Google just told one of the biggest tech companies on the planet: sorry, we can’t give you what you want. According to the Financial Times, Google has imposed limits on Meta’s use of its Gemini AI models because Meta wanted more computing capacity than Google could actually provide. Let that sink in. If Meta can’t get the AI capacity it needs, what does that mean for you?
This isn’t a small hiccup. The restrictions reportedly disrupted and delayed some of Meta’s internal AI projects. And Meta isn’t the only one affected. Several other Google clients got squeezed too, just not as hard. This is the moment AI infrastructure went from “abundant” to “rationed,” and every service business running on someone else’s AI needs to pay attention.
What happened
- Google told Meta around March that it couldn’t meet the full Gemini capacity Meta wanted to purchase.
- Meta’s internal AI projects were disrupted and delayed as a direct result of the shortfall.
- Other Google clients were also affected, though to a lesser extent. Meta got hit hardest because its demand was “exceptionally high.”
- Meta has told staff to be more efficient with AI tokens (the units that measure AI usage) to cope with the restrictions.
- Google and Meta did not respond to requests for comment. Reuters could not immediately verify the report, which cited people familiar with the matter.
The numbers
- $20 billion: Google Cloud revenue in Q1, and CEO Sundar Pichai said computing power constraints prevented even higher growth.
- Google Cloud’s backlog nearly doubled quarter over quarter, meaning demand is far outstripping what Google can deliver.
- Companies continue to spend billions on chips and data centers and still can’t secure enough computing power for AI.
5 things service businesses should take from this
- Single-vendor AI is a liability now. If your entire AI stack runs on one provider’s models and that provider starts rationing, your operations stop. Period. This isn’t theoretical anymore.
- “Unlimited” API access was always a fantasy. Every AI provider has finite compute. The terms of service let them throttle you. Meta found this out the hard way, and your business is way lower on the priority list than Meta.
- Token efficiency matters more than ever. Meta is literally telling employees to use fewer AI tokens. If a company with Meta’s budget is cutting back, you should be auditing your token usage right now.
- Multi-model architectures are no longer a nice-to-have. Building your AI workflows so they can swap between providers (Google, OpenAI, Anthropic, open-source) isn’t over-engineering. It’s risk management.
- Compute scarcity is the new normal. Google Cloud’s backlog nearly doubled in a single quarter. Demand is growing faster than capacity. This squeeze is going to get worse before it gets better.
The hot take
This story isn’t really about Google versus Meta. It’s the beginning of AI supply-chain fragmentation, and it will reshape how every business buys and builds AI. We’re entering an era where your AI provider is also your competitor’s AI provider, and when capacity gets tight, the biggest customers get served first. Small and mid-size service businesses will be the first ones cut. The companies that survive this shift won’t be the ones with the deepest pockets. They’ll be the ones who built vendor-agnostic systems before they were forced to.
The Agency OS play
This week, do one thing: audit every AI model and API your business depends on. Make a simple spreadsheet. Column A is the tool or workflow. Column B is the provider (Google, OpenAI, Anthropic, etc.). Column C is what happens if that provider cuts your capacity by 50% tomorrow. If column C says “we’re stuck,” that’s your priority.
For any critical workflow that’s locked to a single model, start testing alternatives now. If you’re using Gemini for document processing, try running the same tasks through Claude or GPT and compare the output. If you’re using Google’s embeddings, test open-source alternatives like those from Hugging Face. You don’t need to migrate today. You need a backup plan you’ve actually tested, not one that lives in a slide deck.
Finally, look at your token usage. Most businesses are burning through tokens on poorly written prompts, redundant API calls, or workflows that send way more context than they need. Tighten your prompts. Cache responses where you can. Set up monitoring so you know exactly how many tokens each workflow consumes per day. When (not if) your provider tightens the tap, you want to already know where to cut without losing quality. The businesses that treat AI capacity like a scarce resource right now will be the ones still running smoothly when it actually becomes scarce for everyone.