⚡ Key takeaways
- Every layer of the AI stack has strong incumbents and someone moving in from above or below.
- Frontier labs are the most aggressive movers: all of them now sell agents, not just models.
- Model releases in 2026 are spread across many labs; no lab has more than about a fifth of any modality in our data.
- Open-weight models are a falling share of releases but close behind on quality.
- Independent verification and test environments are the thinnest layer, with few credible suppliers.
🗺️ The map
Thesis 1: Every layer has incumbents and an entrant moving in.high confidence
| Layer | Leaders | Challengers | Moving in from |
|---|---|---|---|
| Chips & compute | NVIDIA; Google TPU; AWS Trainium; AMD | CoreWeave, Crusoe, Groq, Cerebras | Labs (own chips, own campuses) |
| Frontier models | OpenAI, Google, Anthropic, Meta, xAI | Chinese closed flagships (Qwen4-Max) | — |
| Open-weight models | Qwen, DeepSeek, Mistral, Kimi, NVIDIA | Thinking Machines (Inkling 2 Small) | Meta moved away from open weights |
| Harnesses | Claude Code, Codex, Gemini CLI and Jules | Cursor, Cognition | Every frontier lab |
| Frameworks | OpenAI Agents SDK, ADK, LangGraph, Microsoft Agent Framework | CrewAI | Labs give them away |
| Runtimes | AgentCore, Vertex Agent Engine, Foundry | LangSmith Deployment, Temporal | Clouds |
| Builders | Copilot Studio, Zapier, n8n, Make | — | Labs (agent builders) |
| Applications | ChatGPT, Gemini and Claude apps | Harvey, Sierra, River AI | Labs (vertical editions) |
| Verification & environments | Few independents | — | Labs (internal gates) |
So what: The layers are not independent markets: the companies with the most capital are moving into the layers next to them.
📊 What the release data shows
Thesis 2: Release activity is spread across many labs, and language models dominate it.medium confidence
🔧 Reproduce this: stack_concentration.py
stack_concentration.py · reads data/model-releases.json
"""stack_concentration.py — how crowded each model layer was in 2026, from the site's own data.
Reads data/model-releases.json (the Model Release Map) and counts, per modality, the
releases dated 2026, how many different labs released something, the top lab's share
and the share of open-weight releases. Standard library only.
"""
import json, os
m = json.load(open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "..", "model-releases.json")))
rows = {}
for x in m["items"]:
if not str(x["date"]).startswith("2026"): continue
for mod in x["mod"]:
r = rows.setdefault(mod, {"n": 0, "labs": {}, "open": 0})
r["n"] += 1; r["labs"][x["lab"]] = r["labs"].get(x["lab"], 0) + 1; r["open"] += 1 if x.get("open") in (1, 2) else 0
print(f"{'modality':8} {'releases':>8} {'labs':>5} {'top lab share':>14} {'open-weight':>11}")
for mod, r in sorted(rows.items(), key=lambda kv: -kv[1]["n"]):
top, k = max(r["labs"].items(), key=lambda kv: kv[1])
print(f"{mod:8} {r['n']:8d} {len(r['labs']):5d} {k / r['n']:13.0%} ({top}) {r['open'] / r['n']:10.0%}")
allr = sum(r["n"] for r in rows.values()); print(f"2026 releases in the map: {allr} across {len({l for r in rows.values() for l in r['labs']})} labs")$ python3 stack_concentration.py
modality releases labs top lab share open-weight
llm 76 14 17% (anthropic) 22%
speech 13 8 15% (elevenlabs) 23%
image 12 6 25% (openai) 25%
video 9 7 22% (google) 11%
world 3 2 67% (google) 0%
2026 releases in the map: 113 across 23 labs
2026 releases by modality (Model Release Map)
| Language models | |
| Speech | |
| Image | |
| Video | |
| World models |
Real script output. Our maps cover some labs more densely than others (Anthropic in particular), so read patterns, not rankings.
So what: No single lab dominates release activity; competition is wide at the model layer and narrows only at the chip layer.
📄 Method and sources
- Layers and company lists come from the site's sourced series (AI Compute Stack, Agentic AI System Designs) and maps.
- Release counts come from the hand-curated Model Release Map; coverage is uneven by lab.
- Layer boundaries are our framing.
Facts are sourced; figures marked ⚠️ come from secondary reports or company statements. Scores, weights and predictions are our judgement and are labelled. Disclosure: Claude, made by Anthropic, helps write this site. Anthropic is one of the companies analysed here, under the same rules as the others.
Model Release Map (site data)
🧭 Foundations: the AI market, Nov 2026
Live data: AI Market Map & Investments · AI Investment Tracker.
The AI stack
You are here
Read
Read