Ted TschoppThe biggest mistake people are making about AI is that they keep asking the same question:
Who has the best model?
Is it OpenAI? Anthropic? Google? Meta? Someone else?
That question matters. But it is not the whole question. And in the long run, it may not even be the most important question.
The better question is this:
Who owns the system where AI actually becomes useful?
Who owns the place where AI sees your work, understands your context, touches your apps, remembers your preferences, respects your permissions, chooses the right model, and then actually does something?
Because that is where the real value of AI is going to show up.
Not just in the model.
Not just in the chatbot.
Not just in the data center.
But in the full stack from the person at the front end all the way to the GPU cluster and power infrastructure at the back end.
That is the AI value chain. And once you see it clearly, a lot of the industry starts to make more sense.
Part 1: The Model Is Not the Product
For the last couple of years, the AI conversation has been dominated by model performance.
Which model scores higher on benchmarks?
Which model writes better code?
Which model has better reasoning?
Which model has the biggest context window?
Which model is cheaper per token?
Those are all real questions. But they are also incomplete questions.
Because most people do not want a model.
They want work done.
They want the email drafted. They want the code fixed. They want the meeting summarized. They want the customer issue resolved. They want the spreadsheet analyzed. They want the weak password fixed. They want the trip booked. They want the contract reviewed. They want the file found, the message sent, the ticket updated, and the decision supported.
The model is one part of that. But it is not the whole product.
The real product is intelligence embedded inside a trusted action system.
That means the model has to be connected to context, tools, data, workflows, permissions, interfaces, and infrastructure.
A model that can talk about work is interesting.
A system that can safely do the work is valuable.
Part 2: The Full AI Stack
To understand where value will flow in AI, you have to map the full stack.
At the very front is the person.
The person has intent. They have a goal. They want something done. They may not know exactly what system needs to be touched or what steps are required. They just know the outcome they want.
Then comes the trusted surface.
That could be a phone, laptop, operating system, browser, chat interface, voice assistant, IDE, CRM, ERP, messaging app, productivity suite, or enterprise portal.
This layer matters because it owns attention. It is where the user expresses intent. It is where the AI receives permission to help.
Then comes the context layer.
This includes the screen, files, messages, calendar, email, photos, browser tabs, documents, contacts, customer records, code repositories, tickets, dashboards, and enterprise data.
AI gets dramatically more useful when it does not require the user to copy and paste everything manually. The more relevant context the system can safely access, the more valuable the AI becomes.
Then comes the identity and permission layer.
Who is the user? What are they allowed to see? What systems can they touch? What can the AI do automatically? What requires approval? What must be logged? What must never leave the device or enterprise boundary?
This is where trust becomes operational.
Then comes the application action layer.
This is the layer where apps expose what they contain and what they can do. A calendar exposes meetings. A CRM exposes accounts and opportunities. A banking app exposes transactions. A project management system exposes tasks. A code platform exposes repositories and pull requests.
Without this layer, AI can advise.
With this layer, AI can act.
Then comes the agent harness.
This is one of the most important layers in the whole stack.
The harness is what turns model intelligence into useful execution. It gathers context, selects tools, plans steps, routes tasks, asks for confirmation, handles errors, checks results, and decides whether to use a local model, a cheaper model, or a frontier model.
The harness is the difference between a chatbot and an agent.
Then comes the model layer.
This includes local models, small specialized models, enterprise models, open-weight models, frontier models, reasoning models, coding models, multimodal models, and domain-specific models.
The model supplies cognition. But the rest of the stack determines whether that cognition becomes useful work.
Then comes the routing and economics layer.
Not every task deserves the most expensive model. Some tasks should run locally. Some should go to a cheap model. Some should go to a specialized model. Some should go to a frontier model. Some should never leave the device. Some should only run inside a private enterprise environment.
This is where AI cost, latency, privacy, quality, and capacity all come together.
Then comes the cloud and private compute layer.
This handles the tasks that are too difficult, too large, too sensitive, or too compute-intensive to run on the device. It includes cloud APIs, private cloud, secure inference, compliance boundaries, observability, logging, and retention policies.
And finally, at the very back of the stack, you have the data center layer.
GPUs. Accelerators. Networking. Memory bandwidth. Storage. Power. Cooling. Model serving. Training clusters. Inference infrastructure.
That is the physical back end of intelligence.
So the AI stack looks like this:
Person → intent → trusted surface → identity → permissions → context → data → app actions → workflow → agent harness → model routing → local inference → private cloud → frontier models → GPUs → data centers → power.
That is the actual value chain.
And once you see that, you stop asking only who has the smartest model.
You start asking who owns the scarce, integrated, trusted control points across the stack.
Part 3: Where Profits Actually Flow
Here is the basic rule of value chains:
Profits flow away from modular commodities and toward integrated differentiation.
If something becomes easy to swap out, it becomes harder to charge a premium for it.
If every company can access five good models through APIs, and those models are “good enough” for most use cases, then raw model access starts to look more modular.
That does not mean models do not matter. They matter enormously.
But if the model becomes interchangeable for a given task, then the profit pool moves somewhere else.
It moves to the product layer.
It moves to the workflow layer.
It moves to the company that owns the customer relationship.
It moves to the system that integrates the model with tools, data, permissions, and action.
This is why Apple is such an important example.
Apple does not make extraordinary profits simply because it sells hardware. Apple captures value because the hardware is integrated with the operating system, the software ecosystem, services, privacy model, developer platform, retail experience, and brand.
The iPhone is not just a device.
It is a trusted surface for digital life.
And in AI, the trusted surface may become even more valuable than before.
Because the AI surface is not just where you click.
It is where the system sees your work, accesses your context, touches your apps, and receives permission to act.
That is an extremely powerful position.
Part 4: The Harness Is Where Agent Differentiation Happens
This is why the harness matters so much.
In AI, people often talk as if the model is the agent.
But that is not right.
The agent is the model plus the harness.
The harness includes tool use, memory, context retrieval, prompts, orchestration, application integrations, safety rules, permissions, routing, observability, and user experience.
Take a coding agent.
The model matters. Of course it does.
But the useful product also needs repository access. It needs to edit files. It needs to run tests. It needs to understand dependencies. It needs to create branches. It needs to open pull requests. It needs to handle errors. It needs to know when to ask the developer for approval. It needs to avoid breaking production systems.
A brilliant model with a weak harness may still be a weak product.
A slightly weaker model with a deeply integrated harness may create more value.
That is the strategic point.
As AI becomes more agentic, the value shifts from “who can generate the best answer?” to “who can reliably complete the task?”
Completion requires integration.
And integration is where differentiation — and profit — tends to live.
Part 5: Why Frontier Labs Want to Move Forward
This creates a strategic imperative for frontier AI labs.
If you are a model provider, you do not want to become just another modular supplier in someone else’s value chain.
You do not want your model to be a component that downstream companies swap in and out based on price.
So you move forward.
You build the chat interface.
Then you add memory.
Then you add tools.
Then you add file handling.
Then you add connectors.
Then you add coding agents.
Then you add enterprise workspaces.
Then you add workflow automation.
And over time, you are no longer just a model provider.
You are becoming a software company.
That means the competitive set of the frontier labs expands. They are not only competing with each other. They are competing with traditional software.
If an AI agent can write code, it competes with parts of the developer tooling stack.
If it can answer customer questions and update records, it competes with customer support software.
If it can summarize meetings, draft documents, analyze spreadsheets, search files, and update workflows, it competes with productivity software.
If it can move across systems and act on behalf of the user, it starts to compete with almost everything.
This is why the future AI industry is not just a model race.
It is a fight over who owns the action layer.
Part 6: Why Platforms Have a Different Advantage
But the frontier labs are not the only players with power.
Platform companies have a different kind of advantage.
They may not always have the best frontier model. But they often own the place where AI becomes useful.
They own the device.
They own the operating system.
They own the app ecosystem.
They own the browser.
They own the productivity suite.
They own the enterprise identity layer.
They own the workflow surface.
They own distribution.
That matters because model capability can be sourced. It can be rented. It can be routed. It can be substituted.
But it is much harder to source a billion trusted devices, a mature operating system, a developer ecosystem, and a daily relationship with users.
This is why the device-plus-cloud model is so important.
Some AI runs locally. Some runs in private cloud. Some runs through frontier models. Some runs through specialized infrastructure.
The user does not want to think about any of that.
The user just wants the system to do the right thing.
That means the company that owns the trusted surface can become the router of intelligence.
It can decide what runs locally, what goes to private cloud, what goes to a frontier model, what stays private, what requires confirmation, and what gets automated.
That is not just a product position.
That is a value-chain control point.
Part 7: The Two Big Bottlenecks
There are at least two major bottlenecks in AI.
The first bottleneck is raw compute.
GPUs. Data centers. Power. Networking. Memory bandwidth. Training clusters. Inference capacity.
This bottleneck is very real. If AI demand keeps growing, compute providers, chip companies, cloud platforms, and data center operators remain incredibly important.
In that world, compute becomes a tax on intelligence.
Every serious AI task burns tokens, GPU time, power, cooling, and infrastructure.
But there is a second bottleneck.
The second bottleneck is the trusted action surface.
Where does AI meet the user?
Where does it see the work?
Where does it touch the apps?
Where does it get permission to act?
Where does it become the default way people interact with software?
That bottleneck is just as important.
If the future of AI is mostly giant models in giant data centers, then the back end captures a massive amount of value.
But if a meaningful amount of useful AI happens through phones, laptops, browsers, operating systems, and workflow platforms, then the economics become more complicated.
The cloud still matters.
But the device, the operating system, the app layer, and the trusted user surface matter too.
The winning strategy may not be cloud-only or device-only.
It may be intelligent routing across the whole stack.
Part 8: Why Apps Need to Become Legible to AI
This also changes what software companies need to build.
For the last couple of years, a lot of AI strategy has been shallow.
Add a chatbot.
Add a button.
Add a summarizer.
Add “AI” to the product page.
That may be useful, but it is not enough.
The deeper change is that software needs to become legible to agents.
That means the app’s data objects must be understandable.
Its actions must be callable.
Its permissions must be clean.
Its workflows must be safe.
Its state must be visible.
Its business logic must be accessible in a controlled way.
In the old world, the user opened the app, learned the interface, clicked around, and completed the task.
In the agentic world, the user asks the system, and the system uses the app on their behalf.
That does not necessarily mean apps disappear.
But it does mean apps become components inside a larger AI operating environment.
The apps that win may not be the ones with the flashiest AI demos.
They may be the ones whose data and actions are clean enough for the AI layer to use safely and reliably.
That is a very different product strategy.
Part 9: Tiering Reveals Scarcity
Pricing tells you where the power is.
When advanced AI is bundled cheaply, people assume intelligence is becoming free.
But when the best models move into usage credits, enterprise tiers, capacity limits, or premium plans, that tells you something important.
It tells you the provider sees that capability as scarce, expensive, differentiated, or strategically important.
This creates a many-tiered AI economy.
At the bottom, you have invisible AI features: autocomplete, summarization, classification, extraction, search, ranking.
Above that, you have copilots that assist humans.
Above that, you have agents that execute bounded tasks.
Above that, you have multi-step agents that move across systems.
Above that, you have high-trust enterprise agents with compliance, auditability, identity, and governance.
And above that, you have frontier reasoning systems reserved for the hardest, most valuable, or most consequential work.
Not every task needs the top model.
So companies will build routing layers.
Use the cheap model when it is enough.
Use the local model when privacy matters.
Use the specialized model when the domain is narrow.
Use the frontier model when the outcome justifies the cost.
This is not just a technical architecture.
It is the economic architecture of AI.
Part 10: Trust Is the Real Constraint
The more AI does real work, the more trust matters.
When AI only writes a paragraph, the risk is limited.
But when AI touches your calendar, email, files, photos, financial records, source code, customer data, contracts, passwords, and enterprise systems, trust becomes central.
In the old software world, trust meant:
Will this app store my data safely?
In the AI agent world, trust means something deeper:
Will this system understand me correctly?
Will it use the right context?
Will it respect my permissions?
Will it avoid exposing sensitive information?
Will it act only when it should?
Will it ask for approval before crossing important boundaries?
Will it stop when human judgment is required?
That is why the trusted surface is so valuable.
The winning AI systems will not just be smart.
They will be trusted enough to act.
And that trust may be one of the hardest things in the entire stack to replicate.
Part 11: The Strategic Conclusion
So the practical conclusion is simple:
Do not analyze AI as a model market.
Analyze it as a full-stack value chain.
The full chain runs from the person to the data center:
Person. Intent. Trusted surface. Identity. Permissions. Context. Data. App objects. App actions. Workflow. Agent harness. Model selection. Local inference. Private cloud. Frontier model. GPU infrastructure. Data center. Power.
Every layer matters.
Every layer can become a profit pool.
Every layer can become a commodity.
Every layer can become a control point.
But the most valuable positions are likely to be where three things overlap:
First, differentiation.
The capability is meaningfully better.
Second, integration.
The system is hard to replicate by assembling generic parts.
Third, customer control.
The company owns the user relationship, the workflow, or the default surface where AI is used.
That is why the model is important, but not sufficient.
The harness may be more important than most people think.
The trusted surface may be the most important layer of all.
And the data center is still underneath everything, collecting rent whenever the task is too large, too complex, or too expensive to run anywhere else.
The companies that win AI will be the companies that understand the whole system.
Not just the model.
Not just the app.
Not just the cloud.
The whole stack.
From the person at the front end to the data center at the back end, and every integration point in between.
Because the real value of AI is not intelligence in isolation.
The real value of AI is intelligence that can see the work, understand the context, respect the rules, choose the right tools, and safely act.
That is where the future profit pools will form.
That is where the strategic fight is happening.
And that is the part of the AI race most people are still missing.
Closing CTA
So the next time you hear someone ask, “Who has the best AI model?” — ask a better question:
Who owns the place where that model becomes useful?
Because that is where the value is.
And that is where the future of AI will be decided.
