Meta spent $140 billion building the AI agent infrastructure, and now the banks might not let its agents through the door.
The Summary
- Meta spent $34.91 billion in Q3 2026 capex, narrowing its full-year forecast to $135-145 billion as it doubles down on AI agent development
- Royal Bank of Canada's CEO is considering blocking Meta's Muse AI agent from accessing customer banking information, citing liability concerns
- Muse has already reached 2 million daily active users, making it Meta's "hero product" just as major institutions begin erecting walls
- The collision between consumer-facing AI agents and enterprise gatekeepers is happening faster than anyone predicted
The Signal
Meta's infrastructure bet is coming into focus. The company burned through $34.91 billion in Q3 alone on data centers and computing power, with a narrowed annual forecast landing at $135-145 billion. That's real money going to actual steel and silicon, not vaporware. Zuckerberg told a podcast earlier this month he was prepared to bet up to $200 billion on AI before he knew the models would even work. Now they work. Muse hit 2 million daily actives.
But here's where it gets interesting. Dave McKay, CEO of Royal Bank of Canada, said publicly he's weighing whether to block Muse from touching customer banking data. Not "we're monitoring the situation." Not "we're in discussions." Blocking. His reasoning: "You have to protect that consumer and then we're liable for that as an organization."
"You have to protect that consumer and then we're liable for that as an organization." — Dave McKay, RBC CEO
This is the access problem no one priced into the AI agent buildout. Meta spent $140 billion creating the pipes and the intelligence. But if agents can't reach your bank account, your medical records, your work email, they're just expensive chatbots. The whole value proposition of an AI agent is that it acts on your behalf across services. If every institution with liability exposure decides to build walls, the agent economy fractures before it starts.
RBC isn't some regional credit union. It's Canada's largest bank, managing $1.7 trillion in assets. When its CEO floats blocking Meta in a Bloomberg interview, he's not threatening, he's testing. He's watching to see who else raises their hand. Because the first bank to block Meta agents might look paranoid. The fifth bank looks like the industry standard.
Key tensions emerging:
- Consumer utility (agents need data access to be useful) versus institutional liability (banks get sued if agents mishandle customer info)
- Meta's scale play (2M daily actives gives negotiating leverage) versus enterprise risk tolerance (one data breach ends careers)
- Open integration (the Web4 promise) versus closed gardens (the Web2 default when lawyers get involved)
The spending uptick was no surprise. Investors dinged Meta's stock 10% last quarter when earnings missed estimates, and 6% the quarter before when capex guidance went up. Wall Street hates uncertainty, and $140 billion is a lot of uncertainty. But Meta's guidance now aligns almost perfectly with analyst expectations of $139.79 billion. The company is signaling: this is the number, we're not going higher, place your bets accordingly.
What no analyst modeled was the access layer blowing up. You can build the smartest agent in the world, but if it can't log into your bank, book your flights, read your email, or pull your medical history, it's a toy. The technical problem is solved. The institutional trust problem is just starting.
The Implication
Watch for two things. First, whether other major banks follow RBC's lead in the next 30 days. If they do, Meta (and OpenAI, and Google) will need to build enterprise trust infrastructure as fast as they built the models. That means audit trails, liability frameworks, and probably insurance products no one has invented yet.
Second, this is a wedge for challenger banks and fintech companies willing to take integration risk in exchange for growth. If incumbents block AI agents, someone will build a bank that doesn't. The question is whether consumers care enough to switch. We're about to find out if "works with my AI agent" is a feature people will move money for.