Open with the price chart: in late 2024, DeepSeek-V3 shipped a 671B-parameter MoE that matched GPT-4-class quality at roughly 1/30th the inference cost, and DeepSeek-R1 followed with reasoning at frontier parity. OpenAI cut GPT-4o pricing twice in the next two quarters, Anthropic released Claude 4 and a Haiku tier priced to compete, and Google's Gemini 2.5 / 3 family undercut both on long-context. 68 published labs now sit under that pricing pressure, with $293B in cumulative disclosed funding — the most capital-intensive category on the platform.
The two-lab gravity well
OpenAI ($193.3B raised, ~$852B valuation, Series E) and Anthropic ($67.6B raised, Series G) absorb most of the western enterprise spend. Mistral AI ($6.3B raised including debt, France) holds the European open-weight position; Black Forest Labs ships frontier image generation from Germany. China runs a parallel stack: Zhipu AI ($1.49B, GLM-4 family) and MiniMax ($2.2B, IPO'd, Hailuo video plus Talkie) compete inside markets the western labs cannot serve directly. Behind them, Cohere ($1.77B Series E ext.), AI21 Labs ($336M Series C), 01.AI ($200M Series A, Yi family), and Reka AI ($180M Series B) defend narrower enterprise and multimodal slices.
The cost compression below
DeepSeek's reported $5.6M training run for V3 — even with the usual caveats about hardware accounting — broke the assumption that frontier-class quality required nine-figure compute spend. Llama 4 from Meta extended the open-weight pressure. Liquid AI ($300M Series B) is going the other direction with state-space architectures aimed at edge and on-device deployment. Skild AI ($2.2B Series C) and Physical Intelligence ($735M Series B) are training models for robotics, where the data bottleneck is the moat, not the GPU budget. 20 disclosed rounds in the trailing 12 months averaged $9.5B — the highest of any NeuronFeed category and an order of magnitude above the platform median.
What 2026 actually tests
Whether scaling laws hold above $1B per training run. Ineffable Intelligence raised $1.1B at seed on a pure scaling thesis. Safe Superintelligence raised $3B Series B for the same. If quality-per-dollar keeps compressing the way DeepSeek and Mistral have shown, the value moves to whoever owns distribution, fine-tuning, and proprietary data. If the next training generation produces another step-change, capital concentration tightens further.