Jul, 134 min to read
Frontend
News · OpenAI backs Merge Labs' seed round to build brain-computer interfaces
Jul, 84 min to read
FrontendOpenAI backs Merge Labs' seed round to build brain-computer interfaces
The investment frames BCIs as an interface problem, with the model layer positioned as the thing that reads intent from noisy neural signals.
The interface, not the device, is the pitch
OpenAI opens its announcement with a claim about computing history rather than neuroscience: "Progress in interfaces enables progress in computing. Each time people gain a more direct way to express intent, technology becomes more powerful and more useful." That sentence tells you how OpenAI is reading this investment. Merge Labs is described as building brain-computer interfaces, but the company is treating the BCI as the newest member of a lineage that runs through keyboards, mice, and touchscreens — a more direct channel for expressing intent.
That framing matters because it separates two things people often blur together. There is the hardware problem of interfacing with the brain safely and at higher bandwidth, which the source attributes to Merge's combination of "biology, devices, and AI." And there is the software problem of turning whatever comes off that hardware into something a machine can act on. OpenAI is explicitly claiming the second layer.
AI as the operating system for noisy signals
The most concrete engineering claim in the announcement is this: "high-bandwidth interfaces will benefit from AI operating systems that can interpret intent, adapt to individuals, and operate reliably with limited and noisy signals." Strip away the frontier language and that is a description of a hard inference problem. Neural signals are sparse, individual-specific, and full of noise; converting them into reliable intent is closer to what large models already do with ambiguous inputs than it is to conventional device firmware.
This is where the "frontend" reading becomes literal. In an ordinary application, the frontend renders state and captures deterministic events — a click, a keystroke. A BCI frontend has no deterministic events. The input is a probability distribution over what the person might mean, adapted per user, degrading and recovering over time. OpenAI is positioning its models as the runtime that resolves that distribution into action. That is a very different job from powering a chat box.
What OpenAI is actually contributing
The announcement names two contributions beyond capital. First, AI to "accelerate research and development including bioengineering, neuroscience and device engineering" — models applied to the lab's own science. Second, and more specific to the interface, OpenAI says it "will collaborate with Merge Labs on scientific foundation models and other frontier tools to accelerate progress." These are distinct bets: one uses AI to build the device faster, the other builds AI into the device's ability to understand its user.
The founding team reflects the same split. The researchers — Mikhail Shapiro, Tyson Aflalo, and Sumner Norman — are credited with pioneering "entirely new approaches to BCI," the biology-and-devices side. The entrepreneurs, including Alex Blania, Sandro Herbig, and Sam Altman "in a personal capacity," cover the company-building side. Notably, Altman's involvement is personal while OpenAI's is corporate, which the source is careful to distinguish.
The implication: OpenAI is treating input methods as within its scope
For teams building on OpenAI's models, the signal here is not that neural interfaces are imminent. It is that OpenAI is defining its territory to include how humans express intent to machines in the first place — not just what happens after a prompt arrives. The company frames BCIs as "a natural, human-centered way for anyone to seamlessly interact with AI," which puts the model at both ends of the interaction: interpreting the input and generating the response.
If that thesis holds, the interface layer stops being a downstream integration detail and becomes something a foundation-model provider wants to own. Anyone designing intent-capture for AI systems — even conventional ones — should read this seed check as a statement about where OpenAI thinks the interpretation of intent belongs, and plan accordingly for a world where the model, not the application, is expected to make ambiguous human input legible.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.