Weekly Neurotech & BCI Digest — November 2, 2026

The first week of November 2026 draws a clear arc: scale is becoming the operative word in BCI. Neuralink is training on 50,000 hours of human intracortical data. Precision Neuroscience has mapped 70 patients' nervous systems end-to-end. Merge Labs is chasing whole-brain coverage through ultrasound. And China is racing to define the standards everyone else will eventually have to meet. Wherever you sit in the stack, the engineering decisions made this month will compound for years.

Research Highlights

Precision Neuroscience's 70-Patient CNS Atlas

On October 8, Precision Neuroscience released a bioRxiv preprint (doi: 10.64898/2026.09.24.754261) describing a functional atlas of the human central nervous system built from high-resolution surface electrophysiology across 70 patients — spanning cortex, brainstem, and spinal cord. The atlas coincides with Precision's 100th patient implant milestone and follows its oversubscribed $250M Series D closed in September.[1] Data were acquired using Precision's Layer 7™ thin-film ECoG array; the manuscript is a bioRxiv preprint and has not been peer reviewed.[2]

Why it matters for engineers: Brainstem and spinal cord recordings in a clinical population are rare and difficult to obtain. If the atlas ships as an open or licensed resource, it would let BCI teams pre-train decoders on anatomically grounded priors — significantly reducing the cold-start problem in motor and speech BCI. Watch the data-access terms closely; structured neural atlases in research vs. commercial licensing have very different engineering implications.

Neuralink VOICE Trial: Real-Time Speech Restoration Demonstrated

Neuralink's VOICE trial (NCT07224256) — a separate feasibility study from the PRIME motor-control trial — targets cortical speech production areas in ALS, stroke, and SCI patients. In September 2026, participant Terry, an ALS patient, spoke in real time through a personalized voice clone derived from pre-ALS recordings. Neuralink's N1 implant (1,024 electrodes, 64 flexible threads, skull-mounted wireless) reads from speech-production cortex; the R1 robot places threads with minimal dural disruption.[3] Neuralink has not published word-error rates or vocabulary size; its public Speech Restoration page reports a median decoding rate of roughly 50 words per minute and says a core goal is to move toward natural-conversation speed.[4]

Academically, UCSF's Chang group separately published in Nature Neuroscience a 253-channel ECoG BCI that decodes speech and gesture simultaneously — offering a peer-reviewed benchmark for comparison when VOICE data eventually surfaces.

Hardware & Devices

Merge Labs × Butterfly Network: Functional Ultrasound as a BCI Modality

Merge Labs (founded January 2026, $252M seed; team includes Tyson Aflalo, Sam Altman, Alex Blania, Mikhail Shapiro) signed a multi-year agreement with Butterfly Network in September to use Butterfly's Ultrasound-on-Chip™ for non-invasive brain imaging and interaction.[5] The deal includes upfront payments, milestone-based license fees, hardware purchase commitments, and ongoing royalties. Separately, Bridge Neurotech (emerged from stealth October 6) is pursuing functional ultrasound imaging (fUSI) via Butterfly's same chip for a distinct non-invasive BCI approach.[6]

Why it matters for engineers: Functional ultrasound offers ~100 µm spatial resolution and ~ms-range temporal sampling of hemodynamic signals — well above fMRI resolution, and compatible with real-time decoding loops in ways that whole-head fMRI is not. The Butterfly chip's CMOS integration means miniaturization paths exist that were implausible with traditional piezo transducer arrays. For signal processing engineers, fUS decoder design is an open field: the hemodynamic response function differs from BOLD, and latency correction strategies from EEG or ECoG don't translate directly. For a closely related view of fUSI's emerging BCI landscape, see the October 12 neurotechnology digest.

Tooling & Datasets

Neuralink's Brain Foundation Model: What 50,000 Hours Buys

Neuralink disclosed that it is pretraining a foundation model on 50,000 hours of intracortical recordings from its Neuralnauts — data collected during everyday BCI use rather than structured calibration sessions.[7] Six participants achieved their fastest-ever cursor-control benchmark scores using the pretrained model. Notably, Neuralink reports the foundation model has not yet beaten single-participant models in live use — a critical distinction between pretraining signal and deployment performance.

The data remain proprietary, but the architecture question — how to pretrain across electrode geometries, inter-session signal drift, and inter-subject variability — is an active open-science problem; EEG foundation-model work on generalizable preprocessing offers a useful public analogue. Public analogues like the Neural Latents Benchmark and MOABB-style public-dataset benchmarking workflows offer partial on-ramps for teams who want to explore similar approaches.

Industry & Ecosystem

China's BCI Standardization Push and Market Signals

China's draft standards roadmap targets at least 40 BCI standards drafted or revised by 2028, covering hardware, data formats, safety, applications, and governance — while ISO/IEC JTC 1/SC 43 is an international BCI standards subcommittee whose secretariat is held by China's SAC.[8] Meanwhile, Chinese neurotechnology activity continues across non-invasive EEG, consumer systems, and fUS BCI.[9]

The combination of a U.S. proposed outbound-investment restriction (Treasury's COINS Act NPRM, with comments due December 7) and China's broader market activity means the structure of global neurotech is being decided now, not in some future commercial phase.

Conclusion

This week's theme is scale meeting architecture. Precision Neuroscience's CNS atlas, Neuralink's 50K-hour pretraining corpus, and the fUS hardware bets from Merge Labs and Bridge Neurotech all point toward a field that is accumulating the data and modality diversity needed to train generalizable decoders — but hasn't yet cracked the deployment gap. The engineering questions that matter right now: How do you fine-tune a foundation model when electrode impedance drifts? Online Bayesian updates for neural drift offer one deployment-oriented direction. How do you validate an fUS decoder when hemodynamic latency varies with task? How do you build for a bifurcated US/China hardware supply chain? The answers will shape who leads the next generation of BCI products.

📄 Paper of the Week: Ho et al., "Toward A Whole-Brain Interface Enabled by High-Resolution Surface Electrophysiology," bioRxiv, 2026. doi: 10.64898/2026.09.24.754261

🛠️ Tool Worth Exploring: Neural Latents Benchmark — the closest public analogue to Neuralink's pretraining problem; structured splits, held-out sessions, and cross-dataset evaluation.

❓ Open Question for Next Week: Neuralink says its foundation model hasn't beaten single-participant models in live use yet. What would a convincing evaluation framework for intracortical foundation models look like — and who should define it?