Conventional wisdom predicted that foundation models—like GPT, Claude, and Gemini—would follow the path of cloud infrastructure: powerful, standardized, and price-driven.
In this logic, as models matured, switching costs would fall. AI would become a plug-and-play commodity: anyone could call an API, wrap a workflow, and scale cheaply.
Investors and analysts expected a race to the bottom—faster inference, cheaper tokens, thinner margins.
The Reality: Context Friction
The data tells the opposite story. In 2025, forward-deployed engineering (FDE) roles grew 800% across major AI firms. Instead of self-serve commoditization, the market is experiencing implementation inflation.
Why? Because models don’t actually “fit” business contexts without human mediation.
AI Model → (FDE bridges) → Business Context → Operational Fit → ROI
Each deployment compounds embedded learning. Implementation teams capture institutional knowledge that can’t be open-sourced or commoditized.
In other words, the scarcity has shifted:
Compute is abundant.
Models are available.
Implementation capability is scarce.
This reverses the expected economic pattern. Infrastructure doesn’t flatten—it stratifies, with deep moats built through tacit deployment knowledge.
2. Value Capture Through Implementation
The Old Model: License Economics
Traditional software monetized through licenses—a simple linear value chain: Customer pays for access to the product (Microsoft, Oracle, SAP).
In that paradigm, value resided in the code—the intellectual property of the product. The vendor’s goal was scalability through standardization: the same product sold to millions of users.
Margins scaled because human labor didn’t.
The New Model: Embedded Economics
In the AI era, this model breaks down. Foundation models aren’t finished products—they’re capability platforms requiring continuous adaptation.
Thus, the new economic structure resembles a hybrid of consulting and software:
AI + FDE Model: Customer → AI Team + FDE Layer + API Access → Outcomes
Each implementation requires configuration, workflow mapping, and ongoing orchestration. The more customized the environment, the higher the value—and the stickier the relationship.
This model produces high margins not from software volume, but from implementation depth.
It’s less “Microsoft-scale” and more “McKinsey-depth.”
Why This Matters
AI vendors now function as transformation partners, not product providers. They sell outcomes, not licenses.
High-touch engagements: Enterprise-scale deployments that require on-site or embedded engineers.
Continuous adaptation: Each new dataset or context reshapes the model’s behavior.
Long-tail monetization: Once integrated, switching vendors is costly—not because of code, but because of accumulated context.
Instead of recurring revenue from user seats, it generates compounding returns from system integration.
This is why AI companies increasingly resemble consulting networks more than software factories—high-value, high-trust, and high-context.
3. The New Moat: Implementation Knowledge
The Traditional Moat
Historically, technology moats came from patents and proprietary code. Companies protected innovation through intellectual property, creating narrow but defensible barriers.
Each deployment enriches the firm’s collective map of how AI behaves in context.
This knowledge is:
Embedded in people (forward-deployed engineers, domain experts).
Captured in workflows (customized orchestration scripts, prompt libraries, pipelines).
Reinforced by relationships (trusted co-development with customers).
It becomes self-protecting: even if the model leaks, the meta-knowledge of how to make it work does not.
That’s the invisible capital of the AI firm—an experiential monopoly.
From Patents to Patterns
The shift is epistemic:
Old moat: defensible IP (patents, algorithms).
New moat: defensible know-how (integration patterns, institutional memory).
This mirrors the industrial age transition from inventors to operators. The firms that scaled weren’t the ones with blueprints—they were the ones that knew how to build, maintain, and replicate factories.
In AI, implementation knowledge is the factory.
4. Strategic Consequences for the Market
a. AI Infrastructure Firms
Providers like OpenAI, Anthropic, and Cohere can’t rely solely on model quality. They must capture and compound implementation data from the field.
Each deployment teaches them how their models perform in varied conditions—creating a closed feedback loop that strengthens both performance and differentiation.
b. Enterprise AI Adopters
For customers, competitive advantage no longer depends on buying the right model but on integrating it effectively.
Implementation velocity becomes a strategic KPI—how fast an organization can contextualize intelligence at scale.
c. Emerging Ecosystem
This dynamic favors ecosystem builders over tool vendors. Those who own both model infrastructure and deployment capability will capture the majority of value.
The story of forward-deployed engineering reveals a deeper truth about the AI economy: we are not in a software market—we’re in a learning market.
The most defensible asset isn’t code or compute; it’s the institutional intelligence that accumulates from embedding systems across diverse, high-friction contexts.
Infrastructure won’t commoditize because context doesn’t.
And as long as humans remain the interpreters of that context, the frontier of AI will be shaped less by model size and more by deployment wisdom.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
