Meta’s $145B AI Bet: From the Iris Chip to Meta Compute and the Agent-Infrastructure Warning

The chip is part of a broader, unusually fast roadmap. In March 2026, Meta announced plans to deploy four new chip generations — MTIA 300, 400, 450, and 500 — by the end of 2027 . According to an internal memo, Iris cleared bug testing in six weeks with no major issues .

Meta's computing power target is 14 gigawatts next year as part of this push . That scale reflects an ambition to move beyond supplementing Nvidia and AMD GPUs toward a more self-sufficient AI silicon strategy.

2. Infrastructure Spending and the Meta Compute Cloud

Meta's 2026 capital expenditure projection is between $115 billion and $145 billion, roughly double the $72 billion it spent in 2025 . Some analysts estimate the final number could reach $135–145 billion . Meta has contracted over 5 GW of data center capacity in just the first half of 2026, and research firm SemiAnalysis projects 2027 capex will be "shockingly high" . CEO Mark Zuckerberg has said Meta plans to build "tens of gigawatts this decade, and hundreds of gigawatts or more over time" .

In July 2026, Meta officially established a new cloud infrastructure division called "Meta Compute" to sell excess AI computing capacity to outside customers . The move, first reported by Bloomberg on July 1, 2026, sent Meta shares up more than 9%, while neocloud rivals CoreWeave and Nebius fell roughly 11–12% . Meta is still deciding whether to offer access to AI models hosted on its infrastructure, raw computing power, or both .

Zuckerberg first signaled the possibility in May 2026, telling shareholders that a cloud business was "definitely on the table" . The company carried $182.9 billion in future lease obligations as of March 31, 2026, and has committed to at least $600 billion in U.S. infrastructure spending by 2028 .

3. Barak Yagour at VB Transform 2026: The Warning on Enterprise Infrastructure

Caveat: VB Transform 2026 took place July 14–15, 2026 at Hotel Nia in Menlo Park . As of this writing, no detailed transcript or news article quoting Barak Yagour's specific remarks from this event has been indexed yet . The exact warning he issued about enterprises needing to urgently rebuild infrastructure for AI agents cannot be independently cited at this moment.

However, available context strongly supports the direction of his message:

• Barak Yagour is Meta's VP of Engineering for AI and Data Infrastructure, leading teams that build and operate Meta's AI and data systems — from recommendation infrastructure to generative AI serving . He was listed as a keynote speaker at VB Transform 2026, whose theme was "Orchestrating AI Autonomy" .
• In a June 2026 keynote at the AI & Data conference at Meta's campus, Yagour said: "Agents are among us, and we are heading for a world where humans move from spending time in the loop to spending time over the loop" . He described agents as not just a feature, but an inversion of how infrastructure itself is built .
• The "Agentic Infrastructure Gap" — the term used by KPMG and cited by VB Transform — describes the reality that only 11% of organizations have successfully incorporated AI agents into their operations .
• Meta's own CEO Mark Zuckerberg told employees in early July 2026 that AI agent development "has not accelerated in the way we expected" , after laying off roughly 8,000 people and reassigning 7,000 engineers to agent-focused teams .

Given Meta's $115B+ infrastructure bet and Yagour's existing public statements on the agentic future, the plausible substance of his VB Transform warning is: current enterprise infrastructure — security, governance, data integration, orchestration — is fundamentally not ready for autonomous AI agents, and organizations need to start rebuilding it now.

What This Means

Meta is simultaneously building a custom chip to reduce its Nvidia dependence, launching a cloud business to monetize its massive infrastructure investment, and publicly confronting the reality that AI agents are not yet living up to expectations. The three themes are connected: the Iris chip and Meta Compute are bets that the infrastructure for AI agents will eventually be enormous, but the current gap between ambition and execution is real — inside Meta and across the enterprise.