Return on Intelligence, Part 8: The New Power Map

Part 8: The New Power Map

This final chapter gathers the whole series into one question: after the bubble breaks, which layers actually own the durable value? If you are arriving here first, the cleanest entry point is still the preface, followed by Part 1.

The market is looking for the next software company, but it should be looking for the next form of the company itself.

That is the deepest point of this series.

The AI bubble is not just a question of whether Nvidia is overvalued, whether OpenAI can justify its infrastructure plans, whether Anthropic can grow into its valuation, whether Microsoft can monetize Copilot, whether Salesforce can defend CRM, or whether every SaaS company can add enough AI features to keep customers paying more.

Those questions matter.

But the bigger question is where power moves when intelligence becomes cheaper, software becomes more fluid and organizations can operate with less human coordination.

The answer may not be the obvious one.

The mature AI economy may not be owned primarily by the companies that made AI famous.

It may be owned by the companies that control compute, devices, identity, data, governance, workflow and execution.

The old map

The current map is still shaped by the software and cloud era.

• At the bottom are chips, data centres, networking, storage and cloud infrastructure.
• Above that are model providers and platform services.
• Above that are SaaS applications.
• Above that are business workflows, users and customers.

In this map, value is often captured by packaged software and cloud platforms. Companies buy applications. Employees use seats. Workflows live inside SaaS products. Data lives in systems of record. Integration connects the stack. The model provider appears as a new intelligence layer inside or beside the existing stack.

This is why the first AI story is so tempting.

Just add AI to the map.

Add models to the cloud. Add copilots to SaaS. Add agents to workflows. Add usage pricing. Add more data centres. Add more chips. Add more automation.

But paradigm shifts do not merely add a layer.

They redraw the map.

The new map

The AI-native map looks different.

• At the bottom are still chips, memory, power, networking and data centres. Physical compute matters. Energy matters. Supply chains matter.
• Above that come devices and operating systems, because much intelligence moves closer to the user, the context and the workflow. Then come models, plural: frontier models, local models, open models, specialized models, private models and embedded models. Then come identity, permissions, policy and governance, because intelligence without control is operationally dangerous.
• Above those sit data and systems of record, because AI is only useful when it can see the right context and act on trusted information. Then come orchestration layers, durable workflows, tool catalogs, API gateways, observability, evaluation and human approval gates. Then come generated workflows and generated software. Then come customer-facing services, internal operations and AI-operated companies.

In this map, the model is important, but it is not the whole system.

The durable power sits where intelligence meets authority.

• Who can see the data?
• Who can decide which model runs?
• Who can grant permission?
• Who can take action?
• Who can observe the result?
• Who can audit the decision?
• Who owns the customer relationship?
• Who owns the workflow?

Those are the questions that define the mature AI economy.

Likely winner: chips, memory, packaging and power-efficient compute

The obvious winner group is still compute.

AI increases demand for specialized processing, memory bandwidth, packaging, networking and power efficiency. Even if centralized API economics weaken, compute does not stop mattering.

In fact, a more distributed AI world may increase the number of places where compute matters:

• frontier training;
• cloud inference;
• enterprise inference;
• local inference;
• edge devices;
• phones;
• laptops;
• cars;
• robots;
• industrial systems;
• data centre networking;
• memory and storage.

The specific winners may shift. Training-heavy demand is not the same as inference-heavy demand. Cloud GPUs are not the same as device NPUs. Memory bottlenecks are not the same as model architecture. Power efficiency becomes more important as AI becomes ambient, but the compute substrate remains strategic.

The caution is valuation and timing.

A company can be structurally important and still overvalued at the wrong point in the cycle.

That was true in dotcom. It will be true in AI.

Likely winner: devices and operating systems

If intelligence moves closer to the user, devices and operating systems become more powerful.

This is why Apple, Microsoft, Google, Android, Samsung and other platform owners matter.

• The device has the context.
• The operating system has the permissions.
• The browser has the activity.
• The phone has the personal life.
• The laptop has the work.

The calendar, inbox, files, photos, messages, contacts, identity, location and app permissions all live close to the user.

A useful mainstream signal here is Apple Intelligence, which is explicitly designed around on-device processing plus Private Cloud Compute. Even if Apple’s particular implementation changes, the architectural direction matters: more intelligence moves closer to the user, the operating system and the permission boundary.

• A remote model API has intelligence.
• The device has life.
• That may be the more durable position.
• The mature AI user may not open a chatbot to ask for help. The help may already be inside the thing they are doing.
• That gives platform owners a strong position.
• They can route models. They can bundle features. They can privilege defaults. They can integrate privacy controls. They can decide which AI capabilities become part of the environment.
• This does not make them unbeatable.

But it gives them a structural advantage over pure model providers.

Likely winner: systems of record

AI needs trusted context and that makes systems of record more important, not less.

Finance records. Customer records. Contracts. Orders. HR records. Identity. Product data. Risk data. Compliance records. Audit trails. Inventory. Usage history. Support history. Pricing rules. Legal obligations.

If AI is going to act, it needs to act on trusted data. The companies that own or govern those records have leverage.

This is why not all SaaS is equally exposed.

• A system of record can become more valuable if it becomes the trusted source for AI-operated workflows.
• A generic workflow layer around a system of record is more vulnerable and the distinction matters.

The mature AI economy may hollow out parts of SaaS while strengthening certain core platforms.

The record survives and the workflow around the record becomes fluid.

Likely winner: identity, permissions and governance

AI without governance is a liability.

As AI systems move from answering questions to taking actions, permissions become central.

• Who is the AI acting for?
• What data can it access?
• What actions can it perform?
• Which systems can it write to?
• Which decisions require human approval?
• How is the action audited?
• How is responsibility assigned?
• How are failures detected?
• How are generated workflows tested?
• How is sensitive data prevented from leaking into the wrong model?

Microsoft is already productizing this layer directly through Copilot Control System, which it describes as a way to secure, govern, manage and measure AI at enterprise scale. That does not prove Microsoft owns the mature AI economy. It does reinforce the broader claim that governance becomes a real product layer, not just an afterthought.

This makes identity and governance strategic.

The mature AI enterprise will need strong control planes. Not just model access. Not just chat interfaces. Not just prompt libraries.

Control planes.

The companies that provide identity, policy, audit, permissions, workflow governance, model routing, evaluation and observability may capture durable value.

This is one reason I think the model API moat is weaker than people assume.

The model can answer.

The governance layer decides whether the answer can become action and action is where business value lives.

Likely winner: workflow orchestration and durable execution

AI agents are unreliable if they are just loops of prompts and tools.

Real business processes need durability, retries, auditability, state, compensation, human approval, error handling, observability and long-running execution.

This makes workflow orchestration important.

An AI-operated company or AI-native enterprise does not simply ask a model to do everything. It needs processes that can survive failure, wait for external events, retry safely, call tools, involve humans, record decisions and produce auditable outcomes.

That is not a chatbot problem. It is an execution problem.

The mature AI economy will need orchestration layers that sit between intent and action.

Those layers may become some of the most important infrastructure in the enterprise.

They will decide how AI work is broken into steps, which tools are called, what state is stored, what humans approve, what happens on failure and how outcomes are measured.

This is where AI becomes operations.

Likely winner: companies with proprietary process knowledge

The best data is not always the largest dataset. Sometimes it is the most specific.

A company that deeply understands its own customers, pricing, risks, workflows, exceptions, supply chain, sales process, product constraints and regulatory environment has an advantage if it can encode that knowledge into AI-operated systems.

This is why domain expertise becomes more valuable, not less.

AI lowers the cost of producing software and output, but it does not automatically know what should exist.

The scarce knowledge becomes:

• what matters;
• what good looks like;
• what can go wrong;
• what exceptions exist;
• what rules are real;
• what customers actually need;
• what risk is acceptable;
• what trade-offs are tolerable.

Companies that can turn proprietary process knowledge into governed AI systems will be stronger than companies that merely buy generic AI features.

This is a hopeful part of the story.

AI does not only advantage the largest platform companies. It can also advantage companies that know their business deeply and can express that knowledge in software, data and workflows.

Pressured group: pure model API companies without distribution

The model labs are important, but pure model API economics are vulnerable.

• If models become cheaper, smaller, local, open and specialized, then centralized API access loses some scarcity value.
• If devices and operating systems absorb everyday AI tasks, model labs lose user surface.
• If enterprises build private routing and governance layers, model labs become suppliers.
• If frontier capability remains expensive, only the hardest tasks justify premium pricing.

The strongest model companies will try to become platforms and they will build consumer products, enterprise tooling, developer ecosystems, agent frameworks, marketplaces, infrastructure partnerships and maybe devices.

Some will succeed, but the weaker ones, especially those without distribution, will struggle. The market currently treats model quality as if it can become durable ownership.

That is the questionable assumption.

Pressured group: thin AI wrappers

Thin AI wrappers are the easiest casualties. A thin wrapper takes a model API, adds a narrow workflow and calls it a company.

Some wrappers are useful. Some will become real businesses by owning data, workflow, distribution or domain trust, but many will be features, not companies.

They will be copied by model providers, bundled into SaaS platforms, replaced by open source, generated internally by customers or undercut by broader tools.

This is normal in a platform shift.

The first wave produces many plausible ideas that do not have durable company-level economics.

When funding tightens, wrappers get exposed.

Pressured group: generic workflow SaaS

SaaS is not dead.

But generic workflow SaaS is exposed.

• If the product is mainly forms, dashboards, routing, approvals, notifications, lightweight records and integrations, AI-generated internal software becomes a threat.
• If the pricing model depends on seat expansion, AI-driven labour compression becomes a threat.
• If the AI features are not differentiated, they become table stakes.
• If the product does not own the system of record, the data, the compliance model, the ecosystem or the customer relationship, it is easier to replace.

The weak SaaS company of the AI era is one that sells a generic workflow and calls it a platform.

The strong SaaS company owns a record, a network, a regulated function, a deep integration layer, a trusted operating surface or a governance position.

The distinction will become obvious after the bubble.

Pressured group: human-heavy services businesses

The AI-operated company is a direct threat to human-heavy service models.

Agencies, consultancies, research shops, outsourcing firms, implementation partners, content operations, support services, sales operations, analytics services and many professional service niches will feel pressure.

The pressure will not be uniform. High-trust, high-judgement, high-relationship work survives longer.

Routine production, analysis, reporting, coordination and execution compress faster.

The dangerous competitor is not always a large incumbent with AI tools. It may be a small AI-operated company with lower costs, faster iteration and enough human expertise at the top.

That changes market structure.

Pressured group: companies with no proprietary workflow or data advantage

AI rewards specificity.

A company with no proprietary data, no deep workflow, no customer trust, no distribution, no system of record, no governance role and no operational insight has little to defend.

This applies to startups and incumbents.

• Putting AI in the pitch does not create a moat.
• Adding AI features does not create a moat.
• Using a frontier model does not create a moat.

The question is what remains when everyone has access to similar intelligence.

If the answer is nothing, the company is exposed.

The new strategic question for businesses

The most important question for businesses is not:

How do we use AI?

That question is too small.

The better question is:

Which parts of our business become more valuable when intelligence is cheap, and which parts were only valuable because intelligence, software or coordination were expensive?

That question cuts deeper.

It forces a company to examine its real moat.

• Do we own data?
• Do we own a workflow?
• Do we own trust?
• Do we own distribution?
• Do we own a regulated position?
• Do we own a system of record?
• Do we know something specific about our customers that others do not?
• Can we encode our process knowledge into software?
• Can we operate with fewer coordination layers?
• Can we build custom internal systems faster than before?
• Can we expose our capabilities as APIs and tools?
• Can we govern AI safely enough to let it act?

These are the questions that matter after the bubble.

The investor question

The investor version is similar. Do not ask only whether a company has AI exposure.

Ask whether it captures AI value.

• Is AI reducing cost or increasing spend?
• Is AI increasing margin or subsidizing usage?
• Is the company selling a durable capability or a temporary wrapper?
• Is the company protected by data, distribution, workflow, governance, compute or trust?
• Is the company exposed to model commoditization?
• Is the company exposed to seat compression?
• Is the company’s capex ahead of demand?
• Is the company the owner of the workflow or a supplier to someone else’s workflow?

During the bubble, exposure is enough. After the bubble, capture matters.

The final prediction

The AI bubble will burst because AI is real.

It will burst because AI is real enough to make models smaller, cheaper and more local. It will burst because AI is real enough to make custom software cheaper. It will burst because AI is real enough to pressure generic SaaS. It will burst because AI is real enough to create AI-operated companies with radically different cost structures. It will burst because AI is real enough to force investors to ask who actually owns the future.

The market has correctly sensed a paradigm shift but it has not yet correctly mapped the ownership layer.

That is the mistake.

When belief changes, the bubble breaks. After that, the real AI economy begins.

It will be useful. It will be profitable. It will be more centralized than many hope. It will be more distributed than some incumbents want. It will be more extractive than the early optimism suggests. It will compress labour, software and coordination. It will create new company forms. It will destroy some moats and create others.

It will not look exactly like the first story. The first story almost never survives contact with the mature paradigm.

I still have those business cards from the dotcom bubble. They remind me that people can stand close to the future and still be wrong about who owns it.

That is where I think we are with AI.

The future is real.

The current map is wrong.

If you want the argument from the top again, return to the preface and read forward through the series as one continuous thesis.