AI, Infrastructure, and the Next Investment Cycle

Summary · 4 min read instead of 52 min

David George, Sarah Wang, Alex Immerman and Santiago Rodriguez are colleagues on a16z's growth team. The team presents its State of Markets deck every year, and George hosts this walk through 25 of its charts.

The team argues the market's rise is backed by earnings rather than higher multiples, and that hyperscaler CapEx of about $780 billion in 2026 still trails demand for compute. They say enterprise AI adoption is early: only 2% of S&P 500 companies track results over time, and power users spend far more than the median. Falling inference costs, routing and fine-tuning are making agents cheaper and improving app margins. They expect public software to split into AI winners and losers, with revenue growth acceleration the test. They close with areas they are watching, including consumer agents, robotics, autonomy, bio, personal health and defense.

Minute by minute

1. 16:53

   Lab revenue outpaces the best software companies. Wang says combined annualized revenue at OpenAI and Anthropic has climbed so fast that the team updates its chart monthly. Estimated net new revenue at the labs has well surpassed that of the best software companies ever built. Rodriguez adds that AI is early both in share of population using it and in share of wallet among users.

2. 18:08

   Enterprise AI results are barely tracked. Immerman cites 69% of S&P 500 companies with live deployments, 30% with quantifiable impact and 2% tracking a metric over time. He says the next stage is deep, recurring workflows. George says the gap between what models can do and how they are used favors application companies, and he cites Revolut's use of Eleven Labs voice for banking workflows.

3. 20:11

   Power users pull away on AI spend. Wang cites YipitData showing median AI vendor spend in the top 1% is roughly eight times that of the top 10%. Inside AI-native portfolio companies, top users spend $7,500 to $9,000 a month against $200 to $400 for the median, over 20 times. George says forward-leaning Fortune 500 companies spend about 1% of headcount cost on AI tools, against as much as 10% at the most AI-native portfolio companies.

4. 22:40

   AI gains show in cost and revenue. Rodriguez says Chime reports cutting cost to serve by over 10% a year for four years, almost 50% compounded, and Shopify's Sidekick lifted the share of customers reaching five orders in 15 days by 8%. Wang says ServiceNow reports over a billion in AI ACV and a 9x rise in agentic deployments. George treats where companies put AI dollars, cost or new products, as a litmus test for their revenue opportunity.

5. 25:38

   Cheaper inference makes agents economical. Immerman ties 14x growth in agent token usage on OpenRouter to multistep tasks. Caching made Hebbia's financial chat workloads 10x cheaper, Databricks' router solves more problems at 35% lower cost than the strongest single model, and Elise's fine-tuned model is 60% cheaper with lower latency. George expects margin improvements at the best app companies.

6. 28:35

   Consumer AI subscriptions are just starting. George cites a survey that just over 2% of U.S. households pay for an AI subscription, with retention that smiles as products improve. Rodriguez compares this with over 200 million households for Amazon Prime and 70 million for Netflix. Wang quotes Josh Elman's view that consumer AI is what people use daily, not what they expense, and says always-on agents will make engagement harder to measure.

7. 31:28

   Agents threaten marketplace advertising profit pools. Immerman says Amazon declined to connect to Muse while Instacart agreed. He argues Amazon gains few new orders from agents, while Instacart, with grocery online penetration still early, stands to gain more. George notes Amazon's advertising business of over $70 billion depends on consumers clicking ads. Rodriguez says reallocated ad spend and lower friction could raise overall consumption.

8. 36:01

   Public software splits into AI winners and losers. Wang says roughly 75% of the public software sample is profitable and only 30% grows at 20% or more. George argues SaaS companies must show revenue acceleration, and he suggests a 10% target. Wang says cybersecurity, observability and vertical software have held up better, and Immerman says the index is back to where it started the year but divided into AI winners and losers.

9. 42:10

   Private markets reach public-market scale. Wang says Stripe's SaaS customer data shows growth accelerating into 2026, which Stripe calls the renaissance. Rodriguez says the top six private companies add up to about $2.4 trillion by last round valuation. George calls the founder the asset class. Carta data shows 58% tender participation, and secondary discounts to last round are basically zero.

10. 48:30

    Next areas: agents, robotics, autonomy, bio. George lists consumer agents, robotics, autonomy, AI and biology, personal health, enterprise diffusion beyond coding, and defense. He says robotics could be larger than LLMs, three to five years earlier. He says Uber and Lyft are about 1% of U.S. miles and could grow at least tenfold with autonomous cars he describes as 14 times safer. Newer defense vendors hold under 5% of military spending.

Published Sep 30, 2026.