A large bank is 18 months into its AI transition. Its coding staff has been reduced through attrition and reassignment. Then new transparency rules arrive, requiring changes to core systems. At the same time, its AI provider announces it is ceasing operations immediately. Now the bank has three problems: code it cannot fully audit, compliance deadlines it cannot meet quickly, and a workforce no longer sized for human fallback. This is not hard to imagine. It’s a normal enterprise transition with one missing assumption: provider continuity. Normally, this would be a hiccup. But when the missing vendor is the system that replaced part of the workforce, its absence becomes a business continuity crisis.
IF Dependency > Financial Viability THEN Who Inherits Control?
What happens if frontier AI becomes critical infrastructure before the companies behind it become financially stable?
Recent reporting and commentary on OpenAI’s leaked financials suggest the issue is not just valuation. It is whether frontier AI’s revenue can scale faster than the cost of serving it. If costs rise with usage, then adoption itself may deepen the instability. The concern isn’t simply that one provider loses money, it’s that the entire stack may depend on subsidies, debt, and future revenue assumptions arriving before the bills do.
The above scenario is not really about one bank. It’s about the hidden assumption inside AI adoption: the provider will still exist when the dependency matures. Once an organization reduces available human fallback, rewrites workflows, and embeds AI into compliance, coding, customer service, logistics, or decision support, vendor failure stops being procurement trouble. Who is responsible for keeping the system alive?
The scenario becomes more plausible if organizations begin reducing hiring, slowing workforce replacement, or reallocating institutional expertise based on expected AI productivity gains. In that case, dependency can form sooner than replacement capacity is tested. A good business continuity plan should catch this, but most continuity plans assume vendors are replaceable, skills are recoverable, and systems can be rolled back. Frontier AI may break all three assumptions while outrunning the planners. Even firms like Deloitte, Accenture, EY, KPMG, and major banks may find their planning assumptions changing faster than governance frameworks can adapt. If AI is also being used to assist that planning, the risk magnifies. A continuity planner can model:
Vendor bankruptcy.
Vendor acquisition.
Data center loss.
Cloud outage.
What is harder to model is:
Loss of institutional skill after AI-assisted workforce transition.
Dependence on model-specific workflows.
Regulatory changes requiring rapid modification of AI-generated systems.
Simultaneous provider failure and organizational dependency.
Continuity standards such as NIST SP 800-34 already expect contingency planning, recovery strategies, testing, and plan maintenance. The gap is not that BCDR ignores vendors. The gap is that frontier AI can erode the internal capacity those plans assume still exists.
The failure is not that the vendor disappears. The failure is that the organization may have already reorganized itself around the vendor’s continued existence. The capital plan assumed substitution, but the recovery plan may require reconstitution. Those are not the same cost. Hiring freezes become recruiting emergencies, attrition becomes institutional memory loss, and savings booked over years may need to be repurchased in months, often at crisis prices.
This wouldn’t be the first time infrastructure outlived its original builders. Clouds survive, utilities and railroads survive, and even telecom survives. Many early innovators do not. Social networking did not disappear when MySpace faded. The question is whether frontier AI will be allowed to fade the same way once enterprises have rebuilt themselves around it. At that point, the question is no longer whether the model was useful. It is who has the legal, financial, and technical authority to keep the dependency layer alive.
Does Microsoft inherit OpenAI?
Does Amazon inherit Anthropic?
Does a sovereign government step in?
Are models nationalized?
Are customers stranded, locked in, or shown the door?
Are service obligations transferred?
I treated the sovereign-dependence side of this in my earlier France/Palantir piece. Here, the same problem moves inside the enterprise: dependency forms first, and control questions arrive later. Recent debate around U.S. export controls, foreign access to advanced models, and Europe’s Palantir dependence shows that AI sovereignty is no longer abstract. It is now a continuity issue.
We may be discussing AI alignment while ignoring a more immediate governance problem: succession. Not “Who controls AI?” but rather, “Who controls it after the original provider cannot?” This is happening while companies and governments race toward sovereign AI, AGI, and superintelligence. In my earlier AGI parity essay, I argued that no major actor may feel able to pause. This is the corporate version of the same trap: dependency forms before restraint, stability, or exit planning catches up. The question is not only who wins. It is who inherits the smoking wreckage.
This entire situation is like selling your smoke detectors on Facebook Marketplace because you’ve never had a house fire, then smelling smoke. Only here, the home may be a bank, hospital, government agency, or your country’s economy.
References
https://www.youtube.com/watch?v=UnLRb-sU-Is Ed Zitron explains OpenAI’s leaked financials, The Tech Report
https://www.youtube.com/watch?v=-Mn-TNLwQys Is the AI Boom About to COLLAPSE?, MS NOW
https://csrc.nist.gov/pubs/sp/800/34/final NIST Standard 800-34 Rev. 1.
https://borealis-traders.com/ai-kill-switch-europe-dependence/ AI Dependence
https://www.youtube.com/watch?v=zMKW-nEv2H0 France 24
https://www.youtube.com/watch?v=giT0ytynSqg, They Keep Silencing Me But I’m Trying to Warn Them!, Diary of a CEO: Geoffrey Hinton
Consulting: Need independent analysis or security support? See AI & Cybersecurity Consulting.