In late August 2026, Tomasz Tunguz published “The Great Segmentation”, arguing that the frontier AI market is no longer competing primarily on price — it’s competing on access. The clearest illustration: Salesforce made Anthropic’s Claude the default AI across its entire CRM and Slack platform, embedding a model selection into the workflow layer before most enterprise customers had formed a considered view about it. Meanwhile, several governments have begun rationing access to the most capable frontier models, negotiating terms that constrain which organizations can use them and on what basis.
Tunguz’s argument is direct. At the frontier, access is the new scarcity. The question is not what you can afford. It’s whether you can get in, and on what terms.
quadrantChart title Enterprise AI Vendor Positioning x-axis Shallow Integration --> Deep Integration y-axis Low Leverage --> High Leverage quadrant-1 Negotiate now quadrant-2 Lock-in risk quadrant-3 Trapped quadrant-4 Evaluate first Evaluation phase: [0.15, 0.85] Controlled pilot: [0.35, 0.6] Platform default embed: [0.85, 0.15] Production dependency: [0.9, 0.05]
The rundown
The dynamics Tunguz describes are already visible in enterprise AI procurement. Model providers are not competing purely on capability anymore — they’re competing on distribution, on embedding, on becoming the default in workflows that enterprises are already running. The Salesforce and Anthropic arrangement is the clearest example: it didn’t give enterprises a model to evaluate. It made that model the presumptive choice across a platform that millions of knowledge workers use daily. Enterprises didn’t select Anthropic for their CRM AI. Salesforce selected it for them.
The government access dimension is a related but separate pressure: frontier model access is becoming a policy instrument, not just a market decision. For organizations in regulated industries, defense-adjacent work, or government contracting, which models are available to them may be answered before they reach procurement.
The combined effect is what Tunguz calls segmentation: organizations that engaged early, at the platform vendor layer, secured access on their own terms. Organizations that are later to those conversations are finding terms already set — and finding that the leverage available during an evaluation is not the same as the leverage available after dependency.
For engineers: you’re increasingly building on infrastructure your organization didn’t choose
The engineering implication of platform-level AI embedding is that the model above you in the stack is not necessarily the one your organization evaluated. When your CRM’s AI is a specific model, when your cloud provider’s AI assistant is a specific model, when your development toolchain has a default model embedded — the integration patterns you write today are model-specific not because you selected that model, but because a platform vendor did.
That creates two practical challenges. The first is portability: code written against a specific embedded model doesn’t migrate cleanly when that model changes, when the vendor relationship shifts, or when your organization wants to move to a different platform. The second is visibility: engineers often don’t know which model they’re calling when they write against an abstracted API. The vendor abstraction hides the dependency.
The engineering response is not complicated: make the dependency explicit. Know which models are embedded in each platform your system integrates with. Document those dependencies as you would document any third-party integration. Where you have a choice of how deeply to integrate, integrate shallowly until you’ve had a real conversation about the vendor relationship at the commercial level.
For business owners: the negotiation window is measured in months, not years
The strategic implication of Tunguz’s thesis is about timing. Once an AI model is embedded in your production workflows, your leverage in that vendor relationship approaches zero. You’re not negotiating from a position of evaluating options — you’re negotiating from a position of dependency, and the vendor knows it.
The window when you have real leverage is the evaluation phase: before integration, before production, before your team’s workflows have been restructured around a specific model’s behavior. That window is compressing because AI adoption is moving fast enough that the pilot-to-production cycle is shorter than most enterprise procurement processes are designed to handle.
The practical move is not to slow down AI adoption. It’s to treat vendor relationship negotiation as part of the evaluation process, not something that happens after you’ve decided. What are the terms if the model changes? How is your data handled? What does exit require, technically and commercially? Those questions are far easier to ask — and get useful answers to — before you’re dependent.
My take
When I was doing contract enterprise architecture work at First American Title beginning in 2009, one of the most striking aspects of the engagement was what 80-plus acquisitions had actually produced at the technology layer. The company had accumulated over 770 applications across 15 subsidiaries — and most of those applications had arrived with their own embedded vendor relationships, already in production, already depended upon. Renegotiating those terms, standardizing on shared infrastructure, deciding what to consolidate and what to retire — that work was among the most expensive and time-consuming parts of the modernization program.
The M&A pattern was clear: each acquisition brought software that was already in use. The leverage to renegotiate the underlying vendor relationships had evaporated with the first production dependency. By the time the architecture team understood what was in the environment, the commercial relationships were already entrenched. The cost of unwinding vendor lock-in that arrived baked-in from acquisitions was enormous and multi-year.
AI adoption via embedded platforms is creating the same compounding dynamic, but at a faster clock speed. When a platform vendor embeds a specific model as the default, every organization using that platform immediately has an AI vendor relationship — whether or not they intended to form one. The question is whether they engage with that relationship explicitly, asking the right questions during the evaluation window, or whether they discover what they’ve implicitly agreed to when something changes — a model deprecation, a pricing shift, a terms revision.
First American spent years managing the downstream consequences of not asking those questions early enough. The window to ask them in the AI vendor context is now, and it’s shorter than most enterprise planning cycles assume.
