The June AI Blackout: What Small Businesses Should Learn About Model Lock-In

The Model Your Workflows Depended On Went Dark

On June 12, 2026, the U.S. government issued an emergency export-control directive that pulled Anthropic's Claude Fable 5 — the most capable AI model on the market — offline for every customer. No warning. No timeline. Just gone.

The model returned this week, wrapped in tighter safeguards, roughly three weeks after the blackout began. In the interim, China's Z.ai released open-weights GLM-5.2 into the vacuum, OpenAI previewed GPT-5.6, and every enterprise using Fable 5 in production got a live stress test of exactly how dependent they were on a single vendor.

New VentureBeat Pulse Research surveyed 145 enterprises across the outage. Two-thirds had already hedged their model strategy before the order came down. The remaining third was all-in on a closed ecosystem when the lights went out. That third is now doing emergency architecture work while everyone else went about their week.

If you run a small business on AI, this is the news event that reframes the entire vendor conversation. Model lock-in stopped being an abstract debate on June 12. It became an operational risk with a date attached.

What Actually Happened, In Plain Terms

Claude Fable 5 launched on June 9 to immediate acclaim and immediate sticker shock — $10 per million input tokens and $50 per million output, roughly triple the previous frontier price. Businesses that had already integrated Anthropic into their workflows started routing high-value work through it. Three days later, the U.S. government issued an emergency export-control directive barring access by foreign nationals. Anthropic had no way to verify nationality in real time. So the model was suspended for everyone.

For roughly three weeks, businesses that had built workflows around Fable 5 had three choices. Fall back to an older Claude model with degraded quality. Rewrite the workflow around a different vendor. Or wait it out and hope the timeline was short.

The businesses that could switch quickly had already done the work. The businesses that couldn't discovered — under time pressure — that "vendor flexibility" is not a feature you add later. It is either designed into your architecture or it is not.

Brian Craig, senior director of architecture at Liberty IT, put it directly on stage at VentureBeat's AI Impact event mid-blackout: "You can't lock in right now in one vendor and even one framework. You need to keep being able to have the flexibility with that backbone to be able to hook into different models, different vendors, depending not so much on who's the flavor of the day, but on what you can feel confident about for the next six months."

That is the entire lesson from June, compressed into two sentences.

The Numbers That Should Change How You Think About Your Setup

The VB Pulse survey is directional — 145 respondents, self-selected, skewing senior and technical. But every question in the survey points the same way, which is what makes the pattern worth paying attention to. Three data points matter for anyone running AI in production.

Two-thirds of enterprises had already hedged before June 12. Fifty-one percent were running a hybrid posture — closed frontier models for general reasoning, open-weight models running on their own infrastructure for specialized execution. Another 16% were actively moving core workflows off closed APIs entirely. Only 32% were still all-in on a closed ecosystem, and even that group is candid about why: the operational overhead of self-hosting still outweighs the savings for them. After June, that calculus has a new variable in it.

Just 1 in 10 enterprises would automatically catch a failing production AI model. Fourteen of the 145 respondents have automated monitoring and alerting running against production systems. Thirty percent rely on humans reviewing critical outputs. Thirty-two percent expect to catch most issues "eventually." Nineteen percent say they would likely hear about a failure from end users first. Eight percent report no systematic visibility into production AI behavior at all. If your AI is generating outputs faster than a human can read them, and your monitoring model is "someone will notice," you are already behind.

Seventy-nine percent of enterprises have already taken a real financial or operational hit from autonomous agents. Most often from shadow AI — unauthorized agentic work run by employees on corporate credit cards, outside anyone's oversight. This is not a hypothetical failure mode. It is happening now, at scale, at the majority of organizations that have any AI in production.

VentureBeat calls the combined pattern the "Control Gap" — the distance between how aggressively enterprises are deploying AI and how little of it they can see, own, or govern. June's blackout turned it into a live stress test. The businesses on the wrong side of the gap felt it immediately.

Why This Hits Small Businesses Harder Than Enterprises

The instinct is to look at a Liberty Mutual or a Morgan Stanley scrambling to reroute traffic and think: sure, but they run hundreds of workflows on AI. My business runs three. The blast radius is smaller. The problem is smaller.

That is exactly backwards.

Enterprises have architects. They have platform teams. They have the budget to build "AI backbones" — Liberty IT runs roughly 50 independently replaceable components spanning security, governance, observability, and orchestration. When Fable 5 went dark, Craig's team routed around it. They had already built the operating layer that made that routing possible.

Small businesses do not have that backbone. When their model vendor goes dark, they either wait, or they rewrite. There is no third option unless someone built the flexibility for them in advance.

The kicker is that small businesses are more vendor-locked, not less. A large enterprise using Claude for one workflow can absorb a three-week gap. A small business running its inbox triage, customer follow-up, and quote generation on the same vendor's API is running its entire operational back office through a single external dependency it does not control.

If that dependency disappears, the business does not degrade gracefully. It stops.

What Model Lock-In Actually Costs (Beyond the Obvious)

The visible cost of vendor lock-in is the one everyone talks about: the vendor raises prices, you have no room to negotiate, you pay. That cost is real, and June added a hard example.

Uber burned through its entire 2026 AI coding budget in four months after Claude Code adoption hit 84% of its roughly 5,000 engineers. Microsoft canceled most internal Claude Code licenses in its Windows and Microsoft 365 division and steered engineers to internal tooling. These are companies with real negotiating power. They still hit the wall.

The less visible costs are worse. Three of them matter for small businesses.

Contextual lock-in. Every workflow you build carries assumptions about the model. The prompts are tuned to how one vendor's model reasons. The tool-use patterns match one vendor's API shape. The memory format assumes one vendor's context window. Switching vendors is not a config change. It is a rewrite of the operational logic you spent months tuning.

Data lock-in. Whatever your agents remember, whatever they have learned about your customers and your operations, whatever context files they have accumulated — if that memory lives inside the vendor's runtime, it moves with the vendor, not with you. When you switch, you start over.

Governance lock-in. The controls you have over what the agent can and cannot do — access policies, escalation rules, approval gates — are typically expressed in the vendor's configuration surface. Those controls do not port. Switching vendors means rebuilding the safety architecture from scratch, under time pressure, while the business is running.

Add these up and the picture gets clearer. The visible cost of lock-in is the bill. The hidden cost is that switching vendors, when you have to, is not a project. It is a rebuild.

What Good Looks Like — and What Bad Looks Like

Bad: Your AI setup is one API key from one vendor, wired into a handful of workflows. Prompts are tuned for that vendor's model. Memory is stored in the vendor's system. When something breaks, you notice because a customer complains or a report is late. Your monitoring model is: "someone will tell me."

If June 12 had targeted your vendor instead of Anthropic, your business would have stopped functioning for three weeks.

Good: Your AI setup separates the runtime from the operating layer. The runtime — the model actually doing the work — is one interchangeable component. You can point it at Anthropic, OpenAI, Google, or a self-hosted open-weight model with a config change, not a rewrite. Your prompts, memory, integrations, and governance policies live one layer above the model. The operating layer is the durable thing you own. The model is a variable you route to.

If June 12 had targeted your primary vendor, your workflows would have shifted to a backup provider automatically, and you would have kept operating.

The difference is architectural. It cannot be added after the vendor goes dark. It has to be built in before you start scaling AI across the business.

Four Steps To Take Before The Next Blackout

The next model outage will happen. It might be another export-control order. It might be a vendor pricing shock. It might be a model deprecation announced with 30 days' notice. You cannot predict which. You can prepare for all of them.

1. Inventory your model dependencies. Which workflows depend on which model, at which vendor? Which of those workflows are business-critical — meaning the business degrades meaningfully if they stop? This should be a one-page document. If it takes longer than that, your architecture is already too tangled to switch cleanly.

2. Test at least one alternate model on your most critical workflow. Not in production. In parallel. Route a copy of the traffic through a second model and compare outputs. You do not need to make the alternate model as good as the primary. You need to know it is good enough to fail over to. If you have never made an alternate model process real work from your business, you do not know what "flexibility" means for your specific setup.

3. Move your memory and context outside the vendor's runtime. Whatever your agents remember about your customers, your workflows, your operational patterns — that data should be portable. It should live in a system you control, in a format you can read and export, with policies you set. If your vendor is holding your operational memory, they are holding you.

4. Build monitoring that would catch a silent failure. Not "someone will notice." Actual instrumentation on production AI outputs: latency, error rate, output quality signals, cost. The 10% of enterprises that have automated monitoring did not build it after June. They built it before, because they treated AI in production the same way they treated any other production system.

None of these steps require a platform team. All of them require intention. The businesses that are calm this week are the ones that did the work before the news event.

Where This Is Heading

Two things are true at once about the AI vendor landscape right now.

The models are getting better. GPT-5.6 is coming. Claude Fable 5 is back. Open-weight models are closing the gap on frontier performance. Any business making a vendor decision today is not making a permanent one. The right vendor for the next 12 months is almost certainly not the right vendor for the 12 months after that.

The vendors are getting more concentrated. The frontier models cost more. The compute footprint required to train them is becoming a national-security concern. Export controls, licensing regimes, and geopolitical constraints are now real inputs into whether a business can access a specific model. This trend is not slowing down.

The businesses that will do well through this are the ones that treat model choice as a rotational decision, not a strategic commitment. They pick the best model for the workload today. They keep the option to pick a different one tomorrow. They invest in the operating layer that makes rotation cheap, and they do not invest in vendor-specific infrastructure that makes rotation expensive.

The businesses that will do badly are the ones that pick a vendor, tune everything to it, and then discover — when the vendor changes prices, terms, or availability — that they are not a customer. They are a hostage.

June 12 was the news event that made this obvious. The next one will not require a news event. It will just happen.

FAQ

Q: Isn't hedging across multiple model vendors expensive and complicated?

Not if it is designed in from the start. The expensive version is building on one vendor for two years and then trying to switch. The cheap version is treating the model as a swappable component from day one. The operating layer — prompts, memory, governance, integrations — lives above the model and does not change when the model does. That architecture costs about the same as single-vendor architecture to build, and roughly nothing to switch vendors in.

Q: What's the difference between "model-agnostic" and just using multiple models?

Using multiple models means you have accounts at three vendors and you route different workloads to different ones. Model-agnostic means your operating layer does not care which vendor a given call goes to. You can change the routing rules without changing your prompts, memory, or business logic. The first is a purchasing pattern. The second is an architectural property. Only the second protects you when a vendor disappears.

Q: My AI is running one workflow. Is this really a risk I need to worry about?

If the workflow matters to the business, yes. The size of your AI footprint does not change the risk of a single-vendor architecture. If your one workflow depends on Vendor X, and Vendor X goes dark for three weeks, that workflow is dark for three weeks. The question is not how many workflows you have. The question is whether the ones you have would survive a vendor outage.

Q: What kind of monitoring do I actually need in production?

At minimum: request latency, error rate, cost per request, and some signal on output quality — even something simple like "did the agent complete the task or escalate?" tracked over time. If you cannot answer "is my AI working correctly right now?" without asking someone, you do not have monitoring. You have hope. The 79% of enterprises that have already taken a hit from autonomous agents mostly discovered the problem after money had already been spent.

Q: How does this connect to shadow AI and unauthorized agent use?

Directly. Shadow AI is the visible symptom of a governance vacuum. Employees start using ChatGPT, Claude, or another vendor on their own because the business has not given them sanctioned infrastructure. That behavior does not just create data leakage risk — it creates lock-in at the individual level. Each employee ends up tied to a specific vendor's product, on a corporate credit card, with no visibility into what they are doing. The June blackout hit shadow AI users hardest, because they had no fallback. See our earlier analysis on the governance crisis hiding in your business for the operational side of this.

Q: Where does small business AI infrastructure fit into all this?

The same architectural principle applies at any scale. You need an operating layer that owns the durable pieces — prompts, memory, integrations, governance — and treats the model as a swappable runtime underneath. That is not enterprise-grade infrastructure. It is table-stakes infrastructure. Our take on why an operating layer matters more than a better runtime covers the specifics of what that looks like for small businesses.

Running AI Without A Vendor Escape Hatch Is A Choice

The businesses that came through June 12 calmly did not get lucky. They made an architectural decision months ago to keep their AI stack model-agnostic. They own their prompts. They own their memory. They own their governance. The model vendors are counterparties, not landlords.

Associates AI is built around that principle. Teammates use whichever model you choose — Anthropic, OpenAI, Google, or a mix — and let you switch without rewriting the workflow. Your memory stays with you. Your governance rules stay with you. When the next blackout hits, you route around it. If you're ready to stop using AI tools and start running a real team of AI coworkers, Associates AI Teammates gives you a no vendor lock-in way to do that. Start building Teammates at associatesai.team.