Meta Pours Up to $145 Billion While Agents Lag: That Gap Is the Market

What Happened

TechCrunch reported on July 2 what Mark Zuckerberg told employees at an internal town hall: AI agent development hasn’t “accelerated in the way” Meta expected. This lands after a year in which the company laid off roughly 8,000 people, about 10% of its workforce, and reassigned 7,000 more into AI-focused groups, including a unit literally named Agent Transformation. Zuckerberg conceded the restructuring wasn’t clean and that its expected benefits “hadn’t come to fruition yet,” predicting improvement over the next three to six months. The money, meanwhile, moves the other way. Meta’s 2026 capex guidance now runs $125 billion to $145 billion, nearly double last year’s spend. And the same week, Google’s and Amazon’s sustainability reports showed the bill underneath the buildout: per TechCrunch’s analysis, Google’s emissions rose 25% year over year and Amazon’s 16%, Google’s Scope 3 emissions have doubled since the 2019 baseline, and Amazon added over 1.2 gigawatts of data center capacity in Q4 2025 alone.

What This Means for Founders

Put the numbers side by side and the gap is the story. Record capital is flowing into infrastructure while the thing that’s supposed to run on it underdelivers inside the best-resourced company on earth. Meta reorganized itself around agents, threw 7,000 people at the problem, and still had to tell staff it isn’t working yet. The wrong read is “agents don’t work.” The right read is that demos work and production doesn’t, and the plumbing between the two hasn’t been built.

That plumbing has three missing pieces. Reliability: agents collapse partway through long tasks, and someone has to find where. Cost: inference bills compound with usage, and the Google/Amazon reports say the energy and manufacturing costs underneath won’t melt away, betting your margins on inference getting radically cheaper is a hope, not a plan. Measurement: without tooling that shows what an agent completed, broke, or hallucinated, no enterprise buyer signs off.

For anyone raising or building in the Valley right now, this is a classic picks-and-shovels setup. The problems big tech can’t solve by reorg are exactly the ones that leak out to startups: evaluation and rollback layers that make agents auditable, cost-control layers that route across models and cache aggressively, observability that turns “it seems to work” into completion rates. And if you’re building an agent itself, the Meta lesson cuts the other way, don’t fight a general-agent war against a company spending $145 billion. Pick one narrow workflow where failure is cheap and completion rates are provable, and own it end to end. VCs burned by agent demos in 2025 now ask for exactly those numbers.

What You Can Do Now

If you sell an agent product, lead with three numbers instead of a demo: task completion rate, human interventions per task, and cost per completed task. If those aren’t on a dashboard yet, that’s this week’s work. If you’re buying agents, don’t inherit Zuckerberg’s three-to-six-month timeline, pick one narrow workflow, run it, and measure completion against cost yourself. Your own numbers will beat any vendor deck.