capability 06 · outcomes engineer

Platform Architecture

Every shared model an enterprise builds eventually has to run on something. Platform architecture is the capability of designing that something so it serves the outcome rather than dictating it.

Removes: understand the technology model · align to how tech is built

By · V-Squared AI

Every shared model an enterprise builds eventually has to run on something. Platform architecture is the capability of designing that something so it serves the outcome rather than dictating it: understanding how technology is built and why, and aligning the build to how the business creates and captures value rather than to the preferences of whoever set it up first.

The most expensive layer to get wrong

It removes the understand-the-technology-model and align-to-how-technology-is-built bottlenecks at the layer where they're most expensive to get wrong, because architecture decisions are the ones you live with longest.

Compound advantage, or cap it

An outcomes engineer with this capability can tell the difference between a platform that compounds advantage and one that caps it.

Architecture decisions are the ones you live with longest.
course in development

The dedicated course is on the way

Platform Architecture is part of the outcomes engineer set; its course is in development. In the meantime, the AI Strategist and Platform Monetization courses cover the adjacent ground.

AI Strategist course
the market is already hiring for this

Design & Product Engineers

This capability removes (G1,F1)/(G1,F3), the technology build-and-align cells design engineers scale up to close. When AI commoditizes the technical work, that removal is the job, not the tooling.

See the roles
the evidence · earnings calls, july 2026

Microsoft published its architecture, and its cost curve.

This capability is about designing the platform so it serves the outcome rather than dictating it. In July 2026 the largest enterprise software vendor gave that instruction to its customers in plain language, then showed the numbers behind it.

peer-reviewedsurvey researchcompany filingsmarket dataearnings call
Microsoftearnings callFY26 Q4 · Jul 29, 2026

Keep the harness separate from the model

Nadella's architectural instruction to enterprises: keep the harness separate from the model so memory, context, and action space sit outside any model family and every model stays substitutable. Use frontier models where they earn it, cheap models where they do not, and train your own when you want neither, because you already hold the outputs, the traces, and the context. He said Microsoft intends to evangelize that design pattern.

Source: Microsoft FY26 Q4 earnings call, July 29, 2026.
what it confirms

This is the difference between a platform that compounds advantage and one that caps it, stated as a design rule. The durable asset is the harness. The model is a substitutable input, and any architecture that treats it otherwise is buying a ceiling.

Microsoftearnings callFY26 Q4 · Jul 29, 2026

The routing ratio, published as a cost curve

A small cyber model outperforming a much larger frontier model at half the cost, with 90% of tasks handled by the small model and 10% escalated. An 89% GPU cost reduction in Dynamics 365 and up to 84% in PowerPoint. 10% lower median token usage on GitHub Copilot. Copilot workload throughput up 4 times since January. Maia 200 at 30% better performance per dollar.

Source: Microsoft FY26 Q4 earnings call, July 29, 2026.
what it confirms

Note what is being optimized. Not capability, but cost per delivered outcome. An architecture that routes everything to the frontier model is paying for a curve the vendor itself stopped riding, and the gap compounds every quarter.

Microsoft and Metaearnings callJul 29, 2026

Single model dependency treated as a continuity risk

Nadella told analysts an enterprise cannot depend on any one model and cannot be subject to the refusal of a single model, framing multi-model architecture as business continuity. Zuckerberg said other companies do not want to rely on a small number of closed labs and that the reliance carries risk.

Sources: Microsoft FY26 Q4 and Meta Q2 2026 earnings calls, July 29, 2026.
what it confirms

Substitutability stopped being a cost optimization and became an availability requirement. That reclassification changes who has to approve the architecture and how it gets justified, which is a business conversation an architect now has to be able to have.

Sources: Alphabet Q2 2026, Microsoft FY26 Q4, Meta Q2 2026, and Amazon Q2 2026 earnings calls and releases, July 2026. Figures are as stated by company executives. Amazon reported after market close on July 30, so Amazon figures come from the release and initial call remarks rather than a full transcript.

the wider evidence · what the buyers already built

Enterprises are running multiple models and finding out what that costs.

The architecture question is no longer theoretical. Buyer-side data shows multi-model as the emerging default, and shows the gap between organizations that know they are exposed and organizations that have designed against it.

peer-reviewedsurvey researchcompany filingsmarket dataearnings call
a16zsurvey research2026 · 100 enterprise CIOs

Multi-model is becoming the default, not the hedge

A survey of 100 enterprise CIOs found 37% running five or more AI models in production, up from 29% a year earlier. Menlo Ventures put enterprise generative AI spend at $37 billion in 2025, more than tripling year over year.

Sources: a16z enterprise CIO survey, 2026; Menlo Ventures, State of Generative AI in the Enterprise, 2025.
what it confirms

Five models in production is an architecture, whether or not anyone designed it. The organizations that chose it deliberately have a routing layer. The ones that arrived at it by accident have five integrations and no way to move work between them.

Enterprise lock-in surveyssurvey research2026

Awareness of the risk is not the same as designing against it

Enterprise surveys reported in 2026 found roughly 81% of enterprise leaders concerned about dependency on a specific AI vendor, with a much smaller share believing they could switch their primary provider without material operational disruption. Reported migration costs run into months of engineering time, driven by rewrites to model-specific prompts, SDK calls, and evaluation frameworks rather than by infrastructure.

Sources: enterprise vendor-dependency surveys reported in 2026. These are vendor and analyst surveys with varying methodology, so the figures are directional.
what it confirms

The rework lives in the layers people do not think of as architecture. Prompts tuned to one model's behavior and evals calibrated to its output distribution are coupling, and they are the reason a swap that looks like a configuration change takes a quarter.

Gartnersurvey research2025 to 2026

The foundation under the platform is the constraint

Gartner research reported that only about 12% of organizations have data of sufficient quality to support AI applications, and forecast that 60% of AI projects unsupported by AI-ready data would be abandoned. Gartner defines AI-ready data as aligned to specific use cases, governed at the asset level, supported by automated pipelines with quality gates, and continuously quality-assured.

Source: Gartner research on AI-ready data, 2025 to 2026.
what it confirms

Architecture decisions are the ones you live with longest, and the data layer is the longest-lived of them. Continuously is the operative word. Data governed at reporting cadences cannot support systems that make decisions hourly.

The CIO and lock-in surveys have samples in the low hundreds and come from vendors and investors with a position in the market, so they are best read as direction of travel rather than as precise measurement.