She Left Google to Build Tech That Could Save Millions w/ Mary Lou Jepsen | EP #142
Summary
Openwater’s core bet is that consumer-electronics economics can turn medical machinery into a shared, software-defined diagnostic-and-therapy platform. Jepsen combines infrared light, ultrasound, electromagnetics, AI and commodity chips into what she calls a “silicon hospital.” Systems that began as room-sized, multimillion-dollar experiments became $10,000 modules she expects to approach $1,000 or smartphone-level cost, making treatment potentially “the cost of a phone call.”
The reported performance is striking, but the evidence spans very different stages and much of the therapeutic work remains preclinical. Jepsen says the optical system measures blood flow 20 times better than any multimillion-dollar MRI or CT result her team found in published literature. In organoids and mice, selected ultrasound frequencies attacked glioblastoma; in a 20-person severe-depression study, nearly half entered remission after brief sessions, but she explicitly describes several other applications as early laboratory work.
The platform’s multi-disease optionality comes from changing frequency, focus and software rather than developing a new molecule for every indication. Jepsen’s analogy is an opera singer breaking one wine glass while leaving everything else untouched: aggressive cancer cells, overfiring neurons and microclots may each respond to different resonances. Her early microclot result was an 80% clearance and a diameter reduction from 8 to 4 microns—material because capillaries are 5–10 microns wide—but she cautions, “this is just lab work.”
Stroke triage is Openwater’s most concrete diagnostic wedge and also its clearest demonstration of regulatory friction. With 151 patients studied at Penn and Brown, Jepsen says the optical device can identify large-vessel occlusion and distinguish mimics such as seizures, creating a case for placing it in ambulances and routing patients directly to thrombectomy-capable hospitals. The FDA requested 10,000 additional patients; at her cited $40,000–$70,000 per trial participant, the validation bill overwhelms a small company even when the prototype works.
Jepsen sees healthcare’s dominant constraint as Eroom’s law: development cycles and costs move opposite to Moore’s law. She cites 26 years and nearly $3 billion for a new drug, and an average 13 years and $658 million merely to win approval for a novel device—rising toward $1.5 billion by reimbursement and standard of care. That structure leaves roughly a million papers on therapeutic infrared light, ultrasound and electromagnetics producing what Diamandis calls “a rounding error to say that almost none” of the technology has made it into people or the healthcare system.
Openwater’s proposed escape hatch is an open-source, for-profit platform financed by a $50 million gift from Ethereum founder Vitalik Buterin. The company opened all 68 patents plus its hardware and software under the AGPL, with different diagnostics and therapeutics envisioned as software on a shared, Android-like base. Jepsen argues shared development and safety data could produce 10–100 times more revenue and margin than any other approach, while manufacturing competition supplies trust: “If we overcharge…they can go to another manufacturer.”
The execution case rests on Jepsen’s repeated record of shrinking “impossible” hardware, while the principal risk is translating demonstrations into scaled clinical evidence. Her path runs from micron-pixel holographic video and early smart-glasses displays to One Laptop per Child, Google, Oculus and finally Openwater. Her operating maxim is that consequential companies should be measured by people reached rather than employee count.
Deep dive
1. A brain-tumor diagnosis turned healthcare access into the mission
While pursuing her physics doctorate at Brown, Jepsen was in a wheelchair, sleeping 20 hours a day, unable to move half her face and eventually unable to subtract. Believing she no longer deserved the doctorate, she called her parents asking to “come home and die.”
A professor focused on her severe headaches and paid for an MRI that found the tumor. The scan required what Jepsen remembers as a roughly 20-by-20-foot shielded room, a large electromagnet, helium cooling and the hospital’s “most expensive room.” Diamandis compares it with equipment of the same size and shape that costs roughly 10 times more today.
She needed one operation and still takes roughly a dozen medications daily. The lasting motivation is less triumphalist than existential: patients often “have to fight for your life,” and surviving forces the question, “We’re here now—what do we want to do with our lives?”
2. Being told “impossible” became Jepsen’s operating signal
Presenting her proposed holographic-video research as a young MIT Media Lab student, Jepsen watched a Nobel laureate stand and dismiss it as “poppycock” that would never work. After retreating to her hotel in tears, she confronted him: saying impossible was insufficient; she wanted him to explain the actual physical obstruction.
Her adviser Steve Benton reframed the attack as jealousy—or evidence that the task was impossible for the critic, who might already have failed at it. In 1987, Jepsen built the world’s first fully computer-generated hologram with micron-size pixels using a Connection Machine, an early parallel supercomputer.
Jepsen and two other students received $4 million from DARPA to commercialize their PhD technology through MicroDisplay. Within a few years they had established manufacturing in Richmond, California, and shipped display hardware resembling Google Glass as early as 1998, with optics supplied by MicroOptical.
As a division CTO at Intel, she challenged its rail-to-rail silicon processes: displays need voltage gradations for grayscale, not merely zero or a fixed voltage. Jepsen says a two-sentence elevator exchange with the CEO exposed a silicon approach that could not deliver the needed gradation, saving hundreds of millions of dollars while making her deeply unpopular internally.
3. One Laptop per Child was a systems redesign, not a cheaper PC
Back at MIT, Jepsen became Nicholas Negroponte’s co-founder and the only other One Laptop per Child employee during its first year. They lived on planes, built the prototype and pursued a $100 computer when comparable laptops could cost $2,000 plus another $2,000 for software.
Millions were eventually produced through a multibillion-dollar nonprofit, open-source effort. The machine was not simply stripped down: Jepsen describes it as the lowest-power and lowest-cost laptop then made, the first with mesh networking, and usable without reading; the team also created keyboards for Amharic and other languages.
Her architecture treated the screen—not the CPU—as the center of the experience. “There could be little green men inside the laptop,” she jokes; if the display remained responsive to a stroke, most electronics could shut down and return within a single-digit number of milliseconds. The laptop offered better resolution than the Apple Retina display while reducing the energy burden for children without dependable power.
Jepsen went to BYD for lithium iron phosphate batteries and conditioned them for about 2,000 charge cycles—10–20 times the contemporary norm in her comparison. A charge lasted a day or two, while hand cranks and small solar panels offered alternatives; she says units remain in use nearly 20 years later.
4. Google and Facebook repeatedly deferred the healthcare moonshot
After OLPC, Jepsen founded Pixel Qi as a fabless display company, using Asia’s manufacturing infrastructure for laptops, tablets, phones and unconventional screens. Sergey Brin recruited the team into Google, where Jepsen arrived with brain-computer-interface and healthcare ideas but was redirected toward consumer-electronics projects for Brin and Larry Page.
Mark reacted enthusiastically to her whiteboard presentation on brain interfaces and healthcare. Once hired, however, she was asked to repair the recently acquired Oculus effort first. Jepsen invented sunglass-display systems she hopes will surface. She started in 2016 and left a year later; although she loved Google’s culture and Sergey, joining Facebook was financially lucrative.
Diamandis’s question about lightweight AR/VR produced a blunt answer: “It’s a matter of will.” After roughly $100 billion of spending, Jepsen finds it surprising how little laboratory technology has shipped and dislikes covering the face with “a giant mask—ski goggles.”
Jepsen’s institutional diagnosis is that executives optimizing advertising click-throughs were being asked to govern unfamiliar physics and hardware. Moonshots, she argues, have often come from focused teams—the Wright brothers, the birth-control pill, or WhatsApp’s roughly 50 people and $19 billion outcome—because politics and headcount are not substitutes for iteration.
5. Openwater applies wave control to cells, blood and neurons
Peter Gabriel supplied both the push and the name. After hearing Jepsen’s plan, he called repeatedly urging her to leave Facebook and build it independently, then wrote about thoughts “flowing like water” and the social challenge of radical transparency. He permitted the Openwater name; Jepsen says he has sweat equity and is also an investor.
The founding premise is that infrared light, ultrasound and electromagnetics penetrate the body, while Moore’s law enables sufficiently fine control over their phase and frequency. Openwater’s early experiments used room-sized systems and large optical tables that floated on air. The objective was to steer waves, make them interfere and selectively resonate with cellular structures.
Its ultrasound module uses an 8-by-8 transducer array and antenna theory to focus energy wherever selected. Jepsen repeatedly uses the opera-singer analogy: match the resonance of one wine glass, break it, and “harm nothing else in the room.”
This selectivity underpins her aspiration to kill cancer without killing healthy tissue, address strokes and pathogens, quiet pathological neuronal activity and eventually address neurodegeneration. Yet the breadth is an ambition, not a single evidentiary claim: she distinguishes strong hospital results, small clinical studies and “early” laboratory work.
6. Eroom’s law makes working prototypes commercially insufficient
Jepsen frames healthcare against a stubborn mortality mix: cardiovascular disease accounts for roughly 30% of deaths in her telling, cancer another 25%, with neurodegeneration, pathogens and chronic diseases taking much of the remainder. Diamandis mentions 55 million people dying globally each year.
Her cited capitalized pathway for a new drug is nearly $3 billion over 26 years. A novel medical device averages about $658 million and 13 years merely to secure FDA approval, then approaches $1.5 billion after reimbursement and adoption as standard of care. She calls these lengthening cycles “Moore’s law spelled backwards”—Eroom’s law.
Patient recruitment compounds the problem at $40,000–$70,000 each. Mental and neurodegenerative indications may require 10,000 or 100,000 participants, while rare-disease economics force enormous prices because a development bill of hundreds of millions must be recovered from only a few thousand patients.
Diamandis says a million scientific papers over 20 years describe infrared light, ultrasound or electromagnetic approaches across hundreds of diseases, yet “it’s a rounding error to say that almost none of this technology has made it into people or the health care system.” Better data and AI, Jepsen argues, could make treatment decisions and regulatory approval safer, but current trials cannot generate that scale quickly.
7. Miniaturization changes both unit economics and experimental velocity
Openwater moved from multimillion-dollar room-scale systems to hospital carts costing roughly $100,000–$500,000, then to compact optical and ultrasound modules priced initially around $10,000. Jepsen expects volume cost near $1,000 or a smartphone, saying the end-state could make each intervention cost about as much as a phone call.
The stroke-detection system uses an optical laser and high-quantum-efficiency camera chips that ship in smartphones and cost about $1 each. The demonstrated holographic module contains eight camera chips and lasers; because the pixels are near the wavelength of light, it can record phase information and reconstruct holographic measurements of blood flow.
Jepsen claims this detects blood flow 20 times better than any multimillion-dollar MRI or CT result her team found in the literature. The technology had already spent four years in hospitals, while the compact version was entering production around the time of the conversation.
The ultrasound console can pair with different 3D-printed mounts, including headsets or body arrays positioned behind the knee. A wearable MRI replacement remains on the back burner while faster-to-ship products are developed, but the larger vision is a “silicon hospital” whose physical platform gains new diagnostic or therapeutic functions through software.
8. Resonance produced provocative cancer and depression results
Jepsen’s cancer hypothesis exploits a mechanical feature she associates with aggressive, metastatic cells: an enlarged nucleus and small cytoplasm caused by rapid DNA replication and growth. Rather than poison the entire body, the team sought frequencies that would resonate with that structure while sparing surrounding neurons and healthy tissue.
Researchers grew 16 glioblastoma types in organoids and swept through multiple octaves and rhythms. They then tested leading settings in mice; Jepsen’s best-described protocol used a two-minute dose, another on day five, a 10% duty cycle and 150 kHz—“the frequency of a fish finder”—at diagnostic ultrasound intensity.
Jepsen says those treatments destroyed tumors and that Charles River autopsies found no damaged healthy cells, unlike chemotherapy, radiation or surgery. The crucial qualification is translational: the work had difficulty entering human trials because of safety requirements, even though glioblastoma itself is rapidly fatal.
In a separate University of Arizona study, Openwater targeted frontal overfiring visible on fMRI in 20 patients with severe depression. Five-minute sessions ran five days in week one and three days in each of the next two weeks; nearly half entered remission and, Jepsen says, remained there. She believes related targeting might address addiction, but presents that as a prospective extension.
9. Stroke triage exposes the cost of clinical validation
Stroke is the world’s second-leading cause of death in the episode’s framing, and large-vessel occlusion leaves only about two hours for intervention. Jepsen says just 5% of United States hospitals can perform the needed thrombectomy, yet ambulances generally deliver patients to the nearest hospital rather than the capable one.
Thrombectomy is fundamentally “a plumbing problem”: clinicians thread a catheter through the carotid artery and retrieve a clot too large for drugs to dissolve reliably. Delay can leave survivors unable to walk, speak or work because brain tissue dies.
In 151 patients studied in cath labs at Penn and Brown, Jepsen says Openwater’s optical unit achieved high specificity and sensitivity for occlusion while identifying mimics such as seizures and measuring capillary flow. Her deployment case is an ambulance diagnostic that chooses the right hospital and alerts its cath lab before arrival.
The FDA’s request for 10,000 more patients converts validation into a potential $400–$700 million exercise at her cited per-patient cost. Diamandis presses the paradox—patients face death or catastrophic disability—while Jepsen’s answer is not to dismiss safety, but to redesign how evidence is generated.
10. Open source is the financing, distribution and trust mechanism
Vitalik Buterin initially contacted Jepsen about COVID and spent a series of Friday-night calls asking questions that consumed her weekends. She cites data from 54,000 veterans suggesting long COVID doubled neurodegenerative risk and raised heart-failure and stroke risk by 173% and 164%, motivating work on blood flow and microclots.
In early laboratory work, ultrasound cleared 80% of what Jepsen calls amyloid microclots and reduced their diameter from 8 to 4 microns. With capillaries only 5–10 microns across, that could determine whether clots pass, but she is explicit: “Again, this is just lab work we’re doing.”
Buterin ultimately gave $50 million using Shiba Inu coin rather than Ethereum, after discussions about opening the platform. Openwater released all 68 patents plus its hardware and software under the AGPL, converting proprietary assets into a base others can manufacture, investigate and extend.
Jepsen envisions different diagnostics and therapeutics as software on a shared platform, with safety evidence and development costs distributed across uses. She attributes 85% of medical-device approval cost to device development. Her investors initially equated open source with charity; her model instead projects 10–100 times more revenue and 10–100 times more margin than any other approach, while OLPC taught her that selling strictly at cost leaves even a transformative system unsustainable.
11. The endpoint is a shared brain instrument, not one approved device
Openwater has begun “writing” to neurons while addressing mental disease, but Jepsen separates that from decoding thoughts, which comes later. She said she showed live at TED—she thinks in 2018—how the system could focus through phantom bone and flesh to roughly one micron, while groups of neurons are the nearer-term target for mental and neurodegenerative disease.
Diamandis extrapolates toward home applications for mood, sleep and other brain-mediated functions. Jepsen’s more immediate case is distributed research: ministries of health could deploy 10,000–100,000 devices, let volunteers join trials from home and potentially own the resulting regulatory approval instead of depending on multinational pharmaceutical companies.
Shared hardware could generate far more safety, efficacy and biological data for AI than today’s tiny studies. Diamandis contrasts one ten-year-old company’s 76 patients with consumer wearables that may be accurate only to plus or minus 25% individually yet become informative across millions: scale itself becomes an experimental asset.
Her founder advice is to “read history,” search 20–50 years backward for abandoned approaches, then recombine them with present capabilities. Measure size by impact rather than headcount, use contract manufacturing instead of owning every factory, and choose work compelling enough that you cannot stop.