In a July 9 essay, Benedict Evans frames the AI token pricing question with structural clarity that most coverage of AI model releases misses entirely: yes, there is a current supply crunch in tokens; no, that crunch is not a stable long-term condition. Every observable dynamic in the market — rapid inference efficiency gains, fierce competition among near-equivalent models, massive capital buildout without natural network effects — points one direction. Foundation model labs are on track to become low-margin commodity infrastructure, the same way cellular carriers did when mobile data traffic rose by orders of magnitude and the value went to applications rather than the pipe.
Evans holds three questions open: Does the frontier keep moving fast enough to stay ahead of pricing pressure? Does meaningful differentiation emerge at the model layer, or does competition commoditize it? Does the semiconductor capex dynamic concentrate the industry to a few players before the commodity phase arrives? His analysis does not answer them definitively — no one can yet — but the structural weight of the evidence points toward commoditization.
For enterprise leaders, the unresolved questions do not change the practical answer. The relevant inference is already available: the advantage will accrue to the application layer, not the infrastructure layer.
quadrantChart title Where AI Stack Value Accumulates x-axis Low Infrastructure Exposure --> High Infrastructure Exposure y-axis Low Application Differentiation --> High Application Differentiation quadrant-1 Differentiated but exposed quadrant-2 Durable advantage quadrant-3 Commodity trap quadrant-4 High risk Model-first strategy: [0.78, 0.32] Undifferentiated AI adoption: [0.72, 0.18] Data plus implementation: [0.22, 0.85] Application-layer focus: [0.18, 0.78] First American Title 2009: [0.6, 0.74]
For engineers: the abstraction layer just moved up
The practical implication for how you build: model choice is becoming a price-performance optimization at specific task types, not a fundamental capability decision. The models are converging. Performance differences between top-tier offerings are narrowing, and the pricing differences are the primary signal worth reading.
What is not converging — and what defines engineering value — is the work done above the model layer. Context engineering: the discipline of designing what information a model receives, in what form, and when. Data pipeline quality: the cleanliness, freshness, and relevance of the inputs flowing into your AI systems. Evaluation harnesses: the rigor of measuring whether your system actually does what it is supposed to do in production conditions. Production engineering: the operational discipline of running AI systems reliably at enterprise scale.
That stack is not a commodity. It requires engineering judgment, domain knowledge, and iteration over real workloads. As the model layer below it commoditizes, the value of engineering well above it increases. The skills that are commoditizing: raw API access and basic prompt composition. The skills that compound in value: system design, data architecture, context engineering, production operations, and evaluation discipline.
For business owners: model selection is the wrong strategic bet
If your current AI strategy is organized around which foundation model you are using, that strategy has a structural problem: it is positioned at the layer that is commoditizing.
GPT-5.6 Sol and Claude and Gemini Ultra are not identical — but the performance gap at most enterprise tasks is already smaller than the implementation gap between two organizations using the same model. The organization that built better data pipelines, better integration architecture, better evaluation discipline, and better human-in-the-loop workflows will outperform the one that chose the better model — even when they are using the same model.
This is why Anthropic, Blackstone, and Goldman Sachs launched Ode on July 15: a $1.5B enterprise AI implementation firm explicitly positioned on the thesis that the next trillion-dollar AI business is implementation, not models. The firm deploys Anthropic engineers directly into client organizations to build bespoke Claude-powered systems. When Blackstone structures a bet at that scale on a specific thesis, that thesis is worth examining carefully.
The practical question for your AI budget: what percentage is going to token cost, and what percentage is going to implementation quality? Most enterprise AI investments are still weighted toward the commodity side — model subscriptions, API access, tool licensing. The organizations building durable advantage are weighted toward the other side.
My take
At First American Title Company — at the time the largest title insurance company in the world, stewarding what was then the world’s largest SQL Server database, roughly 4 TB covering 850 million property records across 100 million U.S. homes — the strategic question was identical to the one Evans is now asking about AI models: where does the sustainable advantage actually live?
The data layer was extraordinary. Nothing like it existed elsewhere at the time. But competitors were accumulating comparable datasets, and the question became inevitable: when the data itself is no longer unique, what is? The answer was the applications. The workflow systems, the integration patterns, the process architecture that made the data useful to a title agent or underwriter at the exact moment of decision. CoreLogic could accumulate property records. Replicating what First American had built on top of them — 770 applications, 20 years of institutional integration, the embedded process knowledge — was a different order of difficulty entirely.
That is the AI story, one layer down. The token is becoming the data pipe. It is becoming available everywhere, at falling cost, from multiple providers with converging capability. The sustainable advantage goes to whoever builds better on top of it: better context, better data, better process design, better governance. That work is hard and slow to replicate. The model subscription is not.
Evans gives the structural frame. The practitioner read is that this transition is already underway. Organizations choosing between AI strategies today are choosing between infrastructure-exposed positions and application-layer positions. The commodity trend makes that choice increasingly consequential over the next two to three years.
