Your AI vendor has owners, competitors, and products of its own. Meta just found out what that means.

Two stories broke a few weeks apart, and almost everyone filed them under different headings. They belong under the same one.

On July 2, the Financial Times reported that OpenAI has proposed handing the US government a 5% stake in the company — worth roughly $42.6 billion at its recent $852 billion valuation. Sam Altman pitched it directly to President Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent, framing it as the best way to share AI’s upside with the public. And he doesn’t want to do it alone: the reporting says Altman wants Anthropic, Google, and Meta to allot a similar 5% each, into a vehicle modeled on the Alaska Permanent Fund — effectively a government-owned slice of the entire American AI industry.

Four days earlier, on June 28, the same paper reported that Google had been rationing Meta’s access to Gemini. Since around March, Google had been telling Meta it could not supply as much capacity as Meta wanted. Meta’s response was to instruct its own engineers to conserve tokens. The shortfall set back several internal AI efforts — including content moderation, where Meta had come to rely on Gemini precisely because it outperformed its homegrown Llama models at flagging harmful posts and scams.

That is the news. Not the valuation, and not the shortage. The fact that Meta — one of the best-capitalized technology companies on earth, running its own frontier lab — got rate-limited by its supplier and had to ask its engineers to use less.

If that can happen to Meta, be clear-eyed about where you sit in the queue.

This is not the risk we wrote about in June

When Fable 5 went from general availability to pulled in 72 hours, we argued that any single model can vanish from under you, and that your architecture should treat the model as a swappable component. That is an availability problem, and the fix is portability: a model-agnostic seam, fallback routing, version pinning.

This is a different problem, and portability alone doesn’t solve it.

The Gemini story isn’t about a model disappearing. Gemini never went down. It stayed up, kept working, and was quietly allocated away from a customer and toward the vendor’s own priorities. In the same window Google was rationing Meta, it agreed to pay SpaceX roughly $920 million a month for about 110,000 Nvidia GPUs — capacity it described as a “bridge” to serve demand for its own Gemini Enterprise product.

Read those two facts next to each other. The supplier was short on compute, was buying more at extraordinary cost, and was routing it to the product it owns rather than the customer who competes with it. Nobody did anything improper. Google behaved exactly as a rational company with its own P&L should behave.

That is the entire lesson: your model vendor is not a neutral utility. It is a company with shareholders, competitors, its own product roadmap, and — if the 5% proposal lands — a government on the cap table. When capacity is scarce, all of those interests get a vote in how it’s allocated. Your workload gets whatever is left.

What this changes operationally

The response isn’t panic, and it isn’t “build your own model.” It’s refusing to be the lowest-priority customer of a supply you can’t see. A few things separate the teams that will absorb the next squeeze from the ones who’ll discover their tier the hard way:

• Know where you sit in the queue — before you need to. Ask your provider directly: is my throughput contractually committed, or best-effort? Best-effort is fine for a prototype and reckless for anything a customer depends on. Most teams have never asked, and assume they’re guaranteed something they aren’t.
• Buy commitment for the critical path, not for everything. Provisioned or committed-throughput pricing costs more per token and is worth it exactly where an outage or a throttle is a business event. Everything else can ride on-demand. The mistake is treating all workloads as equally important — that’s how you end up over-paying and still under-protected.
• Be genuinely multi-vendor where it’s cheap to be. Not two of everything — one live alternative for your highest-value path, tested often enough that you know it works. Meta had Llama and still got hurt, because the fallback wasn’t as good at the specific job. A fallback you haven’t measured on your actual task is a hope, not a plan.
• Make switching a scored decision, not a leap of faith. The reason teams tolerate a bad supply position is that they have no way to know whether the alternative is good enough on their workload. A workload-specific eval set turns “can we move this to another provider?” from a quarter of anxious debate into an answer you get the same afternoon. This is the same discipline that makes model upgrades safe — it just pays off twice.
• Own the layer the vendor can’t ration. Your retrieval corpus, your evaluation set, your guardrails, your prompt and routing logic, your audit trail — none of that is licensed from anyone. It’s the part of the system that keeps its value when the model underneath it changes hands, changes price, or changes priorities. Concentrate your investment there.

What to do this quarter

Three moves, ordered by how little time they take:

1. Find out your actual service tier. One email to your provider or account team: committed or best-effort, what’s the throughput ceiling, and what happens under regional contention. Whatever the answer, you’ll know something today that most of your competitors don’t.
2. Name your single most business-critical AI path — and give it a second source. Just one. Stand up the alternative provider, run your evals against it, and write down the number. If the fallback scores well, you’ve bought real insurance for a day of work.
3. Inventory what you’d keep if you switched vendors tomorrow. Corpus, evals, guardrails, logs, routing. If that list is thin, that’s your roadmap — it’s the only part of the stack whose value is fully yours.

The bigger picture

There’s a version of the 5%-stake story that’s about politics, and a version of the Gemini story that’s about GPUs, and both are interesting. But for anyone actually building on these models, they say the same practical thing: the layer you’re renting is consolidating, and its owners’ interests are getting more complicated — more political, more competitive, more entangled with products that may one day look like yours.

That’s not a reason to bet against the models. They’re extraordinary and getting cheaper, and the smart move is to use them aggressively. It is a reason to be deliberate about what you rent and what you own. Rent the intelligence. Own the corpus, the evaluations, the guardrails, and the judgment about how it all fits together.

Meta had the money, the engineers, and its own frontier lab, and still spent a quarter telling its people to use fewer tokens. The difference between that being an inconvenience and being an outage is entirely in the architecture — and in having asked the uncomfortable questions before the answer mattered.

---

Sources: Financial Times via CNBC and CNN, “OpenAI proposes 5% stake to Trump administration,” July 2, 2026; Financial Times via CNBC, “Google limits Meta’s use of its Gemini AI models,” June 28, 2026; reporting on Google’s SpaceX/xAI GPU capacity agreement, June 2026.