistic context. Its published finding is more careful.
LoRA tuning moved the model measurably closer to a target culture on the Inglehart-Welzel cultural projection. General capability, measured by MMLU, did not fall significantly. That is a positive result. But an independent validation harness, contributed separately, found the shift changes survey answers more than open-ended behaviour. The project states plainly that the next round of work goes at that gap.
This matters because cultural alignment is Tapestry’s stated differentiator. A model that speaks a language with imported values is not sovereign, the project argues. It shows the dial moves on one measure and less on another.
Vietnam and India: intentions, not agreements
The release frames “expanded sovereign AI collaborations with Vietnam and India”. In detail, it is softer. The Vietnam discussions “affirmed an intention to collaborate” on a sovereign model aligned to Vietnamese language and priorities. India’s role is existing rather than new. BharatGen ran one of the two proofs of concept, and a Project Tapestry workshop follows in Mumbai on 15 October.
India’s contribution is substantial in its own right. BharatGen is a government-funded consortium led by IIT Bombay, with roughly ₹1,235 crore of support. It has a 17-billion-parameter multilingual model, with coverage heading for all 22 scheduled Indian languages. Its lead, Prof. Ganesh Ramakrishnan, was named a 2026 Goalkeepers Champion by the Gates Foundation.
What to watch
The broader scepticism about sovereign AI applies here. Researchers argue that sovereignty is a stack of product decisions, not a policy declaration. Others note that even “data never leaves” boundaries can leak fine-tuned weights and adapters.
The economics cut both ways too. Analysts find that open sovereign models are cheap to adopt and expensive to sustain. The adopter inherits the cost of staying current.
Tapestry’s answers are structural. It publishes its weights, its results and its failures, and it separates shared from sovereign layers by design. The next milestone runs to November 2026. The test is whether it adds nodes and moves toward its own base model. That is what turns consortium training from a promising demo into a sovereign AI product.
Editor’s Note
This analysis draws on the 30 September 2026 AI Alliance release and Project Tapestry’s published Milestone Zero documentation. BharatGen‘s public materials and independent research on sovereign AI and federated training inform the context. Where a figure or claim is a company’s own, we say so.