This week we take a look at how AI tools are transforming the Hardware Engineering/ R&D services landscape. Check out this thoughtful compilation of around 37 companies working in this space by Jonathan Beri

Broadly there are 4 lanes of opportunity for AI tools in Hardware Engineering

1. Hardware Design: Massive TAM; Maximum action (14/ 37 mapped startups); $20-25B market size in just tooling - EDA and PCB schematic/ layout design biggest sub-verticals. Underserved areas: everything before the schematic i.e., SOC selection, power/ thermal budgeting and architecture tradeoffs.

2. Firmware Development: Embedded Software tooling is a $20-22B market in 2026 (10% CAGR). 11/ 37 startups here. AI agents for Firmware development is the hottest bucket. Unsolved opportunity: Debugging layer.

3. Edge AI: Fastest growing frontier: $3-5B in 2026 with 10x growth potential. THIS IS THE MISSING CATEGORY in @Jonathan’s map. Edge AI ML Ops and Lifecycle management that solves for hardware fragmentation and software portability is the key.

4. Manufacturing: $8-12B TAM for tooling. Inspection, Design for Manufacturing, Production test generation, Calibration station, yield analytics, field failure analysis are all widely open opportunities for AI startups to target.

Just the tooling layer is a USD 50-65B market. Add services and you are looking at 15x bigger TAM - Global Product Engineering services market is around USD 1.3 trillion and India captures around 10-15% of this at USD 145+ billion.

Opportunity for India

Multiple interesting AI tools have emerged from India in the last couple of years. Personally very excited to see @CraftifAI (@Pratik Sharda @Yashwant) a portfolio company, appearing on this map. CraftifAI is building AI agents cutting across Hardware design, Firmware development and Edge AI.

Other notable Indian startups include: @H2Loop (SLMs for firmware/ drivers), @Maieutic Semiconductors (GenAI copilot for analog IC design), @Autocure Technologies (Layout and routing autonomy), @1Buy.ai (Electronics procurement), @ThirdAI automation (Copilot for semiconductor manufacturing engineers), @Tattvam AI (Accelerate chip design/ EDA iteration loop)

Two biggest opportunities hiding in plain sight for India:

1. Delivering R&D outcomes using AI platforms - somewhat like reimagining Vayavya labs for an AI first/ AI native world.

2. Taking lead in becoming the Edge AI capital of the world (e.g. building the data/ model pipelines for enabling hardware to become smarter)

If you or anyone you know is building in this massive landscape, I would like to chat with you. Please ping/ DM.

Which of these two do you think India can capitalize on in the next 2-3 years?

#MarketMapMondays #AIforHardwareEngineering #ChipDesign #EdgeAI

India has just solved 1 (Planning). The other 2 need massive attention.

Over the weekend, I was chatting with a Unicorn founder. He had recently visited China and was blown away by China’s lead in deep tech. That led me to go down the rabbit-hole of what led this growth. Found interesting research by that explains the 3 key drivers behind China’s growth in Deep Tech:

1. PLANNING: Every 5 year plan lays out clear priority technology industries. This level of focused planning goes all the way back to 1985. Priority industries for 2025-30 cycle: Integrated circuits, Embodied Intelligence, BioManufacturing, Next Generation Batteries, Commercial Space, Domestic Large Aircraft, Low Altitude Equipment, Green Hydrogen, Brain-Computer Interfaces and High End Medical Devices

2. TALENT: “Thousand Talent” program launched in 2008 to attract established Chinese researchers back; 7000 folks had moved by 2018 giving a real push to China’s deep tech dominance.

3. INDUSTRIAL COEVOLUTION: Created overlapping deep-tech industrial ecosystems across 10+ deep tech industries; each industry feeds of the other - e.g. Batteries, Smartphones, Electric Vehicles, Lidar, Drones, Autonomous Vehicles, AI, Semiconductors, Humanoid robots and Industrial robots.

WHAT EXACTLY IS INDUSTRIAL CO-EVOLUTION?

This is about carefully evolving synergy across 4 dimensions:

• Supply - Build domestic suppliers in upstream industries - simplifies sourcing components and enables collaboration to modify/ innovate as required (e.g. Batteries -> Smartphones/ Electric Vehicles, Lidar -> autonomous vehicles, Auto components -> electric vehicles, Semiconductor -> AI)

• Demand - Create a pool of domestic buyers in downstream industries; unlocks demand/ industry revenue for deep tech startups. Tariffs/ Local content requirements to encourage purchase from local startups.

• Technology - Identify and invest in know-how areas that benefit multiple industries - e.g. Poly-silicon production helps both solar photovoltaic cell and semiconductor chip manufacturing. Inverter capability helps for solar, TVs, railways and telecom equipment.

• Scale - If a product can be input for multiple industries, then having all those industries domestically ensures economies of scale for that product. Eg. Lithium battery supply to Consumer electronics, EV and energy storage industries

LEARNING FOR INDIAN DEEP TECH ECOSYSTEM

India’s Deep Tech ecosystem is showing lots of promise. In the last 3.5 years, I have been lucky to work with multiple deep tech founders closely - Inspecity (Space robotics Propulsion), Green Aero (Transonic platform), Armory (Counter drone intelligence), Aspera (Autonomous Cargo planes), Meine Electric (Long duration energy storage), Uncharted AI (Robotics for mining), Mandrake Bio (Gene editing platform), Navana AI (Voice AI), Kubo Care (Elder care) and Cautio (Dash Cam for road safety). The raw ambition of the founders, the depth of IP and the pace of execution all shows that we are sitting on a solid foundation as we look at scaling Deep Tech over the next decade.

To top it, the government has tackled the PLANNING factor with the launch of the RDI scheme - 7 priority sectors of interest identified (e.g. Energy/Climate, Deep Tech, Biotech/ Life Sciences, Digital Economy ) and INR 1 lakh crore funding allocated to back startups and fund managers focused on Deep Tech. This is a great step as we bridge the gap in R&D as a % of GDP (0.6% in India vs. 2-3% for China/ US).

But we still need to address the two other drivers - TALENT and INDUSTRIAL CO-EVOLUTION:

a) Create a plan to attract diaspora talent from all over the world - need a systematic/ structural program to incentivize researchers/ technologists from outside India to move and start deep tech innovation from here. While this has started in trickle over the last few years (e.g. Green Aero Founder Prithwish, Agnikul founder Srinath), we need to make this a flood.

b) Take a ecosystem approach - similar to Chinas industry coevolution - pick a group of industries that can feed off each other and drive synergies across demand, supply, tech and provide scale - e.g. Biotech, MedTech and Healthcare or AI, Semiconductor and Robotics or Manufacturing/ Industrial robotics. Have policies that support the industrial coevolution which will enable a thousand startups to bloom in these areas.

If India could only bu`ild one cluster properly this decade — AI / Semis/ Robotics, or Biotech/ MedTech/ Healthcare AI — which one would you pick and why?

#MarketMapMondays #DeepTech #China #IndustrialCoevolution

Source articles: China's Long Tech Game by Kyle Chan, China's Overlapping tech-industrial Ecosystems

From Physical AI to how AI is shaping construction of the physical world - thanks to an insightful market map by Dawn Capital.

There are broadly 3 phases in Construction: Design, Deliver and Operate. The 3 biggest software sub-categories belong to the Design phase:
1) Computer Aided Design (CAD) - Translates physical idea/ concept to a technical drawing/ design
2) Building Information Modeling (BIM) - Digital representation of a built asset connecting all information (e.g. materials, quantities, systems, schedules and relationships)
3) Mechanical, Electrical and Plumbing (MEP) - Engineering layer that makes a building function (HVAC, power, water, drainage etc)

Multiple other categories sit in the space as a building moves from design to a finished building - e.g. Compliance/ permitting, Construction platforms, Project execution tracking, Asset management. Mostly smaller and less competitive.

Disrupting AEC with AI is not about PROMPT TO IMAGE . The real opportunity is PROMPT to a BUILDABLE SYSTEM. To break incumbent chokehold, founders need to address 4 challenges.

1️⃣ Consistency: Design that is consistent internally (e.g. background, context, site boundaries ) and externally (regulations, guidelines).
2️⃣ Editability: Not just editable drawings but editable systems of work with all interdependencies addressed in BIM.
3️⃣ Collaboration: Multi player workflow across architects, interior designers, structural engineers, MEP engineers, contractors, suppliers, surveyors and authorities with Figma like UX
4️⃣ Performance: 10x better latency / real time performance compared to incumbents .

Solving this is hard, but not IMPOSSIBLE. My bet is EDITABILITY is the hardest to crack and will break the incumbent lock-in. Consistency, Collaboration and Performance are table stakes for most Vertical AI.

Many startups are going after this by picking one point problem at a time. A few Indian founders are also aggressively going after this opportunity:
- Sonar Labs - AI agents for Material Tracking in construction (Vijay Chellaram Dhruv Tekchandani)
- Bandhoo Solutions - AI led Project monitoring, Lifecycle cost management and manpower mobilization (Prashant Gupta)
- PiAxis - AI powered Revit detailing for AEC firms (Swetha Kochar)
- MeltPlan - AI Preconstruction software for General Contractors (Kanav Hasija)

If you are or anyone you know is building in the AI in AEC space, I would like to chat.

I believe India has a structural advantage in this space thanks to a strong domestic industry/ demand without legacy software baggage. This coupled with ability to deliver outcomes with outsourced services can be a game changer. Do you agree?

Source article: https://www.dawncapital.com/content-hub/can-ai-rebuild-the-software-behind-the-physical-world

Even as startups and founders vie to capture this massive opportunity, as a VC you need to think about where THE NEXT BIG BREAKTHROUGH is going to come from. Sapphire Ventures monthly market memo captures this beautifully(link in comments). Five areas look most promising for the next big breakthrough:

1️⃣ Long Horizon Autonomy: Current benchmark (METR time horizon) - Claude Mythos preview completes 50% of software coding tasks working autonomously for 16 hours vs. 9 seconds for GPT-3 in May 2020. Progress here will unlock massive TAM enabling automation of many categories of knowledge work

2️⃣ Continual learning: How do we stop models from catastrophically forgetting and continually learn on the job? Today's models are frozen in time; breakthrough here could accelerate AGI timelines

3️⃣ Recursive Self Improvement: The biggest bottleneck today is building AI. Can we make AI systems build themselves in a continuous loop? This will unshackle progress in AI from human capacity

4️⃣ Physical AI: Physical AI is still waiting for ChatGPT moment. In the last few years the intelligence/ brain aspect has grown faster than the body. Complex problems still waiting to be cracked - manipulation/ dexterity. $5 trillion opportunity waiting for a breakthrough

5️⃣ World Models: How can we take Generative AI beyond the realm of text into spatial intelligence? Imagine a model that can understand the world, plan ahead, imagine consequences of actions before taking them. This is frontier territory and a breakthrough here will impact robotics, AV, gaming, simulation in big ways

WHERE INDIA STANDS

In the last couple of years, we have seen a few startups in India trying to go after these breakthroughs:
- Hexo Labs(Kunal Bhatia and Vignesh Baskaran) solving for Recursive Self Improvment. Check out their SIA paper.
- Quite a few in the Physical AI space - particularly around egocentric data collection and specific robotics use cases (InspeCity in space robotics, Cravv in Physical AI for cooking, Uncharted AI for mining robotics) to name a few

But this is just a tip of the iceberg. Which of these areas do you think an Indian startup will lead the breakthrough? If you or anyone you know is building in these frontier areas - would love to chat and exchange notes. Ping/DM.

#MarketMapMondays #GenerativeAI #WhatNext #FrontierResearch

Source: Sapphire Ventures Monthly Market Memo

2) FP&A/ Forecasting - 6 monthly time to value matters most here

3) AP & expense management - tools that show transaction level savings are most sought after

4) Treasury and Cash - Focus on cash visibility and liquidity

5) Audit and Compliance - security and bottleneck the biggest compliance here

6) Procurement and spend management - High visibility area with interest from across the org

7) Tax - Highest propensity for AI adoption

8) Agentic and Copilot layer - Small, fast growing with both pilot and production answers:

Of these AGENTIC & COPILOT LAYER is the area with maximum interest; albeit with tons of questions around access, approvals, audit trail, and human review.

India historically has been a F&A BPO powerhouse with deep domain understanding. Now we are starting to see AI Native players building on that expertise. We are seeing startups approaching this market either as a product-first play as well as platform led services play. Few examples: Hyperbots Inc. is doing interesting stuff on AP automation. SignalsHQ is building Platform-led-services for Audit/ Tax layer. ClearTax is a leading product-first player in the tax/einvoicing space.

The other big question in the CFO AI stack: who is primed to win - the Incumbents or the AI native competitors?? The jury is yet to be out as there is a strong case for each side to win this war - ownership of data / workflow for the incumbents and AI-native thinking/ speed for the competitors.

Who do you think wins this? share your thoughts in comments. If you are building in this area - would love to chat/ DM.

The map we’re working off comes from Ashish Kakran and team at Seligman Ventures, and it’s one of the more useful ones I’ve seen on this space in a while.

Two things make it worth your time. First, it structures Physical AI around four core layers instead of treating it as one blob called “robotics.” Second, it doesn’t stop at the general purpose robot companies, it goes deep, tracking close to 400 companies across those layers and the verticals built on top of them.

The four layers

Intelligence covers the world models and robot brains, the perception and decision-making stack that lets a machine understand and act in physical space.

Embodiment is the robots and platforms themselves, the hardware that intelligence gets deployed onto.

Deployment and data loops is where robots actually go to work, and where the usage data that improves the next generation of models gets collected.

Infrastructure is the unglamorous but essential layer underneath all three: compute, software, and hardware that the rest of the stack depends on.

Most of the public conversation about Physical AI lives in the first two layers. The real fight for defensibility is probably going to happen in the third and fourth.

Eighteen verticals, and most have no household name yet

Beyond the general purpose robotics companies, the map tracks players across eighteen distinct verticals:

• Warehouse and fulfilment

• Autonomous mobility

• Industrial automation and manufacturing

• Construction and built environment

• Security and defense

• Energy, utilities and inspection

• Marine and ocean robotics

• Agriculture

• Retail, restaurant and service

• Food and beverage processing

• Waste, recycling and sortation

• Cleaning and facility

• Last mile and delivery

• Science and lab automation

• Surgical and interventional robotics

• Healthcare and assistive robotics

• Data center and IT automation

• Space robotics

That’s a wide spread for a category most people still picture as warehouse arms and the occasional viral humanoid demo. Each of these verticals has its own physics, its own regulatory environment, its own buyer, and its own sales cycle. A robot that works in a warehouse doesn’t translate to one that works on a farm or in an operating room. That’s exactly why this space rewards depth over breadth, and why generalist platforms will keep losing ground to teams who’ve picked one domain and gone deep.

Why India could win big here

Here’s the number that stands out to me: 400 companies mapped globally, and India barely registers on this map yet. That’s not a knock on India’s robotics ecosystem, it’s the opportunity. A market this fragmented, spread across eighteen verticals with their own domain requirements, isn’t going to be won by two or three category winners. It’s going to be won vertical by vertical, by founders who understand a specific physical workflow better than anyone else building a horizontal platform ever could.

At Antler India, we’ve made early bets in exactly this pattern. Inspecity is building in space robotics. Uncharted AI is building in mining. Cravv is building in food and cooking. Three completely different domains, but the same underlying thesis: Physical AI in India isn’t one market to win, it’s dozens of narrow, deep ones, each waiting for a founder who knows the domain cold rather than a platform trying to be everything to everyone.

We’re still early. Given the size of the global opportunity and how little of it currently touches India, we want to back a lot more of these bets.

So, where’s the real gap: intelligence, embodiment, or the long tail of verticals nobody’s built for yet? If you’re building in the Robotics Developer Economy or Physical AI more broadly, or you know a founder who is, reach out. I’d love to compare notes.

Check out last week’s piece on the Robotics Developer Economy and the $125B devtools opportunity here!!

In 2020, Tyler Jewell published the Developer-Led Landscape, a map of the $50B+ market of tools sold to, influenced by, or used by software developers. Pull up Tracxn’s Application Development Tools category today and the scale is hard to miss: 1,500+ funded startups, close to $30B in funding, 25+ active unicorns (Databricks, Snowflake, Lovable, Supabase, AppDynamics), and nearly 400 Series C+ companies, before you even count public giants like Atlassian, GitLab, Datadog, MongoDB, and FAANG.

That piece stuck with me for years. I kept waiting for someone to build its equivalent for robotics. Nobody did. Everyone has an opinion on humanoid robots and the BOM (Bill of Materials) on the hardware side. Almost nobody has done the work on the layer underneath them: the tools, data, and intelligence infrastructure that every robot, humanoid or not, gets built on.

So over the last few weeks I went down the rabbit hole and did the work myself. What follows is, as far as I can tell, the most researched piece anyone has published on the robotics developer economy. The map was built from, and stress-tested against, every input I could get my hands on:

• Seven in-depth conversations with robotics founders and operators. One runs 200+ deployed robots with a 70-person engineering team. One has spent two decades across robotics startups and autonomy companies. One is building at the manipulation frontier and tore apart my first draft; his corrections are in here, and they made the thesis sharper. Four portfolio founders building at the cutting edge of Physical AI

• 500+ practitioner comments across five community threads plus a developer survey. This is the unfiltered voice of people fighting ROS at 2am.

• ~230 funded robotics companies with Indian presence, mapped one by one.

The result: a comprehensive map of a market going from $13-20B today to $125-185B by 2035, triangulated from four independent angles and a specific answer to the question I care most about: where does India win? You might be wondering why does it make sense to look at this now:

1. Core robotics developers, the people working on ROS, perception, planning, controls, and robot learning, number around 1 million. Another 1-1.5 million work in adjacent areas like firmware, industrial automation, and embodied models. That’s just 3-5% of the global developer base today (~47M developers), compounding at roughly 20-25% a year. Sustained, that puts the pool near 12-15 million within a decade. Even if AI agents absorb 30-40% of the work, you’re looking at a ~8-10 million-strong robotics developer pool by 2035 - with ability to influence and purchase productivity tools

2. Robotics and Physical AI are inflecting, and the development cycle is the constraint. Today it takes 6-18 months to bring a robot into production. Roland Berger’s 2026 humanoid study names AI architecture as the industry’s primary bottleneck and estimates software and regulation lag hardware by 3-5 years. The hardware is arriving faster than our ability to develop, validate, and operate it.

3. The robotics developer economy today looks like the software developer economy did in the late 2000s. Weak package management, hand-rolled infrastructure, near-zero CI/CD, bespoke observability. A large share of software’s growth over the following decade came from the tooling wave that fixed exactly these gaps. That same wave is coming for robotics.

4. Most of the capital in robotics is concentrated around 2 areas - Simulation and foundation models. Everything in between, the operational middle where robots actually get built, tested, and kept alive, is comparatively unfunded. That asymmetry is the opening that this memo maps.

5. The last decade was about Autonomous Vehicles. The AV-lineage economy (driving simulation, scenario V&V, navigation stacks) that’s mature, capitalized, and consolidated. But the next decade is going to be about the dexterous-manipulation frontier (contact-rich sim, manipulation data, VLA foundation models) that’s pre-revenue, vertically integrating, and still deciding what its categories even are.

PART 2: THE CURRENT LANDSCAPE, AND THE TEN CHALLENGES THAT SLOW EVERY TEAM DOWN

Building a robot used to mean welding a chassis and writing custom code just to turn a wheel. That era is over. But the development lifecycle that replaced it is still brutally inefficient. Before sizing the market, it helps to walk the lifecycle a robot actually goes through, because every tooling category in Part 3 maps onto a phase of it, and every challenge below maps onto a seam between phases.

The lifecycle runs in six phases:

1. Design and build the machine: strategy/PLM, hardware design (CAD, URDF/OpenUSD), sensor and actuator selection, embedded and firmware development (RTOS, motor control, toolchains).

2. Build the robot software: middleware and dev infrastructure (ROS 2 and its commercial layer), perception and state estimation (SLAM, sensor fusion, spatial AI), planning/control/manipulation (MoveIt 2, Nav2, locomotion), logging and observability, and human-robot interfaces.

3. Train the intelligence: simulation and synthetic data, foundation models (VLAs), and the data pipeline that feeds them: teleoperation, egocentric capture, labeling, curation.

4. Validate and secure: unit/integration testing, closed-loop policy evaluation, scenario V&V, hardware-in-the-loop, fault injection, CI/CD, functional safety, and cybersecurity.

5. Deploy and operate: fleet management, OTA updates, remote intervention, telemetry, drift monitoring, configuration and calibration management.

6. Integrate and deliver solutions: the systems-integration and RaaS layer that turns robots into working automation. (A separate services market, roughly 10x larger, tracked outside the dev-tools TAM but impossible to ignore.)

Critically, the lifecycle isn’t linear anymore. The defining structural change of the Physical AI era is the data flywheel: phase 5 feeds phase 3. Deployment experience, every grasp, every failure, every intervention, becomes training data for the next model. Whoever owns that loop owns the moat, which is why (as we’ll see) the frontier labs are all racing to close it inside their own four walls.

The ten biggest challenges

From the primary and secondary research, following are the 10 biggest pain points in the current robotics development stack.

1. Dependency and environment-hell. The single loudest complaint in every community thread. ROS distributions are locked to specific Ubuntu versions; ROS, Gazebo, and packages interact brittly; a two-day fight to get SLAM, Nav2, and Gazebo working on one distro combination (and failing inexplicably on two others) is a normal war story. Practitioners keep multiple parallel machines alive; one reported four bare-metal Ubuntu installs on a single laptop just to keep working setups running. Installing a public codebase from GitHub can take hours. The naming chaos (Gazebo vs. Ignition vs. “new Gazebo”; ROS 1 vs. ROS 2) compounds it. As one commenter put it: robotics has a LOT of DevOps issues.

2. The infrastructure tax. Teams spend more time building dev environments, glue code, and internal tooling than building robots. “Every robotics company I’ve encountered is making their own tools” is a direct quote from the threads, and the operator side corroborates it: production code stays closed because it is the competitive advantage, so nothing accumulates into a shared ecosystem, and every new team pays the tax again from zero.

3. Cross-stack coupling. The deepest structural problem, and the one every practitioner conversation independently converged on: the hardest complexity lives at the interfaces between hardware, firmware, middleware, software, and AI models. There are no stable layer boundaries, so any hardware change cascades through kinematic models, firmware, drivers, and application code. “Why are the books out of date? Hardware changed. Why does the tutorial not work? Hardware changed.” Tools that solve the feedback loops across those interfaces are, in one senior practitioner’s words, “intuitively where the value is.”

4. Middleware architecture limits. ROS is at once the field’s greatest gift and its ceiling. It’s a non-deterministic, tightly coupled monolith: messaging, serialization, process management, and visualization bundled together, none swappable, none best-in-class. The canonical trajectory: teams adopt ROS, progressively replace pieces, and eventually rip it out for MQTT/UDP/custom runtimes. The determinism gap matters most. A deterministic core cascades into better testing, simulation, replay, and safety pathways, which is why challenge 4 sits upstream of challenge 7.

5. Documentation, tutorials, and knowledge rot. Under-maintained docs (academic origins, no commercial DX incentive, a small contributor base), tutorials pinned to dead distro combinations, and fragmented naming that makes even existing answers unfindable. Two nuances to consider

• Commercial vendors are no better (finding a page in a major industrial vendor’s manual is its own ordeal; “enterprise-grade” in robotics has meant locked-down, not well-documented).

• And the deeper cause is that most production robotics code is closed, starving the public flywheel of the examples, Q&A, and blog posts that made web-dev documentation self-sustaining.

6. The debugging and observability gap. Root-causing a failure across layers (was it a sensor fault, a tool mismatch, a driver, bad logic, a model?) is a multi-day ordeal. Telemetry and log pipelines are bespoke; “even visualizing a point cloud from the perspective of time and space is a pain.” Autonomy systems have to diagnose decisions from multiple AI and control processes in real time, and the introspection tooling simply isn’t good enough. Foxglove and Rerun are the bright spots practitioners keep recommending, which tells you how thin the field is.

7. Testing, V&V, and CI/CD, the unbuilt layer. In software, a test is a deterministic assertion. In robotics, outputs depend on physics and stochastic learned policies, so pass/fail becomes a statistical question over many rollouts, the scenario space is effectively infinite, and a failure wrecks hardware or injures someone. The result: robotics’ open-source origins left a near-total absence of unit tests, “demo culture” stands in for verification, and almost every serious company hand-rolls its own evaluation rigs and release pipelines.

8. The sim-to-real gap in unstructured environments. Here the founder feedback forces a split. Driving simulation is largely solved and well capitalized (Applied Intuition at a $15B valuation, Foretellix, NVIDIA): rigid-body vehicle dynamics, road scenario spaces, automotive buyers. Manipulation and unstructured-world simulation is not. Contact-rich physics (deformables, friction, cloth, liquids, cluttered homes, muddy fields) remains the open frontier, which is exactly why manipulation labs lean so heavily on real-world data collection instead. Practitioners agree from the deployment side: current tools are fine for structured, slow-moving industrial settings and break down precisely where the next decade of robotics wants to go. “The IMU isn’t as screwy as it is in real life.”

9. Data scarcity, the real bottleneck. Software developers have effectively infinite text and code to learn from. Robotics developers are starved. Physical-interaction data is expensive to collect, hard to curate, embodiment-specific (you can’t copy-paste experience across robot bodies), and hoarded. Enterprises won’t share operational data; “I cannot share this” is a direct quote from an operator explaining why multi-tenant SaaS is unworkable for him. This scarcity explains downstream symptoms the threads complained about years ago: why general LLMs are unreliable on robotics, why documentation doesn’t compound, and why Roland Berger now names training data as the constraint on leading developers.

10. Structural economics, why the pain persists (and the caveat on everything above). The most important challenge is the meta-challenge: software is a sliver of what a robot costs. In every institutional BOM (Goldman Sachs, BofA, Morgan Stanley), robot production cost is ~100% hardware. A leading robot operator I spoke with runs a cost structure of roughly mechanical 20% / electrical 30% / sensors 40% / misc 10%, no software line, with R&D at 16-17% of burn. Industrial buyers prize never-touch stability over iteration speed (you can’t hotfix a welding line; a missed edge case surfaces as cracked frames five years later). In-house tool-building is fast and rational for a capable team (”an observability pipeline takes 1-2 months at 200 robots”), NVIDIA gives world-class simulation free to sell silicon, and per-robot SaaS pricing struggles at fleet scale. One respected operator’s blunt read: externally sold robotics dev tooling is a hard business, the funded players are struggling, small customers don’t scale with you, and large customers write it themselves. Pain does not automatically equal market size.

PART 3: THE $125B OPPORTUNITY IN 2035

The map and the number

Add up the full software, data, and intelligence stack and the robotics developer economy goes from $13-20B today to $125-185B by 2035. Central CAGR ~22-30%, de-duplicated across layers; hardware, general CAD, and adjacent markets excluded.

Does $125B hold up? Triangulating from four angles

A TAM is only as good as its cross-checks. Let us triangulate this market size:

Angle 1, market-research extrapolation: consistent. IMARC sized robot software at $14.6B in 2022 growing to $43.5B by 2028 (~20% CAGR); rolled forward at a decaying 15-20%, that lands at ~$115-155B by 2035. Allied Market Research has $47.2B by 2030 at 27.3% CAGR; extrapolated at a moderated ~20%, ~$115-120B by 2035. The broad range of $125-185B brackets both, and the implied CAGR (~23-26%) sits inside what the research houses already publish. The 2025 base ($13-20B) is if anything conservative against their current-year figures; the de-duplication explains the gap.

Angle 2, the cost-structure lens: this is the load-bearing assumption, so I’ll name it. Today, software is a small share of what a robot costs: absent from every institutional BOM, ~10-17% of R&D burn at classical-stack companies, under 10% of system cost by veteran-operator estimates. If the total robot economy (hardware plus deployment) reaches $500B-1T by 2035, a $125-185B software/intelligence layer implies software’s value share doubles or triples in a decade. That’s not where the industry is; it’s where the map bets it’s going. The precedent is automotive: McKinsey has automotive software and electronics growing at 4.5% CAGR to $519B by 2035 against ~1% vehicle-market growth, value migrating from bodies to brains in exactly this pattern. Roland Berger (software lagging hardware by 3-5 years; AI architecture as the primary bottleneck) and Morgan Stanley’s brains/bodies/integrators framing institutionalize the same migration thesis. This memo is explicitly a bet on that migration.

Angle 3, the intelligence-cost lens: the AV industry already proved the magnitude. The best-documented robot ever built is the autonomous vehicle, and its numbers show what “software and intelligence” costs when the robot has to work in the open world: McKinsey’s executive survey puts early L4 use cases at $0.6-1.5B in technology development, autonomous trucking at $4B+ of software development, L4/L5 robotaxis at $5B+, with cumulative sector burn of $50-67B through 2030. Single corporates match it (Toyota plus NTT: $3.26B committed to AV AI software alone). Since the hardware of a robotaxi is a modified car, the AV sector shows that for autonomy-grade robots the intelligence stack costs multiples of the hardware program, paid as fixed R&D, invisible on a BOM. That’s the floor under the $30-50B+ intelligence bucket.

Angle 4, per-unit monetization: software revenue is becoming visible. All major humanoid platforms now charge $3,000-$8,000/year in software fees (model updates, fleet dashboards, telemetry, security). On a $50k robot over five years, monetized software is already ~25-45% of hardware price. A rough cross-check on the deployment/ops bucket (Morgan Stanley’s ~13M humanoids by 2035 × $3-8k/yr = $40-100B/yr, before AMR and industrial fleets) actually exceeds my $20-35B range. Maintenance-plus-software runs 10-20% of purchase price annually. The recurring-software layer is real and compounding.

The honest bear/mid/bull. Roughly half the broad TAM rests on two bets: (1) the intelligence layer becomes purchasable third-party revenue rather than internal IP and NVIDIA bundling (~$30-50B of swing), and (2) fleet operations matures into a standardized category rather than staying an in-house module (~$20-35B of swing, the direct objection of my most skeptical operator input, who asked whether “fleet ops” ever becomes a dev-tool category rather than a “fleet deployment and runtime platform,” and noted $25-50/robot/month is hard to justify at scale). So: bear ~$70-90B (modular intelligence stays in-house; ops stays a module), mid ~$125B, bull ~$185B+ (intelligence-as-a-service materializes at AV-like spend intensity). The narrow $60-75B needs no heroic assumptions at all.

Investor’s Perspective: Funded vs. unfunded: where capital is ahead of, and behind, the 2035 TAM

The 2026 funding data sharpens the picture rather than changing it: the top 3 deals took ~86% of robotics-software capital YTD, and top investors are visibly excited only where the outcome looks platform-scale (models, data, generalized autonomy). Classical robotics software infrastructure gets attention only when it’s tied to data, compute, or deployment scale. The barbell is intensifying, which means the middle is getting more open, not less.

PART 4: THE OPPORTUNITY FOR INDIA

India’s strength has historically been software. Given China’s compounding lead in hardware ecosystems (actuators, motors, sensors, batteries, with 50%+ supply-chain overlap between humanoids, automotive, and drones), India has to pick fights where its strengths bind: a 5M+ software-engineer base, the world’s deepest safety-critical QA and DevOps workforce, proven data-operations muscle, and a cost structure built for execution-heavy, always-on work.

The current landscape shows the gap and the opening. Of ~230 funded robotics companies with Indian presence (Tracxn), roughly 75% are full-stack robotics plays and 15% are robotics education; six application sectors (medical, cleaning/field service, logistics, drones, industrial automation, agriculture) account for 80% of the action. Only 8-10% play in the robotics developer economy, the layer this entire memo argues is where a decade of value creation concentrates. As a startup ecosystem, we’re ignoring the market where we have the strongest right to win.

Incorporating the founder and operator feedback, I’ve consolidated the India opportunity into five areas, ordered roughly by nearness of revenue.

1) Systems integration and solutions: the 10x giant, available now

The $30-50B to $120-200B services market is where robotics revenue actually lives today, and it’s structurally an India-shaped business: clients buy an operational outcome, every site has its own edge cases, and delivery rewards engineering depth at a cost structure others can’t match. The operator inputs were unanimous here: integration and solutions have “infinite bandwidth.” OEMs lack the people to convert robots into working factory and warehouse solutions, and demand outstrips supply everywhere. Addverb and GreyOrange are living proof that India can build global integrators; the open question is how many more the ecosystem can mint, and whether AI (which is blurring the product/services line) lets Indian integrators productize their delivery playbooks into higher-margin, repeatable platforms. VC-fundability is genuinely mixed, this is a services-heavy category, but it’s also the layer that generates the deployment data, customer relationships, and domain knowledge that every other India play below feeds on. Think of it as the ecosystem’s cash-flowing anchor tenant.

2) Cybersecurity, certification, and compliance: the most underrated opportunity

This came straight from operator feedback and barely existed in my first draft. Three sub-opportunities.

(i) Robotics-native cybersecurity ($6-10B by 2035): zero-trust, machine identity, and intrusion detection for connected heavy machinery. General OT security is well funded (Claroty, Nozomi), but the robotics-native slice has essentially one seed-stage specialist globally.

(ii) Safety certification and functional-safety engineering: ISO 3691-4, ISO 10218/13849, UL. Certifications take 1-2 years from scratch, they gate every deployment, and Roland Berger flags that existing safety standards were designed for fenced-off automation, not humanoids sharing space with people. New testing and certification regimes are coming, and someone has to run them. A hyperscaler quietly acquiring a software-based safety module (and taking it proprietary) is the tell that this layer has strategic value.

(iii) Export compliance across 20+ country regimes, a pain point operators pay heavily to navigate, with few people worldwide who understand the pathways. India’s right to win is direct lineage: the country already runs the world’s safety-critical QA and compliance for automotive, aerospace, and defence (ISO 26262, DO-178C, MISRA). Robots raise the stakes on exactly that discipline. Accredited certification services, compliance platforms, and safety tooling are viable, defensible, and, crucially, they monetize the industry’s conservatism instead of fighting it.

3) Simulation: own the unstructured frontier

Post the AV/manipulation split, the simulation opportunity for India clarifies into two distinct plays. The mobility/field play: scenario and terrain libraries for unstructured environments, agriculture, mining, construction, disaster response, chaotic traffic. The frontier of field-robotics sim isn’t clean factory floors; it’s exactly the messy conditions India lives with daily, and building those environment libraries is a data-and-engineering grind the West can’t cheaply replicate. It turns a domestic constraint into a globally valuable asset. The manipulation play: contact-rich simulation (deformables, friction, cloth) is the open technical frontier, but be honest about the bar. It needs applied-mechanics and computer-graphics talent, competes with an open-source gravity well (Genesis, MuJoCo) and NVIDIA’s subsidized stack, so the wedge is specialized content, benchmarks, and validation on top of open engines, not a new engine. Duality AI and AuraML are early Indian entrants; the synthetic-data-generation slice (feeding manipulation labs starved for demonstrations) is the nearest revenue.

4) Data infrastructure: the intelligence supply chain, built from India

Physical-interaction data is the scarcest input in robotics, and harvesting it is operationally intensive, talent-heavy, and cost-sensitive, the exact profile that built India’s data-services industry. Repointed from text and images to teleoperation, egocentric capture, and physical-data labeling for embodied AI, that muscle becomes a data-flywheel business serving every manipulation lab at once (Deepen and NeoCambrian are early proof). Two strategic caveats from the research sharpen the play. First, the competition isn’t other BPOs, it’s substitutes: skill-capture gloves, simulation-first synthetic data, and on-the-job deployment flywheels are all rival answers to “cost per useful demonstration,” so an Indian data play has to win on that metric and climb the value chain into curation, quality, and evaluation datasets rather than raw throughput. Second, the enterprise constraint shapes the architecture: operators won’t share data (”I cannot share this”), so the durable version of this business is on-prem/single-tenant data infrastructure and managed operations, not aggregation platforms. Aggregation plays die on the data-hoarding wall; infrastructure plays live inside it.

5) The end-to-end stack for robot software development: the load-bearing middle

The consolidated version of the map’s whitespace, and India’s most natural export. Four connected wedges:

• Testing, V&V, and CI/CD for robots (~$8-12B by 2035; near-total whitespace). The single emptiest square on the map. In software, CI/CD is commoditized; in robotics it barely exists, because every code or model change has to be re-validated against physics, sensor drift, and stochastic policies, so every serious company hand-rolls its release pipeline. This sits directly on India’s largest exportable skill bases: the DevOps/SRE workforce that runs the world’s continuous delivery, plus the safety-critical QA heritage from (b). Simulation-gated, hardware-in-the-loop release pipelines for robots is a category with no robotics-native funded leader to beat. H2LooP (SLMs writing safety-critical firmware with built-in test generation) and KALM are early Indian proof points.

• Middleware, observability, and vertical fleet orchestration (~$10-15B plus $2-4B). Orchestrating multi-vendor fleets for a specific industry (warehousing, manufacturing, agri) is a distributed-systems problem, not a hardware one, and India’s enterprise-SaaS and platform-engineering depth fits it. Botsync’s syncOS is the template. The honest bear case from the operator inputs applies here hardest (in-house builds are cheap; per-robot pricing is tough), so the winning shape is vertical (deep in one industry’s workflows) and outcome-priced, not horizontal seat-based SaaS.

• Deployment and fleet operations / RaaS software (~$20-35B, the biggest infrastructure market in the middle). Keeping thousands of machines alive 24/7 (fleet health, OTA, remote human intervention) is a control-room and always-on-operations problem that rewards India’s track record in cloud infrastructure and managed support. Rapyuta (with deep India engineering) and Cognicept show the opening. Where the “dev tool” framing fails the operator’s test, the managed-operations framing (robot NOC-as-a-service, teleop intervention centers) passes it: services revenue today, software margin as it productizes.

• Embedded and firmware plus AI coding agents (~$8-15B; vastly under-funded). Motor-control firmware, micro-ROS, embedded ML, and the AI-agent layer now emerging on top (CraftifAI, H2LooP, Ignitarium). India’s embedded-systems services heritage (automotive, semiconductor) gives it the raw talent; the agentic wave gives it the product wedge.

The common thread: India wins where the work is operational, safety-conscious, and execution-heavy rather than pure frontier research, and full-stack gravity makes this more true, not less. As the frontier labs vertically integrate, everything they have to buy around their core (data ops, evaluation, compliance, fleet operations, security, integration) grows in proportion. Take the heavyweight solutions layer plus this load-bearing operational middle, and India can build from India for the world, owning the highways of a multi-hundred-billion-dollar robotics economy without needing to win the (China-dominated) body or the (US-dominated, capital-saturated) brain.

So here’s the question I keep coming back to: who builds the unglamorous middle layer of the robotics stack, and do they build it from India?

If you’re building in any of these areas, or you know a founder who is, tag them or DM me. Would love to chat. 💬

With Thanks…

A map like this is only as good as the people willing to argue with it. This one got a lot sharper thanks to a handful of founders and operators who gave me their time, their candour, and in a few cases their sharpest disagreements. My thanks to:

• Arjun Jain, Founder of Fastcode.ai

• Pratik Sharda, Co-Founder and CEO of Craftifai

• Vivek Varadarajan and Pradyumna Vyshnav, Founders of Uncharted AI

• Naveen Arulselvan, CTO of Ati Robotics

• Pranjal Nadhani, Co-Founder and CTO of Cautio

• Paras Savnani, Founder of Robolens

• Gautam Pande, Chief Transformation Officer at GreyOrange

Any errors that survived their input are mine. If you’re building in this space and want to push back on where I landed, I’d welcome it.

Appendix: method and sources (summary)

Primary conversations: with 8 founders/ operators on the entire development stack, challenges and opportunities

Community research: five practitioner threads (500+ comments) spanning robotics development pain points, ROS ecosystem critiques, automation workflow challenges, and a $1M tooling-investment discussion; plus a structured community survey instrument covering 16 challenge dimensions.

Institutional and market research: Goldman Sachs humanoid research (manufacturing cost $30-150k, hardware-led declines); BofA/Morgan Stanley BOM models; Morgan Stanley Global Humanoid Model (brains/bodies/integrators; $4.7T hardware of $5T 2050 TAM; ~13M units by 2035); Roland Berger Humanoid Robots 2026 (software lag 3-5 yrs; AI architecture as primary bottleneck; training data constraint; safety-standards gap); McKinsey AV research (L4 development $0.6-1.5B; robotaxi 5x; $50-67B cumulative burn; auto software to $519B by 2035); IMARC and Allied robot-software market forecasts; STIQ Humanoids; integrator cost benchmarks (programming/engineering 15-40% of deployed cell cost); humanoid TCO analyses ($3-8k/yr software subscriptions).

Frontier-lab analysis: Physical Intelligence (π0/π0.5), Skild AI ($1.4B, 2026), Generalist AI ($400M @ $2B), Rhoda AI ($450M), Dyna Robotics ($120M Series A; DYNA-1), Genesis AI ($105M seed; open-source physics engine), Sunday Robotics ($200M; Skill Capture Glove; ALOHA/UMI lineage), The Bot Company ($300M+ @ $4B+), 1X (world models); 2026 funding-concentration data (top 3 deals ≈ 86% of robotics-software capital YTD).

Figures are directional estimates; ranges de-duplicated across layers. All third-party numbers as reported publicly at time of writing.

Sit with that for a second. The binding constraint on the most capital-rich industry in history is not chips or talent. It is electrons and permits. This week, we’re diving into the most audacious answer anyone has proposed: moving the data centre itself into orbit.

Two sources anchor this edition: Convequity’s deep dive on SpaceX and the orbital AI infrastructure play, and Space Insider’s market map of 66 companies building the orbital data centre ecosystem. Both are linked at the end, and both are worth your time. [h/t: tag Convequity + Space Insider authors]

THE WALL TERRESTRIAL DATA CENTRES ARE HITTING

The demand side is well documented. McKinsey expects global data centre capacity to more than double to ~220 GW by 2030, requiring $6.7 trillion of investment. Hyperscaler IT load is projected to grow roughly 6x by 2035, which means over 100 GW of new capacity has to come from somewhere.

The supply side is where it breaks. Grid interconnection queues stretch for years. Permitting, water, and land fights kill projects before ground is broken. Gas turbines are being deployed at speed, but turbine-blade shortages have created multi-year backlogs. Battery storage helps: with US peak capacity around 1,000 GW running at roughly 50% average utilisation, storing midday solar surplus can nearly double usable dispatchable power. Useful, but nowhere near enough.

Convequity stacks the bottlenecks in order of severity: TSMC logic and packaging (mild), memory (large and growing), lasers for optical interconnects (large), and above all power, which is what killed those 40 projects. Even terrestrial solar, the obvious fix, runs into physics and politics. The atmosphere absorbs a huge share of solar radiation, night imposes an intermittency penalty, the best commercial silicon cells top out around 21-25% efficiency, and tariffs block cheap panel imports anyway.

So where does that leave us? With a relief valve 500 km up.

WHAT AN ORBITAL DATA CENTRE ACTUALLY IS

This is not the satellite industry as we know it, scaled up. SpaceX has filed with the FCC to launch over 1.2 million satellites, more than humanity has ever put in orbit, combined. They form a three-tier constellation that behaves like a single distributed compute fabric:

1️⃣ VLEO (500 km): 800,000 satellites. The edge layer. ~1,000 TOPS each, ~20 ms latency to ground. Real-time inference.

2️⃣ LEO (1,000 km): 300,000 satellites. The regional layer. ~5,000 TOPS each. The bulk of training.

3️⃣ MEO (2,000 km): 118,000 satellites. The cloud layer. ~20,000 TOPS each. Long-running workloads and storage.

Each satellite weighs ~2 tons, and nearly half of that, ~850 kg, is compute payload: radiation-hardened silicon, liquid cooling loops, high-efficiency solar conversion. These are not comms satellites with compute bolted on. They are flying data centres, with the bus designed around the silicon, running 12-24 space-rated chips stitched together by laser links at 100+ Gbps each.

The launch math is what makes it real. One Starship carries 50-75 of these satellites. At that cadence, the full constellation deploys in roughly nine years: 10% by year three, 50% by year six.

WHY THE PHYSICS FAVOURS ORBIT

Four properties of space do the heavy lifting.

The 10x solar advantage. Move a panel to orbit and generation alone improves ~5x: no night, no atmospheric absorption, and multi-junction cells that capture a broader spectrum. Then remove the storage layer entirely, because orbital sunlight is continuous, and you save roughly another 2x at the system level. Compound the two and space solar is ~10x more effective per watt delivered. If Chinese panels cost ~$0.25/W on Earth, the effective orbital figure lands near $0.025/W, excluding launch.

Free cooling, with a twist. Space is cold but there is no air, so chips cannot cool by convection. Heat leaves only by radiating into the ~3 K background of the universe. That sounds like a problem until you look at what Earth charges for cooling: a 1 GW terrestrial cluster draws ~627 MW for silicon and another ~373 MW, nearly 40%, for cooling, conversion, and networking. The orbital fix is elegant. Each satellite is a flat panel: the sun-facing side generates power, the shaded side radiates waste heat. One geometry solves both problems.

Fiber optics without the fiber. Vacuum has zero chromatic dispersion, so laser links can use simple, cheap, low-power direct detection where terrestrial fiber needs expensive coherent transceivers. Light also travels ~47% faster through vacuum than through glass, and optical spectrum is unlicensed. The interconnect is not just relocated. It is physically better.

No grid queue. No permits, no land or water fights, no interconnection wait. The single biggest killer of terrestrial projects simply does not exist in orbit.

Now the economics. With a fully reusable Starship V3, lifting 100 GW takes ~10,000 launches at ~$5M each: $50B, or $500M per GW. Solar adds ~$200M/GW. The non-chip cost of a first-generation space data centre works out to ~$3.3B/GW, against ~$20B/GW of non-IT capex for a terrestrial facility, plus $1B+ in annual opex that orbit never pays.

The chips are the catch. On merchant NVIDIA GPUs, which carry 75-80% gross margins, a space data centre costs ~$33.3B/GW all-in, with IT hardware eating 80% of the total. Convequity maps the descent from there: space-optimised ASICs cut it to ~$16.5B/GW, and vertical integration into your own fab, attacking HBM memory that alone costs $15,840 per high-end GPU, brings it to ~$11.6B/GW. Roughly half the cost of an equivalent terrestrial build. And each orbital GW supports 900 MW+ of usable compute versus ~660 MW on Earth, so revenue per GW nearly doubles too, at ~$20B/yr versus $10-15B.

THE MARKET MAP: 66 COMPANIES, 3 TIERS

Space Insider groups the landscape by position in the value chain:

🔸 Core Builders & Operators: companies whose spacecraft exist to provide compute in orbit, including SpaceX, Axiom Space, and Starcloud

🔸 Essential Enablers: the hardware and software layer, from space-qualified processors to AI accelerators, including NVIDIA, AMD, and HPE

🔸 Strategic Contributors: terrestrial cloud giants forming payload partnerships, including AWS, Google, and IBM

But here’s the part founders should internalise: the incumbents’ head start may not hold. Today’s flagship space laser terminal delivers ~100 Gbps in a 30 kg package costing ~$500,000. The orbital data centre needs 6.4-12.8 terabits in a sub-5 kg terminal at well under $5,000. That is not an upgrade path. That is a different product, and Convequity argues the winning DNA flows from terrestrial AI photonics up to space, not from legacy aerospace down.

The genuinely new white space:

1. Pointing, acquisition, and tracking (PAT) systems, which have no terrestrial equivalent;

2. Free-space optics assemblies replacing simple fiber ferrules;

3. Radiation-tolerant silicon photonics;

4. Inter-orbit relay terminals, a 100% incremental market Convequity sizes at $100M to $1B cumulative;

5. Thermal management for orbit; and in-orbit servicing and assembly.

WHY INDIA COULD WIN BIG

India put an orbiter around Mars for $74M, less than Hollywood spent making Gravity. That frugal engineering heritage, built over decades at ISRO, is precisely the muscle this market rewards. And notice what the white space above has in common: components, photonics, thermal systems, robotics, software. You do not need to launch a constellation to matter in this value chain. You need to build one hard, space-qualified thing exceptionally well and at exceptional cost. That is building from India for the world.

At Antler India, we’ve backed Inspecity, building space robotics: satellite servicing today, orbital data centre assembly next. We think there is room for many more.

To be clear-eyed about it: the whole thesis leans on Starship cadence and solar manufacturing scaling on schedule, and both are genuinely hard. But even if the timeline slips, the bottleneck logic does not change. Power and permits are not getting easier on Earth.

So here’s the question: is orbit the answer to AI’s power crisis, or a moonshot too far?

If you’re building anywhere in this value chain, reply to this post or DM me on LinkedIn - would love to chat! 💬🚀

Sources: Convequity, “SpaceX and the Orbital AI Infrastructure Play” (June 2026); Space Insider, “Market Map: Orbital Data Center”; Morgan Stanley terrestrial cost model via Sapphire VC (May 2026); McKinsey data centre research.

I have been reflecting on this last week, because it reframes every conversation we are having in India about sovereign AI.

The bill is not what you think

The most useful part of the Sapphire breakdown is not the headline. It is the composition.

Compute is only 57% of the cost, about USD 2.85 billion. The main pretraining run is USD 1.25 billion, and here is the detail most people miss: failed runs and ablations add another USD 900 million. Nearly a billion dollars of the budget is experiments that do not ship. That is not waste. That is the price of learning at the frontier.

The other 43%, about USD 2.15 billion, is people, data, safety, and operations. Researcher and engineer compensation alone is USD 900 million. Data licensing and curation is USD 600 million. Safety and red-teaming, USD 350 million.

Read that again. Talent costs as much as the failed experiments. Frontier AI is not a GPU problem. It is a GPU problem, a talent problem, a data problem, and an organizational problem, all at billion-dollar scale, all at once.

Now the India math

Sarvam is India’s most prominent sovereign AI bet. The team has moved fast: two foundational models trained from scratch, deployments reaching 17 million farmers, unicorn status as of June.

Cumulatively, Sarvam has raised around USD 275 million. That includes the USD 234 million first close of its USD 300 million Series B last month, led by HCLTech.

Set that against the Sapphire numbers and the comparison is stark. Sarvam’s entire capital raised, across every round since 2023, is about 5% of one frontier training program. It is less than what a frontier lab spends on researcher salaries for a single model.

I am not pointing this out to diminish what Sarvam has built. Quite the opposite. It makes what they have shipped on that budget genuinely impressive. But it forces an honest question about what “sovereign AI” can mean at this capital gap.

Four questions I keep coming back to

1. Do we have the risk appetite?

Not the capital. India has capital. The question is whether anyone, government or private, is willing to put USD 3-5 billion behind a single program where USD 900 million is expected to burn on failed runs. Frontier labs underwrite failure at that scale as a line item. We never have.

2. Do we have the talent pull?

China answered this gap with the Thousand Talents Plan, bringing back over 7,000 overseas researchers in under a decade. India’s AI diaspora fills the research benches of every major frontier lab, yet we have no systematic mechanism to bring any of it home. What is our version?

3. Or do we refuse to play that game?

The ISRO question. Mangalyaan reached Mars for about USD 74 million against roughly USD 672 million for NASA’s comparable MAVEN mission. Frugal engineering is India’s proven playbook, and DeepSeek showed training costs are not laws of physics. But one caution: cost-efficiency has proven it can catch up to the frontier. It has not yet proven it can define one.

4. Is the answer smaller models, domain by domain?

Maybe sovereignty means owning the models that matter for India: Indic languages, agriculture, public services, financial inclusion. Sarvam’s deployments already point this way. The trade-off: domain models win markets, frontier models win leverage.

What do you think?

I have deliberately left these as questions rather than answers, because I suspect the honest answer combines all four, and because the people best placed to weigh in are the ones building and researching in this space.

So tell me: which path do you think India should take? Is there a cheaper seat at the frontier table that nobody has spotted yet? Or is the whole framing wrong?

The comments are open, and so are my DMs. If you are building in this space, I would love to hear your take.

These aren’t edge cases. They’re what happens when two curves collide: prices going up and quantities exploding.

On the price side, major providers are quietly moving away from generous flat-rate subscriptions toward usage-based billing. GitHub Copilot’s June 2026 pricing changes raised model costs for legacy Pro users by as much as 9x to 18x. Anthropic’s Opus 4.7 shipped with a tokenizer that kept per-token rates identical but generates up to 35% more tokens from the same input. Same sticker price, bigger bill.

On the quantity side, agents changed everything. Token consumption is exploding: OpenRouter reported a 13.5x YoY jump, with roughly 102 trillion tokens consumed in March 2026 alone. Multi-agent systems can burn over 3500x tokens for agentic coding tasks and over 1500x more for a single round of reasoning, according to new research. And while inference costs fall roughly 10x annually for a given level of intelligence, frontier-level pricing has barely moved. The 10x deflation story is true only if you’re happy using last year’s brain.

Here’s the number that should get every CFO’s attention: MMC Ventures estimates 50-80% of enterprise token spend is unnecessary.

This week, we’re dissecting MMC’s excellent market map on the Token Efficiency stack (h/t Advika Jalan, Simon Menashy and Prakriti Roy - full report linked at the end). What follows is my synthesis of their framework, plus where I think the opportunity sits for builders.

Where the tokens actually go

Before the stack, one insight worth internalising: input tokens, not output tokens, drive most of the spend.

This is counterintuitive. Output tokens cost 3-6x more per token because each one requires a full forward pass. But agents are voracious readers. Every loop, every tool call, every retry drags the accumulated context back through the model. MMC calls one version of this the “context avalanche”: after 20 rounds of edits, the model isn’t just reading your latest request, it’s rereading the original prompt, every revision, every abandoned idea. Five separate “make it shorter” requests can cost far more than batching all five changes into one.

Multi-agent setups make it worse. A Writer agent produces a 2,000-token draft, a Reviewer sends back 500 tokens of feedback, the Writer regenerates the whole thing. Without a strict stop condition, they’ll do this 5-10 times over minor issues, and because every round carries the full conversation history, one small disagreement over formatting can burn 50,000+ tokens. We used to think petty human arguments escalated quickly.

Source: Sapphire Ventures Market Memo

The five layers of the token efficiency stack

1. Context and memory management

The fix for the avalanche is feeding agents only the precise information they need: not too little, not too much. Optimisation opportunities exist across the entire stack. Search APIs that return clean, deduplicated context instead of raw webpages. Reusable context infrastructure that turns enterprise knowledge into persistent “company brains.” Tool gateways that trim bloated tool menus (StackOne cuts token use by up to 90% versus naive approaches). Data retrieval layers like Coral that reduce token use by up to 64% on complex tasks. Caching and compression tools like Headroom that collapse context before it hits the model.

There’s a second-order benefit here that I find underrated: an independent context layer makes it far easier to switch models. That matters more than ever (see layer 2). Players: Parallel, Mem0, Headroom, PageIndex, Coral, StackOne. We covered this piece in depth over the last few weeks.

2. Open source and small models

Over 52% of production deployments already run on lightweight or open models, per Weave’s router data. The performance gap is closing fast: z.AI’s GLM 5.2 scores 51 on the Artificial Analysis Intelligence Index versus 55 for GPT-5.5 and 60 for Claude Fable 5. And ensembles change the math entirely. OpenRouter found a budget panel of Gemini 3 Flash, Kimi K2.6 and DeepSeek V4 Pro beat GPT-5.5, came within 1% of Fable 5’s score, and cost half as much.

There’s also a sovereignty angle that got very real this year. The “Fable Fracas” - where US government restrictions forced Anthropic to abruptly cut off access to its frontier models globally - showed that frontier intelligence delivered via API can vanish overnight. Once you have the weights, your capability is yours. Players: Qwen, Kimi, DeepSeek, MiniMax, Mistral AI.

3. Routers and gateways

Most teams default to sending every task to the most expensive frontier model. MMC’s analogy is perfect: it’s lighting a candle with a flamethrower. Weave found 40-60% of coding agent requests are short, structurally simple completions that an open model like Qwen handles at parity with frontier, at roughly 1/40th of the cost.

Routers select the best model for each request, balancing cost, quality, latency and failover. Gateways add the infrastructure around routing: unified endpoints, budget controls, observability, governance. The hard part isn’t picking the cheapest model, it’s knowing when a cheaper model is good enough, which is why live quality scoring and task-specific evals are the real moat here. Players: OpenRouter, Weave, Bloom, Martian.

4. Inference optimisation

Agentic workloads are long-running, asynchronous, context-heavy and dominated by tool use. That demands a different inference stack than interactive chat. Serverless inference wins when traffic is spiky or early-stage because it scales to zero. Dedicated endpoints make sense when traffic is predictable enough to justify reserved GPUs. Self-hosting becomes compelling at high query volumes or for specialised workloads. Players: Together.ai, Baseten, Pipeshift, Doubleword. Our portfolio company Navana runs Voice AI inference on CPUs instead of GPUs for Indian clients: a structural cost advantage in a price-sensitive market.

5. AI FinOps

The layers above are technology. This one is governance and behaviour. The same agent on the same task can vary 30x in cost, because agents can’t predict how much context they’ll accumulate or how many steps they’ll take. Frontier models systematically underestimate their own token usage.

Return on Tokens (ROT, an unfortunate acronym) is the new metric, and the “fully loaded” cost is bigger than most teams realise. Research on agentic software engineering found roughly 60% of token consumption sits in iterative code review and verification, not initial generation. Faros AI found bugs per developer up 54% and code churn up 861% in high-AI-adoption environments, pushing median review time up 5x. Tokens are cheap compared to the engineering hours spent reworking bad output.

On the control side, blunt per-user caps can backfire: one founder noted 5% of employees consume roughly 40% of tokens, and that distribution shifts as adoption grows. Pooled budgets balance experimentation with discipline. A new wave of tooling is emerging here, from existing FinOps and observability players to standalone solutions. Some measure agent costs, margins and ROT. Others enforce real-time spend limits so agents can’t run away with the budget. Players: Paid, Monetise, Faros.

The structural tension

Here’s what makes this space investable: model providers are incentivised to Tokenmaxx. Their revenue scales with consumption. Every optimisation on this list works against their P&L, which means the efficiency layer will be built by third parties, not the labs.

One caveat before you over-rotate on cost cutting: don’t optimise for token cost alone. Aggressive compression loses nuance. Sometimes cheaper intelligence with more context beats an expensive model with less. And not everything needs an agent in the first place: if SQL, rules or templates can handle a task deterministically, reserve the expensive reasoning for work where judgement actually creates value.

The organisations winning here aren’t the ones spending the least on tokens. They’re the ones converting tokens into business outcomes most efficiently.

If you’re building in this space, whether it’s context infrastructure, routing, inference or AI FinOps, I’d love to chat. Reply here or DM me on LinkedIn.

Full MMC report: AI Tokenomics: How to tokenmin while ROImaxxing

Building software keeps getting cheaper. Which means software moats keep getting weaker.

For the last era, software defensibility came down to two things: switching costs and network effects. Network effects were only ever available to a lucky few. And switching costs, the moat most founders actually relied on, are now cracking. New entrants use LLMs to build integrations and migrations fast, so the cost of ripping out an incumbent and moving to a challenger keeps dropping. If a competitor can rebuild your integrations in a weekend, how much is your stickiness really worth?

So the question I keep coming back to is this: in the AI era, how do founders build switching costs that software alone can’t undo?

One answer is to stop competing on software alone and tightly couple it with hardware. Hardware Enabled Software. It’s why I was glad to read a sharp market map of this space by at Triple Point VC, which prompted a lot of this thinking.

What Hardware Enabled Software actually means

Strip away the label and a Hardware Enabled Software company tends to share three traits.

First, it’s vertical software that anchors its intelligence in first-party data gathered by hardware the company deploys itself. Our portfolio company Cautio is a clean example: it puts dashcams in commercial vehicles, and that proprietary data stream powers the fleet safety and management software that operators run their business on. The software is only as good as the data, and the data comes from hardware nobody else owns.

Second, the hardware is usually subsidised to pull software adoption and automate the work downstream. The obvious benefit is lower upfront friction at deployment. The underrated one is time to value: faster installation means the customer sees something useful in days, not quarters.

Third, and this is the part that compounds, the hardware captures data at its physical point of origin. You end up owning scarce context from real-world nodes that nobody can scrape or replicate from a data warehouse. The install drives retention. The data drives the software. Together they form a moat that’s genuinely hard to dislodge, because pulling it out means physically removing devices a CFO has already depreciated.

The below diagram from nicely captures the stack well.

Why now

Here’s the fair objection: IoT has been called the next big thing for over a decade, and it kept underdelivering. The physical go-to-market reality, slow sales cycles, on-site installs, relationship-driven distribution, always reasserted itself. So what’s actually different this time?

Three forces have moved at once.

Hardware costs keep collapsing. LIDAR is down more than 90% over the last decade, pulled along by autonomous driving. Wright’s Law captures the pattern: for every cumulative doubling of units produced, cost falls by a consistent percentage. The robotics and physical AI boom will keep that learning rate running, dragging sensor costs lower across the board. Cheap sensors make persistent coverage of large physical environments economically sane for the first time.

The sensing menu is expanding fast. We’ve gone well past vision. Voice, touch, olfactory, neural interfaces, hyperspectral imaging, photonics, the range of phenomena you can economically sense and act on keeps widening. The design space is enormous and, right now, mostly limited by founder creativity rather than by physics.

Processing that data has collapsed in cost, thanks to AI. The old IoT architecture, sensor to dashboard to human to action, always hit the ceiling of human bandwidth. That ceiling is gone. To make the scale concrete: reading every one of the roughly 100 million CCTV feeds in the US could cost under $300M by 2030, cheaper than renovating the White House. And there’s a deeper point underneath it. As models exhaust the supply of human data for supervised learning, sensor data becomes the scarce raw material that grounds world models and the next generation of agents. Whoever owns the sensor layer owns the training signal.

What we’re seeing across our portfolio

We believe in this thesis enough to back it repeatedly. Of our 50+ AI investments, more than one in five combine software and hardware. A few on the enterprise and prosumer side:

• Cautio builds AI dashcams and software for commercial fleet operators in India.

• Uncharted AI builds autonomous systems that cut field exploration time in mining.

• Kubo Care runs radar-based fall detection for senior living facilities.

• Storefox is a point-of-sale intelligence platform for retail outlets, using audio sensors.

• Aignosis makes AI diagnostic equipment for autism detection in children.

We’ve also backed Craftifai, which builds the platform and harness that accelerates Hardware Enabled Software development itself, a picks-and-shovels bet on the whole category growing.

The consumer side of the thesis

The thesis isn’t only an enterprise story. Some of the most interesting first-party data sits in the home, on the wrist, and at the bedside, and we’ve backed close to five consumer companies built on exactly this pattern: a physical device that captures data nobody else can, feeding software that gets better the longer you use it.

• Cravv is a cooking companion that turns the kitchen into a data source, pairing hardware with software that learns how you actually cook rather than handing you another static recipe app.

• Dusq is a sleep regulation system. The device captures sleep data directly off the body, catches every disruption and responds in seconds, pulling you back into deep sleep before you surface.

• Naptick takes the bedside route to the same problem, a sleeping companion that sits beside you rather than on you, lowering the friction of nightly use while still owning a proprietary stream of sleep data.

• Hify brings sports analytics to everyday consumers, the kind of performance data that used to live only inside pro setups, captured through hardware and made legible to amateurs.

• MyWonder is a screen-free toy for kids, a physical product designed to own the engagement loop and the interaction data, without putting another screen in front of a child.

Different categories, same shape. The hardware is the wedge that gets you into the home and creates the data nobody can copy. The software is what keeps you there.

Where this goes

I’m not claiming hardware is a free pass. The go-to-market is genuinely harder than pure software, installation friction is real, and a bad device sinks the whole stack. But that difficulty is part of the point. The same friction that slows you down also slows down whoever wants to displace you, and you cannot vibe-code the de-installation of a sensor network.

If software is getting commoditised, hardware may be the moat a lot of founders are sleeping on. The surface area for what’s possible here is massive, and we’re early. If you’re building something at this intersection, or you think the thesis is wrong, I’d genuinely like to hear from you.

If you have been anywhere near AI circles over the last few months, you have seen the framing. Jaya Gupta and Ashu Garg of Foundation Capital put a trillion-dollar number on it in their viral piece on the Context Graph. The last generation of enterprise software became trillion-dollar companies by owning what happened: Salesforce for customers, Workday for employees, SAP for operations. The next one, they argue, will own why it happened, the decision traces and judgment that never made it into any system. That “why” is context.

So how do you actually build it? The sharpest map I have seen comes from Miguel Arias and Elisenda Bou, and they make a point I keep coming back to: every founder walks in asking the same question, “how do we make our context usable for agents?” The phrasing is always identical. What sits underneath it never is.

It is not one problem. It is three.

What sorts this space is not company size, it is the kind of context you are dealing with, because that decides which tools are even on the table. Size is a rough shortcut. A solo builder is usually in bucket 1, a startup putting agents on its own data in bucket 2, a large enterprise in bucket 3. But it is the context that sorts you, not your headcount. Glean sells to big enterprises and still works almost entirely on institutional knowledge.

The map breaks into three buckets, each building on the last.

Bucket 1: the agent’s own context. When the consumer is an agent, there is no company stack behind it, but there is still plenty to work with: the operating manual (instruction files like AGENTS.md and CLAUDE.md) plus the accumulated memory, the trace of what was said, decided, and corrected over time. The players here sit on a spectrum sorted by how much they model. At the low end, plain RAG over a pile of files, where Mem0 sits. In the middle, retrieval plus light structure. At the high end, a full knowledge graph or ontology that lets the system infer facts nobody ever wrote down, where Cognee and Zep sit (Zep adds time, so a fact can be “true from, true until”). This is where last week’s memory piece lived.

Bucket 2: institutional knowledge. This is the bucket most people actually mean. The context is not in a clean database, it is scattered across the apps where work happens: Slack, Notion, Drive, Granola, email. The job is to make that mess usable to an agent, and it has turned into a land-grab to be the unifier, the layer everything else connects underneath.

Bucket 3: systems of record. The hard structured stuff: warehouses, CRMs, ledgers, plus the unstructured sprawl, across systems that were never built to talk to each other. This is the hardest bucket, the only one where every problem shows up at once, and the first where governance stops being optional: access control, lineage, audit. As Arias and Bou put it, governance is less a stage than a tax that kicks in with scale.

Surface owners versus neutral players

Cutting across buckets 2 and 3 is a second split worth holding onto.

Surface owners already own a place where context gets created and add connectors to pull in the rest: Slack, Granola and Notion on the unstructured side, Snowflake, Databricks and Datadog on the structured side. Their bet is distribution and native data, because the context already lives in them.

Neutral players own no surface of their own. They index across everything and bet on breadth, permission-awareness, and not competing with the tools you already use. Glean and Atlan are the obvious examples. They differentiate on how much they model, from plain retrieval, to a semantic layer, to a full graph of your company. The short version: surface owners win on data gravity, neutral players win on coverage and trust.

Find your level first

Here is the takeaway I would underline for any founder building in this space. The most common mistake is getting your level wrong in either direction. Overshoot into a bucket-3 ontology when files and retrieval would have carried you, and you are maintaining a knowledge graph to answer questions a vector search could handle. Undershoot, and you starve the agent of the context it needs to do the job. Either way, the application layer on top pays for it. Find your level, then build for it.

For founders, the real question becomes which part of the stack you play in, and what it takes to win the specific market you are in. Two examples from our portfolio at Antler India. Kenko AI is building the end-to-end Vertical AI stack, from system of record to agents, for fitness and wellness businesses in the US. nRev sits in bucket 1 in revenue intelligence, while integrating across buckets 2 and 3.

Where India can win

I see two paths for Indian founders.

The first is going vertical. Pick a market and reimagine the full stack in an AI-native way, from systems of record all the way to agents. We already have more than ten Vertical AI companies in the portfolio building deep, defensible context inside their markets, and this is the path I am most excited about.

The second is going horizontal, owning a specific bucket. Mem0 and Supermemory are doing this in bucket 1. Atlan is the standout India-origin story in bucket 3, a company that began by building India’s national data platform and now positions itself as the context and governance layer for enterprise AI. Bucket 2, the institutional-knowledge unifier, is the biggest open gap I see from India. Whoever cracks it from here has a real shot.

And the deepest moat sits in the context none of these tools fully reach yet: the workarounds, the real escalation paths, the judgment living in people’s heads. It was never written down, which is exactly why it is so hard to capture and so valuable when you do. That, more than any vector store, is where the next durable company gets built.

If you are building a deep vertical play, or a horizontal layer in any of the three buckets, I would love to compare notes.

This week I’m diving into memory, because it’s the least settled of the three and probably the most strategically exposed. A sales agent that forgets last week’s objection isn’t a sales agent. A finance agent that forgets how the company defines ARR will re-litigate the same metric forever. Action got solved first. What the work compounds into did not.

Big credit to for a sharp deep dive that frames this well - check it out here. I’m borrowing his stack and adding an India lens.

The memory stack, in five layers

What I like about Rishi’s framing is that he splits memory into five layers, each of which fails in its own way:

1. Session memory: the live thread the agent is working in right now.

2. Working memory: the task plan, tool outputs, and approvals in flight. Closer to an OS process than a chat log.

3. Semantic memory: the stable facts about a user or account, the stuff a good account manager keeps in their head.

4. Episodic memory: what happened before and how. Past tasks, corrections, and lessons learned the hard way.

5. Governed organizational memory: shared context with permissions, provenance, and an audit trail. The version a security team will actually let you ship.

Most products today are stuck at Layer 1. There’s a lot of climbing left to do.

“Just scale the context window” doesn’t hold

The tempting counterargument is that all of this disappears as context windows grow. Put more in the prompt, let the model read everything. I don’t buy it, on two counts.

First, the economics. Stuffing fifty thousand tokens into every answer is a benchmark stunt, not a runtime. Second, perishability. Context goes stale. What’s true today is often “true as of last quarter,” and a plain store happily preserves the contradictions until someone gets burned.

In the enterprise it gets harder still. You’re not managing one memory, you’re governing an entire context graph: who gets to read, who gets to write, and what actually gets written. That’s a permissioning problem, not a storage problem.

The market map today

Six categories are going after the memory stack, with heavy overlap:

• Memory-native startups: Mem0, Zep/Graphiti, Letta.

• Shared enterprise context: Glean, Dust, Hebbia.

• App context capture: Granola, Otter, Fireflies.

• Agent frameworks and runtimes: LangChain, CrewAI.

• Retrieval and data substrate: Chroma, Pinecone, Weaviate.

• Incumbent platforms: OpenAI, Anthropic, Google, Meta, AWS.

Everyone is attacking from a different direction, which tells you how unresolved this still is. The question worth asking about any company here is the same one: which control point do you own in the write path of what becomes memory?

Wrapping a vector database and calling it memory gets you a seed round. It does not get you defensibility. The real moat sits in one of four places: owning a control point in the write path, spanning multiple layers of the stack, enterprise-grade permissioning and scope management, or owning distribution where the agents and users already are. Ideally you own two of those. The thin wrapper in the middle, with no workflow lock and no governance story, is where the squeeze hits first.

The opportunity in India

Mem0 is the clear breakout from India here, with a USD 25M Series A and a deliberately broad pitch across the memory stack. Supermemory is the other name to watch, backed by a strong set of angels and going after personal knowledge as its wedge.

But two companies don’t make a category. Governed memory for enterprise agents is exactly the kind of hard infrastructure problem we should be good at: deep developer talent, a real appetite for building from India for the world, and a layer that’s still wide open because nobody has earned the trust yet. This is structural advantage hiding in plain sight.

So here’s the question I keep coming back to: who’s building the governed memory layer for enterprise agents, from India, for the world?

If that’s you, I’d love to chat.

As a VC, one of the biggest daily jobs is handling inbound email, and AI has only multiplied the volume. The inbox matters to us in two ways:

👉 You’re hunting for the next outlier. You can’t afford to miss an email

👉 Responsiveness shapes the founder experience. Slow replies cost deals

So this week I went looking at how people are using AI agents to run their inboxes. Came across this simple framing by Menlo Ventures (Source article) (h/t Shawn Carolan & Ryan Hand) mapping the space on 2 dimensions:

1️⃣ Integration: direct-to-server, browser extension, AI-native client, or email superset (a work OS where email is just one layer)

2️⃣ Autonomy: on-command, co-triage (human in the loop), or autopilot

What I found most useful sits after the map: the design choices every winning product has to make on behalf of its ideal customer. No single approach wins everyone, because email behavior varies too much, but a few decisions stood out:

🔸 Trust and security are table stakes. Messages are some of the most sensitive data a company holds, so one breach, or even the perception of careless handling, ends the sales conversation before it starts.

🔸 Speed and simplicity decide adoption. These tools live inside a workflow people touch all day, so the assistant can’t feel like another layer of work. As Superhuman showed, the 100ms rule still decides whether it feels effortless or like extra effort.

🔸 How opinionated should it be out of the box? There’s a real tension between shipping with strong defaults and letting users configure everything. The most ambitious bet is mining a user’s past behavior to personalize deeply without asking much upfront. But trust is earned narrow before it’s earned broad: people won’t hand over wide automation until they’ve watched the product make good calls on small, repeated tasks first.

🔸 How much behavior change does it demand? This is the quiet wedge question. Server-level integrations ask for zero change; a full alternative client asks you to abandon Gmail or Outlook entirely. More surface area unlocks more capability, but habits are sticky and switching costs are real. The winners earn deeper adoption over time instead of demanding it on day one.

🔸 Should the streams be consolidated? If email, WhatsApp, iMessage, Slack and LinkedIn are all just competing for the same attention, why not triage them with one algorithm? Plenty have tried, none have really won, mostly because the UX never felt right for any single medium. Agents that sit on the desktop and intercept the hard-to-integrate platforms might finally crack it.

The line that ties all of this together: “the highest-value action is the one you never had to take. But that only holds if outputs are reliable, since one wrongly sent reply kills trust in a category built entirely on delegation”. In a product built on delegation, reliability is the entire moat.

A few things not on the map but worth chewing on

🔸 VOICE AS A NATIVE INTERFACE: what does email rebuilt around voice look like? Could be a real wedge for India and Southeast Asia.

🔸 B2B vs B2C SPLIT: email isn’t dying with GenZ, far from it. A ZeroBounce survey found 42% of Gen Z name email their favorite work channel, 2x those who picked Slack or WhatsApp, and 60% use it to avoid awkward conversations. Consumer and business inboxes could evolve down very different paths in the world of AI

🔸 AGENT-TO-AGENT EMAIL: what happens when mail is written for agents, not humans? Still email, or a completely new paradigm?

THE OPPORTUNITY FOR FOUNDERS

Email is still the default way to reach someone you don’t know yet, and that isn’t changing. But the client, the autonomy layer, the voice interface: all of it is open to be rebuilt.

If you know a founder or team building here, tag or DM. Would love to chat! 💬

So here is the uncomfortable truth the chip headlines miss: the bottleneck isn’t silicon. It’s electrons.

A data centre can be built in 12 to 18 months. Connecting it to the grid takes five to seven years. That single mismatch is fuelling one of the largest infrastructure cycles of our lifetime. In 2025, data centres pulled 78% of all built-environment venture funding, $4.5B of $5.7B. When a US president invokes the Defense Production Act to classify grid hardware as essential to national defence, as Trump did in April 2026, you know the buildout has stopped being a utilities story and become a national one.

This week I’m dissecting the AI data centre stack, layer by layer, off the back of a sharp market map from Bessemer Venture Partners (h/t Lindsey Li, Brielee Lu, David Cowan and Joshua Hechtman). Then I’m tying it to the opportunity I care about most: the one building in India.

The thesis is simple. The concrete and the GPUs get the attention. The durable value sits in the six enabling layers that make the buildout possible.

THE SIX LAYERS

1. Permitting and site selection

Before a single GPU is racked, a data centre needs permission to exist: zoning, environmental clearances, interconnection agreements, air and water permits, even endangered-species assessments. Each runs on its own clock and its own jurisdiction. Between March 2024 and 2025, 16 developments were delayed or denied on permitting grounds, with local community pushback the leading cause.

McKinsey pegs annual infrastructure permitting spend above $5B, more than a third of the US total, and reckons over $1.5T of infrastructure capital is stuck in the pipeline. Today that work is handled by a handful of large consultancies and more than 50,000 small firms, most with fewer than 15 people, running largely manual processes.

The wedge is AI-native software that compresses this. Players like Paces unify grid, permitting and environmental data into one view and report customers closing roughly three times as many deals, because grid constraints and zoning conflicts surface upfront rather than after capital is committed. What wins here: teams that pull siloed data under one pane of glass and build a flywheel that gets more predictive with every project.

2. Power generation

Power is shifting from the grid to behind the meter. Grid-connected sites are still the largest by project count, but on-site and hybrid generation now account for close to half of all announced capacity. This is the “Bring Your Own Power” movement: generate electricity on-site to sidestep the interconnection queue entirely. Around 50 GW of behind-the-meter gas was announced in 2025 alone. It is now the default way to bring new AI capacity online.

The interesting part is what’s powering it. Boom Supersonic took its jet-engine core and turned it into a 42 MW gas turbine purpose-built for data centres. Others are co-locating batteries behind the meter to charge cheap and discharge at peak. And on the long horizon, fusion sits as the highest-ceiling bet in the stack.

What wins here: modular generation with a low levelised cost of energy and a repeatable deployment playbook. Power equipment has been structurally undersupplied for years, so the companies that redesign around the shortage carry lead-time advantages that are hard to copy.

3. Grid hardware and the middle mile

Every electron that reaches a rack passes through a transformer first, and transformers are the chokepoint nobody priced in. Demand is up 119% since 2019. Lead times from the incumbents have stretched to five years, up from about one before COVID. Switchgear backlogs now run past 60 weeks.

This collides with a step-change in density. Cloud-era cabinets ran 20 to 40 kW. Today’s AI clusters run 500 to 600 kW per rack, and NVIDIA is targeting 1 MW. As the industry standardises on 800V DC, every legacy AC/DC conversion step becomes wasted energy and surplus hardware.

Three openings follow: beat the supply chain, re-architect the transformer itself, and unlock capacity from grid lines already in the ground. Solid-state transformers, built on silicon-carbide semiconductors, collapse what used to be a dozen-plus discrete power systems into one programmable, bidirectional device. On the software side, US grid congestion costs consumers an estimated $11.5B a year, much of it because lines run conservatively below their true limits. Tooling that safely pushes more through existing corridors is, effectively, free capacity.

One detail worth flagging for India: at least one player is already compressing transformer lead times from years to months by manufacturing through contract partners in India. The grid-hardware supply chain is a place India can serve the world, not just itself.

4. Software and orchestration

Generating power is half the problem. The other half is deciding how it gets dispatched and consumed. Most of this still lives in fragmented systems and ageing data-centre management platforms built for an earlier era.

The most defensible opening is here, and it rests on a quiet insight: much of AI’s load is flexible. Training jobs pause and resume, batch inference shifts across regions, low-priority work can absorb latency. Today’s contracts treat all of it as rigid, firm demand, which strands enormous grid capacity. Orchestration software that shifts batchable training to off-peak windows and hands capacity back during grid stress turns that rigidity into a release valve.

Layer on energy economics (utility bills, power-purchase agreements, hedging, battery dispatch) and real-time operational intelligence over the physical facility, and you get the connective tissue of the whole stack: software that spans the IT and physical layers at once and can act on the data, not just chart it. Whoever owns that intersection owns the most valuable real estate in AI infrastructure.

5. Construction, maintenance and labour

The workforce that physically builds this is running dry. Late 2025 saw a shortage of roughly 439,000 construction workers, while peak crew sizes have ballooned from around 750 in the cloud era to 4,000 to 5,000 today. The average data-centre worker is 53, and 60% of providers can’t fill open roles. Labour now delays projects as often as power does.

Robotics and AI software are filling the gap across the full lifecycle: autonomous trenching and earthworks at groundbreaking, drone-based progress and quality monitoring during the build, robotic inspection and maintenance in operation, and robotic disassembly at end-of-life as hyperscalers churn through GPU generations. The most defensible players pair automation with proprietary visibility, predicting failures before they cascade and recovering value from gear that would otherwise be scrapped.

6. Cooling

AI has pushed rack densities to levels the industry has never handled, on track for 50x the power density of cloud-era racks within a year. Air cooling simply can’t dissipate heat fast enough, and overheating eats directly into hardware life and performance per watt.

Liquid cooling has won the argument and now makes up the majority of new AI cooling architectures, split between direct-to-chip cold plates and full immersion. Above the hardware, a software layer is emerging that treats cooling as an active control system, using real-time power draw as an early-warning signal so cooling responds before chips throttle. Cooling has stopped being overhead. It is now part of the performance stack, and increasingly the water and emissions story of an entire campus.

ZOOMING OUT: DATA CENTRES IN SPACE

I spent the last two weeks in San Francisco, and the idea that kept surfacing in every room was the most extreme version of this entire thesis: skip Earth altogether.

If power and cooling are the two hardest constraints on the ground, orbit offers a brute-force answer to both. Unlimited, unfiltered solar, and the vacuum of space as a free, near-infinite heat sink. SpaceX has filed for up to a million data-centre satellites and is openly betting that within a few years, orbit is the cheapest place to run AI compute. Google and SpaceX have reportedly discussed putting compute in orbit; Google’s own test satellites are slated for 2027; NVIDIA has already announced silicon designed for orbital data centres.

The economics are still brutal, and the engineering harder still. This is a decade-out moonshot, not a 2026 line item. But the fact that serious capital is underwriting it tells you how binding the terrestrial bottleneck has become. When the ground is this constrained, the sky stops sounding crazy.

THE INDIA OPPORTUNITY

Here is the stat that frames everything: India generates roughly 20% of the world’s data and holds under 2% of global compute. To make it concrete, a single AWS campus in Indiana has reached 1 GW. India’s entire operational data-centre stock, across every city and operator, sits at about 1.5 GW. One American campus nearly matches a country of 1.4 billion people.

That isn’t a gap. It’s a generational buildout waiting to happen, and the trigger has already been pulled. The market has gone from $385M in 2014 to roughly $10B today, and is projected to reach $36B by 2035, with cumulative capex potentially touching $200B. Three forces are converging at once:

Sovereign compute. Generative AI is expected to add $400B to India’s economy by 2030, and India already runs one of the largest AI user bases on earth. But GPU revenue per MW runs about 10x a traditional data centre, which means India’s AI story is compute-constrained long before it is demand-constrained.

A 20-year tax holiday. The Union Budget 2026-27 introduced a tax holiday running to 2047 for foreign firms delivering global cloud services from India-based data centres. That kind of multi-decade fiscal certainty is rare in any emerging market. Pair it with a structural energy edge, solar at $0.038/kWh, second only to China, against $0.070 in the US, and India becomes a natural landing spot just as the US energy crunch pushes hyperscalers to look beyond North America.

Data localisation. Mandates now require sensitive data to live onshore, yet only about 3% of India’s data is stored domestically today. More local data means more local compute, which pulls in more infrastructure, which enables heavier workloads, which generates still more data. The flywheel feeds itself.

So where does the founder opportunity actually sit? The core infrastructure, the land and shells and megawatts, is a private-equity and sovereign-wealth game. The technology layers are pure venture territory, and they map almost one-to-one onto the six layers above. India’s neo-cloud players already compete on orchestration, thermal management and developer experience, not just space and power. The defensible moat is software, not square footage.

And the risks are real. India’s single biggest execution risk is the time-to-power gap: data centres take two to three years to build, grid upgrades take four to eight. There’s an overcapacity risk if speculative builds run ahead of firm offtake, a water and power-strain risk, and a genuine talent deficit in liquid cooling and data-centre automation. Eyes wide open. But two-to-three-year lead times leave room to course-correct, and the structural drivers aren’t going anywhere.

WHERE WE’RE ALREADY BUILDING

At Antler India, two of our bets already live on the hardest layer of this stack, the energy layer:

🔸 Meine Electric is APAC’s first iron-air long-duration storage innovator, targeting a sub-$0.05/kWh levelised cost with 16 to 24 hours of storage. India’s solar is cheap but intermittent, and the time-to-power gap is the country’s defining constraint. Long-duration storage is precisely how you make that cheap solar dispatchable around the clock and firm a renewable-heavy grid.

🔸 Green Aero builds high-efficiency, hydrogen-ready turbine cores. As behind-the-meter generation becomes the default, the aero-derivative turbine, a jet-engine core repurposed for power, is exactly the kind of modular, firm generation the on-site era will run on. It’s the same playbook Bessemer flags globally, being built from India.

THE BOTTOM LINE

The data centre is the new factory. The factories of the last century were built on steel and railroads; these run on power, cooling and the software that ties them together. The companies that win this cycle won’t just ride the capex boom. They’ll come out the other side as baked-in operational infrastructure with long-duration economics, even after the new-build wave slows.

The global map is clear. The Indian opening is rare. And the question I keep coming back to is which of these six layers produces India’s first $10B infra-tech company: power, cooling, or orchestration?

India is building from India for the world. 🚀

If you’re building in any layer of this stack, in data-centre infrastructure, energy optimisation, cooling, orchestration, or grid hardware, I’d love to hear from you. Drop a DM or tag a founder who is.

Sources: Bessemer Venture Partners (Roadmap: The AI data center stack); Indian Early Stage VC (India’s Data Centre Moment Has Arrived); Sightline Climate, McKinsey, JLL, CBRE, Colliers, Nomura, PwC, Goldman Sachs, NVIDIA, and company announcements.

There’s a stat from OpenRouter this March that should reframe how you think about the AI infrastructure stack.

70% of OpenRouter’s volume today goes to open source models. That’s 70% of nearly 27 trillion tokens in a single week. Roughly 5x growth in three months.

If you’ve been watching the model wars as a two-horse race between OpenAI and Anthropic, this is your wake-up call. The story has moved on. The interesting question isn’t which closed frontier model wins. It’s what gets built on top of the open models that are eating the volume.

And in the agentic era, that question has a very different answer than it did in the chatbot era.

Why open source wins the agent era

A single LLM call is cheap either way. The price difference between Claude Opus and Qwen for one prompt is real, but manageable. Not a business decision.

Now run a fleet of agents.

Each agent in an agentic workflow makes hundreds of sequential model calls. Tool calls. Memory lookups. Reasoning steps. Verification loops. Multi-step plans that branch and retry. A single user request can fan out into 200, 500, sometimes 1,000+ model invocations.

At that volume, the open vs closed gap doesn’t just matter. It widens to 10x-25x.

50 parallel agents on Qwen cost less than 1 agent on Opus.

Read that again, because it’s the entire thesis. The unit economics of agents at scale only work if the underlying model is open source. Closed models will continue to dominate single-turn reasoning and high-stakes calls. But for the long tail of agentic work, the routine fetch-parse-decide-act loop that makes up 80% of agent operations, open source isn’t a preference. It’s the only economically viable path.

Layer on what enterprises actually want when they deploy agents at scale: control over data, control over cost, customization for their workflows, no vendor lock-in. That’s not a wishlist. That’s procurement policy at every Fortune 500.

The conclusion writes itself. Open source models AND open source harnesses are going to own the agentic workload.

The model layer is just the tip of the iceberg

Models are necessary but nowhere near sufficient. The harness around the model, the orchestration logic, the memory layer, the tool calling, the durable execution, the failure handling, is where production agents actually live or die.

And that harness layer is going open source faster than the model layer did.

Early entrants like OpenHands, Cline, and OpenCode proved that open source coding agents could match commercial alternatives on quality. Newer ones like OpenClaw and Hermes are going further. They’re winning vertical adoption, embedding into specific workflows where domain knowledge matters more than raw model capability.

This is the playbook that played out with databases (Postgres beat Oracle in the long run), with operating systems (Linux beat Windows Server in the cloud), and with web infrastructure (Nginx, Kubernetes, all open). It’s playing out now with agent infrastructure.

So where exactly are the opportunities?

The map: Above the AI vs Below the AI

The cleanest framework I’ve come across is from the Basis Set team (h/t to Chang Xu, whose two-part essay on the open source agent stack is required reading - link at the bottom).

They split the agent stack into two layers with very different economics and very different defensibility profiles.

Below the AI: Infrastructure every agent stack needs

This is the layer the models themselves cannot solve. No matter how good GPT-7 or Claude 5 gets, an agent still needs persistent memory, durable storage, reliable tool execution, secure runtime, and observability into what it actually did. These are systems problems, not capability problems.

Security is the youngest category in the stack. Eighteen months ago, “agent security” wasn’t a thing. Today there are well-funded startups on both sides: autonomous security tools that probe and harden agent systems (Runsybil, XBOW, Pensar, Terra) and supply chain security tools that lock down the dependencies agents pull in (Chainguard, FLOX, Socket). Every enterprise deploying agents will eventually need both.

Reliability is where I’d argue the biggest near-term enterprise value sits. Two sub-categories matter: typed interfaces (Boundary, Pydantic) that turn the messy probabilistic output of LLMs into structured data the rest of your system can consume, and durable execution (Temporal, Trigger.dev, Inngest, Restate, DBOS) that handles the long-running, retry-able, crash-resistant workflows agents need. When your agent runs for six hours and the API call fails on hour five, durable execution decides whether you lose everything or pick up from step 47. Prompt chains are not workflow engines. Don’t confuse them.

Agent Native Primitives is the catch-all for the new infrastructure that didn’t exist before agents. Memory and context (mem0, atuin, Entire) so agents can remember what they did yesterday. Storage (Tigris, PgDog, Alien) optimized for the access patterns of autonomous systems. Email (Resend, AgentMail) built for machines that send and parse mail at scale. The principle here: agents need new infra, not retrofitted SaaS. Trying to make agents work with tools built for humans is the friction point that’s holding production deployments back.

Above the AI: The application layer

This is where the audience and the workflow get owned. Above the AI products combine models, harnesses, and UX into something a specific user actually pays for.

Two distinct wedges are emerging.

The Prosumer wedge - I think of this as “OpenArt for Agents.” OpenArt won by making generative image creation accessible to non-engineers: simple UI, opinionated workflows, no need to understand diffusion models. Companies like Dust and OpenClaw are doing the same for agents. The full agentic stack underneath, but packaged for someone who would never write a system prompt. This wedge is going to be huge because the addressable user base is everyone, not just developers.

The Enterprise wedge - or “Red Hat for Agents.” Red Hat won by taking the chaos of Linux distributions and packaging it into something enterprises could buy with a support contract. Same opportunity here. There are now too many models, too many harnesses, too many primitives. CIOs see a bewildering vendor landscape and need someone to put it all together: model serving at scale, model specialization for their domain, integration with their data, SLAs they can sign. Fireworks, Together, Baseten, and Parasail are already pieces of this puzzle on the inference side. The full stack equivalent doesn’t quite exist yet, and that’s the prize.

The India opportunity

Now the part I actually want to talk about.

India has done something interesting and underappreciated. We are already the world’s open source backend. The Linux kernel has heavy Indian contribution. Kubernetes has heavy Indian contribution. Postgres has heavy Indian contribution. Every major open source infrastructure project that powers modern software has a meaningful Indian commit graph.

What India hasn’t done historically is convert that talent into globally distributed dev-tools companies at scale. Postman is the proof point everyone cites because there are so few others. JetBrains, GitHub, MongoDB, HashiCorp, Datadog, Snowflake - the developer infrastructure decacorns of the last 15 years have been built almost entirely outside India.

That’s about to change for one specific reason: the agentic stack is being built from scratch right now. The category leaders haven’t been crowned. The standards haven’t settled. The distribution moats haven’t formed. This is the most open window for new infrastructure companies in fifteen years, and Indian founders have a structural advantage in three dimensions:

1. Engineering depth. The talent that built the open source backend is here.

2. Cost structure. Dev-tools companies built from India can hit profitability at revenue levels that US-based competitors structurally cannot.

3. Global from day one. Open source distribution is borderless. A founder in Bengaluru and a founder in San Francisco have the same GitHub repo, the same Discord, the same developer audience.

And we’re starting to see real action.

Composio: the wedge expanding into the moat

The proof point I want to spend time on is Composio.

Founded in 2023 by Soham Ganatra and Karan Vaidya, both IIT Bombay grads who met at a Physics Olympiad camp and ended up as roommates, Composio sits in the Agent Native Primitives layer. Specifically, the tool calling sub-category. Their product connects AI agents to 1,000+ external tools - GitHub, Salesforce, Slack, Gmail, the long tail of every SaaS application your agent might need to actually do work.

That sounds boring until you realize how much pain it removes. Before Composio, every agent team had to build and maintain their own integration layer: auth flows for every API, retry logic for every endpoint, schema translation for every tool. Composio collapsed that into a single SDK. The result: 27,500+ GitHub stars, 100,000+ developers, $29M raised including a $25M Series A led by Lightspeed in July 2025.

But here’s what makes Composio interesting beyond the numbers. They’re moving up the stack.

Soham’s framing: “Reliability is the biggest pain point getting agents live in production today.”

So they’re expanding from tool calling into reliability. From a wedge into a moat. The deeper insight is that once you own the integration layer, you have proprietary data on which tool calls fail, which patterns work, which workflows are reliable. That data becomes the foundation for a reliability platform that nobody else has the right to build.

This is the playbook. Start with the most painful, most concrete problem in agent development. Win it. Use the position to expand into adjacent problems. Eventually own a stack-level moat.

We had the chance to hear Soham tell this story live at Antler Embark last week. Embark is our US immersion program for Indian founders building AI-first companies for the global market. Hearing Soham walk through the Composio journey - from a $4M seed round with Together Fund and Elevation, through 100,000 developer adoption, to a $25M Series A led by Lightspeed with participation from Guillermo Rauch (Vercel) and Dharmesh Shah (HubSpot) - was a clinic in how Indian founders should think about building global infrastructure companies.

The detail that stuck with me: Soham mentioned that Composio’s positioning has shifted multiple times as the agent market matured, but the underlying technical bet (that integration plumbing would be the bottleneck) was right from day one. Conviction on the wedge, flexibility on the framing. That’s the founder pattern that wins infrastructure markets.

So who’s next?

The talent is here. The TAM is global. Composio is proof that an Indian founder can build a critical layer of the global agent stack and have US-tier venture capital write the checks.

The question is what gets built next.

The Below the AI layers that I think are most undervalued for Indian founders right now:

• Durable execution. Temporal, Trigger.dev, and Inngest are all US companies. The infrastructure DNA needed to build this is exactly what Indian engineers have been doing in payments and fintech for a decade.

• Agent-native memory. Mem0 is the early leader. The space is wide open and the problem is fundamentally a distributed systems problem, which plays to Indian engineering strengths.

• Open source security primitives. Autonomous agent security is a category that didn’t exist 18 months ago. Indian security research talent (the OWASP top 10 has had heavy Indian contribution for years) has not converted into productized companies. That’s a gap.

• Storage and infra for agents. PostgreSQL extensions, vector stores tuned for agent access patterns, file systems that understand agent context. All of this is being built right now.

If you’re building open source agent infrastructure (or thinking about it), reach out. I’d genuinely like to know what you’re seeing in the market that the rest of us are missing.

The window is open. The map exists. The first Indian founder to plant a flag in each of these layers gets the same shot Composio just took.

So who’s next?

Basis Set’s two-part thesis on the Open Source Agent Stack is here: Part 1 | Part 2

If you’re building open source agent infra in India or thinking about it, DM me or drop a comment. Always happy to chat.

#MarketMapMondays is a weekly series where I unpack market maps from leading VCs and ask where Indian founders fit in. Subscribe if you’d like these in your inbox every Monday.

Alphabet is worth ~$3.6T. Microsoft, ~$3.1T.

That’s a 60x gap between the best vertical SaaS company and horizontal incumbents. For two decades, the conventional wisdom was clean: horizontal wins, vertical is a niche. Verticals were too small, too fragmented, too hard to sell into.

AI is about to flip that. Here’s why.

The wallet just got 20x bigger

Menlo Ventures put out a sharp piece this week that reframes the whole opportunity.

Global software spend is roughly $650B. Global labor spend is ~$13T.

Horizontal SaaS captured the software wallet. It built tools that helped people work. Vertical AI is going after something different: the work itself.

When the product does the job rather than supports someone doing the job, the addressable pool isn’t IT budget. It’s the operations budget, the cost of running the business, the line items that show up on the COO’s P&L.

That’s a 10-20x larger pool, and it’s the reason a vertical AI company in claims processing or contract review or radiology can plausibly reach a scale that vertical SaaS never could.

What changed structurally

Three things shifted at once.

Build cost collapsed. What used to need 50 engineers and five years of domain immersion can ship in months. AI compresses the distance between “we understand this industry” and “we have a working product.”

Defensibility moved. The moat is no longer “we built complex software.” It’s owning the workflow, the proprietary data, and the outcome. That’s harder to copy than a feature set.

Pricing decoupled from seats. When AI does the work, you charge per claim processed, per contract reviewed, per design generated. Outcome-based pricing finally has a real substrate. Seat-based SaaS economics start to look quaint.

Why India could win big here

India is sitting on a structural advantage hiding in plain sight.

We have deep domain talent in the exact verticals where AI is now viable: legal, healthcare, financial services, logistics, manufacturing, AEC. Not just engineers, but people who actually understand how claims get adjudicated or how a building gets permitted.

The “building from India for the world” thesis finally has a wedge that doesn’t require a US-based founder. Vertical AI sells on measurable outcomes, not relationships and golf courses. If your product processes claims 10x faster at 1/5th the cost, the buyer doesn’t care where you’re headquartered.

And Indian founders have spent a decade building services + software hybrids for global enterprises. That muscle - knowing how to deliver outcomes, not just ship a SKU - is exactly what vertical AI demands.

We’re already seeing it across the ecosystem: founders attacking specific verticals with AI-native workflows, not horizontal copilots searching for a problem to solve.

The question worth pondering

Which vertical gets its first $100B AI-native company? And does it come out of India?

For 20 years, the answer to “where do the biggest software companies come from” was the same handful of zip codes. Vertical AI doesn’t have that gravity yet. The playing field is flatter than it’s been in a long time.

If you’re building vertical AI, or know a founder who is, my DMs are open. Would love to chat.

But here’s what’s been missing from the conversation: what does AI adoption actually look like from inside a large Indian bank? Not the LinkedIn hot takes. Not the vendor pitch decks. The real, messy, operational truth.

To find out, I kicked off my very first podcast by sitting down with someone who has a ringside view of this: Sameer Shetty, Head of Digital Banking Transformation and Strategic Initiatives at Axis Bank.

Quick background: Sameer and I go back a long way. We both went to IIM Bangalore, were both at McKinsey, and Sameer spent years advising large banks before crossing over to the operator side at Axis. He now leads a team of 2,000 people driving AI and digital for one of India’s largest private banks.

A huge thank you to AIBoomi for making this conversation happen, and to Sameer for being so generous with his time and insights.

The 80/20 rule doesn’t work in banking

This was the first thing that hit me. Sameer was blunt about it.

In consulting, you focus on the 20% of things that drive 80% of impact. In banking? That framework falls apart for two reasons.

One, the “edge case” is the average case. The happy path in a digital banking journey is maybe 50-60% of cases, not 90%. And those edge cases aren’t rare. They are the reason only about 3% of digital consumer journeys in banks end in an actual purchase. Not 30%. Three percent. That number floored me.

Two, you can’t afford a single regulatory divergence. Whether you make one mistake or 10,000 mistakes, the consequence is the same. There’s no room for “we’ll fix it later.”

This is why so many digital investments in banking have struggled to show clear ROI. The tech got better. The apps got better. But your competitor also got better at the same time. It’s a treadmill.

The three buckets of AI impact in banking

Sameer broke down AI’s impact into three clean categories:

1. Cost reduction - the most obvious play, but not the biggest one. Axis has about 1,05,000 employees. Only ~15,000 sit in centralized roles (ops, IT, product, analytics). The rest are in branches and frontline positions. You can’t take half a person out of a six-person branch. So the cost savings from AI in centralized ops are real, but they’re capped by the distributed nature of banking.

2. Revenue growth - this is where the massive unlock is, and also where the biggest question marks remain. If AI frees up five hours of a branch employee’s day, that person can now make more calls, serve more customers, close more business. The cost stays fixed, but the output goes up. Sameer was clear: the real ROI of AI in banking will show up on the revenue line, not the cost line.

3. Risk and compliance - still in early stages. Axis is in the pilot phase for using AI to draft product approval notes, generate first-pass compliance reviews, and route documents more efficiently through committees.

“Zero Ops” - what it actually means

One of the most interesting concepts Sameer unpacked was Axis Bank’s “Zero Ops” vision. It’s aspirational (he was upfront about that), but the direction is clear: take out as much manual operations activity as possible.

A concrete example: car loans. Today, when you apply for a car loan, the bank needs your proforma invoice, own contribution receipt, insurance documents. Every dealer’s proforma invoice looks different. No standard format, no centralized source. Someone has to read each one and enter the data manually.

Axis is now using AI to process these documents, match the data against the application, and flag mismatches to the sales team in real time. The endgame? You walk into a dealer, apply for a loan, get sanctioned, submit your documents, and walk out with the car. Near real-time post-sanction processing.

But Sameer was honest about what “zero” actually means: it won’t be literally zero. There will always be a human checker. The shift is from one person doing 10 files to one person doing 100 files.

The speed trap: 3 weeks to build, 5 months to integrate

This was one of the hardest-hitting insights in the conversation, and every founder building for BFSI needs to hear it.

Sameer shared a pattern they’ve seen across multiple AI projects at Axis:

• Building the AI capability (the bot, the document reader, the model) for the happy case? Three weeks, maybe six.

• Integrating it with the bank’s core systems via APIs, going through security cycles, handling edge cases? Four to five months.

The AI is fast. The plumbing is slow. And that’s not going to change overnight because banks run on legacy core systems that have evolved over decades. If you’re a startup and you think something will take X hours, assume 5X. Sameer said this directly.

What actually makes a startup win with banks

I asked Sameer the question every fintech founder wants answered: how does a ten-person scrappy startup sell to Axis Bank?

His answer was clear and specific:

Build a niche, differentiated capability. That’s your moat. Don’t try to own the end-to-end process. The bank owns that. The bank has the context, the regulatory knowledge, the process ownership. What the bank needs from you is a very specific, hard-to-replicate skill.

Examples he gave: reading bank statements with high accuracy across formats. Handling identity validation where Aadhaar says “Anand Bhai Shah” and PAN says “Anand Shah.” Spotting fraud patterns in CCTV footage across 6,000 branches and 20,000+ ATMs.

If your capability is generic and the next vendor can do the same thing, the bank will run two or three vendors in parallel. But if you’re genuinely the best at one specific thing, and you’re already embedded in the bank’s workflow, replacing you becomes very hard.

His other pieces of advice for founders:

• Security and compliance are non-negotiable from day one. You can’t say “we’ll figure it out in three months.” If there’s even a question mark around regulatory compliance, you won’t get through the door.

• Don’t get enamored by the TAM. The headline number looks massive, but banks don’t pay what you think they pay. The effective TAM shrinks fast once you account for how procurement actually works.

• Integration complexity is real. The technology stack behind banks is diverse, layered, and has evolved over decades. Plan for it.

The ROI question nobody wants to answer honestly

Sameer didn’t dodge the hard question on AI costs. Inference costs are real, and they matter at bank scale. But he shared something important: by switching model architectures and being smart about when to use inference versus cached/precomputed responses, Axis has brought costs down by 50-60% on some use cases. Building your own SLM can bring it down to a tenth of the original cost.

His key insight: for most banking use cases, you don’t need the world’s most powerful LLM. A smaller, cheaper, fine-tuned model that does one thing well is often enough.

The future of work: honest, not rosy

I asked Sameer for his macro view on AI and jobs. He didn’t sugarcoat it.

Certain types of jobs will shrink. That’s a fact. But he drew a useful analogy: when people moved from physical spreadsheets to Excel, the job didn’t vanish. The time spent creating the spreadsheet dropped dramatically, and the person’s output went up. The job changed.

He also made a point I haven’t heard anyone else make as cleanly: a few hundred years ago, there was no entertainment industry because people worked seven days a week with no free time. The Industrial Revolution and electrification created leisure time, which created entirely new industries. AI will do something similar. New categories of work will emerge that we can’t predict today.

His take: net positive, but not painless. The transition will be hard. Certain roles will be displaced. But he’s a believer in human adaptability, and I share that optimism.

Rapid fire highlights

• Worst advice for BFSI founders: Getting seduced by the headline TAM without understanding how much banks actually pay.

• One BFSI buzzword he’d kill: “Synergy.”

• Book that changed his thinking on tech: The Hard Thing About Hard Things by Ben Horowitz.

• Go-to AI tools: Copilot at work; Gemini and Grok at home.

• Bold prediction for AI in banking (5 years): AI will change banking dramatically, but banking itself won’t change. It’s still about trust, sound lending decisions, and getting the right customers. It has been for thousands of years.

My take

This conversation confirmed something I’ve believed for a while: the $50B India AI opportunity is real, but it’s going to be won by founders who understand the problem deeply, not by those with the flashiest demo.

Banking is a $250B+ industry in India. The productivity gap is staring us in the face. Opex ratios going up, not down. Cost-to-income stuck. Employee costs growing 3X in a decade with only 1% productivity improvement.

AI can change this. But only if founders build with the right mental model: niche capability, deep integration, regulatory readiness from day one, and patience for the 5X longer timelines that bank infrastructure demands.

If you’re building AI for BFSI, or if you know a founder who is, I’d love to hear from you. Drop a comment or DM me. This is going to be a big space, and we’re just getting started.

This article is based on Episode 1 of our podcast series exploring India’s $100B tech opportunity. A huge thank you to AIBoomi for partnering on this series, and to Sameer Shetty for an incredibly candid conversation.

Over 200 biological foundation models have been released since 2022. Isomorphic Labs raised $600M. Chai Discovery hit a $1.3B valuation. Bioptimus secured $76M.

Something big is shifting in how we build AI for biology. And this week’s #MarketMapMondays unpacks exactly what.

I came across a really well-done market map by XAnge, put together by Emma Martin, that breaks down how bio foundation models are being applied across protein design, drug discovery, medical imaging, and clinical data. It’s one of the clearest maps I’ve seen on this space, and it got me thinking about what it means for founders, especially those building from India.

Let me walk you through it.

The shift from task-specific to foundational

To understand why this matters, you need to understand what’s changing.

AI in drug discovery isn’t new. Models that predict protein binding, estimate toxicity, or classify medical images have been around for years. They’ve delivered real efficiency gains. But they all share a limitation: each one is trained from scratch to solve a single, well-defined problem. One model per task. Every time.

Bio foundation models work differently. They’re large models pre-trained on massive, diverse biological datasets using self-supervised learning. Instead of learning one task, they learn general patterns and relationships across the data, and can then be adapted to many downstream applications with relatively little additional data.

Think of it as the GPT moment for biology.

Why is this such a big deal? Because biology operates across multiple interconnected levels. Genes, proteins, cells, tissues, organs. And biological data comes in wildly different formats: molecular sequences, medical images, clinical records. A task-specific model trained on one slice of this can’t see the full picture. A foundation model can integrate across these scales and modalities, picking up patterns that span both known and hidden relationships.

Three capabilities make these models particularly powerful. They can generalize to new tasks with few or no examples. They can work across different data types (molecular sequences, images, clinical records) in a single model. And they can generate new molecules, proteins, or antibodies, not just classify existing data.

That last point is worth sitting with. We’re moving from AI that sorts information to AI that creates biology.

Where the money is going

The funding rounds tell you exactly where conviction is building.

Isomorphic Labs, a spin-out of Google DeepMind, raised $600M in its first external round led by Thrive Capital to push foundation models deeper into drug discovery pipelines. Chai Discovery hit a $1.3B valuation off a $130M raise. Their approach uses zero-shot learning to design high-affinity antibodies without extensive target-specific data. Bioptimus secured $76M to build what they’re calling a universal biological foundation model, spanning histology, transcriptomics, and clinical data.

Capital is concentrating around teams that control two things: generalizable models and unique multimodal datasets.

The real moat isn’t the model

Here’s what I find most interesting in Emma’s market map. And it’s counterintuitive if you come from the software world.

The model architectures are increasingly open source. What isn’t replicable is the data.

In biology, the most valuable datasets are built through years of proprietary experiments, clinical partnerships, and exclusive access to patient cohorts. You can’t match them with more compute alone. This is a fundamentally different moat than what we see in most AI verticals, where model architecture or distribution tends to be the differentiator.

Bioptimus is a good example of this data flywheel in action. More partners mean more diverse training data. Better training data improves the model. A better model attracts more partners. The companies commanding the highest valuations aren’t just betting on model architecture. They’re betting on a compounding data advantage that gets harder to replicate over time.

From the map, two company archetypes are emerging:

FM-native players like Bioptimus and Raidium build large multimodal foundation models as their core product. They have full control over architecture and data, which gives them strong long-term defensibility. The trade-off is heavy capital requirements, compute costs, and long development cycles.

FM-adopters and specializers either build foundation models in-house for specific applications or integrate pretrained models into existing workflows. Lower cost, faster deployment, but constrained by the limits of the underlying models they depend on.

Both paths have merit. But if I had to pick where the durable value sits, it’s with the FM-native companies that own their data pipeline end to end.

The biggest unlock: digital twins and virtual patients

Most drugs don’t fail because the science was wrong. They fail late in development because of patient variability, unexpected toxicity, and complex system-level biology that task-specific models were never designed to capture.

This is where multimodal bio foundation models could change the game.

The most promising direction, as Emma’s map highlights, is the development of virtual patient cohorts and digital twin models. These aim to simulate biological systems, tissues, or even groups of patients in silico. The idea is to test hypotheses, explore dose-response effects, identify likely non-responders, and surface safety risks before entering expensive Phase II or III trials.

These models won’t replace clinical trials. But they can help teams fail earlier, cheaper, and with better insight. That’s a massive unlock in an industry where late-stage failures routinely burn hundreds of millions of dollars.

The bottleneck? Multimodal models need paired datasets (omics + images + clinical outcomes linked to the same patients), and those are rare and expensive to collect. Synthetic data generation is one promising workaround, but it’s an active area of research, not a solved problem.

Where does India stand?

This is where it gets personal for me.

India’s pharma market is $60B today, projected to hit $130B by 2030. We rank 3rd globally in drug production by volume. We supply 20% of the world’s generics. Our pharma exports crossed $30B in FY25. The infrastructure, the manufacturing muscle, the regulatory track record: it’s all there.

But here’s the contrast. Tracxn tracks only about 23 AI-in-life-sciences startups in India. For a country with this much pharma depth and AI talent, that number should be 10x higher.

The policy signals are encouraging. The 2026 Union Budget announced the Biopharma SHAKTI initiative with ₹10,000 crore allocated over 5 years to position India as a global biopharmaceutical hub. Another ₹500 crore went to AI Centres of Excellence. Bengaluru, Hyderabad, Pune, and Mumbai are all seeing early-stage computational biotech activity. Peptris Technologies produced India’s first AI-discovered drug candidate (PEPR124, licensed to Revio Therapeutics in 2025). These are real signals.

But most of what I see in India today is applying off-the-shelf AI to existing pharma workflows. That’s useful. It’s not where the transformative value will be built.

The opportunity gap is clear. India has the pharma muscle. It has the data scale from a billion-plus population. It has deep AI and engineering talent. What’s missing is founders going deep on biological foundation models. Building FM-native companies that own their data pipelines, forge clinical partnerships, and compete globally at the model layer.

The wedge is hiding in plain sight. Someone needs to take it.

What I’m watching for

A few things I’ll be tracking in this space:

The first is whether FM-native companies can actually demonstrate clinical traction, not just research publications, but real pharma partnerships with milestone payments. That’s the proof point that separates hype from value.

Second, the data partnership race. The companies that lock in exclusive access to large, diverse, multimodal patient datasets in the next 12-18 months will have structural advantages that are nearly impossible to replicate.

Third, and closest to home: will an Indian founder build a bio foundation model company that competes globally? We have every ingredient. The question is execution and ambition.

If you’re building at this intersection of AI and biology, especially from India, I want to hear from you. Drop a comment, DM me, or tag a founder who should be on my radar.

Market map source: “Bio Foundation Models: A Shift in AI for Biology” by Emma Martin, XAnge (April 2026)

There’s a framework from 1943 that explains where the Consumer AI market is going better than any pitch deck I’ve seen this year. Abraham Maslow never saw a large language model. But his hierarchy of needs, from survival to self-actualization, maps almost perfectly onto how AI companies are building for consumers right now.

I came across a market map last week by Headline’s Matt Brown (Vinay Reddy also collaborated on it) that made this connection explicit. It plots Consumer AI startups against Maslow’s five layers. The moment I saw it, I couldn’t unsee it.

We talk about AI in terms of benchmarks and model sizes. But the consumer companies that actually win? They’re solving for human needs. Full stop.

Here’s how the pyramid breaks down.

Physiological: AI for daily life

The base layer. Meal planning, grocery automation, budget copilots, home energy. Not the kind of stuff that gets standing ovations at demo day. But it’s high-frequency, deeply embedded, and the pain is universal.

Household complexity has exploded over the last decade. Subscriptions, devices, dietary restrictions, kids’ schedules. Most of us are drowning in micro-decisions. AI is now good enough to take a real chunk of that off your plate.

The TAM here is big precisely because the problem is boring.

Safety: AI that protects you

Scams have gotten personal. Phishing emails reference your actual purchase history. Deepfakes cost nothing to produce. Your parents’ phones are targets.

AI fills a gap here that humans can’t cover anymore. Personal data firewalls, health monitoring, consumer cybersecurity, legal and financial copilots. The real shift is from reactive security (you get scammed, then you react) to preventative security (AI catches it before you notice).

Trust is the moat in this layer. The companies that earn it will be hard to displace.

Belonging: AI for connection

I’ll be honest, this layer makes some people uncomfortable. I get why. But the numbers are hard to argue with.

Loneliness is at epidemic levels globally. Social graphs are shrinking in every age cohort. And LLMs have gotten good enough at memory, personality modeling, and emotionally coherent responses that they’re creating something new. Not a replacement for human connection, but a supplement to it. AI companions, community formation, matching based on actual interests rather than algorithms optimized for engagement.

The startups building here are seeing retention numbers that most consumer apps can only dream about. That tells you something about demand.

Esteem: AI for mastery

This is the layer I keep coming back to.

We’ve hit a point where AI can evaluate your work, coach you through a skill, and help you create at a level that used to require years of practice or an expensive tutor. An amateur cook can follow AI guidance to produce genuinely impressive food. A student in a tier-3 city can access the kind of personalized instruction that used to be reserved for prep school kids in metro areas.

AI gives people both capability and confidence. That combination is powerful.

Headline calls one category here “social capital engines,” AI co-creators that help people build a presence and a body of work. I think that framing is right. Esteem isn’t just about feeling good. It’s about having something to show for it.

Self-actualization: AI that helps you become

The top of the pyramid, and the hardest layer to build a company around.

This is where AI moves past efficiency into meaning. Life planning, guided wellness, creative exploration, education pathways that adapt to who you’re becoming rather than who you were. The technical pieces are falling into place: long-horizon reasoning, persistent memory, reflective dialogue. These are the building blocks of coaching and personal transformation, things that used to require expensive human experts.

We’re early. But whoever figures out this layer is going to build something with extraordinary staying power.

Check out this interesting Market Map from Headline VC (Matt Brown & Vinay Reddy)

Why India has an edge

Let me zoom out and make the India case, because when I look at this pyramid through an India lens, the gaps are hard to ignore.

India has 1.4 billion people and structural shortfalls at every layer. Not in theory. In practice, right now:

The physiological layer is wide open. Millions of households are managing complex daily logistics with barely any tooling. AI for meal planning, budgeting, and home management in India is almost untouched.

Safety is urgent. India accounts for roughly 11% of global road accident deaths. Digital fraud is climbing fast. Health monitoring infrastructure is thin once you leave the top 8 cities.

Belonging matters here more than most places. India is one of the youngest countries in the world. A generation that grew up on smartphones but reports increasing social isolation. The demand for connection tools isn’t theoretical.

Esteem is India’s aspiration economy. Tens of millions of people want to learn new skills, build creative portfolios, move up. But quality coaching and tools are expensive and concentrated in a handful of cities. AI can break that geography lock.

Self-actualization is where the need is most acute. India has roughly 0.3 psychiatrists per 100,000 people. Therapy, coaching, personal development, these services barely reach anyone outside the urban upper-middle class. AI changes the supply equation entirely.

At Antler India, we’ve been betting on this thesis. Three of our portfolio companies map onto the pyramid:

Cravv.in is at the Esteem layer. It uses AI to teach people cooking, not just hand them recipes. The difference matters. Recipes are instructions. Mastery is a skill.

Cautio is at Safety. It’s applying AI to road safety, one of India’s most urgent problems and one that most tech companies have ignored.

Wavelength is at Belonging. It’s an AI dating companion that’s approaching the connection problem differently from swipe-based apps.

But we’re barely scratching the surface. I’m looking for founders building India-focused Consumer AI across the full pyramid. The whitespace I’m most interested in:

• AI for education and learning (Esteem / Self-actualization)

• AI for health (Safety / Physiological)

• AI for legal access in India (Safety)

• AI for personal finance (Physiological / Safety)

• AI for mental wellness (Self-actualization)

The playbook for Indian founders here is straightforward: pick a layer, find a painful consumer problem with deep local context, solve it for India, then take it outward.

So what?

Maslow’s hierarchy gives us something that most market maps don’t. A theory of motivation underneath the technology. The Consumer AI winners of this cycle won’t be the companies with the best models. They’ll be the ones that pick a human need and embed into it so deeply that switching away feels like losing a part of your life.

In India, every layer of this pyramid has room for a category-defining company. That’s not hype. It’s arithmetic.

If you’re building in this space, reach out. DM me or leave a comment. I want to hear what you’re working on.

Source article: Market map by Headline (Matt Brown, with Vinay Reddy).