After Generative AI: Why Vertical AI and Physical AI Could Define the Next Decade

For the last few years, the AI conversation has largely revolved around one question:

Who can build the most powerful AI model?

OpenAI, Google, Anthropic, Meta, DeepSeek and others pushed the frontier forward at extraordinary speed. Models learned to write, code, analyse documents, generate images, reason across complex problems and increasingly operate software on behalf of users.

But another transition is now becoming visible.

The next phase of AI may not be defined simply by making models larger.

It may be defined by making intelligence more specialised — and more physical.

Two categories are therefore becoming increasingly important:

Vertical AI — artificial intelligence built deeply around a particular industry, workflow or domain.

And Physical AI — artificial intelligence capable of perceiving the physical environment, reasoning about it and taking action through robots, autonomous vehicles, machines and other embodied systems.

Together, these two shifts could fundamentally reshape the economics of industries ranging from healthcare and finance to manufacturing, logistics, agriculture, travel and construction.

The evolution looks something like this:

Generative AI → AI Agents → Vertical AI → Physical AI → Autonomous Industries

And there are already signs that this transition has started.

The AI market is moving from intelligence to outcomes

AI adoption itself is no longer experimental.

Stanford's 2026 AI Index reports that approximately 88% of surveyed organisations were using AI in 2025, while generative AI was being used in at least one business function by around 70% of organisations. Global corporate AI investment more than doubled during the year.

But enterprises have discovered something important.

Having access to an intelligent model is not the same as having an intelligent business process.

A general-purpose AI can explain an insurance claim.

A useful insurance AI needs to understand the policy, extract documents, assess the claim against thousands of rules, detect anomalies, access existing systems, request missing information and route or approve the claim.

A general-purpose AI can explain airline fares.

A travel-specific AI needs to understand fare families, GDS/NDC content, ticketing rules, schedule changes, penalties, supplier contracts, PNRs, hotel inventory and the commercial logic of an agency.

That difference is the opportunity behind Vertical AI.

Vertical AI: When AI learns the business, not just the language

Traditional enterprise software digitised workflows.

Vertical AI has the potential to perform parts of those workflows.

Instead of building another generic chatbot, companies are increasingly building specialised AI systems for:

healthcare
legal services
banking and insurance
manufacturing
logistics
real estate
education
agriculture
travel
construction
accounting and taxation

The attraction is simple.

The foundation model may eventually become a commodity.

Domain knowledge does not.

The competitive advantage moves towards proprietary workflows, specialised data, regulatory knowledge, integrations, historical transactions and the ability to execute a complete business process.

Bessemer Venture Partners describes this shift as AI moving from traditional "systems of record" towards "systems of action" — software that does not simply store information but acts on it. Its research argues that vertical AI is gaining traction particularly in industries where workflows have historically been manual, language-heavy, multimodal or dependent on expensive professional services.

That is an important distinction.

The biggest Vertical AI businesses may not simply replace existing software.

They could begin replacing portions of the service economy surrounding the software.

The numbers are beginning to support the trend

Menlo Ventures estimates enterprise generative-AI spending reached approximately $37 billion in 2025, up from $11.5 billion in 2024.

More than half went to AI applications rather than underlying infrastructure.

And within applications, spending on Vertical AI reached about $3.5 billion in 2025, almost three times the $1.2 billion estimated for 2024.

Healthcare alone represented roughly $1.5 billion of that Vertical AI spending.

Another revealing statistic is how enterprises are obtaining AI.

Menlo found that 76% of enterprise AI use cases were being purchased rather than built internally in 2025.

That creates a major opportunity for specialised AI product companies.

Enterprises increasingly do not want another AI platform they must spend twelve months configuring.

They want:

"Here is my problem. Connect to my systems, understand my business and solve it."

That is exactly where Vertical AI becomes powerful.

The rise of "Service-as-Software"

SaaS largely sold tools to people.

Vertical AI increasingly sells completed work.

Consider the difference.

Traditional legal software helps a lawyer organise documents.

A Vertical Legal AI can review contracts, identify clauses, compare precedents, prepare research and draft portions of the legal work.

Traditional healthcare software records a doctor's notes.

AI systems can now listen to consultations and generate clinical documentation automatically.

Traditional customer-service software manages tickets.

AI agents increasingly resolve the tickets themselves.

The distinction between a software company and a service provider therefore begins to blur.

Bessemer has highlighted exactly this transition: Vertical AI companies can increasingly resemble intelligent service businesses rather than conventional software products.

This dramatically changes the addressable market.

A SaaS company historically competed for a company's software budget.

A sufficiently capable Vertical AI system can compete for parts of its labour, outsourcing and operational budget as well.

That could make some Vertical AI categories significantly larger than the Vertical SaaS businesses that preceded them.

Then comes the bigger transition: Physical AI

Until recently, most modern AI lived behind a screen.

It could write an email.

Generate software.

Analyse a spreadsheet.

Answer a question.

But it could not pick up the box sitting beside you.

Physical AI changes that.

Physical AI combines AI models with technologies such as:

Computer vision + sensors + spatial intelligence + world models + robotics + actuators + edge computing + simulation.

The result is a machine capable of:

seeing → understanding → planning → acting → learning.

Google DeepMind's Gemini Robotics models, for example, convert vision and language into actions for robots and are increasingly focused on spatial reasoning, planning, dexterity and whole-body control.

NVIDIA is building a similar ecosystem around Isaac robotics software, GR00T robot models, Cosmos world models, Jetson edge computing and Omniverse simulation. ABB, FANUC, KUKA, Yaskawa, Figure and numerous other robotics companies are now building on parts of this ecosystem.

This matters because AI is beginning to cross the boundary between the digital economy and the physical economy.

And the physical economy is vastly larger.

Factories.

Warehouses.

Ports.

Hospitals.

Roads.

Mines.

Construction sites.

Farms.

Airports.

Retail stores.

Energy infrastructure.

These environments represent trillions of dollars of activity that conventional software AI cannot directly automate.

Robotics has already reached significant scale

The transition should not be confused with science fiction.

Industrial robotics is already enormous.

According to the International Federation of Robotics, the worldwide operational stock of industrial robots reached a record 5 million units in 2025.

More than 600,000 new industrial robots were installed during the year, an increase of 11%. IFR expects annual installations to reach approximately 806,000 by 2029.

What AI potentially changes is the flexibility of those machines.

Traditional robots are incredibly effective when the environment is predictable.

Put the component here.

Weld this point.

Move from A to B.

Repeat the same motion 100,000 times.

But real life is messy.

Objects move.

People interfere.

Instructions change.

Products vary.

Unexpected circumstances occur.

AI allows machines to move away from rigid programming towards perception, reasoning and adaptation.

That dramatically increases the number of tasks that can potentially be automated.

McKinsey estimates robotics and Physical AI could create at least $1 trillion in economic value by 2040, with much of the early value concentrated in manufacturing and logistics.

Do not assume the winner will necessarily be a humanoid

Humanoid robots attract enormous attention because they are visually impressive and because our world is designed around the human body.

Doors, stairs, shelves, tools and factories were generally designed for people.

A machine with roughly the same physical structure can theoretically operate in those environments without rebuilding everything.

Companies such as Figure are already testing humanoids in industrial environments. Figure says its earlier Figure 02 deployment participated in production associated with 30,000 vehicles at BMW's Spartanburg plant before its newer Figure 03 returned for more complex logistics work.

But there is an important countertrend.

The first major Physical AI winners may be specialised robots rather than humanoids.

Reuters reported in 2025 that more than 70% of the $2.26 billion invested in robotics companies during the first quarter was going toward specialised, single-function robots rather than general-purpose humanoids. India's Ati Motors was one example, with autonomous machines already performing heavy material movement for industrial customers.

This makes commercial sense.

A specialised machine does not need human-level intelligence.

It only needs to perform one valuable task more safely, consistently or economically than the alternative.

So before we see humanoids cooking dinner in every home, we are much more likely to see AI machines:

moving pallets,

inspecting factories,

sorting packages,

cleaning industrial facilities,

monitoring infrastructure,

operating mining equipment,

assisting surgeons,

harvesting crops,

and transporting materials.

What the world's major economies are doing

The global race is therefore expanding beyond foundation models.

China

China already accounted for 59% of worldwide industrial robot installations in 2025, with approximately 354,000 robots installed during the year.

It is now aggressively pushing embodied intelligence.

In 2026, China's Ministry of Industry and Information Technology launched real-world training programmes for humanoid and embodied AI systems across manufacturing, logistics, healthcare, emergency response and other scenarios, targeting large-scale deployment and the accumulation of real-world robot training data.

China's huge manufacturing ecosystem gives it an important advantage: robots need factories not only as customers, but as training environments.

United States

The US advantage remains extremely strong in AI models, GPUs, software and private-sector capital.

NVIDIA is attempting to provide much of the infrastructure layer for Physical AI.

Google DeepMind is developing robot intelligence models.

Amazon has spent years building one of the world's largest real-world robotics environments.

Companies including Figure, Agility, Apptronik and others are advancing general-purpose robotics.

The US strategy is therefore highly ecosystem-driven: AI models + chips + cloud + robotics startups + industrial customers.

Japan

Japan is attempting to combine its historic robotics leadership with modern generative AI.

In 2026, Japan's METI introduced an AI Robotics Strategy and expanded its GENIAC programme to include development of robotics foundation models and AI-ready manufacturing datasets.

South Korea

South Korea announced plans in August 2026 to invest 2.3 trillion won through 2030 in full-stack humanoid development, alongside significant spending on Physical AI demonstrations across manufacturing, agriculture, logistics, construction, defence and care.

Europe

The European Union has created dedicated programmes for Physical AI and embodied intelligence, explicitly targeting autonomous machines capable of perception, adaptive learning and real-world decision-making.

The strategic direction is becoming difficult to miss.

Governments increasingly view AI not simply as a chatbot technology, but as the intelligence layer for future industries.

Where does India stand?

India enters this transition from an unusual position.

It is not currently leading the world in frontier foundation models.

It also does not yet possess China's robotics manufacturing scale.

But India has several assets that could become extremely valuable during the Vertical AI and Physical AI era:

a huge software engineering base,

deep domain knowledge in global industries,

a rapidly expanding manufacturing sector,

large digital public infrastructure,

a massive domestic market,

strong SaaS capability,

and one of the world's fastest-growing industrial automation markets.

The IndiaAI Mission has an approved outlay of around ₹10,372 crore.

By mid-2026, the government reported that the shared IndiaAI compute ecosystem had expanded beyond 45,000 GPUs, while 20 indigenous foundation-model proposals had been selected. India is simultaneously developing AIKosh datasets, application programmes, skills initiatives and responsible-AI frameworks.

The government has also identified 762 potential AI use cases across 62 ministries, alongside 58 AI Centres of Excellence and hundreds of Data & AI Labs.

This infrastructure is important.

But India's biggest opportunity may not ultimately lie in building another generic ChatGPT.

It may lie one or two layers above the foundation model.

India's opportunity is Vertical AI

India has decades of operational expertise serving industries around the world.

Banking.

Insurance.

Healthcare.

Telecom.

Travel.

Manufacturing.

Supply chain.

Retail.

ERP.

Accounting.

Government.

That expertise can become training data, workflow knowledge and product IP.

Imagine globally exported AI systems built from India for:

Banking: autonomous KYC, underwriting, compliance and fraud investigation.

Healthcare: diagnostics, clinical documentation, hospital operations and medical-image intelligence.

Manufacturing: quality inspection, predictive maintenance and autonomous production optimisation.

Agriculture: crop intelligence, pest identification, irrigation optimisation and autonomous farm equipment.

Travel: AI systems that understand GDS, NDC, hotel inventories, ticketing rules, disruption management and agency operations rather than simply recommending destinations.

Logistics: demand forecasting combined with autonomous warehousing and transportation.

Government: multilingual citizen-service AI built around Indian languages and public infrastructure.

IndiaAI is already supporting domain-specific smaller models covering areas including healthcare and other specialised applications.

This is where India's SaaS and services heritage could potentially evolve into something much more valuable:

India moving from supplying technology talent to owning industry-specific AI products.

Physical AI may be an even bigger opportunity for India

India's robotics position is stronger than many people realise.

The International Federation of Robotics reports that India installed almost 10,500 industrial robots in 2025, 15% more than the previous year.

India is now the sixth-largest market in the world for annual industrial robot installations.

Even more significantly, annual installations increased at an average rate of approximately 27% per year between 2020 and 2025.

Indian companies are emerging as well.

Addverb is developing warehouse automation, collaborative robots, quadrupeds and humanoid systems.

Ati Motors is already deploying autonomous industrial vehicles.

Newer companies are exploring humanoids, inspection robots, drones and autonomous systems.

Government-backed programmes including the National Mission on Interdisciplinary Cyber-Physical Systems have established 25 Technology Innovation Hubs, while the Ministry of Heavy Industries is supporting indigenous industrial robotics development and localisation of components.

India also has a Draft National Strategy on Robotics, and a government Technology Advisory Group met in February 2026 specifically to discuss a strategic roadmap for robotics in India.

There is momentum.

But there is also a gap.

Indian robotics startups raised only a fraction of the capital going into American and Chinese competitors. Research reported in June 2026 indicated Indian robotics startups raised less than 1% of the capital received by US robotics startups during 2025.

This needs attention.

Because Physical AI cannot be created through software expertise alone.

India needs expertise in:

actuators,

motors,

gearboxes,

sensors,

machine vision,

robotics controllers,

edge AI chips,

battery systems,

safety engineering,

mechatronics,

simulation,

and high-precision manufacturing.

In fact, government-backed Indian robotics programmes themselves acknowledge that many critical industrial robot components still have to be imported.

That represents both a weakness and an enormous industrial opportunity.

The most interesting opportunity is where Vertical AI meets Physical AI

The categories will not remain separate for long.

Imagine a manufacturing AI.

Today it analyses production data and recommends changes.

Tomorrow it controls the robots making those changes.

A warehouse AI today predicts inventory movement.

Tomorrow it instructs autonomous machines to physically reorganise the warehouse.

A hospital AI today schedules staff and analyses patient information.

Tomorrow it coordinates service robots transporting medicines and equipment.

An agricultural AI today detects crop disease.

Tomorrow it sends an autonomous machine directly to the affected plants.

A construction AI today analyses the BIM model and project schedule.

Tomorrow it coordinates autonomous equipment on the site.

This is the point at which AI becomes more than software.

It becomes an operating layer for an industry.

And that could eventually become one of the largest technology markets the world has ever created.

Five trends worth watching between now and 2030

The first is the rise of smaller domain models. The best AI for every task will not necessarily be the world's largest model. Cheaper, faster models trained or adapted for specific industries will increasingly handle specialised workflows.

The second is the transition from copilots to autonomous workflows. AI will increasingly be measured by transactions completed, claims processed, orders fulfilled and problems solved—not by how impressive its answers sound.

The third is simulation becoming a strategic asset. Physical AI cannot learn everything through expensive real-world experimentation. Digital twins and synthetic environments will allow millions of robot training scenarios to run before hardware enters a factory.

The fourth is the rise of AI at the edge. Robots, vehicles and machines frequently cannot wait for a distant cloud server before taking an action. More intelligence will therefore move directly onto devices.

And the fifth may be the most important:

proprietary real-world data becomes the new moat.

The Internet provided enormous amounts of text for training language models.

There is no equivalent Internet containing billions of perfectly labelled examples showing robots exactly how to manipulate every object, navigate every factory or recover from every physical mistake.

The organisations that operate large fleets of machines may therefore accumulate one of the most valuable datasets of the next decade:

action data.

What India should focus on now

India does not need to copy the US foundation-model race or China's robotics strategy exactly.

It has a different opportunity.

India can combine its strongest existing advantages:

Software + AI + Domain Expertise + Manufacturing + Digital Infrastructure.

That means prioritising several things.

First, build Vertical AI systems with global ambitions, not another generation of generic AI wrappers.

Second, convert India's factories, ports, hospitals, warehouses and farms into Physical AI testbeds where Indian companies can generate real-world training data.

Third, accelerate localisation of the robotics supply chain—particularly actuators, sensors, motors, controllers, edge systems and precision components.

Fourth, use government and large enterprises as early buyers, not merely providers of grants.

Fifth, develop engineers who understand AI and the physical world simultaneously: software engineers who understand robotics, and mechanical engineers who understand machine learning.

And finally, focus on problems where India's scale itself becomes an advantage.

Agriculture.

Manufacturing.

Healthcare.

Logistics.

Mobility.

Energy.

Defence.

Construction.

Public infrastructure.

These are not small application markets.

They are enormous physical systems waiting for intelligence.

The next AI race will look very different

The first phase of modern AI created machines that could generate.

The second is creating machines that can reason and act across digital systems.

The next phase will create intelligence that understands an industry deeply enough to operate parts of it.

And eventually, that intelligence will acquire hands, wheels, sensors and machines.

That is why Vertical AI and Physical AI should not be viewed as side categories of the current AI boom.

They may represent its next major chapters.

Foundation models will remain enormously important, but increasingly they may resemble infrastructure: the intelligence engines underneath thousands of specialised systems.

The value above that infrastructure will come from understanding the problem.

And the value beyond the screen will come from being able to act on it.

For India, the opportunity is particularly significant.

We spent decades building software for the world.

The next opportunity may be to build intelligent systems that run parts of the world's industries.

And after that?

Machines that can physically run them too.