Tom Tunguz published “A Change in AI Strategy” on October 7, making an argument backed by numbers that are hard to dispute: AI model prices fell 41% over the past year while token usage surged 50%. Supply is outpacing the ability to sustain quality differentiation. That combination describes a commodity market in the making — one where competition has shifted from raw capability to distribution, and where the question of which model to use is becoming less important than the question of who controls where it runs.
His prescribed response follows the diagnosis. In a commodity market, the winning moves are: partner and resell rather than compete on model quality alone; control the user interface where distribution creates the durable moat; and capture routing data so you know which tasks go to which model and at what cost. That’s a strategic playbook for labs — and a pointed warning for enterprise technology leaders who are still treating AI strategy primarily as a model-selection exercise.
mindmap
root((Commodity AI Era))
Models Equalize
Prices fall 41%
Usage surges 50%
Partner and Resell
Leverage open-weight models
Own the Interface
Daily habit is the moat
Capture Routing Data
Know which model wins each task
For the working software engineer: model-specific optimization is a depreciating investment
The technical corollary of model commoditization is that deeply coupling your system to one model’s idiosyncrasies is risky in a way it wasn’t eighteen months ago. Time spent tuning prompts to a particular model’s behavioral quirks, building evaluation pipelines that only work against one API, or integrating model-specific features that don’t translate — all of that work depreciates faster when active price competition is compressing margins across the market.
What doesn’t depreciate: the context layer. The retrieval architecture that ensures the model always has accurate, relevant information from your systems. The evaluation suite that tests for the outcomes that matter for your specific use case, not generic benchmarks. The integration layer that abstracts the model behind a stable interface.
The practical question worth asking internally: if you had to swap the underlying model in 90 days, what would break? If the honest answer is “most of the system,” that’s the architecture risk the commodity market is now rewarding you to address. The model beneath your AI stack is increasingly becoming an infrastructure component, not a product decision.
For business owners and operators: three things Tunguz’s analysis should change
Vendor negotiation posture. Commodity markets favor buyers who can credibly route to alternatives. That leverage only exists if the technical architecture allows routing. If your production AI systems are hard-coded to a single provider, your negotiating position was set the day integration went live — and it has been weakening every month since. The time to establish flexible architecture is before a renewal conversation, not during one.
Interface investment priority. The interface your teams use daily — the AI assistant embedded in their tools, the internal knowledge platform, the workflow surface — is where habits form and organizational value accumulates. In a market where models equalize, whoever controls that daily interface controls the customer relationship at the organizational level. Internal AI tools deserve the same investment consideration as external products, because they follow the same dynamics.
Data capture as a strategic asset. Tunguz’s routing data argument applies directly to enterprise operations. Every AI interaction your organization runs is potentially a data point: which model handled it, at what cost, with what quality result, for what type of task. Organizations building the infrastructure to capture that signal now are building a procurement advantage. Most aren’t.
My take: I’ve seen this pattern play out over 20 years, and it’s always about the interface
When I was at First American Title Insurance, the largest title insurance company in the United States, the company had accumulated over 700 software applications across 15 subsidiaries through 80+ acquisitions over a decade. The technology underneath those applications was a mix of everything — legacy mainframes, third-party SaaS, custom builds from a dozen different eras. That underlying technology changed constantly throughout the years I was there.
What didn’t change was which applications people opened every morning. The interfaces that had become embedded in daily workflow — the ones where institutional habits had formed, where organizational knowledge had settled — those were nearly impossible to replace. Not because the technology was superior. Because the switching cost had accumulated at the human level, quietly, over years of daily use.
The outcome of that pattern mattered significantly. Applications that controlled the daily user relationship had pricing power and organizational gravity that the underlying infrastructure never had. The technology that ran underneath could be renegotiated or replaced. The application the employee opened first every day could not.
Tunguz’s argument maps this exact pattern onto the AI market. The model is becoming the infrastructure. The interface — where your teams work with AI, what they expect it to know about their context, what habits have formed around it — is becoming the real asset.
For enterprise technology leaders, that’s a specific instruction: invest in the interface layer your organization controls, build the context infrastructure that makes your AI surface genuinely useful for your specific situation, and keep the model layer swappable underneath. The model will keep getting cheaper. The daily habit your teams build around an interface won’t.
