The Model Is No Longer the Product. The Workflow Is.

Microsoft Build 2026 opened with agentic infrastructure, not a new frontier model. Google shipped Gemini 3.5 Flash to general availability during a week analysts described as dominated by platform and infrastructure announcements. Three of the most powerful AI companies on Earth just told you the same thing in the same month: the model is no longer the product.

Most of that money is flowing into data centers, networking, runtimes, and orchestration layers. Not into training the next benchmark winner. The capital markets have already priced in the shift. The question is whether builders have.

I think this is the most important strategic reorientation in AI since the launch of ChatGPT in November 2022. And most people building AI products are still optimizing for the wrong layer.

The Stack Gravity Principle

Here is the mental model. Call it the Stack Gravity Principle: in every maturing technology wave, economic value falls downward through the stack, from raw capability toward workflow, distribution, and end-user ownership. It happened in cloud computing. It happened in mobile. It is happening now in AI.

In cloud, Amazon Web Services didn't win because it had the best virtual machines. It won because it owned the runtime, the billing, the developer tools, and the integrations. In mobile, Apple didn't win because it had the best cellular radio. It won because it owned the App Store, the payment rails, and the user relationship.

The Stack Gravity Principle says: when the capability layer commoditizes, the platform layer captures the margin. Gravity pulls value down the stack toward the customer. Builders who resist gravity get stranded at the top, competing on benchmarks while someone else owns the workflow.

The pattern has a predictable trigger. Value falls when the performance gap between top competitors narrows enough that customers stop caring about the difference. That trigger has fired in AI.

Where the Moat Actually Lives Now

The evidence for model-layer commoditization is no longer speculative. It is structural.

Open-source and second-tier proprietary models reached GPT-4-class performance on most enterprise tasks by late 2024, often at 2x to 5x lower inference cost. Their prediction for 2026 is blunt: "domain-trained industrial models become the default." The raw capability gap between the top five model providers has compressed to the point where routing between them based on task and cost is the rational strategy.

Benchmarks have decoupled from production outcomes. Harvard Business School faculty writing in December 2025 emphasized that AI value is now judged by business KPIs: cycle time, error rates, revenue uplift. Not MMLU scores. IBM's enterprise research aligns. The benchmark leaderboard still matters for research. It has stopped mattering for purchasing decisions.

So where is the moat migrating? Three layers down.

Runtimes and orchestration. Only 11% of organizations have moved agentic AI from pilot to production, according to Deloitte's Tech Trends 2026 report. Gartner expects more than 40% of agentic AI projects to fail by 2027 because infrastructure cannot support them. The bottleneck is scheduling, memory, tool use, monitoring, and governance. Not model quality. Microsoft's Rayfin SDK and Fabric-backed runtimes are a direct bet on owning this chokepoint.

Vertical workflow integration. Houlihan Lokey's Q1 2026 report on vertical software states that embedded AI is "one of the most consequential value creation drivers" in the sector. Vertical software's structural advantage is its position as the customer's system of record. Domain-specific data and recurring workflows create a defensible data moat that generic model providers cannot replicate from the outside. Bessemer's ten principles for vertical AI businesses stress the same triad: workflow, proprietary data, and model specialization working together.

Distribution and end-user ownership. More than 70% of existing Google Cloud customers now use Google's AI services. ChatGPT holds roughly 35% monthly active user penetration among U.S. consumers. Gemini reached 26%. These are platform wars fought on integration, habit, and trust. Not on which model scores 2 points higher on a reasoning benchmark.

The contrarian case deserves honest treatment. It is unclear whether model-layer differentiation is permanently diminished or merely in a temporary lull. A sudden capability leap, say a model that is 10x better at autonomous multi-step reasoning, could reset the entire stack overnight. Paul Christiano's analysis of Cursor's Composer 2 argues that as app-layer cloning becomes easier, durable differentiation may actually migrate back down to the model layer. The strongest version of this argument says: platforms and vertical apps are necessary but not sufficient, and the deepest moat ends up back at model plus proprietary training data.

My read on this: both things are true simultaneously, but on different timescales. For the next 18 to 24 months, the platform and vertical layer is where builders can create defensible value fastest. Over a 5-year horizon, the companies that combine vertical workflow ownership with proprietary model capability will be the ones that compound. The mistake is choosing only one layer. The winning move is stacking them.

2031

Three signals inside the same shift

Pull the lens back six years. What does this look like when the dust settles?

The asymmetric bet right now is not on any single model provider. It is on the thesis that AI becomes infrastructure, invisible and embedded, the way databases and APIs are today. Nobody asks which database powers their banking app. They ask whether the app approves their loan in 30 seconds.

By 2031, the companies that matter will be the ones that own the last mile of decision-making in a specific domain. Think about the Costco hot dog principle: Costco doesn't compete on having the best hot dog. It competes on having an integrated system (membership, logistics, real estate, supply chain) that makes the $1.50 hot dog possible and profitable. The hot dog is the model. The system is the moat.

The same flywheel is forming in AI verticals. A legal AI platform that owns the workflow, the document repository, the compliance rules, and the client relationship generates proprietary interaction data every day. That data improves the model. The improved model deepens the workflow. The deeper workflow generates more data. This is a compounding loop that a generic model provider cannot replicate without also building the vertical.

Anthropic's 2026 State of AI Agents Report found that 57% of organizations are already running multi-step agent workflows in production. 81% plan to expand before year-end. Retool's 2026 Build vs. Buy Report shows 35% of enterprises have replaced at least one SaaS tool with software they built themselves, powered by AI developer tools. 78% expect to build more custom internal tools by end of 2026.

This is the shoshin moment. Beginner's mind. The old mental model, "pick the best model and wrap an app around it," is dissolving. The new one: pick the workflow, own the data, treat the model as a replaceable component you optimize continuously. Impermanence applies to model advantages. What looked like a permanent lead in February can be matched by June. What looks like a thin vertical app today can become an unassailable system of record by 2028.

The 70% rule applies here. You don't need perfect clarity on which vertical or which runtime to bet on. You need 70% confidence and fast execution. The cost of waiting for certainty is higher than the cost of being partially wrong and correcting.

McKinsey projects that redesigning workflows with AI could unlock nearly $3 trillion in economic value by 2030. That value will not accrue to model providers proportionally. It will accrue to the platforms and vertical applications that sit between the model and the business outcome. Only cash is real. The rest is accounting. And the cash is moving down the stack.

What to Build This Weekend

Stop optimizing for the model layer. Start optimizing for the workflow layer. Here is how to begin, even if you do not have a CS degree.

Step 1: Pick one workflow you do repeatedly. Invoice processing, content review, lead qualification, customer onboarding. Anything with at least 5 steps that you do more than once a week.

Step 2: Map the decision points. Write down every place in that workflow where a human makes a judgment call. Which of those calls could a model handle at 80% accuracy? That is your insertion point.

Step 3: Build a model-agnostic prototype. Use any orchestration tool. Set up the workflow so the model is a swappable module, not a hardcoded dependency. If you build on OpenAI today, you should be able to swap to Claude or Gemini next month without rewriting your logic.

Step 4: Capture the data. Every interaction your prototype handles generates training signal. Log the inputs, outputs, and corrections. This is your proprietary dataset. After 1,000 interactions, you have something no generic model provider has.

Step 5: Test aggressively and expect breakage. Your first version will fail on edge cases. That is normal. The goal is not perfection. The goal is a working loop: workflow runs, data accumulates, model improves, workflow gets better. First X, then Y, then Z.

The builders who win the next phase of AI will not be the ones who picked the best model in June 2026. They will be the ones who owned a workflow, accumulated proprietary data, and built a compounding system that gets better every week. Gravity is pulling value down the stack. Build where the weight is landing.