---

In my last post, I argued that boards face a growing geopolitical risk in their AI strategy, that a single regulatory action or export control directive could sever access to a mission-critical workflow with zero warning.

If that is true, it raises an immediate question: why are so many enterprises doubling down on deep dependency on a small number of frontier providers?

The answer is economic. And it deserves immediate scrutiny.

---

It isn’t regulation that concerns me most right now.

It’s economics !!!

Many organizations are investing in AI as though the cost of intelligence will remain permanently scarce and permanently expensive. That assumption is beginning to break down, and the enterprises built on it are about to find out what happens when pricing assumptions change mid-cycle.

Think about signing a five-year cloud infrastructure agreement just before hyperscale providers permanently reset the economics of compute. You were not buying a lasting advantage. You were locking yourself into yesterday’s pricing model. Many boards are about to make that same mistake with AI.

There is an important nuance here. Not all AI dependency is equal. An enterprise with an orchestration and governance layer between itself and model providers is in a fundamentally different position from an enterprise that has wired core workflows directly to a single API. Optionality does not require full in-house control. It requires the ability to switch, and someone in your AI value chain must demonstrably own that capability. If you are using a managed services partner to run agentic workflows, the question is not just whether they are using the best model today. It is whether their architecture allows them to substitute models without rebuilding your workflows from scratch.

That distinction matters more than which model is currently running.

---

The Great Commoditization of Intelligence

For the past two years, the dominant assumption has been simple: frontier intelligence would remain scarce, and scarcity would justify premium pricing. That assumption is starting to break.

Across the ecosystem, capable models continue to improve at speed. Open-weight systems are advancing rapidly. Governments increasingly treat AI infrastructure as strategic capability rather than commercial software, which means they are actively funding alternatives to frontier providers.

China’s AI ecosystem illustrates this clearly. Systems such as DeepSeek, alongside models from Alibaba’s Qwen family, show that careful systems engineering and open-weight development can deliver highly competitive performance without relying exclusively on cutting-edge semiconductor supply chains.

Whether this is industrial policy or competitive pressure matters less than the outcome. The economics are shifting.

We have seen this before.

Servers became commodities. Operating systems became commodities. Cloud infrastructure became utilities.

Open source repeatedly compressed margins at the infrastructure layer while expanding value creation above it. AI is following the same trajectory.

That does not mean frontier innovation stops. It almost certainly will not. But it does mean boards need to ask a more uncomfortable question: will your business still pay frontier prices once today’s breakthrough becomes tomorrow’s baseline capability?

---

The Enterprise Contract Trap

Many CFOs are locking in multi-year AI agreements based on today’s economics. Frontier providers argue this is rational. They point to rapid capability gains in reasoning, multimodal systems, and autonomous agents as justification for sustained premium pricing.

They may be right. But that increases the risk rather than reducing it.

If capability leadership shifts every six to twelve months, long-term contracts become structurally fragile. Your architecture will evolve faster than your procurement cycle.

Every enterprise AI contract is implicitly a bet on scarcity. You are betting that the cost of intelligence will remain high enough to justify today’s terms. That is a historically fragile position in any technology that becomes foundational.

Some will argue open-weight systems are not meaningfully cheaper once governance, engineering, security, and compliance are included. That is true today. But it confuses current friction with structural cost.

Managed services compressed complexity in cloud infrastructure; managed inference is beginning to do the same for AI. The direction of travel matters more than the current state.

The honest board conversation is not “open-weight versus closed.” It is: at what capability and cost threshold does open-weight become the rational choice for each of our core workflows, and are we tracking that threshold over time?

Most organizations are not. Most managed service agreements do not surface it either, which means the question has to be owned explicitly at board level.

As intelligence becomes more accessible, advantage shifts away from the model and toward workflows, data, distribution, and execution.

---

The Metric Your Board Isn’t Measuring

Most boards still anchor on cost per million tokens. That is already the wrong unit.

Enterprise AI is moving into multi-agent systems where a single request triggers planning, reasoning, tool use, validation, and repeated self-correction before a result is produced. The user sees one interaction. The system may execute dozens of internal cycles.

This breaks the linear relationship between price and output. Token costs can fall while system costs rise if orchestration is poor.

The KPI that matters is not cost per token. It is cost per completed business outcome.

That requires measuring the full workflow: inference, orchestration, failure handling, human review, and retry logic.

It also exposes a second-order issue most boards do not yet track: whether their AI stack, or their partner’s AI stack, is genuinely provider-agnostic, or quietly optimized around a single vendor’s architecture.

That distinction matters because pricing and capability will not move in lockstep. When they diverge, lock-in becomes visible in operating costs rather than contracts.

An orchestration layer designed for model substitution is structurally more valuable than one optimized for a single provider, even if the single-provider version performs slightly better today.

---

The Geography Arbitrage Myth

A common response from executive teams is straightforward:

“We’ll build a team in Singapore. We’ll route inference through Dubai. We’ll separate legal entities.”

The assumption is that geography solves dependency risk. That assumption is weakening.

Regulators are increasingly focused on ownership, control, technology transfer, and operational influence rather than corporate domicile alone. Major economies are tightening rules around strategic technologies and data flows regardless of where a company is incorporated.

This does not mean global AI is splitting into fully isolated systems. But it is forming partially distinct ecosystems with increasing friction between them. For multinational companies, navigating those boundaries is becoming a capability, not an implementation detail — and it is doubly fragile when the underlying model economics are also in motion.

---

Where Competitive Advantage Actually Lives

None of this implies enterprises should abandon frontier AI systems. Many organizations will continue to rely on them for reliability, compliance, and legal protection. That is rational.

The mistake is assuming this architecture is permanent.

As intelligence becomes more accessible, advantage shifts away from the model and toward the things competitors cannot easily replicate: proprietary workflows, institutional knowledge, high-quality internal data, customer trust, and execution speed.

The same logic applies to AI partners. A partner whose value is deep integration with a single provider’s stack carries a different risk profile from one whose value is workflow design, domain expertise, and cross-model orchestration. Both can be valid choices. But they are not equivalent when infrastructure economics shift.

Models may become interchangeable. Operating systems of work will not.

---

Every Major Technology Shift Follows The Same Pattern

---

The scarce, expensive thing becomes abundant. The companies that treated scarcity as strategy lose. The companies that built differentiated capability on top of abundance win.

The geopolitical risk in the first post can sever AI access overnight. The economic risk in this post can erode AI advantage over time. Both point to the same conclusion: resilience comes from optionality, not dependency.

The question for your board is no longer which AI model to bet on. It is this: what happens to your competitive position when intelligence becomes widely available and competitively priced — and how quickly can you adapt when it does?

The biggest mistake a board can make today is assuming intelligence will remain the scarce asset.

The companies that win this cycle will be the ones that build everything around the moment it stops being one.