Bubble and Build: The 2025 Mad (Machine Learning, AI and Data) Landscape (www.mattturck.com)

FirstMark’s eleventh MAD (Machine Learning, AI & Data) Landscape (2025) is a major redraw that shifts the map from chatbots toward systems that actually do work: agent stacks, tool- and memory-enabled models, governed data planes, and local/on‑device LLM runtimes. The map is simpler (logos cut from ~2,000 to ~1,150) and restructured to show a cleaner flow from data → infra → ML/AI → agents/apps, folding “open source” into the stack because open weights now permeate every layer. The landscape and a searchable interactive site/PDF accompany the report, plus a concise list of 25 trend “ideas” that frame the year. For practitioners and builders the significance is clear: capital is abundant but concentrated, creating fragility via circular financing and customer dependence even as real deployments accelerate. Technically, the frontier moved from bigger transformers to reasoning-focused RL (models that allocate compute to deliberation), agent platforms/tooling, robotics foundation models, and hybrid open/closed stacks. Local inference/NPUs, sovereign AI procurement, and energy/grid constraints are emerging chokepoints shaping datacenter and device strategy. Expect more surgical M&A for talent and capabilities, continued open-weights innovation alongside frontier labs, and a practical pivot from leaderboard wins to real-world benchmarks, red-teaming, and outcome-driven signals.