Where the Next Decade Gets Built

Every few months someone posts a list of categories and everyone nods and nothing changes on Monday morning.

This is the longer version of that conversation. Thirteen directions where the ground is genuinely moving, each with the same treatment: what is actually happening, what the numbers say as of late 2026, how the money gets made, what the next five and ten years look like, and what to watch out for along the way.

A note before starting. This is not a closed list and it is not a set of permissions. There is no shortage of things to build. What has changed is where effort compounds. Put five years into a category with a real scarcity behind it and you end up with an asset. Put the same five years into something a model will do for free in eighteen months and you end up with a feature. The thirteen below are where I see effort compounding, and I have tried to be specific about why rather than asking anyone to take it on faith.

Some of these are stronger than their fans claim. A couple are weaker. I have said which is which.

First, the two mechanics running underneath all thirteen: what work now costs, and how things now get built.

The underlying shift: codified work is getting cheap, everything else is not

The clearest evidence is not a market forecast. It is payroll data.

Stanford’s Digital Economy Lab, working with ADP records covering millions of US workers, published an update in August 2026 to their “Canaries in the Coal Mine” paper. Workers aged 22 to 25 in the most AI-exposed occupations now sit about 19% below their peers in less-exposed fields. A year earlier that gap was 13%. It is widening, not closing.

Two details matter more than the headline. First, the decline comes from reduced hiring, not layoffs. Nobody is being fired. Positions are simply not being opened. Second, the effect concentrates in roles built on codified knowledge, the formal documented kind that can be written down in a manual. Roles built on tacit knowledge, the kind you get from doing the work for fifteen years, show flat or rising employment.

That is the whole thesis in one sentence. If the work can be written down, it is getting cheaper. If it cannot, it is getting more valuable.

Now put the money next to it. Menlo Ventures put enterprise generative AI spend at $37 billion in 2025, up from $11.5 billion in 2024. McKinsey finds 88% of organisations use AI in at least one function, but only 23% are scaling an agentic system in even one function, and roughly 6% qualify as high performers. Gartner’s best case has agentic AI reaching about 30% of enterprise application software revenue by 2035, over $450 billion, up from roughly 2% in 2025.

The gap between 88% using and 23% scaling is where most of the next decade’s companies get built.

And the size of the prize is not the software market. The global services economy runs at roughly $16 trillion a year. The entire software market is around $1 trillion. Software has spent forty years selling tools to people who do services. The bet now is on selling the service itself.

The second shift: the labs became the platform layer, and the build changed underneath everyone

The first shift is about what work costs. This one is about how things get made, and it moved faster than almost anyone planned for.

Software stopped being written and started being delegated

The JetBrains developer survey for May to July 2026 found 90% of professional developers using AI coding agents at work at least weekly, and 68% using them daily. Claude Code came out as the single most-used tool for 31% of developers. Codex went from 3% adoption in January 2026 to 16% by mid-year, roughly a five-fold jump, with awareness rising from 27% to 65% in the same window. GitHub Copilot, the category’s founder, slipped from 29% a year earlier to 21%.

The output numbers are more striking than the adoption numbers. Claude Code was authoring roughly 4% of all public GitHub commits by February 2026, around 135,000 a day, with a single-day peak of 326,000 in March. Roughly half of all code committed to GitHub in early 2026 was AI-generated or substantially AI-assisted.

The important part is not speed. It is the change in unit of work. In 2023 you asked a model to finish a line. In 2026 you hand it an issue, a migration, a test suite, a refactor across forty files, and you review the result. The bottleneck moved from typing to reviewing, and review capacity is now the real constraint on engineering throughput.

That constraint is measurable. Unreviewed AI-generated code carries about 23% higher bug density, and roughly 14.3% of AI-generated snippets contain security vulnerabilities against 9.1% for human-written code. Teams that shipped faster without rebuilding their review and testing discipline are already paying for it.

Prompt tricks became systems engineering

This is the quiet structural change and it maps directly onto the harness argument in section 5.

What now determines whether an agent works in production is context management, permission scoping, sandboxing, tool definitions, audit logging and cost controls. Not phrasing. Every serious agent stack in 2026 looks like infrastructure engineering with a model somewhere in the middle.

The interoperability layer settled faster than most standards ever do. The Model Context Protocol, originally from Anthropic, was donated to the Linux Foundation and has been adopted across Anthropic, OpenAI, Microsoft and Google. Skill definitions converged on a common file standard across multiple agent tools. Agent-to-agent and payment protocols followed, as covered in section 10. For the first time in this cycle there is a plug shape everyone agrees on, which means the integration work you do now is not thrown away when you change model vendors.

The labs are now four things at once

This is what makes the current moment hard to reason about.

They are your supplier, obviously. They are your competitor, because the feature you shipped last quarter may be in the platform next quarter. They are increasingly your landlord: OpenAI grew from 0.2 GW of compute in 2023 to about 1.9 GW by end-2025, with roughly seven Stargate sites in development and over $400 billion committed, plus its own inference silicon developed with Broadcom. Sarah Friar has put a gigawatt-scale site at around $50 billion fully loaded and about three years to build.

And they are now your investor. In December 2025 OpenAI took an equity stake in Thrive Holdings and agreed to embed research and engineering teams inside its portfolio companies. A model provider is now a shareholder in the services businesses being automated with its models.

Meanwhile the enterprise spend has redistributed fast. Menlo Ventures’ work puts Anthropic’s share of enterprise LLM spending at roughly 40% in early 2026, up from 24% in 2024. Revenue figures for both major labs vary widely across sources and months, so I would treat any specific number with caution, but the direction is not in dispute: this is now a multi-vendor enterprise market rather than a single-vendor one.

Platform risk, stated plainly

In June 2026 OpenAI announced it was winding down Agent Builder and Evals, with availability ending 30 November 2026, pointing developers to the Agents SDK instead. That is entirely normal platform behaviour and not a criticism. It is simply the thing every founder building on someone else’s abstraction should price in.

The lesson is specific. Build on the primitives, not the conveniences. Own your evaluation sets, your data, your customer relationship and your process definitions. Treat the vendor’s orchestration layer as replaceable, because it is.

Startups are being built structurally differently, and there is now real data on it

A Harvard Business School and INSEAD working paper by Hyunjin Kim and Rembrand Koning, published June 2026, analysed close to 50,000 Y Combinator and PitchBook-listed venture-backed startups founded between 2020 and 2024. Against non-AI startups in the same industry and cohort, AI-native firms are:

• 25% smaller in headcount
• 13% higher in engineer share
• Roughly 15% lower in both entry-level workers and managers
• Raising about 20% more capital per employee, with higher valuation per employee

Flatter, more senior, more technical. The middle layer of coordination shrank because the coordination itself got absorbed. Notably, these firms raise comparable funding and reach comparable valuations. They just do it with fewer people.

Gartner projects that by 2030, AI-native development platforms will lead 80% of organisations to evolve large software engineering teams into smaller AI-augmented ones, with 60% running smaller teams at scale as soon as 2029, up from 15% in 2026.

The financing pattern changed to match. Revenue per employee, which used to appear in Series B diligence, now shows up in seed conversations. Seed deal counts are down roughly a quarter year on year while individual rounds got larger, with more than half of seed dollars flowing into rounds of $10 million or more. Funding has gone K-shaped: a handful of very large AI rounds absorbed roughly two-thirds of global venture dollars in Q1 2026.

What this actually does to the thirteen

Three consequences, and they run through every category in this piece.

The codebase stopped being a moat. Building software was the barrier for thirty years. It is not any more. Which is precisely why the thirteen categories below are organised around scarcities that code cannot manufacture: power, atoms, permission, proprietary records, trust, accountability.

Small teams can now attack big categories. A four-person team with good tooling can credibly build a domain harness, a vertical agent, or the software layer for a services roll-up. What it cannot manufacture is the industry relationship, the licence, the data or the workforce. So the leverage sits with operators who understand a domain and now have engineering capacity they never had before.

Speed of building went up, speed of trust did not. Regulated approval, safety certification, clinical validation, grid interconnection, procurement cycles: none of these got faster. That gap between build speed and trust speed is where most of the durable value in the next decade will sit, because it is the one thing the tooling cannot compress.

1. AI-native service firms

What is happening

Investors stopped waiting for founders to build this and started buying the businesses themselves.

General Catalyst expanded its Creation Fund from $800 million to $1.5 billion. Thrive Capital set up Thrive Holdings with over $1 billion, and in December 2025 got OpenAI to take an equity stake and embed research and engineering teams directly inside the portfolio companies. 8VC, Lightspeed, Khosla and Bessemer are all writing cheques into the category. Total disclosed deployment is above $3 billion.

The results so far are real, if early:

• Long Lake, a homeowner-association management business incubated by General Catalyst, raised roughly $670 million and reached $100 million of EBITDA in under two years.
• Crescendo, an AI-native contact centre, hit a $500 million valuation with margins reported at roughly four times a traditional contact centre.
• Crete Professionals Alliance, rebranded to Current in June 2026, grew past $300 million in annual revenue across more than 30 accounting firms.
• Beacon raised a $225 million Series C in June 2026, taking it past half a billion in twelve months, while acquiring at roughly one company per week.

The mechanics

The arbitrage is specific. Buy a services business at 4 to 6 times EBITDA with 5% to 15% operating margins. Automate 30% to 70% of the repetitive work: intake, scheduling, document handling, reconciliation, first-line support, quality checks. Push margin toward 35%. You make money twice, once on the margin expansion and once on the multiple the market pays for a software-like margin profile.

The hard part is not the AI. It is that you have bought a company full of people whose jobs you are about to change. Integration debt is real. Culture is real. The technical build is maybe 20% of the work.

Five years

Consolidation in accounting, legal support, insurance broking, property management, IT managed services, dental and veterinary practices, medical billing, staffing. Expect several platforms above $1 billion in revenue. Expect at least two high-profile blow-ups where the automation did not land and the debt did.

Ten years

The category splits. Some of these become genuine operating companies with durable local moats. Others turn out to be leveraged roll-ups that used AI as the pitch and rate arbitrage as the actual engine. The differentiator will be whether the acquirer built one operating system across all acquisitions or just bolted a chatbot onto twelve different back offices.

What to watch

Circular financing. When the model provider takes equity in the holding company that buys the customers of the model, the reported success is partly an accounting artefact. Ask any of these platforms what their margins look like if they had to pay list price for inference.

2. Offline businesses

What is happening

Roughly 52.3% of US employer businesses are run by owners aged 55 or older. Estimates of the value about to change hands range from McKinsey’s $5 trillion in small business assets over a decade to broader figures around $10 trillion. In HVAC, plumbing and electrical specifically, 45% to 55% of owners are over 60.

Private equity has already moved. Nearly 800 HVAC, plumbing and electrical companies have been acquired by PE since 2022, per PitchBook data cited by the Wall Street Journal and the American Investment Council.

The honest counterpoint

The “silver tsunami will hand you cheap businesses” thesis has been wrong for fifteen straight years, and it is worth saying so plainly. The search fund acquisition rate fell from 86% for the 2007–2010 cohort to roughly 48% for funds launched between 2021 and 2024. It got harder to buy a business, not easier. Average owner retirement age moved from about 65 to about 71. Federal Reserve research suggests sale decisions track health and personal circumstances rather than age.

So the businesses are not flooding the market. The opportunity is in the ones that never get listed, and in operating them better than the last owner did.

There is also a plumbing problem in the literal sense of intermediation. The IBBA has around 2,900 members in the US. Total active advisers, including non-members, is maybe 10,000 to 15,000. The gap between businesses needing a structured exit and people qualified to run one is the widest in a generation.

Why offline holds up

An agent can schedule the technician, quote the job, order the part, chase the invoice and handle the follow-up. It cannot climb into the crawlspace. The physical constraint is the moat, and it is the only kind of moat that does not depreciate when the next model ships.

Five to ten years

Margins in trades businesses improve materially as dispatch, quoting, procurement and collections get automated. The winners will be regional operators with 10 to 50 locations, a single operating system, and real technician retention. Technician wages go up, not down, because the scarce input is the person with hands and a licence.

For India specifically

The demographic story is different. India does not have a retiring-owner wave of the same shape. What it has is extreme fragmentation in facility management, cold chain, diagnostics, equipment servicing, industrial maintenance, and construction subcontracting. The consolidation opportunity here is organising unorganised work, not buying out boomers.

3. Distribution: media, community, audience

What is happening

The creator economy is somewhere between $250 billion and $390 billion depending on whose definition you use, with forecasts running to $650 billion by 2030 and past $1 trillion by 2032. US creator advertising spend more than doubled in three years, from $13.9 billion in 2021 to $29.5 billion in 2024 per IAB.

Meanwhile, in a 2026 survey of 3,000 UK and US creators, 94% reported using AI in their content work.

Read those two facts together. The cost of producing content has collapsed. The supply is effectively infinite. Attention did not expand to match.

The counter-signal nobody expected

The backlash is measurable. iHeartMedia’s internal research found roughly 90% of its listeners want media made by humans, including listeners who use AI tools themselves. Research in the Journal of Business Research found that when people believe emotional marketing copy was written by AI, they judge it as less authentic and show weaker purchase intent, even when the text is otherwise identical. The Association of National Advertisers named two words of the year for 2026: “authenticity” and “agentic AI”, which is about as clean a summary of the moment as you will get.

The mechanics

When production cost goes to zero, the only scarce asset is permission. An email list, a WhatsApp community, a Telegram group, a paying membership, a podcast audience that shows up. These are assets because the audience has agreed to hear from you, and that agreement cannot be generated.

The second mechanic is agentic discovery. As buyers increasingly get answers from models instead of blue links, being the source a model cites becomes the new distribution. That is not keyword SEO. It is being sufficiently specific, structured and authoritative that a retrieval system has reason to pull you.

Five years

Owned audience becomes a balance sheet item in the way brand once was. Expect acquisitions of newsletters, communities and niche media by operating companies who want a demand channel that does not depend on an ad auction.

Ten years

Distribution and product merge. The companies that own the audience will manufacture into it rather than the reverse. This is already how several consumer categories work.

What to watch

Audiences are rented until they are portable. If your entire community lives inside a platform whose algorithm you do not control, you have influence, not distribution.

4. Proprietary datasets

What is happening

Licensed data has become a real market. Estimates put AI company spend on licensed data above $4 billion in 2025. The single largest publicly reported deal is News Corp with OpenAI, reported at around $250 million over five years. Reddit disclosed roughly $203 million in aggregate data licensing contract value around its IPO. Anthropic’s $1.5 billion settlement with authors received final court approval in July 2026, working out to roughly $3,000 per book across about 500,000 works, the largest copyright settlement in US history.

Two structural shifts are more interesting than the headline numbers.

Shift one: from dumps to feeds. Attribution and live-access deals went from 2 in 2023, to 11 in 2024, to 18 in 2025, with 34 projected for 2026. Labs increasingly want an ongoing grounding feed, not a one-time archive.

Shift two: operational data became an asset class. In August 2026, Google won Spirit Airlines’ de-identified enterprise data at a bankruptcy auction for $10 million, outbidding Mercor at $7.5 million. That is the first public price signal for a company’s operating history as a standalone AI asset.

Why operational data specifically

Agents need to learn from real work with real outcomes. Support tickets with resolutions. CRM records with won and lost reasons. Project histories with actual overruns. Maintenance logs with the failure that followed. Claims files with the settlement. None of this exists on the open web, and no amount of synthetic generation produces the messy causal structure of a real operation.

The mechanics

Three things determine whether your data is worth money:

• Scarcity. Can a lab reconstruct it from other sources? If yes, the price is zero.
• Provenance. No serious buyer signs in 2026 without a clean chain of rights. The EU AI Act requires general-purpose model developers to publish summaries of training content, which pushes the diligence upstream.
• Outcome labels. Raw records are cheap. Records paired with what happened next are expensive.

The practical implication for anyone operating a business today: instrument your workflow to capture outcomes, not just transactions. Most companies log what was done. Very few log what worked.

The uncomfortable truth

Most data holders will never get a deal. There is one Wall Street Journal and one Associated Press. A mid-size publisher or a mid-size company has neither scarcity nor leverage. For those owners, the realistic paths are collective licensing through marketplaces, or using the data internally to build something rather than selling it.

Five to ten years

Data marketplaces mature for the long tail. Sector-specific data cooperatives emerge in health, insurance, logistics and industrial maintenance, where no single player has enough volume but twenty together do. Pricing shifts from one-time to per-query grounding fees.

5. Domain-specific harnesses

This is the least understood item on the list and possibly the most valuable.

What a harness actually is

A harness is not the model and not the agent. It is the environment the agent runs inside: the tool definitions, the permission scopes, the memory and retrieval layer, the evaluation set, the escalation rules, the human approval gates, the audit trail, and the rollback path when something goes wrong.

Why it matters

The 2026 Stanford AI Index reports agent task success on the OSWorld benchmark jumping from about 12% to roughly 66%, with real-world task success around 77.3%. And yet autonomous agent deployment across business functions remains in single digits.

That gap is not a model problem. It is a harness problem. A 77% success rate is excellent in a benchmark and unacceptable in accounts payable. What converts one to the other is the machinery around the model: knowing which 23% to catch, routing it to a human, and learning from the correction.

IDC predicts a 15% productivity loss by 2027 for companies that fail to establish AI-ready data foundations. That is the same observation from the cost side.

The technical state of play

MCP has become the default connectivity layer between agents and enterprise systems, now governed by the vendor-neutral Agentic AI Foundation. Agent SDKs from the major labs ship as production infrastructure with memory, tool use and orchestration built in. The field has moved from prompt engineering to context engineering, which is a plain way of saying the hard part is deciding what the model sees, not how you phrase the request.

The business

You own the process definition, the eval set and the exception queue for one industry’s core work. Bridge inspection reporting. Clinical prior authorisation. Trade finance document checking. Mining haul plan compliance. Reinsurance treaty review.

The commercial beauty of this position is that model progress helps you and does not threaten you. When a better model ships, your harness gets more accurate and your margins improve. You swap the engine; you keep the chassis, the road and the driver’s licence.

Five to ten years

Harnesses become the real switching cost in enterprise AI. Ten years out, buying a harness for your industry looks like buying an ERP did in 2005: expensive, sticky, and impossible to rip out without a two-year programme.

6. Robotics and physical AI

What is happening

2025 was the first year humanoid robotics produced verifiable commercial results rather than demo videos.

Global humanoid shipments reached roughly 18,000 units in 2025, up an estimated 508%, with Chinese manufacturers accounting for close to 90% of volume. Market revenue was about $4.9 billion in 2025 and roughly $6.2 billion in 2026, which is small. Funding was not: robotics startups raised about $14 billion in 2025 per Crunchbase, the highest year on record, and Dealroom tracked $8.7 billion of humanoid-specific venture funding through July 2026.

The large rounds: Figure AI raised over $1 billion at a $39 billion valuation in September 2025, Skild AI $1.4 billion in January 2026, NEURA Robotics up to $1.4 billion in June 2026, Apptronik $935 million in total. Mobileye agreed to acquire Mentee Robotics for $900 million in January 2026. Agility Robotics went the SPAC route with Churchill Capital XI in June 2026 for over $620 million in expected proceeds.

The one deployment result worth studying

Figure’s pilot at BMW Spartanburg: more than 30,000 vehicles produced, over 1,250 operating hours, more than 90,000 parts handled over ten months. Figure’s BotQ facility had delivered over 350 Figure 03 units by January 2026 and moved from one robot per day to one per hour.

That is a real manufacturing ramp. It is also, in absolute terms, tiny. Pilot deployments are still 10 to 50 units per major customer.

Where the money actually is

Not in humanoids. Robotic foundation models and general-purpose robots together took 77.6% of disclosed physical AI capital over the twelve months to July 2026. Humanoid form factors took $1.49 billion across five deals.

And underneath both sits a supply chain almost nobody is funding at the same intensity: harmonic drives and cycloidal reducers, high-torque-density actuators, force and tactile sensing, battery packs sized for duty cycles rather than range, teleoperation rigs for data collection, safety certification, and field service networks.

Price points tell you where this goes. Unitree’s R1 is priced at $5,900 for pre-order and the G1 at $16,000. Industrial humanoids sit in the $80,000 to $250,000 band. Most enterprise deployments use robotics-as-a-service pricing, which means the real comparison is cost per shift against human labour, not sticker price.

China’s MIIT humanoid action plan targets 100,000 units deployed by 2027, which would exceed the rest of the world’s installed base combined.

Five years

Structured, single-task deployments scale in automotive, warehousing and logistics. Wheeled and stationary manipulators outnumber bipeds by a wide margin because bipedalism is expensive and usually unnecessary. Data collection becomes a business in its own right.

Ten years

Either general-purpose manipulation works and this becomes one of the largest industries on earth, or it plateaus at task-specific automation and the humanoid valuations look like 2021 in hindsight. Both outcomes are live. The honest answer is that the bottleneck is manipulation reliability in unstructured environments, and nobody has solved it yet.

For India

The opportunity is not building the robot. It is components, contract manufacturing, motion control software, integration and service. India already does precision machining and electronics assembly at scale. Actuator and reducer manufacturing is a natural adjacency, and it is currently a near-monopoly held by a handful of Japanese and Chinese suppliers.

7. Physical products with a fan following

What is happening

Two forces are colliding.

First, generic products are about to get price-shopped by software. When an agent is told to buy trail-running shoes under a set price arriving by Friday, unbranded commodity goods compete on a spec sheet. Margin compression is the obvious outcome.

Second, consumers are pulling hard in the opposite direction on anything that feels human. Razorfish research found 82% of consumers would prefer a human customer service representative over an AI chatbot, and that exclusivity, early access and VIP treatment now outperform discounts and points as loyalty drivers. Lippincott’s 2026 trend work frames the opening precisely: treat physical goods not as commodities but as collectible moments, crafted for display, social currency and emotional resonance.

The mechanics

The loop is simple and hard to execute. Gross margin funds community. Community lowers customer acquisition cost. Low CAC funds better product. Better product deepens community.

The operational levers are drops and scarcity, membership tiers, involving the community in product decisions the way LEGO Ideas does, repair and resale programmes that extend the relationship past the transaction, and physical spaces that turn the brand into a destination.

Why this is the actual defence against agentic commerce

An agent asked for “shoes” runs a comparison. An agent asked for your brand by name runs a purchase. Brand demand is the only instruction that survives intermediation. Everything else is a spec sheet.

Five to ten years

Expect a barbell. Very large commodity manufacturers optimised for agent-mediated procurement at low margin, and a long tail of high-margin brands with real communities. The middle, which is where most consumer brands live today, gets squeezed hardest.

8. Compute and energy

What is happening

This is the largest private capital deployment of our lifetimes, and the constraint has moved from chips to electricity.

The four US hyperscalers guided to as much as $630 billion of capital expenditure in 2026, against roughly $388 billion in 2025. Amazon alone signalled around $200 billion, Google $175 to $185 billion. BloombergNEF puts capex for the 14 largest publicly owned data centre operators near $750 billion in 2026 against just under $450 billion the prior year. Dell’Oro has total data centre capex on pace to exceed $1 trillion in 2026.

Over 23 GW of data centre capacity was under construction globally at the end of September 2025, about three quarters of it in the US.

The physics that decides everything

A conventional data centre rack draws 10 to 15 kW. An AI training rack with current accelerators draws 40 to 70 kW, and next-generation configurations push toward 100 kW. Facilities are shrinking in footprint while exploding in electrical load. Land is cheap. Megawatts are not.

The US Department of Energy projects the grid needs roughly 100 GW of new capacity by 2030, about half of it data-centre driven. BCG estimates a US data centre power shortfall that could exceed 45 GW by 2030. Deloitte projects data centre power demand reaching 176 GW by 2035, about five times 2024 levels.

The neocloud layer

CoreWeave reported $5.13 billion of 2025 revenue with a backlog that ballooned past $66 billion and has since been reported near $100 billion, on roughly 250,000 deployed GPUs. Meta’s total commitment to CoreWeave reached approximately $35 billion through 2032. Nebius signed a $17.4 billion GPU-as-a-service contract with Microsoft and a separate deal with Meta reported at $27 billion. Microsoft has committed over $60 billion across Nscale, Nebius, CoreWeave, IREN and Lambda, effectively renting capacity rather than building all of it.

Both CoreWeave and Nebius have each contracted roughly 3.5 GW of power. The vast majority is not yet energised. CoreWeave is targeting 1.7 GW active by end-2026, Nebius 800 MW to 1 GW. The gap between contracted power and live power is where this entire category’s risk sits.

Where the durable businesses are

Not in reselling GPUs. In the inputs that are physically scarce:

• Grid interconnection queue positions and land with existing substation capacity
• Transformers, switchgear, medium-voltage equipment (lead times here are the real constraint)
• Liquid cooling, CDUs, thermal engineering, commissioning
• Behind-the-meter generation, PPAs, storage
• Long-duration operations and maintenance contracts

These are unglamorous and they are where the margins will persist after the GPU cycle normalises.

The India picture, concretely

India’s operational data centre stock reached roughly 1.6 to 1.8 GW by mid-2026. H1 2026 added 258 MW of new IT capacity, up 59% year on year per Savills. Vacancy hit a record low of 2.8%. JLL projects capacity reaching 6 GW by 2029, requiring around $110 billion of capital. CBRE has put cumulative investment commitments above $126 billion, with $56.4 billion in 2025 alone.

Policy moved too. Budget 2026–27 introduced a 20-year tax holiday running to 2047 for foreign cloud providers, plus a 15% cost-based safe harbour margin for Indian data centre entities serving related foreign companies. On the power side, India is expanding nuclear capacity from 8.78 GW toward 22.38 GW by 2031–32, with a stated target of 100 GW by 2047, against a 500 GW renewable target before 2030.

Growth is concentrated in Mumbai, Chennai, Hyderabad, Pune, Noida and Visakhapatnam. Chennai’s advantage is eight subsea cable landings. The binding constraints are the same everywhere: power evacuation, water, and how fast a state can actually energise a connection.

What to watch

Circular financing. Chip vendors investing in the clouds that buy their chips, labs taking equity in the infrastructure they rent. GPU depreciation schedules assume a useful life that may not survive two more hardware generations. Neither CoreWeave nor Nebius is profitable, and both fund expansion with debt and dilution. If inference demand grows more slowly than the contracted build, the losses land on whoever holds the lease.

9. Health, longevity and care

What is happening

This is the largest category on the list by absolute market size and the least dependent on any AI thesis being right.

WHO projects that by 2030, one in six people worldwide will be aged 60 or over. The global long-term care market sits at roughly $1.2 to $1.27 trillion. Broader elderly care definitions put the market above $2 trillion in 2026. Longevity clinics specifically are a smaller but fast-growing $6.02 billion in 2026, projected to $9.55 billion by 2030.

And the supply side is in trouble. Turnover in long-term care facilities exceeds 50% annually in some regions. Caregiver shortages are the most-cited operational constraint across every survey in the category.

Where AI actually lands

Documentation, first. Clinical documentation agents reduce documentation time by 30% to 42% and save clinicians up to 66 minutes per day. That is not a marginal gain, it is roughly an extra hour of clinical capacity per clinician per day in a system where clinician hours are the binding constraint. Abridge reached a $5.3 billion valuation on essentially this.

But the money in this category is not in another documentation app. It is in care delivery capacity. Home care operations. Diagnostics throughput. Chronic disease panels managed at scale. Rehabilitation. Staffing and training pipelines.

Five years

Care shifts decisively toward the home because facilities cannot staff themselves. Remote monitoring becomes standard for chronic conditions. Reimbursement models catch up unevenly and that lag is where operators make or lose money.

Ten years

Prevention gets paid for. Today almost all health spending is triggered by illness. As biomarker panels get cheap and longitudinal data accumulates, risk-based pricing for healthspan becomes commercially viable. This is a slow-moving change that will look sudden in retrospect.

For India

The gap is enormous and structural. Ageing plus nuclear families plus almost no formal elder care infrastructure plus very low insurance penetration for long-term care. There is no equivalent of Home Instead operating at national scale. Diagnostics is consolidating fast; care delivery is not. Anyone who can build a trustworthy, trained, retained caregiver workforce with real supervision has a decade of runway.

10. Marketplaces and social networks, for people and agents

What is happening

The plumbing for agents to transact got built in about eighteen months.

• x402 (Coinbase) revives the HTTP 402 status code for stablecoin payments over HTTP. As of May 2026 the Dune dashboard attributed roughly 130 million all-time transactions, dominated by Base and Solana and settled almost entirely in USDC, with annualised volume estimated around $600 million as of Q1 2026. Stripe integrated it on Base in February 2026; Cloudflare and AWS Bedrock support it.
• ACP (OpenAI and Stripe) standardises agent-to-merchant checkout using shared payment tokens bound to a specific merchant, amount and time window, single use.
• AP2 (Google, 60-plus partners) handles authorisation with cryptographically signed mandates proving a human approved a specific transaction intent.
• TAP (Visa, launched October 2025 with Cloudflare) signs agent identity into HTTP headers for merchant verification. Mastercard’s Agent Pay works similarly.
• MPP (Stripe and Tempo, launched March 2026) uses a sessions model where an agent pre-authorises a spending limit and streams micropayments.

Demand is showing up on the retail side. Adobe Analytics measured a 4,700% year-on-year jump in generative-AI referral traffic to US retail sites between July 2024 and July 2025.

The business

Marketplaces have always made money on liquidity, trust and dispute resolution. Agent-era marketplaces need three additional things: machine-readable inventory and structured product data, verifiable agent identity, and refund and dispute semantics that work when no human was present at purchase.

Whoever solves trust for agent transactions in a specific vertical, freight capacity, industrial spares, API services, compute itself, gets to charge a toll.

The risks are already documented

Security researchers have shown x402’s metadata fields travel in plaintext to the payment server and the facilitator before settlement, with no sanitisation, and can carry personal data. Free-riding attacks against deployed configurations have been demonstrated. Fraud liability rules built around a human tapping a phone do not map cleanly onto software buying at 3am.

Human networks, separately

People networks do not disappear. They get more valuable, because verification and belonging cannot be synthesised. The growth areas are small, verified, high-context networks: professional guilds, alumni groups, industry-specific communities with real identity checks. The open social web with unverified accounts is heading toward being unusable as a discovery layer.

Five to ten years

Protocol consolidation, probably around two or three winners at different layers rather than one. Agent-to-agent procurement becomes normal in B2B before it becomes normal in consumer. The consumer version arrives when liability is settled, not when the technology is ready.

11. Real assets

What is happening

McKinsey estimates a cumulative $106 trillion of infrastructure investment needed through 2040, with energy and power alone requiring around $23 trillion. PwC’s longer-horizon number is $151.1 trillion through 2050. The G20 Global Infrastructure Hub has long estimated a roughly $15 trillion gap between projected and required investment by 2040.

Private capital is stepping in. Global infrastructure fundraising reached a record near $200 billion in 2025, surpassing the previous 2022 high. Annual data centre investment specifically will reach around $215 billion in 2026, more than double 2024 levels. Asia-Pacific accounts for more than half of global infrastructure spending.

Why this belongs on a list about AI

Because AI turned out to be a real-assets business. Every frontier model needs land with power, fibre routes, water rights, cooling, and somebody to operate it for twenty years. The abstraction layer sits on concrete.

Beyond data centres: industrial sheds near consumption centres, cold chain, logistics parks, grid-adjacent land, transmission corridors, water treatment, and the long tail of equipment that gets rented rather than bought.

The mechanics

These are cash-yielding, inflation-linked, and structurally hard to replicate because the scarce input is permission. A competitor with the same capital cannot conjure an interconnection agreement or a right of way. That is a different kind of moat from a software feature, and a much more durable one.

Five to ten years

Returns compress as capital floods in, which is what always happens. The alpha moves from owning assets to developing them, meaning the ability to take land through permitting, power and construction to an operating asset. That is an execution capability, not a capital capability.

For India

This is the most direct opportunity on the list for anyone already in infrastructure. The national build across highways, metros, water, power evacuation and industrial corridors is running simultaneously with a data centre build requiring $110 billion by 2029. The scarce skill is not capital. It is land aggregation, statutory clearance, power tie-up and delivery discipline, which are exactly the capabilities Indian infrastructure firms already have and rarely monetise as a product.

12. Vertical agents

What is happening

Gartner expects 40% of enterprise applications to ship with task-specific AI agents by the end of 2026, up from under 5% in 2025. Domain-specific agents in BFSI, healthcare, legal and engineering are the fastest-growing segment, with one forecast putting the growth rate at about 62.7% annually through 2030.

The revenue is real and concentrated:

• Harvey crossed $300 million ARR in legal
• Sierra crossed $150 million in customer support
• Abridge reached a $5.3 billion valuation in healthcare documentation
• EvenUp doubled to roughly $2 billion in personal injury law
• Salesforce Agentforce reached approximately $800 million ARR, up 169% year on year per FY2026 updates, though still a small fraction of Salesforce’s $30 billion-plus total

Where they work and where they do not

They work in domains with structured workflows, repeatable decisions, clear compliance boundaries and enough transaction volume to justify the build. BFSI leads precisely because compliance forced data discipline years ago. The constraint that slowed cloud adoption in banking now accelerates agent ROI, because governed data is what agents need to function.

They fail where there is no evaluation set, no exception handling and no data foundation. Forrester predicts enterprises will defer roughly 25% of planned AI spending into 2027 as financial rigour kills proofs of concept that never had a measurement plan.

The pricing shift

Seats are dying as a pricing unit for work that agents do. Usage-based pricing is mainstream. Outcome-based pricing is where buyers say they want to go, and the practical landing zone is hybrid: a base fee covering a volume of tasks, then a per-task or per-outcome rate above it. Something like a monthly platform fee including a task allowance, with metered overage, is what most contracts actually look like once legal has been through them.

Five years

Vertical agents eat the workflow layer of vertical SaaS. The a16z framing of a 10x expansion in addressable revenue per customer is aggressive but directionally right, because you are no longer selling a tool to a team, you are selling the team’s output.

Ten years

Consolidation. Most vertical agent companies get acquired by the incumbent whose workflow they automated, because the incumbent owns the system of record and the customer relationship. The exceptions will be those that became the system of record themselves.

13. Security

What is happening

Gartner’s 2Q26 forecast puts worldwide information security spending at $248.9 billion in 2026, up 12.7% in constant currency, reaching $372.6 billion by 2030. Within that, AI-amplified security goes from around $49 billion in 2026 to $204 billion by 2030, and it is the only accelerating curve in the entire forecast.

Securing AI itself, as distinct from using AI to secure other things, is forecast to reach almost $4.8 billion in 2027. AI application security is the largest segment at around $851 million, with AI usage control at $749 million growing fastest at 73%, and AI gateways at 70.9%.

The imbalance is the story: enterprises are spending roughly 17 times more on AI-powered security tools than on securing the AI those tools depend on.

The specific threat surface

Gartner predicts that by 2029, over half of successful cyberattacks on AI agents will exploit access control weaknesses and prompt injection.

Identity is the centre of it. CyberArk’s research puts the ratio at roughly 82 machine identities for every human one, many holding privileged access nobody monitors. And agent identities are not like traditional service accounts. A service account is provisioned once with fixed scope and predictable behaviour. An agent acquires credentials at runtime based on what it decides it needs mid-task, orchestrates other agents creating access paths nobody designed, inherits entitlements from whoever invoked it, and behaves differently between runs. “Normal” becomes a moving target, which breaks the entire anomaly-detection model.

The architectural response is short-lived, task-scoped credentials replacing long-lived ones, plus policy-driven authorisation for machine actors. Gartner lists IAM adaptation for AI agents as a top-six 2026 trend.

Separately, post-quantum cryptography moved from research to procurement. Forrester expects quantum security to exceed 5% of IT security budgets in 2026, which means cryptographic discovery, migration planning and algorithm testing are billable work starting now.

Five to ten years

Agent identity and authorisation becomes a category as large as human IAM is today. Runtime protection for agents, AI gateways and usage control consolidate into platforms. Post-quantum migration becomes a decade-long services business with a hard 2030 deadline behind it.

The common thread: six scarcities worth building against

Strip away the sector labels and thirteen categories collapse into six scarcities. This is the useful part of the list, because these six generate far more than thirteen businesses. Any combination of them is a starting point.

• Power and land (compute and energy, real assets)
• Atoms and hands (offline businesses, robotics, physical products, care)
• Permission (distribution, communities, licences, interconnection agreements)
• Proprietary records (datasets, operational history, outcome labels)
• Trust and the right to act (security, marketplaces, agent identity)
• Accountability (harnesses, vertical agents, anything where somebody must sign off)

Every one of these is expensive to acquire and slow to copy. Which is exactly why they are worth the years it takes.

The practical version is a single question worth asking about anything you are building:

Would a competent competitor with the same model, the same money and six months be able to replicate this?

If yes, you have built a feature, and the right move is to keep going until you have built the thing underneath it. If no, you have built an asset. The question is not a filter for rejecting ideas. It is a test for knowing how much further you need to go.

What is repricing, and where that effort should move instead

Nothing in this section is a dead end. Each of these is effort that used to compound and now does not, which means the same energy redirected slightly produces much more.

• Thin wrappers. A prompt and a UI gets shipped as a platform feature or cloned in a fortnight. The same team, pointed at the harness underneath (the evals, the permissions, the exception handling for one industry), builds something nobody can clone.
• Seat-based pricing. You cannot sell 50 seats to a team that will be 12 people. Move the unit to tasks or outcomes and the revenue per customer usually goes up, not down.
• Traffic without an owned audience. Search-sourced traffic monetised by display ads is structurally impaired. The same content effort, pointed at a list, a community or a membership, becomes a durable asset.
• Services priced by headcount-hours. Clients now know roughly what the work costs to produce. Price the outcome instead and the margin conversation changes completely.
• Selling data that is already public. Aggregated public data has no licensing value. Operational data with outcome labels does, and most companies are already generating it and discarding it.
• Pilots without evaluation sets. Roughly a quarter of planned enterprise AI spend is being deferred because pilots could not prove anything. Build the measurement first and you become the vendor that survives the budget review.

Where to put the next three years

A few things follow from the data rather than from opinion.

Start where you already have unfair access. These thirteen categories all reward incumbency of some kind: an existing customer base, an operating history worth licensing, land, licences, a technician workforce, regulatory relationships. Very few of them reward starting cold. If you are inside an industry today, your advantage is the process knowledge and the data exhaust you are probably throwing away.

Instrument outcomes now, even if you have no monetisation plan for the data. The Spirit Airlines auction priced one company’s operational history at $10 million in a bankruptcy. Yours is worth nothing until it is structured, consented and paired with results. That work takes two years and cannot be backfilled.

Assume the model layer is not where you win. It commoditises on a roughly annual cycle. Build so that a better model makes you more profitable rather than less necessary.

[output truncated at 50000 of 52100 characters. Pass a larger max_chars (default 50000) to see more, or use read_page with a ref_id to focus on a smaller section.]

Pick two, not thirteen. The interesting companies of the next decade will sit at intersections rather than inside single categories. A proprietary dataset plus a vertical agent. Compute plus a real-assets play. An offline services network plus a distribution channel it owns. Care delivery plus vertical agents. Single-category businesses are getting crowded; the intersections are almost empty.

The short version

More is being built now than at any point in my working life. The build-out running through data centres, grids, robots, care systems and industrial capacity is the largest capital deployment most of us will see, and the overwhelming majority of it needs people who understand a specific domain deeply, not people who understand models.

That is the encouraging part of the data. Tacit knowledge went up in value, not down. Ten or twenty years spent learning how an industry actually works just became the scarce input, provided it gets combined with the new tooling rather than defended against it.

The window on all thirteen is open, and on most of them it stays open for a decade. The only thing that closes quickly is the part where you were early.

Data in this piece is drawn from Stanford HAI’s 2026 AI Index and Digital Economy Lab, the Kim and Koning AI-Native Firms working paper (HBS/INSEAD, 2026), Gartner, Forrester, IDC, McKinsey, Menlo Ventures, JetBrains’ developer survey, BloombergNEF, Dell’Oro, PitchBook, Crunchbase, Dealroom, CBRE, JLL, Savills, IAB, WHO and company disclosures, as reported through September 2026. Market sizing figures vary widely by definition and source; where estimates conflicted I have said so rather than picking the most flattering number.