Foundation Models — The frontier labs
Fable 5 · data curated through July 2026
Layer 3 of five in the Five Layer AI Cake. 39 companies, 63 events.
What this layer is
The model itself — architecture, weights, capabilities. The narrow funnel the whole stack aims at.
Layer 3 of five. Below it sits Cloud & Compute, which it pays. Above it sits Tooling & Orchestration, which pays it.
39 of the atlas's 103 companies capture value here. 26 of them also capture value on another layer, which makes them vertical integrators; 13 compete on this layer alone.
Vertical integrators on this layer (26)
These companies span two or more layers, so they hold margin at every floor they occupy.
| Company | Founded | Also on | Influence | Why it is here |
|---|---|---|---|---|
| Microsoft Azure | 2008 | Cloud & Compute · Applications | 10/10 | OpenAI infra. Copilot everywhere. |
| OpenAI | 2015 | Applications · Tooling & Orchestration | 10/10 | ChatGPT, GPT-4, Sora, o-series. Defined the category. |
| Anthropic | 2021 | Applications · Tooling & Orchestration | 9/10 | Claude, MCP, constitutional AI. Safety-first frontier lab. |
| Google Cloud | 2008 | Cloud & Compute · Silicon · Applications | 9/10 | TPUs + Gemini + Workspace. Only full-stack vertical. |
| Hugging Face | 2016 | Tooling & Orchestration | 8/10 | The GitHub of ML. 1M+ models hosted. |
| Meta AI | 2013 | Tooling & Orchestration | 8/10 | Llama, PyTorch — the open-weight anchor. |
| SpaceX·xAI | 2002 | Applications | 8/10 | Public markets priced a model lab inside a rocket company. Owns Cursor now. |
| Databricks | 2013 | Cloud & Compute · Tooling & Orchestration | 7/10 | Acquired MosaicML ($1.3B). Enterprise AI plumbing. |
| Waymo | 2009 | Applications | 7/10 | The first physical AI at scale. 10M+ rides/month by 2025. |
| Apple Intelligence | 2024 | Applications | 6/10 | On-device Foundation Models + Vision Pro. Privacy-first, slow, consumer-scale. |
| ByteDance Doubao | 2023 | Applications | 6/10 | Doubao 1.5 Pro. TikTok's parent quietly shipped a frontier Chinese model. |
| Midjourney | 2021 | Applications | 6/10 | Image gen leader; profitable without VC. |
| Perplexity (co.) | 2022 | Applications | 6/10 | Builds own search index + fine-tunes. Vertical from day one. |
| Physical Intelligence (π) | 2024 | Applications | 6/10 | π₀ — general-purpose robot foundation model. Levine + Hausman + ex-Google Brain. |
| xAI | 2023 | Applications | 6/10 | Grok. Colossus supercluster in Memphis. |
| Baidu ERNIE | 2019 | Applications | 5/10 | ERNIE 4.5 + Bot. Anchor of China's home-field LLM ecosystem. |
| ElevenLabs | 2022 | Applications | 5/10 | Voice cloning + TTS. Owns audio. |
| Krutrim | 2023 | Applications | 5/10 | Ola's AI lab. Open-sourced IndicLLM 7B/12B. Bhavish Aggarwal-backed. |
| Moonshot (Kimi) | 2023 | Applications | 5/10 | Kimi k1.5 — China's first reasoning-model breakout. Million-token context. |
| NVIDIA GR00T · Isaac | 2024 | Tooling & Orchestration | 5/10 | Humanoid foundation model platform + Isaac simulation. Robotics dev platform. |
| Runway | 2018 | Applications | 5/10 | Gen-2/Gen-3 video models. Hollywood’s AI vendor. |
| Wayve | 2017 | Applications | 5/10 | End-to-end learned driving + GAIA world model. UK AV champion. |
| Figure | 2022 | Applications | 4/10 | Humanoid robots. Ex-OpenAI Helix model. |
| MiniMax | 2021 | Applications | 4/10 | MiniMax-01 with MoE attention. Hailuo video gen. |
| Suno | 2022 | Applications | 4/10 | Music generation. Current RIAA defendant. |
| Inflection | 2022 | Applications | 3/10 | Pi assistant. Team acqui-hired by Microsoft (2024). |
Pure plays on this layer (13)
| Company | Founded | Also on | Influence | Why it is here |
|---|---|---|---|---|
| Google DeepMind | 2010 | — | 9/10 | AlphaGo → Gemini. Research arm of Google AI. |
| Alibaba Qwen | 2023 | — | 7/10 | Qwen3 72B+. Aggressively open-weight, fastest-closing competitor to US labs. |
| AI4Bharat | 2019 | — | 6/10 | IIT Madras research. IndicNLP, IndicBERT, IndicTrans. The academic spine of Indic AI. |
| DeepSeek | 2023 | — | 6/10 | Shocked the market Jan 2025 with R1. MoE efficiency. |
| Sarvam AI | 2023 | — | 6/10 | Indic-first foundation models. Sarvam-1, Sarvam-2B. Lightspeed + Khosla-backed. |
| Mistral | 2023 | — | 5/10 | European frontier lab. Open-weight pragmatism. |
| Safe Superintelligence | 2024 | — | 5/10 | Sutskever’s post-OpenAI lab. Zero products, $30B valuation. |
| World Labs | 2024 | — | 5/10 | Fei-Fei Li's spatial-intelligence lab. $230M seed 2024, the highest-profile launch. |
| Cohere | 2019 | — | 4/10 | Enterprise-focused. RAG-first positioning. |
| Skild AI | 2023 | — | 4/10 | General-purpose embodied foundation model. $300M raised 2024. |
| Thinking Machines | 2024 | — | 4/10 | Mira Murati’s lab. Post-OpenAI brain drain continues. |
| Zhipu AI | 2019 | — | 4/10 | GLM-4. Tsinghua spin-off; hit by US entity list 2025. |
| AI21 Labs | 2017 | — | 3/10 | Jamba, early GPT-3 competitor. |
What happened here (63 events)
| Year | Event | Magnitude | What happened |
|---|---|---|---|
| 1950 | Theseus (Shannon) | 2 | Shannon's relay mouse learns a maze at Bell Labs — machine learning shown in hardware, six years before AI has a name. |
| 1956 | Dartmouth Workshop | 3 | The term "artificial intelligence" is coined. |
| 1958 | Perceptron (Rosenblatt) | 2 | First trainable neural network. |
| 1986 | Backprop popularized | 2 | Rumelhart, Hinton, Williams. Key algorithm returns. |
| 1997 | LSTM (Hochreiter, Schmidhuber) | 1 | Sequence modeling foundation for 15 years. |
| 2010 | DeepMind founded | 2 | Hassabis + Legg bet on AGI-scale RL. (Google DeepMind) |
| 2012 | AlexNet crushes ImageNet | 3 | GPU + CNN + data. The deep-learning era begins. |
| 2013 | Meta AI (FAIR) founded | 1 | LeCun joins Facebook. Open research ethos. (Meta AI) |
| 2013 | Word2Vec | 1 | Mikolov et al. — vector semantics. The prehistory of embeddings. |
| 2014 | GANs (Goodfellow) | 2 | Generative models take flight. |
| 2015 | OpenAI founded | 3 | Non-profit mission to ensure AGI benefits all humanity. (OpenAI) |
| 2017 | "Attention Is All You Need" | 3 | Transformer paper. Rewrote the field. |
| 2018 | BERT + GPT-1 | 2 | Pretrain-then-fine-tune becomes default. (OpenAI) |
| 2018 | AlphaFold 1 | 1 | Structural biology becomes tractable. Nobel bait. (Google DeepMind) |
| 2019 | GPT-2 — "too dangerous to release" | 2 | Scaling laws hint. Partial release controversy. (OpenAI) |
| 2020 | GPT-3 (175B) | 3 | Few-shot learning. The first time the world noticed. (OpenAI) |
| 2020 | AlphaFold 2 cracks CASP14 | 2 | "A solution to the protein folding problem." (Google DeepMind) |
| 2020 | Scaling Laws for Neural LMs | 2 | Kaplan et al. — the paper that justified every subsequent billion. (OpenAI) |
| 2021 | Anthropic founded | 2 | Amodei siblings + ex-OpenAI team. Safety-first. (Anthropic) |
| 2022 | Stable Diffusion open-sourced | 2 | Open-weight image gen explodes creativity (and lawsuits). |
| 2022 | JEPA (LeCun) | 1 | Joint-Embedding Predictive Architecture. LeCun's quiet bet against pure transformers. (Meta AI) |
| 2023 | GPT-4 | 3 | Genuinely useful across domains. Multimodal. (OpenAI) |
| 2023 | Claude 2 public release | 2 | 100K context window — a new frontier. (Anthropic) |
| 2023 | Llama + Llama 2 open-weights | 3 | Meta torpedoes the closed-model moat. (Meta AI) |
| 2023 | Gemini 1.0 | 2 | Google finally ships. Multimodal from birth. (Google DeepMind) |
| 2023 | xAI + Mistral founded | 1 | Musk and ex-Meta French team enter the race. (xAI) |
| 2023 | Amazon invests $4B in Anthropic | 2 | Google follows with $2B. Frontier labs become cloud-anchored duopolies. (AWS) |
| 2023 | Mamba / State-space models | 1 | Gu + Dao — linear-time sequence models. The first credible transformer alternative. |
| 2023 | Qwen + Baidu ERNIE launch | 1 | China's frontier-lab race accelerates. Qwen opens up aggressively. (Alibaba Qwen) |
| 2024 | Claude 3 + GPT-4o + Gemini 1.5 | 3 | Capability parity at the frontier. Long context becomes normal. (Anthropic) |
| 2024 | OpenAI o1 — reasoning chains | 3 | Test-time compute = new scaling axis. (OpenAI) |
| 2024 | Meta Llama 3 (405B open) | 3 | Frontier-class weights released freely. Zuck doubles down. (Meta AI) |
| 2024 | Inflection team → Microsoft | 1 | Regulatory workaround for the modern era. (Inflection) |
| 2024 | Ilya Sutskever launches SSI | 2 | Raises $1B at $5B pre-revenue. Superintelligence as a product. (Safe Superintelligence) |
| 2024 | xAI raises $6B at $24B | 1 | Musk assembles 100K H100s in Memphis in months. (xAI) |
| 2024 | V-JEPA 1.5 (self-supervised video) | 1 | LeCun's roadmap to autonomous machine intelligence, one video frame at a time. (Meta AI) |
| 2024 | "Situational Awareness" published | 2 | Leopold Aschenbrenner's essay reframes AGI as sovereign infrastructure. Shifts Capitol Hill. |
| 2024 | Physical Intelligence ships π₀ | 3 | First generalist robot foundation model. VLA policies cross form factors — the GPT-3 moment for embodiment. (Physical Intelligence (π)) |
| 2025 | DeepSeek R1 shocks the market (Jan) | 3 | GPT-4-class reasoning, open weights, trained on a shoestring. NVIDIA drops 17% in a day. (DeepSeek) |
| 2025 | Claude 3.7 + Opus 4 | 2 | Agentic coding capabilities become table stakes. (Anthropic) |
| 2025 | Mira Murati → Thinking Machines | 1 | OpenAI brain drain accelerates. (Thinking Machines) |
| 2025 | Anthropic $3.5B at $61.5B | 2 | Valuation triples in 12 months. Enterprise Claude adoption compounds. (Anthropic) |
| 2025 | OpenAI $40B round ($300B+ valuation) | 3 | Largest private funding round in history. SoftBank anchor. (OpenAI) |
| 2025 | V-JEPA 2 + Genie 2 | 1 | World-model architectures go multi-modal + controllable. Parallel track to LLM scaling. (Meta AI) |
| 2025 | Moonshot Kimi k1.5 reasoning breakout | 1 | First Chinese reasoning-tier model. Million-token context beats US offerings. (Moonshot (Kimi)) |
| 2026 | Claude Opus 4.7 (1M ctx) | 2 | Long-context reasoning becomes practical default. (Anthropic) |
| 2026 | Genie 3 mainstream · video world models | 2 | Agents trained in generated environments. Sim-to-real pipeline becomes default. (Google DeepMind) |
| 2026 | ByteDance Doubao 1.5 Pro hits parity | 1 | TikTok's parent closes the gap. Open-weight strategy erodes US closed moats. (ByteDance Doubao) |
| 2026 | Washington switches off a frontier model | 3 | Commerce export-control order forces Fable 5 and Mythos 5 dark worldwide — the first government recall of a commercial frontier model. (Anthropic) |
| 2026 | Claude Fable 5 + Mythos tier ships | 2 | A Mythos-class model on general sale — for exactly three days. (Anthropic) |
| 2026 | GPT-5.6 launches to a government-approved list | 2 | Twenty partners vetted by Washington; OpenAI says gating 'shouldn't become the default.' GA July 9. (OpenAI) |
| 2026 | Fable 5 returns; Sonnet 5 ships | 2 | Order lifted after 18 days; Sonnet 5 lands near-Opus agentic performance at Sonnet prices. (Anthropic) |
| 2026 | SpaceX IPOs at $1.77T on an AI story | 2 | Largest IPO ever ($75B raised); ~90% of the pitch was the xAI subsidiary. (SpaceX·xAI) |
| 2026 | Gemini 3.5 + Omni at I/O; Ultra cut to $200 | 2 | Distribution beats margin — the price cut said the quiet part. (Google DeepMind) |
| 2026 | DeepSeek takes outside money at last | 2 | $7.4B first external round; Tencent and CATL took lock-ups and zero votes — the state fund kept the voting rights. (DeepSeek) |
| 2026 | Grok 4.5, trained alongside Cursor | 2 | First visible product of the Cursor deal, three weeks after closing. (xAI) |
| 2026 | EU softens the AI Act, keeps the GPAI hammer | 1 | High-risk deadlines pushed 16 months; 3%-of-turnover GPAI fines still start August 2. |
| 2026 | Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures | 2 | Detectors of alignment failures screen deployed language models and score alignment benchmarks without task-specific fine-tuning. (Meta AI) |
| 2026 | ViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generation | 2 | Few-step autoregressive video diffusion enables low-latency streaming generation using representation distribution matching. |
| 2026 | MOPD-Router: Rethinking Teacher Routing in Multi-Teacher On-Policy Distillation | 2 | Multi-teacher on-policy distillation (MOPD) integrates specialized capabilities into a single student, but existing practice typically ha... |
| 2026 | Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks | 2 | Merging low-rank adapters (LoRAs) promises to eliminate the overhead of swapping task-specific weights at inference time. However, existi... |
| 2026 | Disaggregated Quantization: Specializing LLM Prefill and Decode | 2 | Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact wei... |
| 2026 | Recursive Harness Distillation across Agents for Robot Manipulation | 2 | A central goal in robotics is to enable manipulation across changing tasks and environments. Vision-language-action (VLA) models provide ... |
Claims that rest on this layer
- Layers are value capture, not operations — Companies sit where they capture value, not where they operate. That single rule is what makes the cake a map instead of a directory.
The rest of the stack
- 1 · Silicon — GPUs, TPUs, ASICs
- 2 · Cloud & Compute — Hyperscalers, neoclouds
- 3 · Foundation Models — you are here
- 4 · Tooling & Orchestration — SDKs, agents, frameworks
- 5 · Applications — End-user products
Or read the full index, or open the interactive view filtered to this layer.
The Five Layer AI Cake is published by ConeApex, the trading name of Nirukta Digital XR Technologies Private Limited (CIN U62090TS2025PTC198275), Hyderabad, Telangana, India. The dataset is licensed CC BY 4.0 — cite it freely. See llms.txt for how to attribute it.