Thoughts on Healthcare Markets & Technology Podcast

Garmin just launched a $200 screenless health band with no subscription. That one decision puts the entire wearable subscription model under structural pressure.

The sensor stack inside every biometric band costs $15-25 to build at scale. The subscription was never about hardware. It was about the intelligence layer. Garmin just commoditized both.

WHOOP is valued at ~$3B on ~1M paid subs at $239/yr. Oura has sold ~2.5M rings. Both face a buyer pool that is expanding fast but skewing price-sensitive, driven by the Hemsworth and Bryan Johnson effect.

The real question is not which...

CMS’s draft MTF guidance is the most detailed operational document in the entire IRA drug pricing rollout, and almost everyone is going to misread the 14-day clock.

The Medicare Transaction Facilitator is not the payer. It is a data layer. The underlying claim pays the usual way. The MTF identifies eligible events and ships claim-level data to the manufacturer.

Two modules: Data Module is mandatory for everyone. Payment Module is optional. You can opt out of the payment switch. You cannot opt out of reporting. That asymmetry is the spine of the whole design.

CMS proposed a $260M imaging cut for CY 2027. It only works because two characters on a claim line, modifier PO, reliably identify the target setting. That is interoperability doing real work.

The cut applies only to noncontrast imaging in APCs 5521-5524 at grandfathered off-campus departments billing PO. Rural sole-community hospitals are exempt. On-campus departments are not touched. Scope matters.

The utilization case: volume up 38% and utilization per FFS beneficiary up 67% from 2016-2025, while FFS enrollment fell 17%. CMS pegs 2027 savings at $260M and $8.5B over a decade.

The price transparency side of the package...

CMS wants sub-second Medicare claims adjudication. The statutory payment floor is still 13 days. Those two facts don’t cancel each other. They coexist, and that’s the whole story.

The ClaimsCore RFI contemplates replacing FISS, MCS, the DME system, and the Common Working File with a commercial cloud platform. That’s a draft acquisition concept, not an award. Nobody has replaced Medicare’s shared systems.

A live bridge, MAP, already runs in limited parallel with the legacy system. Standing it up required preserving continuity with 5 downstream repositories. Even a thin slice of new volume needed every ol...

CMS prior auth FHIR APIs are not a proposed rule. They are a January 1, 2027 hard deadline for MA plans, Medicaid managed care, CHIP, and federally facilitated exchange issuers.

The decision clocks already went live in 2026: 72 hours for expedited requests, 7 calendar days for standard. Specific denial reasons required. Public reporting of PA metrics started this year.

Four APIs due January 2027: Patient Access, Provider Access, Payer-to-Payer, and Prior Authorization. The PA API must accept submissions and return structured approvals, denials with reasons, or info requests.

The APIs solve transport. They do not find the clinical...

Traditional MIPS has a death date. CY 2029 is the cutoff. MVPs become the only reporting option for non-APM clinicians after 2028. Three performance years to rebuild.

93% of MIPS reporters earned a positive adjustment in 2017-2019. The largest was 1.88%. Everyone passed, nobody got paid, every practice paid real overhead to get there.

The new core measure rule is the sharpest edge. Miss the required core measure without attestation and you get zero out of ten quality points. A ten-point crater at 75-point threshold flips bonuses to penalties.

43 of the 180 surviving quality measures change specification in...

Hospitals are sitting on a recurring, contractual, uncorrelated cash flow buried in ‘other operating revenue.’ Nobody finances it. Nobody prices it. That is the opportunity.

It is health system data licensing revenue: deals with pharma, RWE vendors, registries, and now AI labs paying top dollar for clinical text. At big AMCs this runs tens of millions a year, well below potential.

The template is Royalty Pharma. Buy forward rights to the stream from institutions that need capital now more than they need checks over 15 years. The spread lives in the modeling gap.

Three forc...

Neutral atom quantum computers are raising billions. Here is what health systems and pharma boards are not being told about what they can actually do.

The Rydberg blockade switches atom interactions from millihertz to hundreds of megahertz with a laser pulse. 11-12 orders of magnitude. That is the gate. That is the whole trick.

Gate fidelity is now in the mid-four-nines range with theoretical headroom to five nines. All the remaining gap is engineering: laser noise, atom temperature, stray fields. The physics has room left.

Three real buckets for healthcare: quantum chemistry for...

Payers and providers sign one contract. Then each builds separate software to interpret it. Two codebases. Two answers. That gap is where denials live.

The industry response: buy better AI. Providers use it to fight denials. Payers use it to generate them. Both sides get better weapons. The battlefield never changes.

This is a coordination problem, not a tech problem. Card networks solved it for payments. Central counterparties solved it for securities. Healthcare reimbursement has the same structural signature.

The fix is a neutral exchange where both sides pre-approve one executable version of...

CMS flagged ~$93.5B in improper payments across 5 programs. A Texas court just vacated the rule designed to claw it back. This is not a compliance story. It is a market structure story.

Medicare Advantage hit 6.09% improper payment rate, within target. Sounds fine. But that measure captures documentation gaps, not systematic risk-score inflation worth tens of billions more.

RADV was supposed to fix that. The 2023 Final Rule would have turned it into a multibillion-dollar recovery engine. Plans sued. Northern District of Texas vacated key portions on Sept 25, 2025.

Here is the insight: litigation that disrupts...

Anthropic launched a science research platform AND announced its own drug programs on the same day. The toolmaker opened a mine next to its customers’ mines.

Claude Science connects 60-plus scientific databases, runs code on HPC clusters, and keeps data on your own infrastructure. Reproducibility is baked in, every result traces to source code.

The conflict: Anthropic sells to pharma and also runs its own preclinical programs. Official answer is they only chase neglected diseases. Disease interest is not static, and that boundary is softer than it sounds.

BCG data on AI-discovered molecules: Ph...

A scoliosis surgery. Surgeon earned $8,016. The assistant standing next to him got $196,566. That is a 24x multiple in favor of the person holding the retractor.

It is not a one-off. Texas prostatectomy: surgeon $1,843, assistant $50,456. NJ breast reconstruction: surgeon $2,707, assistant $111,000. Same law, same arbitration channel, same result.

The mechanism: No Surprises Act arbitration. Assistants stay out of network on purpose. Arbitrators evaluate their claims in isolation, with no visibility into what the surgeon on the same case collected.

One Wisconsin spinal fusion was split into 11 separate arbitration filings. They won all 11. Total payout: ~$196k...

FDA is formally reviewing BPC-157, TB-500, MOTS-c, Semax, and 3 other gray market peptides at its July 2026 advisory committee meeting. The wellness clinic industry is watching closely.

These 7 peptides have no USP monographs, no FDA-approved drug status, and no path to legal 503A compounding until FDA acts. Most were previously in a do-not-compound category. Removal from that list is NOT a green light.

Across all 7 substances, FDA’s background documents flag the same problems: weak chemical characterization, no modern human clinical trials, unresolved immunogenicity risk, and manufacturing quality that is largely undocumented.

TB-500 is WA...

Pathway Labs cleared the FDA for six structural heart conditions off one ECG. The $8.5M seed round is the least interesting part of this announcement.

The model trained on 700k+ ECG-echo pairs at NewYork-Presbyterian. In a 3,200-ECG head-to-head, it hit 77% accuracy vs 64% for cardiologists. That gap is real.

Then it landed on OpenEvidence, used by 750k+ verified clinicians. That distribution solves the problem that kills most cardiac AI: getting a busy doctor to actually open the tool.

The multicondition framing also quietly ends the single-indication race. Matching six FDA clearances means assembling six...

Hospitals pay vendors 6-7 figures yearly for software a frustrated clinician could have built over a weekend. The clinician usually spec’d it out anyway.

Agentic coding tools just flipped the constraint. The gap was never clinical expertise - it was engineering bandwidth. Now that gap is mostly gone.

Academic medicine already has the perfect learning model: tumor boards, M&M conferences, grand rounds. An AI coding academy should borrow that, not bootcamp culture.

The binding constraint now is ownership and incentives. If the inside deal is a black box, talented clinicians route ar...

Every provider org, RCM vendor, and clearinghouse rebuilds the same prior auth logic from the same payer PDFs. In isolation. Every time. That is the infrastructure problem hiding in plain sight.

Payer coverage policy is conditional logic: inputs are codes, diagnoses, durations, site of care. Outputs are auth required, document these things first. It behaves like software maintained with no version control.

CMS mandates a Prior Authorization API by Jan 1, 2027 for MA, Medicaid, CHIP, and FFE plans. The Da Vinci standards define the pipes. Nobody is building the content layer that makes the rules inside...

Children learn without labels, without feedback, and on almost no data. Healthcare AI still struggles with all three. Here is why the gap is a product opportunity.

In 1996, eight-month-olds extracted word boundaries from a two-minute speech stream with no cues except raw statistics. Same machinery works on shapes and tones. It is self-supervised learning on a banana-sized power budget.

Clinical labeled data is brutal to get. But continuous glucose traces, fetal monitoring strips, ICU telemetry: oceans of unlabeled sequence mostly discarded. Self-supervised pretraining on one modality turns 1,000 labels into 50,000.

Kids also do not...

Midjourney’s scanner went viral. The most-shared posts claimed its AI rebuilds your scan in real time. The company said there is no AI in the scanner at all. The feed inverted the most important fact.

The imaging runs on Butterfly Network’s chip, 40 modules per prototype, under a deal worth up to $74M over 5 years. BFLY jumped 52-56% in a day. The headline said Midjourney. The business story is Butterfly.

Current prototype: ~20 min, not 60 sec. About a dozen people scanned. Team of roughly 9. Sixty seconds is the goal. MRI comparison is launch-deck framing, not a pu...

A self-funded employer spending $50M/yr on health claims often cannot produce its own claims data. The TPA holds it and charges for access. The employer is the fiduciary. This is the baseline absurdity.

The contracts make it worse. TPA shared-savings fees run 20-30% of repricing. Audit rights are written to find nothing. Data ownership clauses say the TPA owns data the employer paid to generate.

The benefits consultant is supposed to fix this. Except a large share of consultant pay comes from carriers and PBMs through overrides and retention bonuses the employer cannot see.<...

SpaceX at $1.77T. Anthropic near $965B. OpenAI targeting $1T. Three filings just warped the entire listing market, including for healthcare.

80% of global VC went to AI in Q1 2026. Four names took $188B of a record $300B quarter. Every healthcare company was fighting over the scraps.

A trillion-dollar IPO does not compete with a $1B health tech float. It makes a $1B float invisible to generalist investors, for better or worse.

The real signal is not the IPO calendar. If late-stage private rounds clear at flat-to-up valuations, the listing window is opening. Confidence flows...

Tampa General cut early sepsis deaths 68% and saved 886 lives using Palantir. The algorithm is not the reason why.

The system pulls Epic data, lab feeds, and device telemetry into one patient object model and scores every patient every 15 minutes. But the model just flags. The rapid response team and the antibiotic clock do the saving.

Every failed sepsis alert tool died the same way: fires to a bedside nurse managing 6 patients, joins 40 other interruptions, gets dismissed. Override rates hit 95%. Tampa General fixed this by routing alerts to a dedicated response team, not the bedside.

Healthcare employs 1 in 9 Americans. Anthropic just named it as both the biggest AI automation target AND the designated landing zone for everyone else’s displaced workers. Same industry. Opposite fates.

The back office is the kill zone. Autonomous coding vendors report 85-95% of charts processed without human touch. Ambient documentation went from demo to procurement line item in 30 months. This isn’t hype. It’s already in the attrition numbers.

Anthropic’s policy framework literally proposes enhanced benefits for displaced workers who opt into healthcare and elder care roles. A frontier AI lab is designing the safe...

CMS just mandated that 43 states verify Medicaid work requirements by Jan 1, 2027. Almost none of them have built the software to do it.

Georgia spent $92M on its work requirement pilot. More than $2 in administration for every $1 in actual medical care. 4,300 enrollees against 246,000 eligible.

This is not a workforce policy. Most Medicaid adults already work or qualify for an exemption. The mandate is a documentation and data-matching problem at scale.

The 90/10 federal match means states pay only 10 cents per dollar of development. That rewrites the unit economics for every vendor in this space.

The administration is building a regulatory pathway for AI systems that diagnose and prescribe with no physician in the loop. Utah already has a pilot running. The Medical Licensing Board asked for an immediate suspension.

A Nature Medicine study gave chatbots 1,200 real patient simulations. Correct diagnosis rate: 34%. They were worse than Google at guiding people to the right decision. These same models pass medical licensing exams.

The self-driving car comparison is everywhere. But robotaxis are geofenced - Waymo refuses the hard routes. The entrepreneurs are pitching a general-purpose AI that handles everything. That is Level 5...

Mayo Clinic will OWN the frontier healthcare AI model it is building with Microsoft. Microsoft will distribute it through Azure. That ownership split is the entire story.

Mayo brings 54M+ de-identified patient records, 5.8B+ images, and years of longitudinal linkage infrastructure. Microsoft brings compute, engineers, and Azure distribution. Neither could do this alone.

The benchmark driving the superintelligence language: Microsoft’s MAI-DxO hit 85.5% on 304 hard NEJM cases. Physicians averaged 20%. Catch: physicians were not allowed to look anything up. Not quite the same test.

Microsoft already owns Nuance DAX, the ambient scribe in thousands of...

A YC company grew 17x and hit seven figures in ARR in under a year by not selling software. It does the fax, referral, scheduling, and collections work instead.

15-25% of specialty referrals never get scheduled at all. Not lost to competition. Lost to intake paperwork. One client had 4,800 referrals sitting in Epic for 4 months.

Payers denied more claims in 2025 than 2024. Up to 65% of denials are never appealed. Not because practices win, but because the appeal costs more than the claim.

The comp is Tennr: $101M Series C, $605M valuation, 10M docs/month...

Smart Therapy LLC billed $14M in phantom autism therapy over 5 years. The state had all the data to catch it. Nobody ran the queries.

How it worked: teenage relatives billed as clinicians, parents paid $300-$1,500/mo per kid to enroll, forged supervisor sign-offs, billing maximums every single day, transport vendors on the same payroll.

EIDBI claims in Minnesota went from $600K in 2018 to $400M by 2025. 14 Medicaid programs flagged high risk. One Optum-run tightening cut ~$165M in spending across those programs in a couple quarters.

Every signal was loud and public: billing the cap...

VERVE-102 just published Phase 1b data in NEJM. Single IV infusion. 88% PCSK9 reduction. 62% LDL-C reduction. 18-month durability. No serious adverse events in 35 patients.

But the core problem being solved is not efficacy. Statins work. PCSK9 inhibitors work. The problem: roughly half of patients prescribed lipid-lowering therapy stop taking it within a year. Every year. Consistently.

Adenine base editing does not cut DNA. It converts a single nucleotide letter in the PCSK9 gene in liver cells. No double-strand break. The target mimics a natural human variant already field-tested by evolution.

The 18-month durability signal...

The real hospital AI regulator is not the FDA. It is the Joint Commission, which accredits 80% of US hospitals and controls access to 60% of their revenue.

In September 2025 the Joint Commission plus CHAI released AI governance guidance. Once it enters accreditation standards, hospitals operationalize it the same way they operationalize infection control. Mandatory committees, audits, evidence binders.

That shifts the AI procurement gate from innovation officers to legal, compliance, and clinical risk. Vendors without documentation mapping to the guidance get filtered before pricing conversations happen.

The core insight: every AI deployment is now...

HHS shipped an LLM workflow for federal audits without an RFP. No vendor circus. No multi-year authorization. That is the actual story from May 21.

Single audits from every entity spending over $1M in federal funds sit in the Federal Audit Clearinghouse. Most were never read. HHS is now reading them with ChatGPT and threatening funding cuts for chronic deficiencies.

Three months earlier, CMS dropped the CRUSH RFI: provider enrollment reform, lab fraud, DMEPOS moratorium, AI-assisted coding review, and a direct question about using AI in the insurance marketplaces. This is a coordinated policy signal, not...

Forus raised $160M at a $1B valuation on ~$50M in annualized revenue. That is a 20x multiple at Series B. Here is why Thrive and General Catalyst wrote that check.

Healthcare is a transaction routing economy, not a care delivery industry. Every prior auth, every benefits check, every specialty pharmacy assignment is a transaction. The intermediary stack exists entirely because of that fragmentation.

Specialty drugs changed the math. A therapy at $60K-$3M per patient makes every coordination failure expensive. Manufacturers now pay real money for faster time to therapy. The transaction layer became economically...

California pays RNs $148K nominally. After cost-of-living adjustment it drops to third. Oregon wins. Minnesota is second. Hawaii, nominally second, falls dead last.

Adjusted NP pay: Oklahoma is first at roughly $148K annualized. Iowa second. Kansas third. California does not crack the top ten. The coastal sweep inverts completely.

Indiana has the highest cost-adjusted PA wage of any state for any role in the BLS May 2024 dataset. $76.61 per hour adjusted. California PAs, nominally highest, land at $53.78 adjusted.

That gap between Indiana and California PAs is about $23 per hour. Annualized, it is a $48K...

Q1 2026 produced four hardware-driven digital health mega-rounds in one quarter. Whoop at $10.1B. Neuralink at $9.7B. BrainCo at $1.3B unicorn birth. eMed at $2B. Either hardware is back or this is the top.

Context: Peloton went from $171 to under $4. Lululemon wrote off Mirror entirely. Fitbit got swallowed by Google. The consensus by late 2024 was that consumer health hardware without a regulatory moat was getting bundled into irrelevance by Apple Watch.

What changed: Whoop and Oura are not hardware companies in any traditional sense. The device is the acquisition cost for a subscription and a compounding...

CMS just reimposed a nationwide home health agency moratorium. LA County has 12-15% of all HHAs while holding 3% of Medicare beneficiaries. Ohio found 18 agencies at one address. The pay-and-chase loop is running again.

Here is the absurd part. GPS-verified visit data is already being collected at every Medicaid home health visit under the 21st Century Cures Act mandate. Six data elements, including location, captured at the moment of service.

That data sits in aggregator databases. It gets used for post-pay audits. It does not feed into the claim adjudication engine before payment goes out the...

Q1 2026 managed care earnings looked like a recovery. They were not. Here is what actually happened across UHC, CVS, Elevance, Humana, Centene, and Molina.

Cost trend did not reverse. Hospital costs are still up. Senior utilization is still elevated. Specialty drug spend is still ugly. Plans just stopped waiting for the trend to fix itself.

What changed: benefit design tightened, networks narrowed, prior auth volume rose, pharmacy controls got harder, and entire product lines and counties got exited. Less coverage, not cheaper care.

The Molina $96m impairment on its MA-PD exit is the...

The CEA’s six hundred billion dollars drug savings projection is built on data the report’s own footnotes say don’t measure net prices. The policy is net-price-based. The data are gross-price-based.

Why did seventeen pharma CEOs sign voluntary MFN deals without anyone forcing them? One sentence in CMS guidance: GENEROUS supplemental rebates won’t affect Medicaid Best Price or three hundred forty billion ceilings.

Without that carve-out, every MFN rebate resets Best Price, cascading through three hundred forty billion across sixty six billion dollars in annual safety-net drug purchases. No CFO signs that deal. Th...

$2.58B across 10 healthcare VC rounds in 60 days. The pattern is clear: capital is concentrating on scarce inputs, not apps.

Sanofi is writing equity into Earendil after a $2.56B deal. Regeneron put $200M into TriNetX for exclusive multi-omic data rights. Pharma is not renting access anymore. It is buying it.

Earendil at $787M. WHOOP at $575M with Abbott and Mayo on the cap table. Beeline at $300M with BMS assets and the SpringWorks team. Every check buys something nobody else can copy fast.

The market is not frozen. It just stopped paying for vague...

Quick Links: Knowledge Base, Podcast, and Social

Knowledge Base — search and filter every article and podcast episode by topic, section, and keyword: kb.onhealthcare.tech

Listen to the Podcast — every article is also available as an audio episode. Free subscribers get the public episodes; paid subscribers get the full archive including subscriber-only episodes. Listen on Apple Podcasts, Spotify, or browse all episodes on the Substack Podcast page.

For paid subscribers — your subscription unlocks the entire research archive (538+ deep-dives), every paid podcast episode, and full search inside the Knowledge Base. To listen to paid episod...

Quick Links: Knowledge Base, Podcast, and Social

Knowledge Base — search and filter every article and podcast episode by topic, section, and keyword: kb.onhealthcare.tech

Listen to the Podcast — every article is also available as an audio episode. Free subscribers get the public episodes; paid subscribers get the full archive including subscriber-only episodes. Listen on Apple Podcasts, Spotify, or browse all episodes on the Substack Podcast page.

For paid subscribers — your subscription unlocks the entire research archive (538+ deep-dives), every paid podcast episode, and full search inside the Knowledge Base. To listen to paid episod...

Quick Links: Knowledge Base, Podcast, and Social

Knowledge Base — search and filter every article and podcast episode by topic, section, and keyword: kb.onhealthcare.tech

Listen to the Podcast — every article is also available as an audio episode. Free subscribers get the public episodes; paid subscribers get the full archive including subscriber-only episodes. Listen on Apple Podcasts, Spotify, or browse all episodes on the Substack Podcast page.

For paid subscribers — your subscription unlocks the entire research archive (538+ deep-dives), every paid podcast episode, and full search inside the Knowledge Base. To listen to paid episod...