On August 31, Tom Tunguz published “The Price of Entry to the Frontier” — a short piece that names something most enterprise AI coverage is working around without saying directly. The frontier AI market is segmenting, and the segmentation is not by capability tier or price tier. It is by access.
Two developments drove the argument. Salesforce designated Anthropic as the default AI embedded in its platform — the world’s largest CRM, plus Slack. For Salesforce’s enterprise customer base, Claude is now the embedded AI present by default in the workflows they run every day. Simultaneously, governments have begun rationing access to the most capable AI models through export controls and allocation programs that limit which organizations and jurisdictions can use them at all.
Tunguz’s conclusion: at the frontier, access — not price — is the new scarcity.
quadrantChart title Enterprise AI: Capability vs. Access Difficulty x-axis Easy Access --> Rationed Access y-axis Lower Capability --> Frontier Capability quadrant-1 Negotiate before you build quadrant-2 Open alternatives emerging quadrant-3 Commodity tier quadrant-4 Off the list Rationed frontier models: [0.82, 0.88] SaaS-bundled AI: [0.58, 0.68] Enterprise API tier: [0.48, 0.74] Open-weight local models: [0.20, 0.46] Commodity AI APIs: [0.16, 0.32]
For engineers: access constraints are an architectural dependency
If your team is building on frontier model APIs and your organization does not have an enterprise agreement that guarantees capacity, you have an architectural dependency that does not appear on your architecture diagram.
Rate limiting at the commodity tier is a technical problem with technical solutions: caching, fallback models, queue management, retry logic. That’s a solvable infrastructure problem, and solving it is table stakes.
Rationed access at the frontier tier is a different class of problem. If your enterprise agreement does not give you a guaranteed allocation, and frontier capacity is being rationed by the provider or government programs, you may not be able to purchase your way to more capacity. The throughput isn’t available for purchase — it’s allocated by relationship or designation.
The practical implication for any team building on frontier models: evaluate your API tier and your organization’s standing with the vendor with the same rigor you would apply to a primary data store or core infrastructure partner. What throughput does your agreement actually guarantee? What happens to your service if the vendor reallocates capacity toward their own products or toward preferred enterprise accounts? What’s your exit path if access becomes untenable?
These are not hypothetical questions. They are the questions that surface during incident postmortems and vendor renewal negotiations — usually after you’ve built three years of business logic on top of a model API.
For business owners: bundled AI is a vendor decision you may not have made deliberately
The Salesforce-Anthropic deal deserves more attention than it has received in enterprise circles. Salesforce made Anthropic the default AI for the world’s largest CRM. That decision affects every Salesforce customer using AI features in their instance — which, in 2026, is most of them.
If your company runs on Salesforce, your de facto AI vendor for CRM-layer automation is now Anthropic, regardless of whether your technology team evaluated Anthropic directly. That’s not a criticism of the choice. But it is a vendor decision that happened in your stack without going through your vendor evaluation process.
The pattern Tunguz identifies is the broader version of this: enterprise SaaS platforms are embedding AI partners, making that partnership the path of least resistance for customers, and in doing so making vendor decisions on their customers’ behalf. That’s bundling. It has a long history in enterprise software, and the evaluation framework is the same: understand what you’re getting, understand what you’re not getting, and negotiate the terms before you’re dependent on the relationship.
Two practical steps worth taking now: first, identify which of your enterprise SaaS platforms have embedded an AI vendor by default, and audit what data those AI components can access within your instance. Second, map which of your internal workflows are running on those bundled AI components versus AI vendors your team selected directly. The goal is a clear AI dependency map — not to avoid the bundled AI, but to know what you’re depending on.
My take
When I was at First American Title, the organization had acquired over 80 companies in a decade. The result was 770 applications, dozens of embedded vendor relationships, and an architecture where replacing any given platform required understanding a web of integrations that had grown organically over years of acquisitions.
Some of those vendor relationships were contractually difficult to exit. Others were technically difficult — systems had been built against specific APIs, data formats, and vendor-specific behaviors that lacked clean abstraction layers above them. The window to negotiate those vendor relationships well had been years earlier, at the point of initial adoption. By the time I arrived, the leverage was gone. The cost of replacing a deeply embedded vendor exceeded the cost of accepting unfavorable terms.
With AI vendors, most enterprise organizations are at that early window right now. The Salesforce-Anthropic deal is a clear signal that AI vendor selection is shifting from a direct enterprise decision to an indirect one — made by the SaaS platforms you depend on, embedded in ways that compound over time.
Tunguz is right that access is becoming the scarce variable at the frontier. The follow-on question for enterprise teams is straightforward: are you negotiating your AI vendor relationships while you still have leverage, or are you discovering what you agreed to after you’ve already built on it?
