On October 9, 2026, Sabi raised a $50 million seed round led by Khosla Ventures to build a non-invasive brain-computer interface (BCI) in the form factor of a baseball cap: a custom EEG chip, a dataset of 100,000+ hours of labeled neural recordings, a claimed 50x power-efficiency advantage over the industry-standard TI ADS1299 chip (per the company's internal benchmark), and a prototype targeted for CES 2027.
It's a bold bet on one side of the most important divide in neurotechnology: non-invasive versus implantable BCIs. This article maps the spectrum, explains the physics, and lays out what "imagined speech decoding" really means — plus what investors should diligence before writing checks.
The BCI spectrum
Invasive BCIs place electrodes directly on or in the brain. They capture high-resolution signals from individual neurons, enabling precise control — moving cursors, robotic arms, even decoding attempted speech in paralyzed patients. The cost: brain surgery, with its risks, regulatory burden, and near-zero consumer appeal.
Non-invasive BCIs read signals from outside the skull — typically EEG (electrical activity via scalp electrodes), fNIRS (blood-oxygen changes), or related modalities. No surgery, but the signal must travel through skin, skull, and cerebrospinal fluid, each of which smears and attenuates it.
Between them sit minimally invasive approaches (e.g., endovascular stentrodes threaded through blood vessels), trading some signal quality for avoiding open-brain surgery.
Sabi is betting that non-invasive is where the mass market lives — because a baseball cap has roughly a billion times the addressable market of a neurosurgical implant.
The physics bottleneck
The fundamental challenge of scalp EEG is signal attenuation. Electrical signals from the cortex spread as they pass through tissue, so by the time they reach scalp electrodes, the spatial resolution is coarse — centimeters, not the millimeters or single neurons of implants. Hair, sweat, and motion add noise.
This is why claims in this space need careful parsing. A custom chip with 50x better power efficiency than the TI ADS1299 (the workhorse EEG front-end chip) is a real engineering achievement if it holds up — power is the binding constraint on wearable neural devices. But better sensing doesn't change the physics of what the skull lets through. The question is always: given the signal available at the scalp, what can you actually decode?
That said, the field has genuine tailwinds: better electrodes (dry-contact, in-ear, tattoo-like), better algorithms (deep learning on large neural datasets — Sabi's 100,000+ hours of labeled recordings is exactly the kind of moat that matters), and better chips. The trajectory is real even if any single claim deserves scrutiny.
What "imagined speech decoding" really means today
Headlines about decoding "imagined speech" deserve a decoder of their own. Current capabilities vary enormously by modality:
- Implantable systems have decoded attempted speech in clinical trial participants with impressive (though still limited) accuracy — real science, narrow population.
- Non-invasive systems have demonstrated coarse classification: distinguishing a small vocabulary of imagined words, or detecting speech vs. silence, under controlled conditions. This is genuine progress, but it's closer to "neural button presses" than telepathy.
The gap between "we classified 10 imagined words in a lab" and "you can think-type at conversational speed" is where most hype lives. Investors and journalists should ask for the vocabulary size, the accuracy, the conditions, and whether results replicate across users without per-person calibration.
Why wearability is the commercialization unlock
Even modest non-invasive capability becomes commercially interesting if the form factor disappears into daily life. A cap you wear anyway is a fundamentally different product than a headset you strap on for a "brain session." The input-methods opportunity is the sleeper: as AI agents become the primary computing interface, hands-free, silent neural input — even at low bandwidth — could be transformative for accessibility first, then for everyone.
What investors should diligence
- Dataset scale and labeling quality. Neural decoding is a data problem. 100k+ hours of labeled recordings is a serious asset — ask how it was labeled, how diverse the subjects are, and whether models generalize to new users.
- Independent validation. Internal benchmarks against the TI ADS1299 are a starting point, not proof. Ask for third-party or peer-reviewed results.
- The regulatory path. A wellness device (focus tracking, meditation) faces a radically lighter burden than a medical device (diagnosing conditions, controlling prosthetics). Know which lane the company is in — and whether its ambitions require switching lanes.
- The actual decode task. Vocabulary size, accuracy, latency, calibration requirements. Insist on specifics.
Startup white space
Beyond the core BCI hardware race: neural input methods for AI agents, accessibility-first applications (where even low-bandwidth control is life-changing), developer tooling for neural data, and hybrid systems combining EEG with other signals (eye tracking, EMG) to get usable input today rather than waiting for physics to cooperate.
Why it matters for founders: If you're building in neurotech, pick your lane deliberately — wellness vs. medical is a regulatory chasm, and form factor is destiny. The most defensible moat is proprietary neural data at scale, not any single chip.
Why it matters for investors: Non-invasive BCI is moving from science project to venture-backable, but the diligence bar must stay high: datasets, independent validation, and honest decode-task specs. The winners will likely be data companies wearing hardware costumes.
Angeled.com connects founders with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



