Butterfly Network

The moat is silicon, and a substantial share of the clinical intelligence running on it belongs to other people.

The scarce asset is the chip. Replacing piezoelectric crystals with a semiconductor is what lets one handheld device image the whole body, and it is what competitors cannot copy without their own fabrication programme. Strip every model out and a differentiated imaging device remains, which is the test this index applies.

The company's own inference is nonetheless real and in one case genuinely novel. The blind sweep gestational age tool cleared in early 2026 is the substantive example, because it removes the requirement that the operator know how to obtain standard views. That is not an incremental measurement aid, it is a model absorbing the skill that previously gated the examination.

What holds the grade at C is the partner ecosystem. Thirty developers deliver their own applications on this hardware, four of them holding their own regulatory clearances. That is a deliberate and sensible strategy, and it means a meaningful portion of what a clinician actually does with artificial intelligence on a Butterfly device was built and cleared by somebody else. This is the same pattern graded at C for Huma and Arcadia in this index, where the platform holds the differentiating asset and the intelligence is partly a marketplace.

The enterprise workflow platform pulls the same direction, since documentation, compliance and billing workflow is software rather than inference.

Ask which shipped tools are the company's own models, and what share of clinical use runs on partner applications.

The clinician remains the acquirer and the interpreter, and one cleared tool moves a real piece of expertise into the software.

The ordinary posture is assistive. Guidance features help obtain a usable image, automated measurement tools calculate from what was acquired, and a clinician reviews and decides. Nothing diagnoses autonomously and nothing acts on a patient.

The blind sweep gestational age tool is the interesting case and it is worth stating precisely what it changes. Conventionally, estimating gestational age requires an operator who knows how to obtain specific standard views, which is the skill that takes training. A blind sweep removes that requirement: the operator sweeps, and the model does the rest. The clinician still decides what to do with the number, so oversight of the decision is intact. What has moved is the competence required to produce the input, and the person holding the probe may now be unable to judge whether the model's answer is plausible.

That is a defensible and valuable trade, particularly in settings with no sonographer, and it is the explicit reason such a tool matters in global health. It also means the usual safeguard, an expert who would notice a wrong answer, is precisely the thing the tool was built to do without.

What is not described is what the tool does when the sweep is inadequate, whether it refuses to produce a result rather than producing a poor one, or what confidence information reaches the operator.

Graded B.

Ask what the tool does with an inadequate sweep and what confidence signal the operator sees.

More technical disclosure than most, and none of it is model performance.

The hardware and the imaging approach are described with real substance. Semiconductor based imaging replacing piezoelectric crystals is a specific architectural claim, generational chip development is discussed publicly in financial disclosure, and the company has published white papers on named scanning techniques covering their clinical utility and how they change examination protocols. A technical evaluator can form a genuine picture of how the device produces an image.

Regulatory clearance also creates a transparency floor that unregulated software in this index does not have. A cleared device has a public summary describing intended use, and the performance data behind it was examined by a regulator, so an assessment exists even where the vendor has not marketed it.

What is missing is everything a buyer would use to compare. No model card, no accuracy, sensitivity or agreement figure for the gestational age tool against a reference standard, no error distribution, no validation population, no retraining cadence and no drift monitoring account were located on any public surface.

The cloud update mechanism raises a version transparency question that nothing addresses. Capability arrives on devices already in the field through software updates, so the model a clinician used last month may not be the model they use today, and no published version history or change communication for the models was located.

Graded C.

Ask for the clearance summaries by number, accuracy against a reference standard, and how model version changes are communicated to users.

The third party application layer is disclosed unusually well and everything underneath it is not.

The partner ecosystem is the strong half and it is genuinely rare. Thirty developers are counted publicly, their regulatory progress is reported in financial disclosure, and four are stated to hold their own clearances with one holding a breakthrough designation. A clinician using a partner application is using software built by a named third party that has passed its own regulatory examination. Most vendors in this index will not name a single model dependency; this one reports on the state of its ecosystem quarterly.

The hardware chain is partly visible for an unusual reason. Because the company designs its own imaging silicon and licenses that technology onward to other companies through a separate business line, its position in the chain is documented in financial filings in a way a component buyer's would not be. Chip development milestones are discussed publicly.

What remains closed is the middle. No cloud or infrastructure provider is named, no sub processor register was located, no fabrication partner is identified for the silicon, and nothing describes the provenance of the company's own models, including what the gestational age tool was built from or whether any external model provider is involved.

A dependency question specific to the partner model also goes unaddressed. If a partner application is withdrawn, loses clearance or ceases trading, what happens to a customer relying on it is not described.

Graded C.

Ask who provides cloud infrastructure, what the company's own models are built from, and what happens when a partner application is withdrawn.

A regulatory first, named institutional deployments with quantified results, and audited financials behind all of it.

The strongest item is the clearance itself. Being the first company to obtain clearance for a blind sweep tool means a regulator examined performance data for a model that estimates gestational age without the operator acquiring standard views, and permitted marketing on that basis. Management stated it deliberately took a longer validation path and treated the software with the seriousness of a device given the clinical consequences. Regulatory clearance is a weaker signal than a randomised trial and a considerably stronger one than a case study, and very little in this index has it for the artificial intelligence specifically rather than for the hardware.

Customer evidence is named and quantified. The University of Rochester reported a 116 percent increase in point of care ultrasound revenue following a systemwide deployment, which is an operational rather than clinical outcome but is stated with an identifiable organisation attached. Indiana University School of Medicine adopted a one probe per student model, and nearly 1,000 probes were sold across six institutions in a single year.

Scale is checkable because the company is publicly listed, so revenue, growth, cash position and burn are filed rather than asserted. Few vendors graded here can be verified that way.

What holds it below the top grade is the clinical layer. Company material refers to clinical publications showing workflow and cost benefits without citing them, no accuracy figure for the gestational age tool was located outside the regulatory process, and no patient outcome study surfaced.

Ask for the published clinical literature by citation, and for the accuracy of the gestational age tool against a reference standard.

A cloud archive of clinical images is exactly the asset an imaging model needs, and nothing published says whether it is used that way.

Studies acquired on these devices are stored in the company's cloud, which is how the archive, sharing and workflow features function and is a reasonable design. It also means the company holds a large and growing corpus of ultrasound imaging with associated clinical context, accumulated across hospitals, medical schools, veterinary practices and global health programmes.

The question that follows is unavoidable and unanswered. Nothing located states whether customer acquired images are used to train, tune or evaluate the company's models, whether they inform partner applications in the developer ecosystem, whether any such use is confined within a customer boundary, or whether a customer may decline while still using the product. The partner ecosystem sharpens it, because thirty external developers building applications on this platform need imaging data from somewhere, and the platform is the obvious source.

Third party commentary raises a related concern about archives residing in the vendor cloud and the migration difficulty that creates. That is a commercial lock in argument rather than a stewardship one, but it rests on the same fact.

The acquisition endpoint adds a second surface, since images are captured on phones and tablets before reaching the cloud and nothing describes local retention.

Graded C: the storage design is coherent and the governance of what is stored is undisclosed.

Ask whether customer images train any model, whether partners can access them, and whether a customer can decline.

One credential is held and every contractual term is private.

The located evidence is an information security management certification, described as obtained to support enterprise adoption, which is a recognised standard requiring external audit of a security management system rather than a self assessment. That is more than many records in this index hold and it is the right kind of credential for a company running a cloud archive.

Business associate status is the correct posture, because clinical studies acquired on these devices are stored in the company's cloud on behalf of the healthcare organisations that acquired them.

What is absent is the agreement. No template business associate agreement, breach notification window, liability cap position, audit rights statement or data return provision was located.

Two features of this product raise obligations that a purely enterprise platform would not. Acquisition happens on a clinician's phone or tablet, frequently a personal device, so protected health information is created on endpoints outside any hospital's management, and nothing describes what is cached locally or what happens when a clinician leaves. And the individual sales channel means many devices are bought by single clinicians and small practices through commerce rather than procurement, where an agreement may never be negotiated at all and the terms of service become the whole arrangement.

Graded C.

Ask for the template agreement, what is retained on the acquisition device, and what governs individual purchasers who never sign an enterprise contract.

One audited credential, stated plainly, with nothing else on the public record.

The located evidence is certification under the recognised information security management standard, described as obtained specifically to support enterprise adoption and the scalability of the cloud platform. That is an externally audited management system rather than a self attestation, and naming the business reason for obtaining it is more candid than most.

What is absent is the rest of the artefact set a health system procurement process requests. No service organisation control report at either assurance level was located, no healthcare specific assurance framework certification, no penetration testing cadence, no vulnerability disclosure programme and no trust centre or portal where a buyer could retrieve current attestations without asking.

The device layer is a security surface in its own right and the record is silent on it. These are connected medical devices pairing with personal phones and tablets over consumer networks, in hospitals, in vehicles, in patients' homes and in low resource clinics abroad. Firmware update integrity, device pairing and authentication, and what happens to a lost or stolen probe are security questions distinct from cloud security, and nothing located addresses any of them.

The partner ecosystem adds a third surface, since thirty external developers deliver applications onto this platform and nothing describes what security review they pass or what data they can reach.

Graded C.

Ask for a service organisation control report, the device and firmware security model, and what security review partner applications undergo.

The most substantive device regulatory position in this index, covering both the hardware and the models running on it.

The imaging system itself is cleared, with the current generation obtaining clearance in January 2024 and carrying a European conformity mark alongside it. That is expected for a diagnostic imaging device and is table stakes.

What earns the grade is the artificial intelligence. In early 2026 the company became the first to obtain clearance for a blind sweep tool, estimating fetal gestational age from a sweep requiring no expert view acquisition. Management described taking a deliberately longer validation route and treating the software with the seriousness of a device because of the clinical consequences, which is a choice rather than an obligation: a great deal of software in this index is positioned to stay outside device regulation, and this company walked into it. A first in class clearance also means there was no predicate to lean on.

The partner ecosystem extends the position rather than diluting it. Four developers building on this platform hold their own clearances and one holds a breakthrough designation, so the applications a clinician reaches through the marketplace are regulated in their own right rather than riding on the host's status.

One structural feature deserves attention. New capability reaches devices already in use through cloud delivered software updates, which is efficient and means the regulatory state of a given probe depends on what version it is running and which market it is in. The company noted the tool applies in cleared markets, so global health deployment and domestic deployment are not the same regulatory situation.

Ask which specific tools are cleared under which numbers, and how version state maps to regulatory status across markets.

The deployment strategy makes population generalisation the central safety question, and nothing published addresses it.

Ultrasound is operator and patient dependent in ways other imaging is not. Image quality varies with body habitus, and a model trained predominantly on one population may degrade on another for reasons that have nothing to do with the algorithm and everything to do with what reaches the transducer.

The gestational age tool concentrates the risk. Fetal biometry based dating rests on growth references, and growth patterns differ across populations, with a substantial literature on whether a single international standard or population specific references should be used. A model estimating gestational age has some reference embedded in it, explicitly or through its training data, and which one it is determines whether the estimate is accurate for the woman being scanned. Nothing located states the training population, the reference used, or performance by maternal characteristics.

The consequence is concrete rather than abstract. Gestational age drives decisions about delivery timing, growth restriction assessment and viability, so a systematic dating error shifts clinical management. And the company is explicitly rolling this tool into global health markets, meaning the populations furthest from a typical development cohort may be the ones relying on it most, precisely because no sonographer is available to sanity check the result.

A dedicated pass located no bias testing, no subgroup performance data, no fairness review and no model documentation.

Ask what population the gestational age model was trained and validated on, which growth reference it embeds, and how it performs across maternal body habitus.

Recourse is undefined, and clearance makes the question sharper rather than settling it.

A dedicated pass located no indemnification position, no warranty covering model output, no accuracy guarantee, no service credit regime and no described route to dispute a result.

The exposure here is clinical and direct. A gestational age estimate drives decisions about delivery timing, growth assessment and viability. If a model produces a materially wrong date, the harm is to a pregnancy, and it may not surface for weeks. The tool is designed for use by operators who do not have the expertise to independently judge the answer, which is its purpose and also means the usual backstop is absent by design.

Regulatory clearance is sometimes read as resolving liability and it does not. Clearance establishes that a regulator accepted the evidence for a stated intended use. It says nothing about who bears the loss when a cleared device produces a wrong answer within that intended use, and that allocation lives entirely in contract.

The commercial structure complicates it further in a way most records in this lane avoid. Many devices are bought by individual clinicians and small practices through a commerce channel where no negotiated agreement exists and standard terms govern. Partner applications add a third party: when an application from the developer ecosystem produces a wrong result on this hardware, whether the host, the developer, or neither carries responsibility is not described anywhere located.

Graded D.

Ask what liability attaches to cleared tool output, what the standard terms say for commerce purchasers, and how responsibility is allocated for partner applications.

Interoperability exists, is priced separately, and is the subject of the sharpest external criticism of this platform.

The workflow platform relaunched in November 2025 is built around exactly this problem. Its stated purpose is to address undocumented examinations, compliance risk and administrative burden in point of care ultrasound programmes, which is a description of images taken on handheld devices never reaching the record. Selling that as a distinct enterprise product is an admission that the base configuration does not solve it, and also evidence the company understands the gap.

The commercial structure is the issue. Interchange of images in the standard medical format is not included in any membership tier and must be bought separately, so a clinician who buys a device and a subscription cannot by default get studies into the hospital archive. For an imaging device that is a significant boundary, because an examination that does not reach the record is difficult to bill, difficult to audit and unavailable to the next clinician.

Third party comparisons raise a related architectural criticism, that images route through the vendor's cloud rather than passing directly to a hospital archive over the institution's own network, which competitors selling direct connectivity present as a differentiator. That commentary comes from parties selling alternatives and is recorded as their argument rather than as a finding, but the underlying architecture is not disputed.

Nothing located describes record system integrations by name, whether ordering and results flow bidirectionally, or how examinations reconcile to an order.

Ask what imaging interchange costs, whether direct archive connectivity is possible without the vendor cloud, and which record systems are integrated.

The shape is clear and the specifics are not.

What is established: acquisition runs on a clinician's phone or tablet, studies move to the company's cloud, and capability updates reach fielded devices through that cloud. There is no indication of an on premises or customer hosted option, which is a coherent architectural choice for a consumer form factor device and a real constraint for organisations whose policy requires one.

The open questions are the usual ones and none is answered. Whether customer data occupies dedicated or shared tenancy is not stated, no infrastructure provider is named, no region selection is described, and no recovery objective or availability commitment was located.

Residency deserves more weight here than for a domestic only platform. The company sells across the Americas, Asia and Europe, targets global health programmes explicitly, and holds a European conformity mark, so it operates in jurisdictions with materially different data protection regimes. Nothing located states where images are stored for a customer in any given market.

One deployment characteristic is genuinely distinctive and cuts both ways. Because imaging is done in software, the company can deliver new clinical capability to devices already purchased without any hardware change, which is a real advantage. It also means the device in a clinician's hand is defined by a remote software state the clinician does not control, and the same channel that adds capability is the one that enforces the membership condition under which a lapsed subscription disables saving, advanced modes and artificial intelligence.

Ask where images reside by market, what availability commitment exists, and what happens to stored studies when a subscription lapses.

Prices are published, which almost nothing else in this index manages, and the structure behind them carries a consequence a buyer should understand before signing.

What is public is substantial. Device list prices are visible, with the current generation around 3,899 dollars and the earlier model lower, sold through a direct commerce channel as well as through sales. Membership tiers are named and priced, at roughly 299 dollars a year for the core tier and 420 dollars for the advanced tier, with a one time multi year option available for a single user on a single probe. A prospective buyer can therefore build a five year cost model unaided, which is the practical test this axis is really asking.

The structure is where the diligence sits. A membership is required to activate the device, and if it lapses the probe reverts to live viewing only, with no saving of studies, no advanced modes and no artificial intelligence. For any documented clinical use the subscription is therefore permanent rather than optional, and the honest way to read the device price is as an entry fee rather than a purchase. Licensing is per user in the individual tiers, so a shared probe in a three clinician practice can triple the recurring cost while the hardware count stays at one.

Two material costs sit outside the published ladder. Interchange of images in the standard medical format is not included in any tier and is bought separately, and the enterprise workflow platform is sold on negotiated contracts, with a first seven figure total contract value deal disclosed and no rate card published.

Graded B: genuine published pricing for the individual buyer, negotiated and undisclosed for the enterprise one.

Ask for enterprise platform pricing, the cost of imaging interchange, and whether licensing is per user or per device at scale.

One device covering what normally takes several, sold into a wider range of settings than anything else in this lane.

The clinical breadth follows from the hardware. A single semiconductor probe images across depths that conventionally require separate transducers, so abdominal, lung, cardiac, musculoskeletal, vascular and obstetric scanning run on one device. That is a real coverage claim rather than a marketing one, and the company's own commercial argument rests on it: buying four conventional probes to reach whole body coverage costs several times more.

Setting coverage is unusually wide and includes places most vendors in this index never reach. Hospitals and health systems are the core, but medical education is a distinct and growing channel with one probe per student programmes, veterinary practice is a named revenue channel, global health programmes are explicitly targeted, and home and community care entered commercial deployment in 2026 with an initial statewide rollout. A tool that works in a teaching hospital and in a low resource clinic is covering genuinely different ground.

The obstetric capability deserves particular note in that context, because a gestational age estimate obtainable without expert scanning skill is most valuable precisely where expert scanning is unavailable.

What holds it at B is depth per specialty. Handheld imaging does not replace a cart based system for every indication, and nothing located establishes where the boundary sits, which specialties are served to diagnostic standard rather than for triage, or how the partner applications distribute across specialties.

Ask which indications are supported to diagnostic rather than screening standard, and how partner applications cover specialty gaps.