Andreessen Horowitz Closes $15B Mega-Fund With Dedicated AI Infrastructure Slice

Andreessen Horowitz closed its largest-ever fundraise in January 2026, raising more than $15 billion across six funds. The allocation tells you exactly where the firm thinks defensibility is going to live in this cycle: $1.7 billion is dedicated to AI infrastructure (compute, data platforms, foundational tooling) and a parallel $1.7 billion sits in an Apps fund focused on AI-powered consumer and enterprise products. Kleiner Perkins announced a $3.5 billion AI-only fund around the same window — one of the largest single-firm AI vehicles ever raised.

What the capital is chasing

The 2026 thesis pattern across top-tier funds is consistent. The thin GPT-wrapper era is over and pre-seed investors have sharpened up dramatically. What's getting funded now:

• Vertical foundation models in domains where general-purpose models are weak — radiology imaging, microbial biology, materials science, robotics, code generation for specific languages and frameworks.
• Agent infrastructure and middleware — search and retrieval, tool catalogs, observability, evaluation platforms, agent identity and trust.
• AI-native services businesses that look more like operating companies than software — done-for-you legal, accounting, customer success, with AI doing the work and humans doing the QA.
• Deep tech and hardware — robotics, edge inference silicon, energy-efficient training, the physical layer.

Vertical concentration in the data

a16z's 2025 funding mix gives a useful shape: roughly 40% healthcare, 25% infrastructure, 20% vertical copilots, 15% entertainment and logistics. Foundation-model bets — once the headline category — now go to a small set of firms with the technical and capital depth to underwrite them: Radical Ventures, Khosla Ventures, a16z, Sequoia, and a handful of strategic investors.

Founder reality in 2026

For founders raising right now, the practitioner consensus across recent YC batches and seed rounds is harsh but consistent: investors want to see a moat that doesn't evaporate when the next OpenAI release ships. That usually means proprietary data, a real workflow embedded in a specific industry, fine-tuning or post-training that creates measurable defensibility, or a services-as-software motion that compounds over time. "We have a better prompt" is no longer an answer. Neither is "we have a better UX over GPT-4o."

"The moat conversation has changed. It used to be 'what's defensible against competitors.' Now it's 'what's defensible against your underlying model being deprecated by a Tuesday release.'"

The good news: search infrastructure, agent middleware, and vertical workflow products are exactly the categories where the capital is. The bad news: positioning into any of them requires technical depth that a year of vibe-coded prototypes won't substitute for.