Pawel Rzeszucinski is Senior Director of Data and AI at WebPros.getty​On June 13, 2026, Anthropic's customers around the world lost access to two of its most capable models overnight, not because of a bug, a price change or a competitor's move but because the U.S. Commerce Department ordered it. Anthropic was forced to disable Fable 5 and Mythos 5 for every non-American user, including its own foreign employees, citing an unspecified national security concern, as Fortune reported. Within weeks, OpenAI had to limit its own GPT-5.6 access under similar government pressure. For any organization that builds mission-critical workflows on top of proprietary models from main AI labs, the lesson landed unexpectedly: A frontier model is not infrastructure you control; it is a service someone else can switch off.This is the moment enterprise AI sovereignty stops being a theoretical governance debate and becomes a practical survival skill. From my perspective, the businesses that treat model access as a utility they can always renegotiate are making a category error. The ones that treat it as a strategic dependency, on par with energy supply or internet connections, are the ones positioning themselves correctly for what comes next.When Access Becomes A Geopolitical Lever​For most of the past three years, the assumption underpinning enterprise AI strategy was simple: Rent the best model from a handful of U.S. labs, accept the cost and treat capability as the only variable that mattered. That assumption is breaking down on two fronts at once. The first is access risk, made vivid by the Fable and Mythos shutdown. The second, less visible but arguably more consequential, is that the capability gap justifying dependence on closed frontier labs has nearly disappeared.Zhipu's GLM 5.2, released under an unrestricted MIT license, is the clearest evidence of this. According to Artificial Analysis, "GLM 5.2 is the new leading open weights model" on its Intelligence Index, and on FrontierSWE, a benchmark built for realistic long-horizon coding tasks, it lands within a single percentage point of Claude Opus 4.8 while undercutting it on price by roughly fivefold, as detailed by Avenchat's benchmark review. Security researchers at Semgrep went further, finding that GLM 5.2 with no scaffolding at all outperformed Claude Code on a reasoning-heavy vulnerability detection task. The gap between open-weight and closed-frontier models, once measured in years, is now measured in low months.The Rise Of Intelligence Per Dollar​That shift has produced what CNBC calls the defining metric of 2026: intelligence per dollar. As enterprises strain under frontier token spend while government restrictions cloud access to the labs producing it, Gabe Pereyra of Harvey told CNBC he has "been consistently surprised by how quickly the open source has caught up." A model that is free to download, fine-tune and run on an enterprise's own servers is no longer a compromise. Increasingly, it is the more rational choice, even before you factor in the risk that a closed provider's access can be revoked by policy rather than by your own decision.None of this means frontier models become irrelevant. Cutting-edge research, the hardest mathematical reasoning and genuinely novel coding problems will likely still call for the most capable, most expensive systems available, used sparingly as a planner or a reviewer rather than as a default workhorse. But that is precisely the point: Frontier intelligence is becoming a specialist tool, not a general utility. Most enterprise workloads—document processing, classification, customer support, structured data extraction—do not need a generalist genius. They need a specialist that knows the job cold. Research compiled by SuperAnnotate and corroborated by peer-reviewed studies on arXiv shows that fine-tuned small models, sometimes with only a few hundred labeled examples, consistently match or exceed much larger general-purpose models on narrow, well-defined tasks while running at a fraction of the cost. We are watching the AI stack stratify into two tiers: an expensive, rarely used frontier tier and a cheap, owned, fine-tuned tier doing the actual volume of work.Building An Enterprise AI Sovereignty FrameworkTreating this as a one-off model selection decision misses the point. Here is how to build sovereignty into the way your organization operates.1. Build in-house fine-tuning capability now. Waiting until access is disrupted to learn how to fine-tune an open model is the wrong order of operations. Start with a single, well-bounded workflow—support ticket classification or contract clause extraction are good candidates—and use it to build institutional muscle before you need it under pressure.2. Adopt a tiered model architecture. Reserve frontier, closed models for the small number of tasks that genuinely require them, such as fundamental research or the hardest coding problems, and route the bulk of token volume to fine-tuned, open-weight models you control. This is not a downgrade; it is a deliberate allocation of spend to where capability actually matters.3. Map every workflow's dependency on a single external provider. Audit which business-critical processes would break if a given model lost access overnight. Treat any workflow with no fallback as a single point of failure, not a convenience.4. Normalize open-weight experimentation across the organization. Sovereignty cannot be a project run by a small infrastructure team while everyone else defaults to whichever closed model is most familiar. Put a clear open-weights usage policy in place, naming trusted providers and approved models, and actively encourage employees across functions to experiment with them on real work, not just pilot projects.ConclusionThe Fable and Mythos shutdown will not be the last time a government, a regulator or a vendor's own commercial calculus removes access to a model an enterprise has come to depend on. What has changed is that, for the first time, businesses do not have to accept that risk as the cost of staying competitive. Open-weight models have closed the capability gap enough that owning your AI stack is increasingly the more economical, more defensible choice.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?