Who builds, funds and shapes AI.
The labs, models, software and investors behind frontier AI, the people who lead them and the essays that explain the field. Organised by channel, each linked to its source.
People
16 blocksResearchers, founders and executives whose decisions shape the field: labs, chips, data and the debate about where it goes.
- Sam AltmanCEO of OpenAI. Product cadence, distribution, and capital narrative for the lab that still sets consumer AI defaults.openai.com(opens in a new tab)
- Elon MuskControls the SpaceXAI / X / Grok stack. Real-time distribution plus training cluster ambition in one operator.x.ai(opens in a new tab)
- Mira MuratiFormer OpenAI CTO; CEO of Thinking Machines Lab. Customization and steerable systems over pure scale theater.thinkingmachines.ai(opens in a new tab)
- Andrej KarpathyEx-Tesla AI / OpenAI. Eureka Labs and public teaching that still train how builders actually learn deep learning.karpathy.ai(opens in a new tab)
- Ilya SutskeverOpenAI co-founder; now SSI. Singular bet on safe superintelligence, research taste with maximal stakes.ssi.inc(opens in a new tab)
- Dario AmodeiCEO of Anthropic. Safety-first lab framing, Claude product line, and a clear alternative power center to OpenAI.anthropic.com(opens in a new tab)
- Demis HassabisCEO of Google DeepMind. Research spine behind Gemini and the long game on science-grade AI systems.deepmind.google(opens in a new tab)
- Arthur MenschCo-founder and CEO of Mistral AI. European frontier ambition built around deployable open models and enterprise control.mistral.ai(opens in a new tab)
- Jensen HuangCEO of NVIDIA. Owns the scarce layer, accelerators, networking, and CUDA lock-in every frontier lab rents.nvidia.com(opens in a new tab)
- Leopold AschenbrennerAuthor of Situational Awareness. Timeline, power, and capital framing that still structures serious AGI debate.situational-awareness.ai(opens in a new tab)
- Alexandr WangFounder of Scale AI. Data, eval, and labeling infrastructure that turned into strategic leverage across labs and defense.scale.com(opens in a new tab)
- Aravind SrinivasCo-founder and CEO of Perplexity. Turned cited answers into a consumer search wedge and a direct distribution fight with Google.perplexity.ai(opens in a new tab)
- Liang WenfengFounder of DeepSeek. Quant-funded open weights that repriced assumptions about frontier cost curves.deepseek.com(opens in a new tab)
- Zhilin YangFounder and CEO of Moonshot AI. Long-context research and the Kimi line turned a Chinese assistant into an open-frontier model stack.moonshot.ai(opens in a new tab)
- David HolzFounder of Midjourney. Small team, strong aesthetic taste, image defaults set outside the big-lab PR cycle.midjourney.com(opens in a new tab)
- Peter SteinbergerPSPDFKit founder turned agentic coding operator. Practical signal on what multi-agent dev workflows actually survive contact with shipping.steipete.com(opens in a new tab)
Models
11 blocksThe main model families. Each takes a different approach to price, openness and distribution.
- ChatGPTOpenAI’s consumer and work surface for the GPT line. Default interface habit for non-specialists, distribution as moat.chatgpt.com(opens in a new tab)
- ClaudeAnthropic’s model family. Strong long-context and coding posture; preferred stack for a lot of serious agent work.claude.ai(opens in a new tab)
- GrokSpaceXAI’s model line, wired into X. Real-time retrieval bias and a different product culture than the SF lab consensus.x.ai(opens in a new tab)
- GeminiGoogle DeepMind’s frontier family. Multimodal + Search/Workspace distribution, infrastructure few others can match.gemini.google.com(opens in a new tab)
- MistralMistral AI’s open and commercial model family behind le Chat. A European deployment option with strong enterprise and self-host posture.mistral.ai(opens in a new tab)
- PerplexityAnswer engine with citations as the product. Search replacement wedge; model choice secondary to workflow lock-in.perplexity.ai(opens in a new tab)
- LlamaMeta’s open-weight line. Sets the baseline for self-host, fine-tune, and “good enough offline” deployments.llama.com(opens in a new tab)
- DeepSeekOpen and API models that punch above spend. Reasoning and coding SKUs that force Western labs to justify price.deepseek.com(opens in a new tab)
- KimiMoonshot AI’s model family. K3 brings open weights, native multimodality, and 1M context to long-horizon coding and knowledge work.kimi.com(opens in a new tab)
- QwenAlibaba’s open-weight model family, spanning edge-sized releases to Qwen3.8 Max. Broad modality coverage and serious self-host economics.qwen.ai(opens in a new tab)
- GLMZ.ai’s open-weight model line. GLM-5.3 pushes long-horizon coding and agent work, with cyber capability now part of the risk surface.z.ai(opens in a new tab)
Companies
17 blocksLabs, infrastructure, and the platforms they ship through. Who owns which layer of the stack.
- OpenAIFrontier lab behind ChatGPT and the GPT API. Still the reference customer acquisition engine for generative AI.openai.com(opens in a new tab)
- AnthropicClaude lab. Enterprise and developer trust narrative; constitutional/safety brand as commercial differentiator.anthropic.com(opens in a new tab)
- Safe Superintelligence Inc.Ilya Sutskever’s straight-shot lab with one goal: safe superintelligence. No product sprawl, research conviction at maximal stakes.ssi.inc(opens in a new tab)
- SpaceXAIFormerly xAI; now branded under the SpaceX orbit. Grok, large training clusters, and tight coupling to X distribution.x.ai(opens in a new tab)
- Google DeepMindGoogle’s consolidated research + product AI org. Gemini, infra, and the academic talent density of the old DeepMind core.deepmind.google(opens in a new tab)
- Mistral AIParis frontier lab. Open-model credibility plus enterprise deployment control, Europe’s clearest independent counterweight.mistral.ai(opens in a new tab)
- NVIDIAPicks and shovels with pricing power. GPUs, interconnect, software stack, the bottleneck every trainer queues on.nvidia.com(opens in a new tab)
- MetaOpen-weight counterweight via Llama. Social distribution plus a deliberate strategy to commoditize the model layer.ai.meta.com(opens in a new tab)
- MicrosoftOpenAI’s primary cloud and distribution partner. Copilot surface area across Office, Windows, and GitHub.microsoft.com(opens in a new tab)
- Scale AIData engine and eval infrastructure for frontier training and government work. Bottleneck vendor, not a chat app.scale.com(opens in a new tab)
- DeepSeekHangzhou lab funded by quant profits. Open releases that compressed the perceived cost of competitive reasoning models.deepseek.com(opens in a new tab)
- Moonshot AIBeijing lab behind Kimi. K3 puts open-weight, native multimodal, million-token agent work in the frontier conversation.moonshot.ai(opens in a new tab)
- Alibaba / QwenAlibaba’s model team and cloud distribution stack. Qwen spans small local releases to open Max-class weights at frontier scale.qwen.ai(opens in a new tab)
- Z.aiLab behind the GLM family. Open-weight coding and agent models that keep China’s frontier field broader than DeepSeek alone.z.ai(opens in a new tab)
- Thinking Machines LabMurati’s post-OpenAI lab. Heavy seed capital; thesis on controllable, customizable systems rather than one global chatbot.thinkingmachines.ai(opens in a new tab)
- PerplexitySearch-native AI company. Monetizes answers and research workflow, not raw token dumps.perplexity.ai(opens in a new tab)
- Hugging FaceDefault hub for weights, datasets, and open ML plumbing. Where open models actually get distributed and debated.huggingface.co(opens in a new tab)
Software
7 blocksProducts built on the models, including tools we use in our own work.
- HiggsfieldAI-native creative suite for image and video. Creator-facing production loop, camera control and multi-model generation in one surface.higgsfield.ai(opens in a new tab)
- CursorAI-first code editor. Became the default agentic IDE habit for a large slice of serious builders.cursor.com(opens in a new tab)
- NotebookLMGoogle’s source-grounded research notebook. Strong when the job is “reason only over these docs,” including audio overviews.notebooklm.google.com(opens in a new tab)
- MidjourneyImage generation with a distinct aesthetic lane. Proof a focused product can lead without hyperscaler theatrics.midjourney.com(opens in a new tab)
- ElevenLabsSpeech synthesis, cloning, and dubbing. Infrastructure for voice agents and media pipelines that need believable audio.elevenlabs.io(opens in a new tab)
- ApertusSwiss public LLM effort (ETH/EPFL orbit). Sovereignty and multilingual fit, relevant if you care about CH-local stack choices.swiss-ai.org(opens in a new tab)
- GrokipediaCommunity wiki for models, labs, and AI landscape notes. Fast orientation layer when the map keeps moving.grokipedia.com(opens in a new tab)
Investors
5 blocksInvestors in AI, and the theses behind what they fund.
- Y CombinatorBatch machine for early AI-native startups. Still the densest network effect at pre-seed/seed for technical founders.ycombinator.com(opens in a new tab)
- a16zLarge multi-stage firm with loud AI infra and apps thesis. Capital plus narrative, shapes what gets permission to exist.a16z.com(opens in a new tab)
- Sequoia CapitalClassic multi-stage power. Deep history in defining platforms; still a reference check on serious AI rounds.sequoiacap.com(opens in a new tab)
- Founders FundThiel-orbit firm. Concentrated bets, contrarian posture, comfortable with hard-tech and frontier ambition.foundersfund.com(opens in a new tab)
- Thrive CapitalGrowth investor with outsized OpenAI and internet-software exposure. Quiet compounder behind several defining rounds.thrivecap.com(opens in a new tab)
Insights
3 blocksEssays that still frame the debate, and useful context for the news.
- Situational AwarenessAschenbrenner on AGI timelines, industrial policy, and power. Still the long memo serious people argue with.situational-awareness.ai(opens in a new tab)
- The Last QuestionAsimov, 1956. Entropy, computation, and the end state of asking machines for answers, useful myth for the current buildout.users.ece.cmu.edu(opens in a new tab)
- What 2026 Looks LikeKokotajlo’s scenario sketch of near-term AI. Reference point for how forecasting aged once the year arrived.lesswrong.com(opens in a new tab)