The Deployment Layer

“The oil companies did not lose to the energy companies. They lost to whoever owned the distribution.”

I. The Wrong Question

There is a question that almost nobody in enterprise technology is asking correctly right now.

The question is not: which AI model is the best?

The question is: who owns the deployment relationship?

These are not the same question. They are, in fact, almost opposite questions. And the difference between asking one versus the other is the difference between building a business and building a feature.

The consensus in 2026 is roughly this: frontier AI labs are the defining companies of the decade, whoever has the best model wins, and enterprises should pick a lab partner and build on top. This consensus is understandable. It is also, I believe, structurally incomplete in a way that will be obvious in three years and profitable to recognize now.

This memo is about why.

II. Models Are Already a Commodity

Let me state something plainly that most people are still reluctant to say out loud: AI models are a commodity.

Not immediately. Not uniformly. But directionally and irreversibly, the trajectory is clear.

The evidence is already in front of us. New entrants have demonstrated that models of comparable capability can be built at a fraction of the cost and compute that established frontier labs had spent years insisting was necessary. Open-source models closed the gap faster than almost any serious analyst predicted. Token prices have been falling quarter over quarter. The cost per million tokens is moving in one direction only.

This is not a criticism of the frontier labs. It is a structural observation about what happens to every technology that becomes sufficiently well-understood. Processing power became a commodity. Storage became a commodity. Cloud compute became a commodity. The model is next.

The more important insight is this: commoditization of the model layer does not destroy value in AI. It redistributes it. The question is redistributed to whom?

The frontier labs are not losing because they are building bad products. They are losing strategic position because they are building the wrong layer to own. Token maximization is a rational short-term strategy and a poor long-term one. Enterprises do not have loyalty to a model. They have loyalty to outcomes. The moment a cheaper model produces the same outcome, the switch happens. And it is happening.

III. The Financial Arm Analogy

In the 1970s and 1980s, the large industrial conglomerates in India and across Asia began building something that looked strange at the time. Alongside their manufacturing operations and distribution networks, they built financial arms. Not banks exactly. Not pure investment companies. But structured entities that allowed the parent group to deploy capital, access credit, fund expansion, and manage liquidity in ways that could not be outsourced to external institutions without surrendering strategic control.

At the time, the obvious question was: why is a cement company building a financial arm? Why is a textile group doing capital markets work?

The answer, which seems obvious in retrospect, is that capital deployment is too consequential to leave entirely to people who do not understand your business. The external bank does not know your order book. The external fund manager does not understand your seasonal cash flows. The people who can manage your financial relationships with true precision are the people who live inside the business. The financial arm was not about becoming a bank. It was about owning the strategic layer that everything else depended on.

Capital was the resource that defined industrial-era competitive advantage. Intelligence, specifically the operational intelligence derived from your own data and deployed into your own decisions, is the resource that defines the current era. And the lesson from the financial arm is this: you cannot fully outsource a resource that is strategic to your survival.

IV. Every Organization Will Need an AI Services Arm

Within the next several years, the following statement will seem as obvious as the financial arm analogy above: every serious organization needs an internal AI deployment capability.

Not a vendor relationship. Not an API subscription. An arm.

The reason this has not happened yet at scale is the same reason the financial arm took time to emerge: it requires admitting that a core function is too important to be fully managed by outsiders. Most organizations have not yet internalized that AI deployment is that kind of function. They will. And when they do, the companies that built this capability early will have an advantage that latecomers will find very difficult to close.

The wrapper startups are not the answer. They can switch models, which is good. But they have no domain depth, which is fatal. They are interchangeable. They will be squeezed from above by the large embedded platforms that expand into every workflow, and from below by the commoditizing models that reduce their core differentiation. The thin wrapper's existential question is always: why would the enterprise not just get this from the platform vendor they already depend on?

The frontier AI labs are not the answer either. Their incentive is to sell tokens. Your incentive is to solve problems. These objectives are not aligned. They are, in many contexts, directly opposed. Token maximization and outcome optimization point in different directions.

V. The Token Arbitrage: Why Large Embedded SaaS Is the Most Important Trade in Enterprise Technology

Now I want to make a second argument that I believe is one of the most consequential observations about enterprise technology in this moment, and one that the market has not yet fully priced.

Consider the enterprise customer bases of the large embedded SaaS companies. The numbers are staggering:

These are not software companies that happened to get large. These are distribution networks. And in an agentic AI world, distribution is the entire game.

Here is the logic. Frontier AI labs are competing intensely on model quality while simultaneously facing commoditization pressure from open source below and market share war among themselves. In this environment, they are increasingly motivated to pursue volume over margin. The path to volume is wholesale distribution through companies that already own enterprise relationships at scale.

The large SaaS companies are positioned to be exactly this: wholesale token buyers and retail intelligence sellers.

The math is clean. A dominant ERP vendor buys tokens at institutional wholesale rates negotiated against a volume commitment that no individual enterprise could match. It deploys those tokens through AI agents embedded in workflows that hundreds of thousands of companies depend on, in processes that cannot be easily switched off. It charges those companies for outcomes, automation, and intelligence. The margin sits with the platform. The token cost is a COGS line that compresses over time as competition among frontier labs intensifies. The switching cost for the enterprise customer is enormous because the workflow integration is structural, not peripheral.

This is already beginning. The CEOs of the largest enterprise SaaS companies have publicly stated that agent-driven revenue is a meaningful growth vector. The translation is direct: we have the distribution, we will buy the commodity, and we will sell the intelligence. The business model is flipping from seat-based licensing to consumption-based intelligence delivery.

This is the most important structural trade in enterprise technology right now, and it is hiding in plain sight.

VI. The Structural Problem with the Middle

If large embedded SaaS is the winner at scale, and if deep vertical specialists are the winner in high-stakes domains, then what happens to everyone else?

The companies that escape this squeeze share one characteristic: deep domain knowledge combined with deployment depth. They are not selling AI. They are embedded in a function that cannot be replicated by a general-purpose platform.

VII. Vertical Supply Chain Positioning

There is a concept I return to consistently in how I think about our own business and how I advise the organizations we work with. I call it vertical supply chain positioning.

The question for any technology company is: where do you sit in the chain, and can the chain function without you?

The ideal position is one where both the layer above and the layer below depend on your continued presence. Not because you are the only theoretical option, but because the switching cost in practice, given the integration depth, the institutional knowledge accumulated, and the operational risk of change, makes replacement implausible under normal circumstances.

This is distinct from simply being large or being contractually locked in. Lock-in is about contracts. Supply chain depth is about capability. One is fragile and contestable. The other compounds over time.

“The frontier AI labs are not a threat to a company in this position because the model is a replaceable input. You switch from one frontier lab to another in a configuration file, not in a restructuring. That is the correct relationship to have with a commodity supplier.”

You treat it with respect. You negotiate hard on price. You maintain portability. And you never allow it to become load-bearing in your architecture in a way that cannot be reversed.

The large embedded SaaS companies own this position at the enterprise workflow layer. The specialized deep-integration players own it in the verticals that require genuine domain expertise. Everything in the middle is under structural pressure.

VIII. What This Means If You Are Running an Organization Today

If you are a CXO reading this, the practical implications are direct.

You are going to make AI decisions in the next 18 months that will determine your competitive position for the next decade. Most of those decisions will appear to be procurement decisions. They are actually architectural decisions. And the difference matters enormously.

The question is not which AI vendor to sign with. The question is whether you are building internal deployment capability that accumulates knowledge about your specific operations, or whether you are renting intelligence from a vendor who knows nothing about you and will reprice whenever the market allows.

The financial arm analogy is precise. You would not allow a bank you do not control to manage the entire capital structure of your business. You should not allow an AI vendor you do not deeply integrate to manage the entire intelligence layer of your business.

IX. A Note on Timing

One of the consistent errors in technology strategy is confusing the right idea with the right moment. The internet was the right idea in 1994. It was also very difficult to build durable value on in 1994. When matters as much as what.

I believe the moment for the AI services arm thesis is now. Here is why.

X. Where SVECTOR Stands in This Architecture

I want to be direct about how I think about our company in relation to this thesis.

SVECTOR is not an AI company in the sense that the frontier labs are AI companies. We use models the way an aircraft manufacturer uses engines: as a necessary and important input to a system whose purpose is something entirely different. Calling SVECTOR an AI company because we use AI is like calling an aircraft manufacturer an engine company because planes have engines. The engine makes flight possible. It is not what flight is for.

Our purpose is decision infrastructure for organizations that operate in complex, high-stakes environments: industrial operations, financial systems, defence and security, enterprise governance.

We build the systems that organizations rely on to understand what is happening, decide what to do, and act on it with precision.

In the architecture described in this memo, we occupy a specific and deliberate position. We are not a large embedded SaaS company with a million customers. We are a vertical supply chain player with deep deployment depth in sectors where the cost of getting the intelligence layer wrong is measured in operational failures, systemic financial exposure, or security consequences that cannot be undone.

Below us, model providers are commodity inputs that we select, negotiate, and replace based on performance and price. Above us, the enterprises and governments we serve cannot operate their intelligence layer without our deployment depth. That is the position I intend to hold and to deepen.

The AI services arm thesis is not an external market observation for us. It is a description of what we are building. And we are building it in the sectors where the consequences of getting it wrong are highest, which means the value of getting it right is highest, which means the structural position compounds fastest.

Siddharth Shah

Founder & CEO, SVECTOR