Software Is Being Eaten — by AI
In 2011, Marc Andreessen said "software is eating the world." In 2026, the world is eating back.
GPT-4 lets someone without an engineering team build a working SaaS in a weekend. Cursor, Claude Code, and Copilot multiply senior engineer output by 3-5x. Open-source model inference costs drop 90% annually[1].
The result: the marginal cost of software approaches zero.
If anyone can use AI to build a rehab tracking app, a PROM collection tool, or a patient education platform in three days — then these things are no longer valuable in themselves.
This isn't hypothetical. SWORD Health ($4B valuation) spent seven years building its AI motion recognition engine[2]. Today, open-source MediaPipe Pose achieves >98% accuracy in compensatory movement detection[3]. MedBridge launched phone-camera-only motion analysis in 2025[4] — no additional hardware required.
When the tool becomes free, the tool is not the moat.
Thin Software vs. Thick Software
Let's define these terms.
Thin Software is an interface layer. It presents existing data in a better way, or executes existing processes more conveniently. Dashboards, forms, workflow automation, most SaaS — all thin software.
Characteristics of thin software:
- Data sources are public or substitutable
- Core logic can be rewritten by AI in a week
- Competitors only need a better UI or a lower price
- User loyalty comes from habit, not irreplaceability
Thick Software is a system deeply integrated with proprietary data sources. It doesn't just present data — it generates data that others cannot obtain, then trains models that only it can train.
Characteristics of thick software:
- Data sources are proprietary, physical, non-replicable
- Core value lies in the data flow, not the code
- Competitors cannot catch up with better software because the bottleneck is data acquisition
- User loyalty comes from this system knowing things other systems don't
The "Thin Software" Graveyard in Healthcare AI
Over the past five years, hundreds of AI products appeared in digital health. Most are thin software:
- AI triage chatbots: Trained on public medical literature; any LLM can do it
- Rehab motion recognition: MediaPipe/YOLO Pose is open-source and accurate enough
- Patient education platforms: Content from public guidelines; AI generates multilingual versions instantly
- Scheduling/workflow tools: Standard CRUD apps, replicable by Cursor in three days
These products aren't bad — they're undefendable. When AI drives software development costs toward zero, anyone can build a functionally equivalent alternative. Price competition begins. Margins vanish.
Data Is the Moat — But It Depends on Which Data
"Data is the new oil" has been said to death. But most people miss a critical distinction:
Replicable data ≠ moat. Public medical literature, FDA databases, PubMed papers — any AI company can scrape them. Models trained on this data can be matched by anyone.
Non-replicable data = moat. Data generated from the physical world, requiring specific hardware to collect, existing only in specific clinical contexts — this is the real barrier.
Specifically:
| Data Type | Replicable? | Moat? |
|---|---|---|
| PubMed literature | Anyone can scrape | No |
| Generic wearables (steps, HR) | Apple Watch has it | No |
| Phone-based motion capture (ROM) | MediaPipe is open-source | No |
| PROM questionnaire data | Requires patient relationships + process | Weak |
| Implantable sensor tissue force data | Requires implanting specific hardware | Strong |
| Sensor + PROM + motion cross-correlation | Requires the full stack | Strongest |
Implantable Sensors = The Foundation of Thick Software
Now connect the two concepts.
Data from implantable sensors is currently the most non-replicable data type in digital health. Here's why:
- Physical barrier: You need a regulatory-cleared medical device, implanted in a human body, before you can start collecting data. No shortcuts, no API, no open-source alternative.
- Clinical relationship barrier: A surgeon must choose to use your implant. This isn't a download decision — it's a medical decision affecting a patient's lifetime.
- Time barrier: Every data point comes from a real surgery and a real recovery process. You cannot accelerate training data generation.
- Integration barrier: A single sensor's data answers only one question ("how much force is the tissue bearing?"). Only when combined with motion data, PROMs, and clinical records can it answer the complete clinical question.
This is why sensors are the foundation of thick software: they create a data channel that only you can access. Any AI model built on this channel — prediction models, alert systems, personalized rehab protocols — automatically inherits this moat.
The Trend Persona IQ Started
Zimmer Biomet's Persona IQ is a milestone in orthopedic digitization — the only commercially available smart orthopedic implant today[5]. It achieved something no one had done before: giving surgeons visibility into patient recovery before the follow-up visit.
Through cumulative daily gait data, Persona IQ can now[6]:
- Predict venous thromboembolism (VTE) risk — abnormal gait pattern changes detected before clinical symptoms appear
- Flag early signals of periprosthetic joint infection (PJI) — sudden activity drops, deteriorating gait symmetry
- Recall at-risk patients before their scheduled visit — when data shows anomalies, the system alerts even between appointments
This is a massive step forward. Traditionally, surgeons only learned about patient status at follow-up visits. Persona IQ turned "passive waiting" into "active monitoring," taking the critical first step in orthopedic digital transformation.
But Persona IQ also defines the boundary of this generation's technology: it measures motion and can detect symptomatic changes, but cannot yet capture asymptomatic tissue-level changes. A patient may feel fine and walk normally, while the implant-bone interface is already developing abnormal stress distribution — this "surface normal, interior deteriorating" scenario is invisible to accelerometers and gyroscopes.
The question the next generation of implantable sensors must answer: what is happening inside the tissue before the patient feels anything?
This is precisely the domain of passive LC force sensors — measuring not motion, but force. Not "how the knee moves," but "how much stress the tissue bears." CardioMEMS already proved this technology pathway in cardiology — the CHAMPION trial showed a 37% reduction in heart failure hospitalizations[7]. By 2026, CardioMEMS has been implanted in over 100,000 patients, Abbott's heart failure division continues growing at 12% annually, and the FDA approved a next-generation reader in February 2026[8]. This isn't a historical precedent — it's an accelerating market. Discovery R brings the same physics into orthopedics — not to replace Persona IQ, but to add the layer it cannot yet see.
De Novo's Thick Software Stack
Connecting all the threads, our strategy isn't "build the best AI" or "build the best app." It's to build an irreplaceable data flow:
Layer 1 — Force: Discovery R implantable LC sensor → real-time tissue interface force data. No phone app or wearable can substitute.
Layer 2 — Motion: Phone CV + wearables → ROM, gait, compensatory movements. This layer's technology is substitutable — but only gains complete meaning when integrated with Layer 1.
Layer 3 — Experience: Automated PROM collection → patient-reported pain, function, quality of life. Digitization reduces collection cost, meeting CMS 2028 compliance.
Layer 4 — Prediction: ML models consuming all three layers → predicting who needs intervention, when, and how.
Any competitor can replicate Layers 2, 3, and 4 in software. But without Layer 1's sensor data, their models are like a world missing one dimension — they can see the surface but not the interior.
What "Full-Stack AI" Really Means
We discussed "full-stack AI" in How AI Is Changing Orthopedic Care. Now we can define it more precisely:
Full-stack AI doesn't mean "do everything yourself." It means "in the complete chain from data source to clinical decision, at least one link is non-replicable."
If your AI system only consumes public data → you're building thin software. If your AI system consumes proprietary data → you're building thick software.
The sensor is the non-replicable link. Software is the pipeline that turns this link into clinical value.
The moat is not in the software. The moat is in the data flow. Software is the channel that directs the moat in the right direction.
Conclusion: AI Won't Kill Healthcare Software Companies
What AI will kill is healthcare software companies that only have software — products that wrap public data in a nice interface.
What AI will amplify is companies that own proprietary data sources — because AI drives the cost of "turning data into insight" toward zero, while the cost of "acquiring proprietary data" hasn't decreased.
When software is free, data is the product. When AI is free, the sensor is the moat.
References
-
Stanford HAI. AI Index Report 2025 — AI model costs and performance trends. Link
-
SWORD Health acquires Kaia Health for $285M. MobiHealthNews. 2026. Link
-
Towards Intelligent Assessment in Personalized Physiotherapy with Computer Vision. Sensors. 2025. PMC
-
MedBridge Motion Analysis — phone-camera-based movement assessment. Link
-
Zimmer Biomet shares smart knee data at AAOS 2024. MedTech Dive. Link
-
Persona IQ sensor ROM correlation with in-office measurements. Journal of Clinical Orthopaedics and Trauma. 2026. ScienceDirect
-
Abraham WT, et al. Sustained efficacy of pulmonary artery pressure to guide adjustment of chronic heart failure therapy: CHAMPION trial. The Lancet. 2016;387(10017):453-461. PubMed
-
Abbott Wins FDA Approval for Updated Heart Failure Monitoring Device (CardioMEMS HERO). MedTech Dive. February 2026. Link