Neuralink Brain Data Analysis: Processes and Insights

• Neuralink's implant records electrical signals from 1,024 electrodes in the motor cortex, detects neuron spikes on its chip, and sends the data wirelessly to a computer.
• Machine learning decoders turn those signals into cursor movement, clicks, and typing, and the median participant now reaches about 10 bits per second.
• In 2026, Neuralink pretrained its decoders on over 50,000 hours of participants' neural data, which cut calibration time for some users.

Recording the signals

Neuralink's N1 implant records brain activity through flexible threads in the motor cortex: 64 threads with 16 electrodes each in early implants, and 128 thinner threads with 8 electrodes each in later ones, for 1,024 channels in both. The electrodes pick up voltage changes from nearby neurons. The implant amplifies these microvolt-level signals and detects spikes on its chip before sending the data over Bluetooth to a nearby computer or phone.

This gives far more channels than most earlier brain-computer interfaces, which used fewer, larger electrodes. In human trials, spikes appear on screen soon after surgery. Neuralink's roadmap called for about 3,000 channels in 2026, and a 3,000-channel implant was disclosed in September 2026.

Finding the spikes

A spike is a brief voltage peak that shows a neuron firing. The processing has to separate those peaks from noise caused by heartbeat, muscle activity, and the implant moving slightly in the brain. What remains is a record of when neurons near each electrode fired, which the decoder uses as its input.

Decoding intent

Decoding is where machine learning models translate firing patterns into what the user intends, such as moving a cursor up and to the right. Models are trained during calibration sessions, in which the participant imagines movements while the software learns which patterns go with which actions. The models also have to keep adapting, because the signals change from day to day.

Neuralink measures performance in bits per second (BPS), a measure of how much information a user can send. Able-bodied mouse users average about 8 to 10 BPS, and the median Neuralink participant is now about 10 BPS, with a record of 11.32 BPS, according to Neuralink's October 2026 update.

The same update says users averaged 55 minutes a week of calibration, and that its 2026 decoder update, pretrained on over 50,000 hours of neural data, cut that to as little as 10 minutes a week for some users. Decoders also handle gaming controls, typing, and, in the CONVOY study, a robotic arm, as participant Alex Conley has shown.

What the trials have shown

PRIME Study participants averaged about 50 hours a week of independent use, and some topped 100 hours (Summer 2025 update). UK patient Paul controlled a computer hours after surgery and has played games like Dawn of War. Participants use Neuralink's Webgrid game as a standard test of cursor speed, and their scores have improved as the decoders improved.

Neuralink reported zero serious device-related adverse events as of January 2026.

What comes next

Neuralink's roadmap targets about 10,000 channels in 2027 and 25,000 in 2028. More channels give the decoders more information, which should allow faster and more precise control. Neuralink also uses simulations of motor cortex activity to test decoder designs before they reach participants.

This article was first published in October 2025 and corrected in October 2026; the latest patient count is on the patient table.