Neuralink’s N1 implant places 1,024 electrodes on 64 ultrafine threads into the motor cortex, where they record the spiking of individual neurons. Those spikes are meaningless until a machine-learning decoder maps them to intent — a cursor vector, a keystroke, a spoken phoneme, or a robotic joint command. The decoder is not a fixed program; it is a statistical model trained on each user’s own neural patterns, and it must be recalibrated because neural signals are non-stationary: the same intention produces subtly different spiking day to day. In Neuralink’s first participant, Noland Arbaugh, many threads retracted from the cortex in the weeks after his January 2024 implant, cutting the usable signal; the company recovered performance largely through software and decoding-algorithm changes rather than reoperation. That episode is the clearest proof that in a brain-computer interface, the model — not just the hardware — carries the function.
As of 2026, Neuralink’s clinical footprint spans several programs. PRIME (motor control, U.S.) has been extended internationally as CAN-PRIME (Canada, from late 2024), GB-PRIME (United Kingdom), and UAE-PRIME with Cleveland Clinic Abu Dhabi. CONVOY tests control of an investigational assistive robotic arm, with at least one PRIME participant enrolled. VOICE targets speech restoration and received an FDA “Breakthrough Device” designation in 2025; Blindsight, a visual-cortex stimulation implant, received the same designation in September 2024. Secondary trackers put implants at roughly a dozen by late 2025 and around 21 enrolled participants by early 2026, though Neuralink has not published a single official running total. In June 2025 the company closed a $650 million Series E at a reported $9 billion pre-money valuation, led by ARK Invest, Sequoia, and Founders Fund. Neuralink also uses machine learning and computer vision in its R1 surgical robot, which threads the electrodes with sub-millimeter precision.
Where does the decoding actually run today? On the participant’s own device — the Neuralink app on a paired laptop or phone in the person’s home. That architecture is not incidental; it is the future of the field, and it points directly at why local, owner-controlled AI compute matters. First, decoders drift and need frequent personalized retraining. A home GPU appliance can retrain a user’s decoder overnight on that day’s neural data, so the model wakes up as sharp as the person, without shipping raw brain signals to a distant server.
Second, latency and reliability are existential when an implant is your voice or your hands. Someone who depends on a decoder to communicate cannot have their speech throttled by a congested internet link or silenced by a cloud outage. Third, neural data is now legally sensitive: Colorado’s HB 24-1058 (signed 2024) and California’s SB 1223 (signed September 2024, effective January 2025) both classify neural data as sensitive personal information, Montana’s law took effect in October 2025, and internationally Chile adopted a neurorights constitutional amendment in 2021 while UNESCO advanced a global neurotechnology recommendation. Keeping the model and the data in the home is the cleanest way to honor those protections.
Fourth, real assistance is multimodal — fusing neural signals with cameras, eye tracking, EMG, and smart-home control — and that fusion is compute-hungry and latency-sensitive, which favors a local core over the cloud. Fifth, the same home cortex that runs a communication decoder can drive an assistive robot arm or exoskeleton, exactly the direction CONVOY points. Safeera’s Googie is built for this role: sixteen RTX PRO 6000 Blackwell GPUs, liquid-cooled and quiet enough for a living room, giving a BCI user a private, always-on, switch-off-able place to train and run the model their independence depends on.