Physical AI (the application of AI to machines that perceive, reason, and act in the physical world) has become a defining theme over the last 2-3 years. Robotics as a category has raised a record ~$42B in 2025, ~9% of global venture funding. The momentum continued in 2026 - Skild AI raised $1.4B, Apptronik expanded its Series A to more than $935M, and Bedrock Robotics raised a $270M Series B.

Aside from $, the stronger signal is the level of deployments we’ve seen in the market. Figure says its robots worked 10-hour shifts at BMW and contributed to 30K+ vehicles (Figure). Agility Robotics says Digit has moved more than 10K+ totes at GXO, the logistics provider (Agility). Mind Robotics is targeting manufacturing, Dexory is building warehouse intelligence, Watney Robotics is working in data centers, and Multiply Labs is automating cell-therapy manufacturing. Robots aren’t just going commercial, but also vertical.

Why this moment feels different

Robotics has produced impressive demonstrations for decades (throwback to Sony’s companion dog - ERS 1000). What changed is the convergence of several curves:

• Foundation models: Give robots general knowledge, semantic understanding, and reasoning capabilities → reduce need to program every task or environment from scratch

• Vision-language-action (VLA) models: Connect what a robot sees + what a person asks to execute on physical action to complete tasks

• World models: Help robots predict how objects, people, and environments will respond before they act → enabling better planning over longer and more complex tasks

• Simulation and synthetic data: Allow robots to practice / train at scale and encounter edge scenarios → reduce cost, time, and risk of collecting every experience in the real world

• Edge inference and on-device compute $: Moves decision-making onto the robots → allowing for improved speed, reliability, privacy, and ops in environments with limited connectivity (e.g., construction, mining, space?)

• Sensors, actuators, and hardware economics: Better perception, more precise movement, and falling cost of components → making robots more capable and affordable for commercial deployments

With that, the key shift now is that the bottleneck is moving from raw intelligence to deployment. When talking to founders, a lot of the focus is now being centered around reliability, safety, workflow integration, field service, and customer ROI. A robot that works 90% of the time may be an extraordinary demo, but an unusable production system.

That insight shaped our investment in Bedrock Robotics. In construction, autonomy is shifting from building smarter individual machines (e.g., tele-ops) to building the operating system for the entire jobsite. Drawing on their experience at Waymo, Bedrock’s team is building an autonomy platform that runs across existing equipment, learns from real production environments, and helps customers complete work more safely and efficiently.

The next data flywheel

The first AI era was built mainly on static, human-generated corpora. Physical AI adds continuous experience: sensor streams, human demonstrations, tele-ops, robot trajectories, interventions, failures, recoveries, and simulation-generated scenarios - a lot driven from the key unlocks noted above.

Therefore, the deployed machine becomes both a product and a data-collection node. As mentioned by Rebecca over at Union Square Ventures, immediate control happens at the edge. Training, simulation, analytics, and fleet management happen centrally. Each deployment can improve the rest of the fleet.

Similar to what conversations have been about in enterprise applications, the moat isn’t a model checkpoint, but more so about proprietary operational experience, plus the distribution and infrastructure required to turn that experience into better policies. As base models become more accessible, experience at the edge compounds.

Mapping the market

The stack spans models, robot policies, simulation and fleet infrastructure, hardware, applications, and field operations. Frontier labs are taking different approaches. Google DeepMind is productizing embodied reasoning through Gemini Robotics. OpenAI has rebuilt an internal team spanning the model stack, hardware, and software.

Players across the industry are pushing to build autonomy, but with differing approaches to the architecture. The core debates are generality versus specialization, horizontal versus vertical, model-first versus application-first, and data-first versus hardware-first.

Where value will accrue and what I’m watching

My current view is that value will concentrate around four control points:

• Distribution into mission-critical workflows. The winners will be embedded in high-value customer workflows and integrated into existing systems, not simply offering the best autonomy model. Customers ultimately optimize for uptime, throughput, safety, and completed work, not intelligence in isolation. Distribution and workflow ownership become the path to durable adoption.

• Proprietary experience loops. The moat comes from closed-loop learning across deployments. The most valuable data connects observations, actions, and outcomes across environments, particularly around interventions, failures, recoveries, and edge cases. Scale alone is not enough. Differentiated, high-quality operational data compounds with every deployment.

• Deployment infrastructure and operations. As heterogeneous robotic fleets become commonplace, a horizontal operating layer will emerge for fleet management, observability, evaluation, remote intervention, safety, compliance, and orchestration across machines. Just as cloud infrastructure abstracted distributed computing, robotics infrastructure can abstract deployment and operations across diverse hardware platforms.

• Full-stack deployment economics. Companies that control the software stack, from autonomy models and runtime to hardware integration and field operations, can optimize cost per productive hour. Commercial models such as Robotics-as-a-Service, leasing, and outcome-based pricing convert upfront capital expenditures into predictable operating costs while aligning vendor incentives with customer ROI.

The future favors vertical systems that own deployment and data, alongside a smaller number of horizontal model and infrastructure platforms that become standards.

The next phase will be less about viral demos and more about operating metrics. Robotics will be evaluated based on key metrics – i.e., intervention rates, cost per productive hour, fleet-learning speed, safety validation, data ownership, field service, etc.

The robots I imagined in 2008 looked like finished products. The reality is more interesting. Physical AI is emerging as a stack, an ecosystem, and a new operating model for industries where work happens through atoms rather than pixels. The winners will turn intelligence into reliable deployment, then turn deployment into data, learning, and better economics.

If you’re building at any of these control points - or have any feedback - I’d love to hear from you!

For most of the LLM era, reinforcement learning had exactly one job: alignment – i.e., training the model to conform to human preferences, safety guidelines, and conversational formats.

Techniques such as Reinforcement Learning from Human Feedback (RLHF) and supervised fine-tuning transformed raw foundation models into products people could actually use. Over the past few years, models learned to follow instructions, avoid harmful outputs, and mimic human preferences. Reinforcement learning was primarily a post-training mechanism -- a way to improve behavior rather than expand core capabilities.

This paradigm is changing. Reinforcement learning is moving from alignment to execution -- shifting from techniques that merely refine a model’s responses to systems that enable models to act, learn, and improve within real-world environments.

Historically, the structural moat was access to large, static datasets (a la Scale AI, Mercor, Surge, etc.). Today, the emerging moat is ownership of the interaction loop itself: the environment where agents take actions, receive feedback, and learn from outcomes. The most valuable platforms are increasingly the ones that control this loop.

As we continue evaluating the enterprise applications layer, this is where the puck is moving.

From Planner to Executor

Modern AI systems can be thought of as operating across three layers:

• Knowledge: Traditional LLMs optimized for language understanding, retrieval, and content generation.

• Planner: Reasoning models that decompose complex problems, generate multi-step strategies, and self-correct on paper. Although initially novel, this reasoning has remained largely syntactic/determinstic -- they understand the structural logic of a plan, but operate in a vacuum. Models construct plans without any real-world validation of whether they actually work.

• Executor: Agents trained through interaction with a live environment, discovering and refining optimal policies based on real-world outcomes rather than predicted ones.

The transition from planner to executor is where most of the difficulty and enterprise value lies.

Even the strongest reasoning models cannot reliably predict optimal action sequences without interacting with the systems around them. Whether the task is managing inventory, processing insurance claims, or resolving a customer support ticket, success depends on navigating a dynamic environment whose state changes over time.

Reasoning alone is insufficient. To close the gap, an executing model needs two things:

1. An environment in which it can act.

2. A reward signal that verifies whether those actions succeeded.

This is exactly what reinforcement learning provides. Unlike supervised learning, which simply mimics past historical examples, RL learns through active interaction. The agent takes actions, observes consequences, receives feedback, and updates its behavior accordingly.

Every interaction generates a trajectory. Every trajectory produces pristine, new training data. Every iteration improves the policy. Better policies drive more usage, creating even more trajectories.

The loop becomes the asset.

Why This Favors Vertical Apps (Claude Who?)

The fundamental rule of reinforcement learning is simple: it thrives when reward signals are verifiable and domain-specific, and collapses when they’re generic, ambiguous, or easy to game.

A horizontal consumer agent can easily master booking flights or buying shoes. The reward is binary and public: Did the transaction clear? Was a receipt sent out?

Enterprise workflows are more complex. A horizontal model cannot natively verify if a complex contract redline will survive opposing counsel, or if a medical billing code satisfies a highly localized payer policy. These reward signals are not open-source; they are buried deep inside institutional context.

The reward function lives exclusively inside the application.

Specialized vertical software providers (e.g., Crosby, Harvey, Abridge, etc.) host the workflow, capture the action, record the downstream outcome, and register the immediate system feedback. Over thousands of tasks, these operational traces form a compounding, proprietary learning machine. The workflow owner implicitly becomes the owner of the reward model.

This is the exact layout underpinning Factory’s recently published “Software Factory” architecture. Software development is reframed entirely as a closed loop where agents write code and immediately stress-test it against isolated “digital twins” of enterprise infrastructure.

+1 point for the application layer! Massive economic value will accrue to the specialized platforms that own the highest-fidelity, domain-specific execution loops.

Mapping the Infrastructure Stack

When analyzing the emergence of these reinforcement learning loops, different infrastructure players are approaching the technical stack across three distinct layers (some companies I’m fired up about below):

1. Compute & Foundations

• Prime Intellect: Aggregating global decentralized compute and building a full open-source post-training stack -- providing the training harnesses, asynchronous RL trainers, and open environments needed to build self-improving agents.

2. RL Environments & Execution

• Applied Compute: Providing the mission-critical foundation and continuous learning infrastructure for enterprises to host proprietary data, build specialized models, and safely manage massive agent workforces.

• Trajectory: Providing the essential logging, observability, and evaluation infrastructure needed to map complex, multi-step agent behaviors and capture clean behavioral traces.

• Raindrop: Agent observability, evaluation, and trajectory intelligence infrastructure. Captures behavioral traces, detects failures, and transforms production interactions into actionable learning signals for continuously improving agent systems.

3. Optimized Inference

• Baseten: Building the highly specialized runtime and serving architecture required to deploy, scale, and rapidly iterate on custom, post-trained models.

For a deeper dive, here’s a solid, highly detailed market map from the team over at Sapphire Ventures (shoutout to their blog and golf analogies for helping me better understand the space)

The Inevitable Convergence

As is always true with infrastructure profit pools, these clean boundaries could start to blur.

Because an optimal RL loop requires a model to continuously cycle between running rollouts inside a sandbox, calculating rewards, and serving production traffic, the gravity of data and dollars is forcing aggressive architectural convergence. Hearing buzz about RL environment players building out highly optimized model-serving infrastructure -- effectively moving downstream to secure compute and host the production inference for the very models they help post-train.

When the training loop is the runtime, whoever owns the environment ultimately dictates where the inference spend flows.

Applications change. Frameworks change. Cloud providers come and go. But every product ultimately depends on a system that stores, retrieves, and serves data. Whether you’re a two-person startup or a F500 enterprise, building a consumer app or a defense system, somewhere in the stack there is a database.

And in many cases, it’s the piece of infrastructure that becomes most important and most painful as you scale.

For most of computing history, databases were viewed as plumbing - critical, but largely solved. Enterprises bought Oracle. Startups deployed MySQL. The goal was simple: set it up, keep it running, and avoid touching it whenever possible.

Then three platform shifts reshaped the market.

First, data exploded. Mobile applications, SaaS software, and billions of connected devices dramatically increased the amount of information software needed to store and process. Databases built for the previous era began to show their limits.

Second, the cloud turned infrastructure into software. Instead of buying servers, companies rented compute on demand. A market long dominated by Oracle, IBM, and SAP opened up to a new generation of cloud-native database companies, each optimizing for a different combination of scale, performance, simplicity, and cost.

Third, AI introduced entirely new workloads. Training models requires infrastructure capable of processing massive datasets. Running them in production requires semantic search, low-latency retrieval, vector storage, and entirely new data architectures. AI created requirements that much of the existing database stack was never designed to handle.

Underneath these shifts is a quieter trend that may prove even more important: databases have become developer platforms.

A decade ago, standing up a production-grade database required infrastructure expertise, configuration, and operational overhead. Today a developer can provision a managed database in minutes with authentication, storage, APIs, observability, and scaling built in.

The result is a dramatic compression in the distance between idea and product. Small teams can build software that previously required entire engineering organizations. Internal tools that once took months now take days. AI coding tools can generate applications and deploy them directly onto modern database platforms.

This evolution matters because databases sit at the center of every major trend in software: cloud computing, open source, developer tools, data infrastructure, and now AI.

This breakdown is a tour through eight eras of database evolution, each emerging to solve a limitation created by the one before it. By the end, the modern infrastructure stack should feel less like a collection of buzzwords and more like what it actually is: forty years of the industry debating how to store information reliably, retrieve it efficiently, and make it accessible to the people and systems that need it.

A debate that AI has made dramatically more important, and one that remains far from settled.

What Is a Database?

A database is a system for storing, retrieving, and managing information.

Every application you use depends on one. When you log in, a database verifies your credentials. When you make a purchase, a database records the transaction. When you receive a recommendation, a database is queried to determine what’s most relevant to you.

At its core, every database has two jobs:

1. Write data reliably.

2. Read data back quickly.

Everything else is an optimization around those two objectives.

The Fundamental Jobs of a Database

Most of database history is really the story of one problem: transactional workloads and analytical workloads want fundamentally different architectures.

OLTP (Online Transaction Processing)

OLTP databases power live applications. They handle individual reads and writes in milliseconds and prioritize speed, consistency, and availability.

Examples include:

• A user logging in

• A payment being processed

• Inventory updating after a purchase

The defining characteristic is simple: a very large number of small transactions where correctness matters.

OLAP (Online Analytical Processing)

OLAP databases answer business questions. They scan and aggregate large amounts of historical data to generate insights.

Examples include:

• Revenue by region last quarter

• Product engagement trends

• Customer churn over the past year

The defining characteristic is the opposite of OLTP: fewer queries, but significantly larger and more computationally intensive ones.

This distinction matters because the architectures that excel at OLTP often perform poorly at OLAP, and vice versa. Storage formats, indexing strategies, compute architectures, and optimization techniques are frequently designed around one workload at the expense of the other.

For most of computing history, companies solved this problem by running two separate systems: an operational database for applications and an analytical database for reporting. Data was continuously moved between them through pipelines and ETL processes.

The tension between those two worlds, and the industry’s repeated attempts to unify them, is the engine behind nearly every major database innovation that follows.

The Evolution of Databases

Each era of database innovation emerged in response to a constraint created by the previous one. Some optimized for consistency. Others optimized for scale, performance, developer experience, or analytical power.

Yet nearly all of them revolved around the same unresolved question:

Can a single system serve both operational and analytical workloads, or will those worlds remain separate?

The history of databases is, in many ways, the history of that debate.

Supabase: Building the Backend for AI-Native Apps

If PostgreSQL won the database wars, Supabase is trying to win the platform layer built on top of it.

PostgreSQL (a.k.a. “Postgres”) is one of the world’s most widely used open-source databases. As mentioned, at the core, a database stores and organizes application data - e.g., user accounts, transactions, product catalogs, chat messages. Over the past decade, Postgres has become the default choice for many modern software teams because of its reliability, flexibility, and large ecosystem of extensions.

Supabase started by making Postgres easier to deploy and manage in the cloud. Rather than forcing developers to assemble dozens of separate services around a database, it packaged the common backend components most applications need into a single platform.

Supabase started as a managed Postgres offering and has expanded into a full backend platform that includes authentication, storage, real-time subscriptions, edge functions, autogenerated APIs, and native vector search through pgvector. What we’re hearing from developers is that they increasingly want a complete backend, not just a database —> a single platform that handles the most common application requirements out of the box.

The company sits at the intersection of two powerful trends: open-source infrastructure and AI-assisted software development. Many AI coding tools and app generators now deploy directly onto Supabase by default, making it a natural backend for the growing population of developers building with AI. Unlike Firebase, which relies on proprietary abstractions and a NoSQL data model, Supabase is built directly on PostgreSQL. Developers get the benefits of a managed platform while retaining the ability to self-host, migrate, or access the underlying database whenever they choose.

In many ways, Supabase is doing for Postgres what Shopify did for commerce: taking a powerful underlying system and wrapping it with the tools, workflows, and developer experience needed to build production applications quickly. The database remains the foundation, but the value increasingly comes from the platform layered on top.

As Supabase scales following its recent funding round at an $10B pre-money valuation, the central question is whether it can preserve the simplicity that fueled its adoption while supporting increasingly complex production workloads. PostgreSQL remains one of the most trusted and widely deployed databases in the world, but enterprise-scale applications introduce challenges around global replication, connection pooling, multi-region deployments, and horizontal scaling.

Curious to see how Supabase evolves its architecture to handle sharding, distributed workloads, and the growing demands of AI-native applications. As AI-generated software matures from prototypes into mission-critical systems, developers will need infrastructure that combines ease of use with enterprise-grade reliability. If Supabase can navigate that transition while maintaining its open-source ethos and developer-first experience, it has a credible path to becoming the default backend platform for the next generation of software development.

All of that said, one of the reasons I’ve been so excited to follow Supabase is the team behind it. Paul Copplestone has been one of the most thoughtful, developer-first founders I’ve met. His relentless focus on open source, product quality, and empowering developers has been a major driver of the company’s success. I’m excited to see what he and the broader Supabase team continue to build in the years ahead.

When I first started recruiting for venture roles ~2 years ago, infrastructure software felt like a daunting black box. Over the past couple of years I’ve learned a lot, but there’s still plenty I’m figuring out — which is why I’m starting this series.

Infra Breakdown is my attempt to deconstruct the different layers of the infrastructure stack, aiming to (i) crystalize my own perspective and (ii) hopefully share a few helpful insights with you along the way.

For those less familiar with the space, many of today’s most valuable companies aren’t building the consumer applications we use daily (like Airbnb or Netflix). Instead, they are building the fundamental pipes and platforms that those applications run on.

The scale of this sector is often underestimated. In fact, the public market value of infrastructure software assets is roughly 3x that of application software – currently valued at $4.9T on $590B of total revenue (hopefully a bit higher by the time you read this… unless SaaSpocalypse has its way).

To put it in perspective, there are many companies that sit underneath the visible software economy: Databricks powers data pipelines, Snowflake runs analytics workloads, Supabase runs application backends, Stripe runs payments… the list goes on.

These are the “picks and shovels” that enable millions of other products to exist. Yet, despite powering huge portions of the internet, infrastructure remains one of the least understood parts of the technology stack.

This series aims to break it down, layer by layer.

Why Infrastructure Software Matters Now More Than Ever

Three structural shifts are expanding the importance of infrastructure software.

1. Application / software creation is accelerating: AI tools, low-code platforms, and developer frameworks (from the Lovable / Claude Code’s of the world to platforms like Canva) are dramatically lowering the cost of building software. The number of applications being created is exploding – 72% of developers who have tried AI use it every day; GitHub reported ~986M commits (+25% y/y)

2. Developers no longer build everything themselves: A decade ago, building an application required provisioning servers, managing databases, setting up authentication, and operating infrastructure. Today, developers increasingly assemble products from managed services (e.g., Supabase, Neon, etc.).

3. Infrastructure platforms capture compounding value: The companies providing these foundational layers benefit from powerful dynamics: Every new app increases demand for infrastructure → Developer ecosystems compound over time → Switching costs rise as systems scale.

This is why infrastructure software companies frequently become some of the most valuable companies in technology.

The Evolution of the Infrastructure Stack

Most modern software systems are often described as a layered stack. A few years ago, this idea was popularized as the “modern data stack” thesis: a clean set of modular layers where each company owned a specific function (fun Reddit thread on the rise and fall of the modern data stack).

The logic behind this framing was fair: software infrastructure should behave like Lego blocks (i.e., cloud data warehouses + SaaS tooling replacing monolithic platforms) – teams could assemble best-of-breed tools at each layer, and startups could win by owning a single, well-defined layer of the stack.

From Static Stacks to AI Gravity Wells

In practice, the software market rarely stays neatly layered for long. Microsoft is the classic example of this as they initially found their wedge in infrastructure with Windows (‘80s - ’90s) → expanded into the application layer with Office (‘90s - early ’00s) → and then scaled cloud / developer services with Azure (2010s). And they’ve kept pushing on since with acquisition / expansion of Teams, Dynamic 365, GitHub, etc.

Microsoft didn’t just stay in its lane; it expanded until it owned the entire road.

Instead of thinking about software as a static stack, it is more accurate to view it as a series of platform control planes sitting at critical points in the developer workflow. In the AI era, this transition has accelerated. The “Modern Data Stack” was a useful map for an earlier phase of the ecosystem, but the winners of this next era aren’t just modular layers — they are gravity wells.

In the current AI infrastructure landscape, the stack is collapsing into consolidated power centers:

• The Model as the Operating System: We are seeing a pivot where the LLM is no longer just a component; it is becoming the central orchestrator for compute, memory, and logic.

• Workflow Integration over Modularity: Enterprises are moving away from “best-of-breed” point solutions (i.e., separate vector DB + separate monitoring tool + separate prompt manager) in favor of integrated platforms that own the end-to-end lifecycle.

• The Data Moat: The “gravity” is generated by where the data lives. If a platform owns the context, it eventually owns the application.

The real architecture of the AI era looks less like a ladder and more like a set of overlapping platforms that expand until they dominate the entire developer workflow. To win today isn’t to be a better layer, but moreso becoming the platform that the rest of the ecosystem is forced to orbit.

In this series, we’ll be breaking down the critical points of the AI infrastructure stack.

Next up: Databases.

AI health coaches aim to solve this challenge and represent the next abstraction layer in wellness: shifting the value proposition from passive tracking to dynamic guidance.

This shift is not merely a trend; it is the result of three converging forces:

1. Interface Evolution: LLMs are replacing static dashboards with conversational dialogue, turning “data viewing” into “problem-solving.”

2. Biometric Liquidity: Sensor density (e.g., Continuous Glucose Monitors, high-fidelity wearables, etc.) has reached a usability threshold where continuous, multimodal data is finally reliable.

3. The Optimization Mindset: A cultural pivot from episodic, reactive care toward proactive, continuous performance and longevity.

As a result, we’re seeing longstanding fitness players pushing to productize the “coach” concept inside existing consumer surfaces like wearables and subscriptions (e.g., Fitbit Premium’s AI coach powered by Gemini).

In this post, I’ll aim to break down:

(i) what the current capital cycle within healthtech tells us about the opportunity,

(ii) why this moment represents a structural inflection point,

(iii) how the AI health coach landscape maps today, and

(iv) where durable value is most likely to accrue and the risks that could undermine it.

The Capital Cycle in Digital Health

Since 2021, >$80B of venture funding has flowed into health tech companies, initially propelled by COVID-era demand for telehealth, home diagnostics, and remote monitoring. After a notable pullback from pandemic-era peaks, we’re seeing digital health funding rebound in 2025 w/ Rock Health reporting that U.S. digital health startup funding reached $14.2B in 2025, up 35% y/y, with health AI taking a majority share.

I enjoyed a recent LinkedIn post from Julie Yoo (General Partner at a16z), who noted that just a few years ago, “calling yourself a consumer health company in front of VCs was a virtual death sentence.”

Today, that “death sentence” has evolved into a relevant investment thesis. We are seeing a massive shift in capital allocation as AI-enabled companies captured ~54% of total funding dollars in 2025, up from ~37% the prior year; these companies also commanded a ~19% premium on average deal size.

This funding profile reflects a wider shift in investor priorities over time:

• 2020-2022: Capital flowed primarily to telehealth access, remote diagnostics, and point solutions driven by the pandemic demand.

• 2023-2024: Post COVID, there was a period of repricing and selectivity, with emphasis on clinical outcomes and sustainable unit economics.

• 2025-present: A renewed focus on AI and intelligence layers that synthesize existing data into contextual decisions, not just data collection (represented with investments like Abridge’s $316M Series E for ambient clinical reasoning, Oura’s $900M Series E for longitudinal longevity data, Hippocratic AI’s $126M Series C for generative nurse agents, and Function Health’s $298M Series B for consumer-led AI lab synthesis).

This backdrop establishes a strong foundation for the emergence of AI health coaches as a new category of personalized wellness infrastructure.

The Inflection Point: Why Now

Similar to many other data-rich categories, AI unlocks (in health care and other categories) aren’t emerging because the underlying data is “new”. They are emerging because the interface, infrastructure, and incentives have finally aligned.

These structural shifts make this moment different from prior waves of digital health.

1. The Interface Shift (from Dashboards to Dialogue): Health data has existed for years, but interpretation hasn’t scaled linearly. LLMs transform static dashboards into conversational, contextual guidance (as exemplified by the snip of Oura’s UX and user flow below). Instead of reviewing graphs and guessing what to do, users can ask direct questions and receive personalized recommendations in real time.

2. Data Liquidity + Sensor Maturity: Wearables, continuous glucose monitors (CGMs), and health platforms now generate continuous, multimodal data across sleep, recovery, activity, glucose, stress, temperature, etc. Subscription-based wearables have normalized longitudinal tracking. What changed is not the volume of data, but the ability to synthesize it. AI models can now integrate disparate signals into coherent, personalized daily plans at scale – something that previously required human coaching and input.

3. Economics + Cultural Pull Toward Optimization: Consumers are moving away from episodic care toward a lifestyle of proactive optimization, a trend led by younger cohorts who prioritize continuous mental well-being (as seen in share of digital health adoption by cohorts below). Historically, human-led coaching has been plagued by poor scalability. AI solves this by enabling high-frequency personalization without a proportional increase in labor costs (note: cost delta not as relevant relative to other labor intensive industries). This unlocks a hybrid model where AI manages daily engagement and humans handle high-stakes escalation – driving superior unit economics while maintaining patient trust.

Combined, these forces create a structural inflection point: AI health coaches are not just possible, but economically and behaviorally viable at scale.

The Landscape: From Features to Infrastructure

The AI health coach category is maturing into a fragmented ecosystem rather than a monolithic market. Differentiation is now driven by three primary vectors: data ingestion methods (biometric sensors vs. self-reported logs), clinical depth (wellness vs. regulated interventions), and proactivity (passive insights vs. automated behavior loops).

Below is a current view of the market map, with representative companies by maturity and category:

Where Durable Value Accrues

As we’ve seen in other healthtech categories (e.g., Ambient AI), health coaching (at least today) will not be a single breakout winner. The value will concentrate around control points where data, distribution, and measurable outcomes reinforce each other over time. Two pools stand out as structurally advantaged:

Sensor-Native Platforms (Consumer)

The most durable position in AI health coaching sits with companies that already own: (i) continuous biometric data, (ii) a daily engagement surface, and (iii) a subscription relationship.

Players like Google/Fitbit, Oura, WHOOP, and Garmin are not starting from zero. They already collect multi-year datasets across sleep, HRV, strain, temperature, and activity. AI coaching becomes a layer that interprets what they already capture, creating four compounding advantages:

1. Longitudinal personalization: Multi-year datasets improve recommendation quality and increase switching costs as the model understands individual baselines (e.g., stress response and recovery dynamics).

2. Zero marginal distribution cost: Adding an AI coach to an existing subscription base (as seen with Google’s 2026 expansion of the Gemini-powered Fitbit Coach) is a retention and ARPU expansion lever rather than a standalone monetization gamble.

3. Habit embedding: AI guidance delivered at the moment of wake-up or pre-workout has structural behavioral leverage because the hardware is already part of the user’s daily routine.

4. Expansion to the enterprise: These companies are pivoting toward workforce resilience dashboards (ex. WHOOP x Lean Business Services, Oura x DoD) using AI as a population-level triage tool to flag rising burnout or illness trends across a team before they lead to absenteeism.

Note: While I am bullish on sensor-native platforms as the primary unlock for effective AI coaching, it is important to distinguish between a technological catalyst vs. an investable category at this specific maturity stage. From a risk/reward perspective, to get excited about investing in these assets, I’d personally be watching out for: [i] the next horizon of growth for these platforms (e.g., broader clinical / enterprise adoption), [ii] valuation and exit math (are these assets priced to perfection?), and [iii] data commoditization as biometric sensors become ubiquitous.

Outcome-Validated, Employer & Payer Platforms (Enterprise)

The second pool sits where the money actually lives: with employers, insurers, and risk-bearing providers. Companies like Omada, Lark, and Hims & Hers (minus the current regulatory ?s) win by proving ROI, not just engagement.

1. ROI drives budget durability: Unlike consumer apps that churn when a user loses interest, platforms that move the needle on glycohemoglobin levels or weight loss unlock multi-year enterprise contracts.

2. Hybrid model improves margin: AI now allows these platforms to automate 90% of daily interactions while reserving high-cost human experts for clinical escalation. This preserves medical trust while radically improving unit economics.

3. Integration into care pathways: Platforms embedded in employer benefit stacks or care management workflows become harder to displace than standalone wellness apps.

The Strategic Risks: The “Bear Case”

While the capital for AI health coaching is expanding, the path to a defensible category winner is fraught with structural headwinds:

• Coaching becomes commoditized. If foundation models make conversational intelligence ubiquitous, differentiation collapses. Every app can bolt on a “coach.” Distribution wins, standalone platforms struggle.

• Platform incumbents absorb the category. Apple, Google, or major ecosystem players could integrate AI coaching natively into health data hubs, compressing margins for independent apps.

• Engagement fatigue sets in. Health apps historically face churn once novelty fades. If AI advice feels repetitive or non-differentiated, retention may erode.

• Regulatory gravity increases. As AI recommendations approach medical territory, liability and compliance burdens rise. The closer coaching gets to clinical decision-making, the tighter scrutiny becomes.

What’s Next?

As the “noise” of raw data settles, the “signal” of AI-driven intervention is becoming the new standard for healthtech. Moving forward, the indicators of category maturity will be:

1. “Closed-Loop” Breakthroughs: Integrations where an AI coach doesn’t just suggest an action but executes it (e.g., adjusting your smart thermostat or grocery delivery based on recovery and glucose spikes).

2. Hardware Consolidation: Whether independent players (e.g., Oura, WHOOP) can maintain their independence or become acquisition targets for Big Tech looking to buy underlying biometric data liquidity.

The winner won’t be the “bot” with the best personality; it will be the platform that most seamlessly integrates into the fabric of a user’s daily life, turning biological signals into effortless longevity.