Sabi wants to turn deliberate thoughts into AI commands through an everyday cap. The funding is large and the product is still in development. Here is what the money should produce, and what the evidence shows so far.
The funding
Source: Sabi company website (sabi.com)
On October 9, 2026, Sabi, a non-invasive brain-computer interface startup, announced a $50 million seed round led by Khosla Ventures. Accel, Initialized Capital, and DST Global also took part, among others. Former OpenAI chief product officer Kevin Weil is on the investor list. Forbes reports the company is based in Palo Alto.
The idea is simple to state. Today, when you want AI to do something, you type or speak. Sabi wants a cap that picks up brain signals when you deliberately form an instruction and turns them into input a computer can receive.
Nobody can buy the product yet. That makes $50 million at the seed stage a large check for an unproven device.
What the money should turn into evidence
Where the $50M is going, and the open questions it has to answer
The question for a seed-stage company is which uncertainties the money can resolve. For Sabi, the main ones are:
- Can the chip capture signals consistently?
- Can the model adapt to different people?
- Can the cap be comfortable enough to wear?
- Will people want to keep using it?
The announcement says the funds will go toward a larger neural dataset, finishing a custom chip, building early devices, and growing the team. These are the steps between a research prototype and a usable product.
The possible use cases are easy to picture. Your hands might be busy. You might be somewhere you can't talk out loud. But these are potential uses, not proven ones. Only real users can show whether a cap beats typing or speaking.
The team: research is not the product
MindEye and the Sabi cap differ in data, equipment and task
Sabi has two cofounders: CEO Rahul Chhabra and CTO Atmadeep Banerjee. Banerjee is a co-first author of MindEye, a NeurIPS 2023 paper.
That background matters, but it needs context. MindEye used fMRI (functional magnetic resonance imaging) data to retrieve and reconstruct images. It shows the team can combine brain data with machine learning.
Sabi's cap uses EEG (electroencephalography) and is meant for everyday wear. The data, the equipment, and the task are all different. The research says something about the team's ability. It says nothing about how the cap performs.
The same applies to the investor list. It shows Sabi has backing. It doesn't show how much each firm invested, and it doesn't mean any investor has independently validated the product.
How the technology is supposed to work
Sabi's stated three-stage design and the open question at each stage
The system can be broken into three stages.
1. Capture the signal. EEG records electrical signals linked to brain activity. According to the company, Sabi puts custom chips and tiny sensors inside a cap that reads signals through hair. "Non-invasive" here means no surgery, no conductive gel, and no direct scalp contact. This is the company's stated design.
Everyday wear is the hard part. Blinking, facial muscles, and body movement can all distort EEG recordings. Researchers call this interference "artifacts," and handling them is a known signal-processing problem. That is a general challenge for EEG, not a reported flaw in Sabi's system.
2. Decode the intention. Sabi feeds the signals into what it calls a Brain Foundation Model. Its website describes the sequence: sensors collect data, the data trains a model, and the model maps signals to outputs.
The practical questions are specific:
- How long does a new user need to calibrate?
- Does a model trained today still work when the same person puts the cap on tomorrow?
- What happens when fit, hairstyle, or activity changes?
These answers decide whether the device works outside a lab.
3. Hand the result to AI. Once the model thinks it has recognized an intention, when should the system act? A sensible design would show a candidate command and let the user confirm or fix it first. A language model can make a command read smoothly, but smooth output doesn't mean it kept what the user meant. The input layer needs a way to correct and cancel.
The 77 percent figure
The 77% figure and the conditions that limit how it can be read
This is the number most likely to be quoted, so it needs care.
In the Forbes report, the CEO said an early, wired, helmet-like prototype reached up to 77 percent transcription accuracy in lab conditions, for users who had used it for more than two weeks.
It is a specific figure, not a vague claim. But it comes with conditions:
- Lab setting
- Early, wired, helmet-style prototype, not the cap
- Users with more than two weeks of practice
- An "up to" result, not an average
- A founder's statement to a journalist, without a published test protocol
It cannot be read as the everyday accuracy of a wireless cap. On its own, it also doesn't tell us whether the product is useful. We would need to know how accuracy was measured, whether vocabulary was limited, how many people were tested, and how failures were recorded.
What would make it a business
Counting the whole interaction, and the reinforcing loop that would still need numbers
A first product needs one task people want to repeat. "Everyone needs a more natural way to talk to AI" is a vision, not a use case. The narrower question is where the cap saves time, cuts errors, or makes something possible that wasn't before.
That means counting the whole interaction: putting the cap on, calibrating, waiting, confirming, and fixing mistakes. Fast recognition won't help much if setup takes minutes every time.
Public materials don't show pricing, paying customers, or unit economics. Hardware sales, software subscriptions, or developer access are all possible paths. For now, they are hypotheses.
On defensibility, the question is whether the chip, the data, and the user experience reinforce each other. Better capture could yield better training data. A better model could shorten calibration. A better experience could keep people wearing it. Each of these is a "could" that needs numbers behind it.
A large dataset is not automatically a moat. It matters whether the data covers different users and conditions, whether it reduces errors, and whether data permissions allow long-term training and product use.
The tests that matter
Six repeatable tests a credible evaluation would include
The gap between a good demo and a usable product is closed by repeated testing, not a single showing. A credible evaluation would:
- Time a new user's first setup
- Bring users back the next day to see if they must recalibrate
- Let users move normally and record how long the device works and when it fails
- Measure correct commands per minute and the time needed to fix one error
- Track false activations, when the system acts without the user meaning to give a command
- Have an independent team reproduce results under clearly described conditions
One smooth minute in a demo says little about the rest of the day.
What to watch next
Three questions for Sabi's planned CES 2027 prototype showing
Sabi says it plans to show early interaction prototypes at CES 2027. When that demo happens, three questions matter more than how impressive it looks:
- How quickly can a new user get started?
- How many errors occur in ordinary conditions?
- After correcting those errors, is it still easier than typing or speaking?
Sabi has a direction worth following and the money to pursue it. Whether it becomes a product people keep wearing depends on evidence that doesn't exist publicly yet.
