Neurosoft Bets on a "Cortical Foundation Model" — and Taps Mila for the Data Problem

Neurosoft Bioelectronics and Mila announced on September 24, 2026 that the Geneva-based neurotechnology company has joined the Montreal research institute's industry network, with the stated goal of accelerating AI-powered neural interfaces for patients with severe neurological conditions. The centerpiece of the effort is what Neurosoft calls a cortical foundation model — a large model trained on neural recordings rather than text. Both organizations are unusually candid that the main obstacle is not architecture but data: brain recordings, they say, are orders of magnitude scarcer than the text corpora that trained today's language models.

What was actually announced

The announcement is a research partnership, not a product, a funding round or a clinical milestone. Neurosoft gains access to Mila's AI research community and what Mila describes as a network of close to 2,000 researchers and professionals. Mila frames the tie-up as part of its AI and Health focus area.

No financial terms were disclosed. The release does not say how many researchers will work on the project, whether Mila takes any equity or licensing position, how long the partnership runs, or what specific deliverables either side has committed to. It also does not name a target date for a model, a dataset, a publication or a clinical readout.

Notably, the collaboration is not starting from zero. Neurosoft neuroscientist Nicolas Gninenko says the company has an ongoing collaboration with Professor Guillaume Lajoie's lab at Mila, which he characterizes as one of the most productive scientific partnerships Neurosoft has engaged in. Joining the industry network, in his framing, formalizes and broadens work already underway. That is a more modest and more credible claim than most partnership announcements make, and it is worth reading the release as an expansion of scope rather than a new beginning.

The hardware angle: surface recording, not penetration

Neurosoft's stated approach is to build soft, stretchable electrode arrays that conform to the surface of the cortex and capture activity across brain networks without penetrating tissue. The company says the implantation uses a minimally invasive procedure rather than open surgery. It describes itself as an EPFL spin-out, based in Geneva and New York, and says it has moved its technology from the lab into first-in-human clinical studies in the United States and Europe.

That design choice carries a real trade-off that the announcement does not address directly. Surface arrays that sit on the cortex generally trade single-neuron resolution for broader spatial coverage and a lower surgical risk profile. Penetrating electrodes have historically been the route to high-bandwidth decoding of intended movement or speech. Neurosoft's pitch is essentially that coverage plus scale plus better algorithms can substitute for depth — that recording from entire networks safely and routinely produces enough high-quality data to train models that translate neural activity into movement, speech and perception. The release asserts this goal; it does not present decoding results, bandwidth figures, electrode counts, patient numbers or any comparison against existing approaches. None of that was disclosed.

The clinical detail is similarly thin. "First-in-human clinical studies" is a broad description that can cover anything from a handful of acute intraoperative recordings to chronic implants. Study size, endpoints, duration, regulatory pathway and results are not in the announcement.

Why data scarcity is the real story

The most substantive part of the release is its framing of the technical problem. Neurosoft says it is training a cortical foundation model on large-scale neural recordings, explicitly analogous to how large language models are trained on text. It then immediately concedes the disanalogy: brain data remains many orders of magnitude scarcer than text.

That is the crux. The scaling playbook that produced modern language models assumed a near-inexhaustible supply of cheap, weakly structured training data. Neural recordings are the opposite — expensive to collect, gated by surgery and ethics review, limited to small patient cohorts, and noisy in ways that vary by individual, electrode placement, session and time. Signals also drift as tissue responds to the implant. A model that generalizes across patients, conditions and clinical settings has to absorb all of that variability from comparatively tiny datasets.

Mila's answer, as stated, is sample-efficient learning: algorithms that learn more from less data, and models that encode prior knowledge about the brain into how they learn. Executive Vice President Stéphane Létourneau calls building AI that understands the brain "one of the most demanding tests we can set for sample-efficient learning," and says the partnership gives Mila's NeuroAI researchers a path from foundational algorithms toward clinical impact. This is a defensible research bet — strong inductive biases and architectural priors are the standard response when data is the binding constraint — but it is also a bet that runs against the grain of the scaling-first orthodoxy that made foundation models work in the first place. Whether the analogy to language models survives contact with a data regime that small is an open scientific question, not a settled one.

Neurosoft CEO and co-founder Nicolas Vachicouras argues that a cortical foundation model could do for brain-computer interfaces roughly what large language models did for natural language processing, enabling capabilities that scale across patients, conditions and clinical contexts beyond what current algorithms can reach. He positions the company's ambition as moving from a platform that records the brain to one that interprets it. No evidence for that capability was presented in the announcement.

Where this sits in Mila's partnership strategy

The deal fits a visible pattern. Mila, the nonprofit founded by Yoshua Bengio and created through a partnership between Université de Montréal and McGill University, says it is supported by the Government of Canada, the Government of Quebec and more than 150 industrial partners. Its recent announcements include an open-source AI initiative with Mozilla, a bilateral research partnership with Germany's DFKI, and a world-models collaboration. Adding a clinical neurotechnology company extends the same model: an academic institute supplying research depth to industry partners who bring proprietary data and hardware.

For engineers and founders, the interesting question is who ends up owning what. Foundation-model partnerships between nonprofits and startups typically involve negotiated positions on data access, model weights, publication rights and downstream commercialization. None of that was described here. In a domain where the training data comes from implanted patients, those governance questions — consent scope, data reuse across studies, whether trained representations can move between institutions — are not administrative footnotes. They are likely to shape how fast the science can actually move.

What to watch

• Whether either party publishes decoding results, dataset sizes or electrode specifications that would let outside researchers judge whether surface recording supplies enough bandwidth for a useful cortical foundation model.
• Any disclosure of the data governance terms: who controls the neural recordings, whether model weights or representations are shared, and what patient consent covers with respect to model training.
• Progress reports from Neurosoft's first-in-human studies in the US and Europe, including cohort size, implant duration and whether signal quality holds up over time.
• Peer-reviewed output from the existing collaboration with Guillaume Lajoie's lab, which would indicate whether the sample-efficiency approach is producing transferable methods or one-off results.
• Whether other BCI companies adopt the foundation-model framing, and whether any shared or pooled neural dataset emerges — the most plausible route around the scarcity bottleneck that both partners identify.