Sabi raises $50 million for a brain-sensing baseball cap

By Inside BCI · October 11, 2026

Khosla Ventures has led a $50 million seed round in Sabi, a California startup building a baseball cap meant to turn words a wearer imagines, and trained mental commands, into text and instructions for AI assistants. Forbes, citing a person familiar with the deal, put the valuation at $600 million, and reported separately that the cap will not be available until 2027. The cap has no price and no peer-reviewed accuracy data, and its first public showing is planned for CES in Las Vegas in January.

In April, when it came out of stealth, Sabi said its first product would be a thought-to-text beanie available by the end of 2026. The cap now comes first.

The round

Accel, Initialized Capital, former OpenAI chief product officer Kevin Weil, DST Global, Collaborative Fund and Ascend joined Khosla Ventures in the round, announced on 9 October. Sabi has not confirmed the valuation. The money will go on expanding its neural dataset, finishing its custom chip, manufacturing early devices and hiring in neuroscience, hardware and machine learning, and Sabi is preparing units for people on its early-access waitlist.

Vinod Khosla, the firm’s founder, argues that a billion keyboards cannot be replaced with a billion brain surgeries. He expects early versions to make mistakes and improve over time: “The error rate may not be zero in the beginning.”

A useful comparison is Merge Labs, the brain-computer interface company co-founded by Sam Altman, which came out of stealth in January, backed by OpenAI, with a seed round of $252 million at a reported $850 million valuation. Merge describes itself as a research lab working on a horizon of decades, and says its first products will help patients with injury or disease. Sabi is aiming at a consumer device for everyday use with AI. Neither has shipped a product.

What Sabi is building

Sabi says the cap looks like an ordinary baseball cap from the outside, with sensors inside built around its own chip for electroencephalography, or EEG, which records the brain’s electrical activity at the scalp. The sensors are designed to work through hair, without gel, surgery or direct contact with the skin, and Sabi says the chip, still in development, uses little enough power for the electronics to fit inside ordinary headwear. The cap is designed to pair wirelessly with a phone or computer.

Sabi is developing three forms of interaction. Words a wearer deliberately imagines could appear as text on a connected screen. A trained mental command could become an instruction for an AI assistant, such as drafting a message. And the system could anticipate likely keystrokes or on-screen actions and offer them for the user to confirm. Sabi plans to show early versions of all three at CES.

Decoding is meant to run on what Sabi calls its Brain Foundation Model, a large AI model that Sabi says was trained on more than 100,000 hours of labelled recordings taken while people imagined words and performed set mental tasks, the largest disclosed dataset of its kind by the company’s count. To build it, Sabi hired more than 100 contractors to wear headsets while reading, speaking, thinking and listening. It has not said whether any recordings were made with the cap itself.

What has changed since April

In April, Sabi described a beanie carrying 70,000 to 100,000 miniature sensors, with a baseball cap version to follow, and a target of about 30 words per minute. Khosla Ventures was already backing the company, alongside Accel, Initialized Capital and Kevin Weil, but no amount was given.

The funding now has a size and three more named backers, and the training figure is unchanged at about 100,000 hours. The October announcement gives no sensor count, describing each sensor as one to five millimetres across, and no typing speed. Forbes puts the prototype at up to 100,000 sensors.

The evidence so far

Rahul Chhabra, Sabi’s co-founder and chief executive, says an early, wired, helmet-like prototype in the laboratory can transcribe people’s thoughts with up to 77 per cent accuracy for those who have used it for more than two weeks, and that accuracy rises as the model learns each user. Sabi has not said how accuracy was measured, over what vocabulary, or what a random guess would score.

Nicholas Hatsopoulos, a University of Chicago professor who studies neural decoding for brain-computer interfaces, says signals filtered through the skull and scalp are of lower quality. Writing a good AI prompt already takes effort, he says, and he doubts one can be drawn reliably from the brain.

The founders

Chhabra, the chief executive, worked on AI research at Stanford. Atmadeep Banerjee, the chief technology officer, is co-first author of MindEye, a 2023 paper at NeurIPS, a leading AI research conference. The two were roommates at BITS Pilani in India and founded Sabi in May 2024. Sabi says its team includes hardware, machine-learning and design leaders from Kernel, Apple, Microsoft, Meta, Adobe and Nike, and neuroscientists from Stanford.

MindEye decoded images people were looking at from fMRI, which tracks changes in blood oxygen inside a large scanner. Sabi’s cap aims to decode words people imagine from EEG at the scalp, so the earlier result does not show what the cap can decode.

Neural data and California law

Sabi says neural data is encrypted from the sensor onwards and that its system is designed to decode signals while the data stays encrypted, so that its servers do not receive a user’s unencrypted neural signals. It is also building controls over when the cap records, which applications can receive commands, and which actions need confirmation. Text and commands passed to an AI assistant would still be readable by whoever runs that assistant.

California has treated neural data as sensitive personal information under its consumer privacy law since January 2025, for the businesses that law covers. From 1 January 2027, AB 1883 also bars employers from using AI-enabled workplace surveillance tools to collect employees’ neural data, with exceptions that include tools used to ensure safety. An employer that issued Sabi caps to California staff as a keyboard replacement could fall within it.

Not disclosed

Sabi has not disclosed a price, a date when customers can buy the cap, or how many units it will send to waitlist members. It has not published peer-reviewed results, a typing speed, or whether it still targets 30 words per minute. It has not said whether any decoded text has yet come from the cap itself, how long a new user must train, how many sensors the production cap will carry, how its 100,000 hours were counted, how its encrypted decoding works, how it would respond to a law enforcement request for neural data, or which regulatory route, if any, the cap will take.

What to watch

Whether the CES showing in January turns imagined words into text on a cap worn outside the laboratory, and at what speed and accuracy.

Whether Sabi publishes peer-reviewed results or technical detail on its dataset and encrypted decoding.

Whether Sabi names a price and ship date, and when early-access units reach the waitlist.

Whether employers pilot the cap, and how they treat California’s AB 1883 from 1 January 2027.

Sources

Primary:

Coverage:

Cross-reference to prior InsideBCI coverage:

market Sabi Khosla-Ventures Vinod-Khosla Rahul-Chhabra Atmadeep-Banerjee Accel Initialized-Capital Kevin-Weil DST-Global EEG consumer-BCI non-invasive-BCI imagined-speech brain-foundation-model CES-2027 Merge-Labs AB-1883 neural-data funding