Large Language Models

Capital is concentrating in a few frontier labs, not broadening

Velocity is cooling (-0.36), with 21 deals in 28 days, but $20.5B of capital, so average deal size is huge while deal count shrinks. Weekly flow swings from $31.4B across 13 deals (Aug 10) and $19.5B (Sep 7) to near-zero weeks in mid-September and late September, then $15B across just 3 deals the week of Oct 19. A $120B growth round was flagged, and Series D+ rounds (5 deals) account for $29B versus about $3M in seed across 12 deals. Nvidia (11 deals), Google, and OpenAI (6 each) are the most active investors, so strategic capital rather than classic VC is setting the pace.

Anthropic's revenue reportedly hit $16.3B in H1 2026 versus $1.1B a year earlier, while it spent nearly $9B on compute and disclosed more than $600B in future infrastructure commitments, with up to another $150B under negotiation. It plans to funnel 5% of IPO shares into a fund for job-market harms, and Eric Schmidt reportedly holds 1%. However, delayed OpenAI and Anthropic IPOs are making LPs cautious about Q4 deployment.

Why it matters · The IPO timing of OpenAI and Anthropic is the main valuation catalyst, and its delay is freezing downstream VC allocation decisions.

New releases emphasize cost and speed rather than raw capability. OpenAI's GPT-6.1 Terra matches GPT-6 Astra at one tenth the cost and 3x the speed, and MiniMax M3.1 Flash is 3x faster and 70% cheaper for agents. Google's Gemini 4 Carbon stresses long-horizon reasoning and cybersecurity, and DeepSeek V4.1-Flash is getting a dedicated inference stack. Price-performance is becoming the competitive axis.

Why it matters · Rapid per-token price collapse compresses margins for model resellers and rewards labs with the cheapest compute.

OpenAI models and Codex are now available through Google Cloud's Model Garden and Marketplace, and customers can draw down existing cloud commitments to buy them. AWS's CEO said customers wanted OpenAI models next to their data, and Google sells a competitor's models alongside Gemini. Exclusive cloud partnerships such as Microsoft-OpenAI are eroding in favor of multi-cloud availability.

Why it matters · Model labs gain reach across all hyperscalers, while clouds compete on being the neutral marketplace rather than on a house model.

OpenAI's Dots (ChatGPT Pro, from $100/month) and Meta's Muse ($20 and $100 tiers) launched as personal agents that handle multistep tasks, though authentication still blocks purchases. Codex and Claude Code are described as standard in every company's toolbox, and Claude for Excel points to Office work as the next agent surface. Meta's Muse also suspended sharing after a data-exposure issue, showing that reliability and safety gate adoption.

Why it matters · Paid agent tiers at $100/month are establishing a new ARPU ceiling, but trust and security failures can stall adoption.