Meta infrastructure VP warns of 20-month window to rebuild for AI agents
Meta VP of Engineering Barak Yagour outlined a major overhaul of the company's data infrastructure, driven by an identified need to accommodate a 30x increase in agentic AI queries. The transition replaces batch ETL processing with real-time streaming architectures and introduces schema-aware storage to optimize GPU utilization for recommendation and reasoning systems.
Key Takeaways
- Agentic queries at Meta grew 30x in six months, outpacing the capacity of infrastructure built over the last two decades.
- Data systems are shifting from batch ETL to real-time streaming to support low-latency ranking and recommendation pipelines.
- Meta is targeting 500 million queries per second and a petabyte per second of throughput for training data reads.
- New trusted data environments use real-time masking and tracing to allow agents to explore data while maintaining human-level governance.
- Storage is being redesigned as schema-aware to pull specific columns and time ranges, aiming to eliminate idle GPU capacity.
Why It Matters
The traditional compute-to-user ratio is effectively dead. As individual engineers begin spawning dozens of autonomous agents, legacy infrastructure cannot scale with the resulting exponential load. For the streaming industry, this shift indicates that pattern-matching algorithms are being replaced by reasoning models that require full behavioral history rather than summarized signals. To maintain competitive recommendation engines, platforms must migrate from 24-hour batch cycles to streaming architectures that can react to user intent in milliseconds. Watch for whether Meta successfully open-sources these "agent-aware" infrastructure components to set a new standard for enterprise data governance.
Additional Context
The urgency in Meta’s timeline corresponds with a broader industry shift toward AI-centric hardware and cloud services. Per Reuters in July 2026, Meta CEO Mark Zuckerberg cautioned that while the company is spending up to $145 billion on AI infrastructure this year, the initial progress of its internal 'agentic development' has been slower than expected. This friction has prompted a strategic pivot; Bloomberg reported on July 1, 2026, that Meta is developing 'Meta Compute,' a business line to rent out excess AI computing power and sell API access to its Muse Spark models. This move directly challenges the cloud dominance of Amazon and Microsoft by monetizing the same infrastructure Yagour is now rebuilding. On the hardware front, Meta is scaling production to feed these new agent-aware systems. According to data from iDevice in July 2026, Meta and partner EssilorLuxottica have sold over 9 million lifetime units of AI-enabled glasses, including the Ray-Ban Meta Gen 2. The upcoming Gen 3 model, expected at Meta Connect in September 2026, is rumored to significantly increase 'Live AI' endurance, further driving the high-frequency agentic queries that Yagour’s team must now support at exabyte scale. Combined with human-level internet traffic being surpassed by automated traffic—reaching 53% in 2026 per Imperva—the pressure to move code from human-readable SQL to agent-optimized interfaces has moved from a theoretical exercise to a core engineering requirement for the next 20 months.
Read full article at venturebeat.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source