Tomasz Tunguz published Microsoft Resells the Frontier in late July, and it is worth reading before your next cloud AI vendor evaluation. The headline numbers: Azure grew 43% last quarter. Google Cloud grew 82%. Most analysis treated this as a product-quality story — better models, better developer experience, better documentation. Tunguz’s read is more structural: it is a cost-floor story.
Google owns both layers of its AI stack. Gemini is a Google model, trained and served on Google infrastructure. The TPU chips running that inference are designed and manufactured inside Alphabet. Every dollar of margin on an AI API call stays inside the company. Microsoft’s path is different: it resells OpenAI’s models and runs workloads on NVIDIA GPUs. Two external vendors set the cost floor for two critical layers. That structural difference, Tunguz argues, is what the growth rate divergence reflects — and what is likely to compound over time.
flowchart TD A[Enterprise AI budget] A --> AZ[Azure] A --> GC[Google Cloud] AZ --> O[OpenAI models<br/>third-party cost floor] AZ --> N[NVIDIA compute<br/>third-party cost floor] O --> AZX[Margin exits<br/>at both layers] N --> AZX GC --> GM[Gemini models<br/>Google-owned] GC --> TP[TPU chips<br/>Google-owned] GM --> GCX[Full-stack margin<br/>retained] TP --> GCX class AZX warn class GCX good classDef good fill:#163a26,stroke:#44cc77,color:#d7ffe6; classDef bad fill:#3a1620,stroke:#ff5555,color:#ffd9d9; classDef warn fill:#3a2e16,stroke:#ffaa33,color:#ffe9c7; classDef accent fill:#15233b,stroke:#4488ff,color:#dce9ff;
For the working software engineer
If you are building on cloud AI APIs today, the practical question Tunguz’s analysis raises is about pricing trajectory. Google’s owned stack gives it the structural ability to reduce API costs as its chip and model economics improve, without negotiating with a third party. When Google’s TPU manufacturing gets more efficient, Google captures that improvement directly. When Microsoft’s NVIDIA compute costs change, that change is mediated by a supply relationship Microsoft does not control.
This shows up in real cost curves. Google Cloud AI pricing has declined on a reasonably consistent schedule over the last two years. Azure OpenAI pricing has moved less predictably, dependent on OpenAI’s commercial decisions and NVIDIA’s hardware supply dynamics.
For engineers building applications where cost-per-token is a meaningful variable — document processing at volume, high-frequency classification, retrieval pipelines — this is not theoretical. A 25–30% cost reduction changes the viability of features that currently do not pencil out. The vendor whose cost floor is structurally more likely to fall is directly relevant to architecture decisions being made today.
The practical engineering response is to evaluate your application’s portability. Teams that built directly against provider-specific endpoint formats — Azure OpenAI endpoint schemas, Google Vertex-specific response structures — have less flexibility when pricing conditions shift. Teams that built against a model-agnostic abstraction layer can redirect workloads without a full rewrite. That abstraction adds a small upfront cost. Not having it when you need it adds a much larger one.
For business owners and operators
Most enterprise cloud AI procurement frameworks compare vendors at a point in time: current model benchmarks, SLA terms, integration with existing systems, and existing vendor relationships. The question of whose pricing is structurally more likely to fall over a three-year deployment horizon rarely appears in an evaluation rubric.
That is a gap worth closing. If your AI workloads are material to your cost structure — or you expect them to be as usage scales — understanding whether your vendor owns its supply chain matters. Google’s vertical integration from chip to model gives it a fundamentally different cost dynamic than a company that must buy compute from NVIDIA and model capabilities from OpenAI. The growth rate difference Tunguz documents is partly a reflection of that structural advantage.
The practical move is not to migrate away from Azure based on this analysis alone. Azure has real advantages: existing Microsoft 365 integration, enterprise support infrastructure, and vendor relationships that reduce procurement friction. Those are meaningful, especially for organizations already embedded in the Microsoft ecosystem. The move is to be honest about what you are actually buying — and to make sure your architecture gives you optionality if the pricing environment shifts.
The organizations most exposed are those that have signed multi-year enterprise AI agreements at current pricing without modeling what happens when usage scales significantly, and without building any flexibility into their architecture. That combination is worth examining now, before volume makes it a hard constraint.
My take
At First American Title, the largest title-insurance company in the US, I ran architecture and integration for a 900-person engineering division operating 770 applications across more than 15 subsidiaries. Part of that work involved evaluating vendor relationships at a scale where the unit economics of each relationship were genuinely material — where small pricing differentials, multiplied by actual volume, produced budget-line implications that never appeared in any RFP response.
The pattern I saw consistently: vendor selections made on current capability — benchmark performance, feature lists, sales relationships — repeatedly failed to surface the cost structure underneath. Some vendors had reseller arrangements that created cost floors they could not negotiate around. That structural limitation was invisible in any comparison document. It only appeared when we modeled the volume we were actually going to run at, which was typically 10–20 times the initial pilot volume.
Tunguz is describing the same structural issue at the cloud AI layer. The question most enterprise procurement teams are asking — “which model performs better on our benchmark?” — is important. It is also incomplete. The question they should also be asking is: who controls this vendor’s cost floor, and what happens to our unit economics when we scale? That question has a structurally different answer for Azure and for Google Cloud. The answer is not circumstantial — it is embedded in the ownership of the underlying infrastructure.
The organizations that will navigate the next few years of cloud AI pricing well are the ones that asked that question before they committed to an architecture, not after.
