Andreessen Horowitz closed its Machine Age Fund at $1.1 billion committed to AI infrastructure — chips, memory, networking, storage, data centers, and robotics. The fund arrives eighteen months after the initial AI deployment wave, when model builders burned through hyperscaler credits and venture operators started pricing their own silicon.

The firm announced the close Friday through a blog post. No allocation breakdown by category was disclosed. The fund marks a16z's first dedicated infrastructure vehicle separate from its broader AI funds, which have historically blended application-layer and infrastructure bets. The $1.1 billion figure puts it behind Sequoia's separate infrastructure efforts but ahead of most multi-stage firms attempting similar picks-and-shovels plays. The timing is intentional: NVIDIA's H100 lead times remain above 90 days, and custom ASIC design cycles now justify venture checks at the foundry-partnership stage.

This matters because the infrastructure layer is where margin compression happens first and where vertical integration creates the next decade's cost advantages. Hyperscalers are already designing their own training chips — Google with TPUs, Amazon with Trainium, Microsoft with Maia. The venture-backable opportunity sits in specialized accelerators for inference, memory-bandwidth solutions that bypass HBM bottlenecks, and networking fabrics that reduce multi-node training overhead. A $1.1 billion pool allows a16z to lead or co-lead deals in companies building alternatives to the NVIDIA-TSMC stack, which is the only venture-scale bet left in a market where foundation models are capital-intensive and application layers commoditize quickly.

The fund also signals where a16z expects the next 24 to 36 months of AI infrastructure spending to concentrate. Data center capacity remains the binding constraint — U.S. power-grid upgrades for AI-specific facilities are running 18 to 30 months behind demand, and co-location providers are pre-leasing space to hyperscalers at rates that price out most mid-market AI labs. Robotics inclusion in the mandate is notable: embodied AI requires edge inference, which means custom silicon designed for power efficiency rather than raw FLOPS. That market is still pre-scale, but the companies that solve it will sell hardware at margin, not rent compute at cost.

Allocators should track three follow-on signals. First, whether a16z leads any ASIC or networking deals in Q2 2025 — those term sheets will clarify whether the firm is buying into existing roadmaps or funding new architectures. Second, data center co-investment announcements with infrastructure operators or utilities, which would indicate the firm is moving beyond component bets into vertical plays. Third, any portfolio company pivots from inference APIs to in-house silicon, which has already happened twice in the past 90 days among AI application companies a16z backs. Those pivots are expensive and suggest margin pressure is worse than public comments indicate.

The Machine Age Fund closes while inference costs are falling 40% year-over-year and training costs remain flat. The gap is the arbitrage.