An AI system can know a remarkable amount about heart disease and still have no idea what your heart is doing right now.
That gap gets less attention than the models themselves. We talk about medical exam scores, diagnostic reasoning, and whether a chatbot can explain a confusing lab result. But every useful answer depends on something more basic: collecting the right information about the person asking.
Butterfly Network is betting ultrasound can help close that gap.
The company is best known for a handheld ultrasound probe that connects to a mobile device. Underneath that product is a semiconductor platform that Butterfly is opening up to developers and other device makers. Its ambition reaches beyond making an existing scan more portable. It wants to make imaging data easier to collect, build around, and use throughout care.
In an interview with The Neuron, Steven Cashman, Butterfly’s chief business officer, described the opportunity this way: “Fundamentally what we do is we bring different data to the table earlier or more frequently than you would have used ultrasound for.”
That's a useful way to think about the next phase of medical AI. Better algorithms matter. So does giving those algorithms something useful to examine.
An image collected during an appointment can help answer a question about that moment. Measurements collected over time could help reveal how something is changing. Imaging available during a procedure can help a clinician act on what is happening in front of them.
Each requires hardware, software, and a workflow that makes collecting the data practical.
Butterfly’s pitch is to supply more of that foundation.
Its Ultrasound-on-Chip platform uses microscopic electromechanical elements on a semiconductor chip. Through Butterfly Embedded, partners can incorporate that technology into new devices. Through Butterfly Garden, developers can build software around the company’s probes.
Conventional ultrasound already uses digital processing. Butterfly’s argument is about the compact, programmable semiconductor architecture and what developers can build with it.
That distinction matters because “more data” can mean several different things. There is the information used to train a model, the measurements that model receives when someone uses it, and the software connections that make those measurements available.
Butterfly is working on the collection and access side of that equation. Whether the resulting information supports a useful medical decision depends on the application.
Cashman came to that opportunity from the other direction.
Before joining Butterfly, he ran Caption Health, which developed AI for cardiac ultrasound. He needed imaging hardware that could work with the software his team was building. According to Cashman, he paid Butterfly to develop the SDK that became the foundation for Butterfly Garden.
“I wanted to put my AI on their chip and on their software,” he said.
There is a practical lesson in that origin story. A promising medical algorithm still needs a way to acquire usable information and reach the people who need it.
AI can help with acquisition, too. The FDA’s review of Caption Guidance describes software that guides users toward standard cardiac ultrasound views. That is a distinct task from interpreting an image or diagnosing a condition: helping someone collect an appropriate image in the first place.
Butterfly’s current developer tools give applications access to captured images and controls such as depth and gain. Developers can build acquisition guidance, measurements, and interpretation tools around those capabilities.
The opportunity extends across the process, from positioning a probe to presenting a result a clinician can use.
And it is starting to show up in products beyond Butterfly’s own handheld device.
On October 1, Butterfly and Mendaera announced that Mendaera’s Focalist system is commercially available with Butterfly imaging technology. Focalist combines handheld robotics, real-time ultrasound, and guidance software for needle procedures. The companies describe it as the first Butterfly Embedded partner technology to reach the market.
That is a concrete example of imaging becoming part of an action. A clinician needs to place an instrument accurately, and the system provides information and guidance during the procedure.
The announcement describes a path toward further AI-enabled interpretation and robotic action. Those future capabilities should be distinguished from what the product does today. Still, the underlying direction is clear: medical devices can become more useful when they can collect information about their immediate surroundings.
Midjourney offers a much stranger-looking example of the same idea.
Yes, that Midjourney.
Its medical division is developing a scanner that uses ultrasound transmitted through water to reconstruct images of the body. Cashman described founder David Holz’s approach as: “Let’s pipe the data into the GPU farm.”
The connection is easier to understand once you look past the company’s image-generation reputation. Reconstructing an image from physical measurements presents a substantial computational problem.
Midjourney says the images in its announcement came from real volunteers. Its initial product is intended for body-composition analysis, rather than diagnosing or treating disease. The company’s updated FAQ places its first opening target in early 2028 and says performance validation is still underway.
Those limits matter. An ambitious scanner demonstration does not establish a replacement for existing diagnostic imaging.
But the project illustrates what Butterfly’s strategy could enable: another company taking its imaging technology and building a different collection system, user experience, and computational pipeline around it.
This is where Cashman brings in Nvidia.
“I think of Butterfly a lot like the NVIDIA of medical imaging or medical device technology,” he said.
The useful part of that comparison is the platform model. Hardware becomes more valuable when other developers have tools to build applications around it. Butterfly wants its chip and software to support products its own team would never develop alone.
That ambition still has to earn its place in healthcare. Partners need evidence that their applications work, a viable route to market, and a reason for clinicians or consumers to use them. A medical platform succeeds through those products and their outcomes.
Cashman sees the potential payoff in moving imaging earlier in care.
In the interview, he described checking a patient’s heart or lungs near the beginning of an emergency-department visit, bringing imaging closer to the initial assessment. Longer term, he envisions more opportunities to monitor changes outside traditional imaging appointments.
The appeal is straightforward. Information available sooner could help people reach an appropriate decision sooner.
The harder question is which information helps, for whom, and under what circumstances. More frequent measurement creates value only when the measurements are reliable and the resulting decisions improve care. It also needs to fit into the work clinicians already do.
As our coverage of AI navigating healthcare’s workflows has shown, getting useful technology into the care process presents its own substantial challenge.
Safety and interpretation remain part of that process. Ultrasound does not use ionizing radiation, but the FDA notes that it can produce biological effects and recommends prudent use. Different applications require their own evidence. Access to a scan also does not automatically make a consumer chatbot a validated diagnostic tool.
Cashman acknowledges uncertainty about how imaging information will eventually reach consumers. His broader argument is that patients could make more informed decisions when they can better understand what is happening inside their bodies.
That possibility gives the technology stakes beyond a smaller probe or a clever algorithm.
The Neuron has followed the growing effort to make AI useful in science and healthcare. Butterfly’s story adds a physical dimension to that effort: building the instruments that let software observe something it otherwise could not.
Cashman calls the coming years “the decade of ultrasound.”
The more immediate test is whether Butterfly and its partners can collect useful information at the right time, interpret it reliably, and turn it into a better decision.
For medical AI, that work begins before the model answers. It begins with what the system can see.
Related reading
- AI Agents Are Starting to Navigate Healthcare’s Messiest Workflows — Explores the operational work required to connect patients with care.
- The AI Battle Reshaping How Doctors Document Patient Visits — Examines another way AI turns information collected during care into usable records.
- OpenAI’s Science Week, Explained — Places this story within the broader effort to build AI tools that deliver useful results in science and health.
