The same systems that scaled for three billion humans are choking on queries from a few thousand agents.
The Summary
- Meta's agentic queries grew 30x in six months, breaking infrastructure assumptions built over two decades
- Automated traffic hit 51% of total internet traffic in 2025 and is growing 8x faster than human traffic
- Three core systems are failing at once: capacity planning, identity management, and request velocity
- Enterprise AI sovereignty requires control of the full stack, from GPUs to connectors, not just the model
The Signal
Barak Yagour, Meta's VP of Engineering for data infrastructure, opened his VB Transform talk wearing Ray-Ban Meta AI glasses, a physical reminder that agents are already here. His message was sharper than the product demo: companies have maybe 20 months to rebuild infrastructure that assumed humans were the only users who mattered.
The numbers make the case. Meta's agentic query load increased 30x in a single half. Not 30 percent. Thirty times. That's the kind of exponential that doesn't flatten out or give you breathing room. It compounds until something breaks.
"What happens to the infrastructure we've spent years building when agents and not humans become the main consumers of that. That's the world we're stepping into."
Yagour identified three failure modes hitting at once. First, capacity planning built for predictable human behavior can't handle agent query patterns that spike without warning. Second, identity systems designed around usernames and permissions don't know what to do when the requester is an autonomous process acting on behalf of someone who's asleep. Third, velocity assumptions collapse when an agent can fire off 10,000 requests in the time it takes a human to type one search query.
The broader data supports Meta's internal experience:
- Bot traffic crossed 51% of total internet traffic in 2025
- AI-driven traffic is growing 8x faster than human traffic
- The inversion happened faster than most infrastructure teams could adapt
Rachad Alao, Cohere's VP of product engineering, added a sovereignty angle that connects to Yagour's infrastructure argument. For banks, hospitals, and governments building agentic systems, control can't stop at the model layer. You need governance over where data lives, which jurisdiction processes it, and whether you can swap vendors without rebuilding from scratch.
Alao's definition of AI sovereignty: tight control over the entire stack. That includes GPUs and private cloud infrastructure, the governance layer routing requests between models, and the connectors and search tools that let agents touch enterprise data. Downloading an open model or running behind a firewall isn't enough when the real risk is losing control of how agents interact with systems you're responsible for.
The timing pressure comes from both directions. Infrastructure built for humans is already breaking under agent load. But building new infrastructure without sovereignty guarantees means locking yourself into dependencies you can't control when regulations tighten or vendor priorities shift.
The Implication
If you're running infrastructure at scale, Yagour's 20-month window is the clock. Capacity, identity, and velocity systems need rethinking before agent load goes from 30x to 300x. If you're building agents for enterprises handling regulated data, Alao's sovereignty checklist is the requirement doc. Can you move your data. Can you change vendors. Can you prove where processing happened. If the answer is no, you're building on someone else's foundation.
The companies that figure this out in the next year will have working systems when agent traffic becomes the majority. The ones that wait will spend 2027 explaining downtime.