Deep Learning

Mega-rounds at trillion-dollar valuations are swamping everything else

Capital is concentrating in a few frontier-lab raises. A $30B growth round at a $1.4T valuation [47] and a $120B growth round [8] account for most of the $22.1B captured in the 28-day window and the $15B week of Oct 19. Meanwhile weekly deal counts have fallen from 16 to 4, and velocity is cooling at -0.40. The stage mix shows $33B in series D+ across only 7 deals, against 13 seed deals totaling $46M. Anthropic's disclosed $600B+ in future infrastructure commitments [19][28] shows what the money is being raised to fund.

Anthropic's revenue reportedly reached $16.3B in H1 2026, up from $1.1B a year earlier, while it spent nearly $9B on compute [28]. Its infrastructure obligations exceed $600B, with another $150B under negotiation [19]. At the same time, delayed OpenAI and Anthropic IPOs are cited as a source of VC market uncertainty [27], and an IPO round signal has appeared [22].

Why it matters · Public-market readiness now depends on whether revenue growth can outrun compute obligations, which sets exit timing for the whole ecosystem.

Google Cloud is now selling OpenAI models and Codex through Model Garden alongside Gemini [20][25][11], and AWS says customers wanted OpenAI models where their data already lives [18]. This shows that cloud channels, not model exclusivity, are becoming the moat. Alongside it, MiniMax's M3.1 Flash claims roughly 3x speed and 70% lower cost than US frontier models [42][46], and French software firms moved to Qwen after US access limits [31].

Why it matters · Pricing power at the model layer is eroding, which favors the clouds and cheap open-weight providers over single-model vendors.

OpenAI's Dots and Meta's Muse launched on the same day [44][40] and handled complicated multistep tasks, though both stumbled on authentication [38][39][43]. GPT-6.1 Terra offers computer use at one tenth the cost of GPT-6 Astra [2]. Meta's Muse also had a data-exposure incident via shared links [12].

Why it matters · Agent reliability and security, not raw capability, will decide which assistants earn payment and account access.

Google's Genie 4 real-time world model [1] and Harness², which lets agents recursively evolve their own prompts, tools and memory [33], point to a shift toward generative environments and self-improvement. Gemini 4 Carbon emphasizes long-horizon reasoning and cybersecurity [0]. OpenAI's purchases of Orchestra Bio and another biotech startup [10][24][26] show labs pushing into science.

Why it matters · Value is moving from chat models toward simulation and autonomous research loops, which opens new categories beyond language.