AI Intelligence · 2026 Refreshed Jul 13
01 Stack Landscape 02 Labs Briefing 03 Frontier Board 04 Robotics 05 Robotics Stack

Field briefing · Frontier AI labs · 13 July 2026 · fact-checked edition

One consensus, four doctrines

What 22 frontier AI labs across the US, Europe, Canada, and China are actually building — where they contradict one another, where they quietly agree, and who is betting more than a billion dollars that large language models are a dead end.

22 labs profiled 4 competing doctrines $2B+ raised on non-LLM architectures since 2024 61% of OpenRouter tokens served by Chinese open weights (May 2026) 555,000 GPUs at the largest single training site

§1

Executive summary

Mid-2026 has produced a strange industry: near-total agreement on the product and near-total disagreement on the destination. Every major lab now sells the same thing — an agent that plans, uses tools, and completes work — while the people building them openly dispute whether the technology underneath can ever become general intelligence.

1.1

Agents are the one product thesis everyone shares

OpenAI's GPT-5.6 (public July 9) ships a multi-agent mode that runs concurrent subagents in one request; Anthropic's Opus 4.8 runs hundreds of parallel subagents in "dynamic workflows"; Moonshot's Kimi K2.6 scales an Agent Swarm to 300 sub-agents; ByteDance pitches Seed 2.1 as a model measured by whether it can carry a real task from requirement to validated result. Nobody's flagship pitch in 2026 is a chatbot.

1.2

Open-vs-closed no longer tracks geography or ideology — it tracks business model

Chinese labs plus Mistral and Ai2 flooded the market with permissively licensed weights: by May 2026, Chinese open models served roughly 61% of tokens on OpenRouter, and Qwen passed one billion cumulative Hugging Face downloads — overtaking Llama as the most-downloaded open family. Meanwhile Meta, the original open-weights champion, went proprietary with Muse Spark in the same year Baidu shipped 100+ open releases.

1.3

Dissent against LLMs is now capitalized at frontier scale

For the first time, "LLMs won't get there" is a funded position, not a blog post: Yann LeCun's AMI Labs raised a $1.03B seed at $3.5B pre-money (Europe's largest seed ever), Fei-Fei Li's World Labs holds a reported ~$1B for spatial intelligence, and Chollet's Ndea and Friston's Verses round out the counter-bet — with Sutskever's $3B SSI half-in as the "new recipes, still deep learning" hedge.

1.4

Compute has decoupled from doctrine

SpaceX acquired xAI outright in February (all-stock, ~$1.25T combined) and became a landlord: Colossus 1 — 220,000 GPUs and 300MW — is rented to rival Anthropic, Colossus 2 capacity is being marketed too, and Grok's future inside the merged company is an open question. Infrastructure is becoming its own business that pays out no matter whose theory of intelligence wins.

1.5

Timelines have not converged — capital is funding contradictory physics

Dario Amodei's public position expects systems broadly better than humans at most cognitive tasks by 2026–27; LeCun says decades and a different architecture entirely. Both raised billions within months of each other. The market is not pricing a consensus; it is buying every side of the argument.

Eight months that redrew the map

December 2025 – July 2026 · events color-coded by the doctrine of the lab making them

Dec 2025

AMI Labs launches in Paris — LeCun leaves Meta to build JEPA world models

Jan 2026

Colossus hits 2 GW — 555K GPUs, $18B, largest single AI site

Thinking Machines' $50B round collapses; senior staff return to OpenAI

Feb 2026

SpaceX acquires xAI — all-stock, ~$1.25T combined entity

Doubao passes 200M DAU as Seed 2.0 ships; Cohere reports $240M ARR

Mar 2026

AMI raises $1.03B — Europe's largest-ever seed round

1.4 GW Paris AI campus announced by MGX, Nvidia, Bpifrance & Mistral

Apr 2026

Meta ships Muse Spark (Apr 8) — first closed flagship; lands 4th on benchmarks

DeepSeek V4 preview (Apr 24) — $3.48/M output tokens; Kimi K2.6 ships Agent Swarm

May 2026

Gemini 3.5 Pro at I/O (May 19) — 2M-token context, Deep Think

Microsoft unveils 7 MAI models at Build — "long-term self-sufficiency"

Chinese open weights hit 61% of OpenRouter tokens

Jun 2026

Mistral in talks at €20B — up from €11.7B in nine months

Kimi K2.7-Code (Jun 12) and Seed 2.1 (Jun 23) ship as agent models

Jul 2026

GPT-5.6 goes public (Jul 9) — Luna/Terra/Sol tiers, after a government safety review

Muse Spark 1.1 + Meta Model API; xAI rebrands as SpaceXAI

§2

The lab atlas: what 22 labs are actually building

Each lab is tagged with the doctrine it behaves by — not what its press releases say. Figures are from public reporting as of July 2026 and should be treated as directional.

SCALE BELIEVERS — the current recipe, pushed harder, gets there HEDGED BETS — LLMs necessary but not sufficient COMMODITY & SOVEREIGNTY — the model is a means, not the end DISSENTERS — LLMs are the wrong substrate for intelligence

The atlas at a glance

22 labs by the doctrine they behave by (this report's placement)

Scale believersOpenAI · Anthropic · Meta · xAI · MSFT · T. Machines 6
Hedged betsDeepMind · SSI 2
Commodity & sovereigntyMistral · Cohere · Ai2 · DeepSeek · Qwen · Kimi · GLM · ByteDance · Baidu · Tencent 10
DissentersAMI · World Labs · Ndea · Verses 4

The largest camp by number of labs — commodity & sovereignty — is also the one that makes the least money per token. The smallest camp — the hedged middle — contains the most-distributed model on Earth and the best-funded company with no product.

United States — the frontier six

Closed frontier · massive compute
OpenAISCALE

SAN FRANCISCO · GPT-5.6 (LUNA / TERRA / SOL) · STARGATE BUILDOUT

Building: a "super-assistant" and the infrastructure of agentic work.

GPT-5.6 went public July 9, 2026 in three tiers — Luna (fast), Terra (balanced), Sol (most capable) — after an unusual gated launch: at the US government's request, it shipped first to ~20 trusted partners behind a safety review. Its multi-agent beta runs concurrent subagents in a single request, and ChatGPT Work turns connected apps and files into documents, spreadsheets, and presentations. The API language has shifted from completions to delegation.

AnthropicSCALE

SAN FRANCISCO · OPUS 4.8 / SONNET 5 · AGENT CONTROL PLANE

Building: the trusted agent layer for enterprises and governments.

Bets that interpretability and safety research convert into durable advantage in regulated workloads. Opus 4.8's "dynamic workflows" plan a task and run hundreds of parallel subagents in one session, with verification before reporting back; Claude now embeds in Slack (and reportedly soon Microsoft Teams), and enterprise admins get model-level entitlements and spend alerts — the control-plane pitch made concrete. Notably rents 220,000 GPUs and 300MW of inference capacity on xAI's Colossus 1.

Google DeepMindHEDGED

LONDON / MOUNTAIN VIEW · GEMINI 3.5 PRO · ISOMORPHIC LABS

Building: Gemini everywhere, world models underneath, and a drug-discovery engine.

Gemini 3.5 Pro — announced at I/O on May 19, 2026, general availability July — carries a 2M-token context window (the largest of any production frontier model), a built-in Deep Think reasoning mode, and parallel tool orchestration, riding unmatched distribution across Search, Workspace, and 3B+ Android devices. Yet Hassabis publicly argues LLMs "can't understand reality," pushing a convergence of LLMs, world models, and AlphaGo-style planning. Spinoff Isomorphic ($600M raised; Lilly, Novartis, J&J deals) targets first clinical trials of AI-designed drugs by end-2026.

Meta Superintelligence LabsSCALE

MENLO PARK · MUSE SPARK 1.1 · $14.3B SCALE AI DEAL

Building: proprietary "personal superintelligence" — a full reversal of the open-Llama era.

After Llama 4 disappointed, Zuckerberg paid $14.3B for half of Scale AI and installed Alexandr Wang as chief AI officer. Muse Spark (April 8, 2026) is Meta's first closed frontier model — deployed instantly across 3B+ users on Facebook, Instagram, WhatsApp, and Ray-Ban glasses, but a benchmark disappointment: fourth on the Artificial Analysis index behind GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6 at launch. Muse Spark 1.1 and the Meta Model API (July 9) began monetizing it anyway; the thesis is that value accrues to data manufacturing and distribution, not open weights.

xAI → SpaceXAISCALE

MEMPHIS / HAWTHORNE · GROK · COLOSSUS 2: 555K GPUS / 2 GW

Building: the biggest computer on Earth — and renting much of it out.

SpaceX acquired xAI in February 2026 (all-stock, ~$1.25T combined); Musk wound down xAI as a standalone company in May, and the unit rebranded SpaceXAI in July. Colossus 2 — 555,000 GPUs, $18B of hardware, 2GW — is the world's largest single-site installation, but the deeper strategy is landlord: Colossus 1 (220K GPUs, 300MW) is leased to Anthropic, Colossus 2 capacity is being marketed to outside tenants including Google, and orbital data centers are the stated long game. Whether Grok keeps frontier ambitions inside SpaceX is openly questioned.

Microsoft AI (MAI)SCALE

REDMOND · 7 MAI MODELS · MAIA SILICON · FAIRWATER DCS

Building: full model self-sufficiency after the OpenAI renegotiation.

Suleyman's MAI Superintelligence Team shipped seven in-house models at Build 2026 — MAI-Thinking-1 (reasoning), MAI-Code-1-Flash (code), MAI-Image-2.5, MAI-Voice-2, MAI-Transcribe-1.5 and Flash variants — trained from scratch with, Microsoft stresses, no distillation from anyone else's models. MAI-Thinking-1 claims blind-test parity with Claude Sonnet 4.6 and a match with Opus 4.6 on a major coding benchmark. The frame: "long-term self-sufficiency" and a "humanist superintelligence" that complements rather than replaces people.

United States — the pure research bets

Small teams · concentrated theses
Safe Superintelligence (SSI)HEDGED

PALO ALTO / TEL AVIV · $3B RAISED · $32B VALUATION · NO PRODUCT

Building: one thing — safe superintelligence — and nothing else until then.

Sutskever's small lab has raised $3B across two rounds ($1B in September 2024, $2B at a $32B valuation in April 2025, backed by Greenoaks, a16z, Alphabet, and Nvidia) with no product, no papers, and no revenue: the market's purest bet on a founder's research taste. His public framing: the age of scaling is over and a "research era" has begun, where algorithmic innovation — not more GPU-hours on the same recipe — drives progress.

Thinking Machines LabSCALE

SAN FRANCISCO · TINKER · $2B SEED AT $12B · $50B TALKS COLLAPSED

Building: frontier models plus open tooling — the inverse of SSI's straight shot.

Murati raised a record $2B seed at $12B (July 2025, a16z-led) and shipped Tinker, a fine-tuning API for open models, that October. Then the strain showed: talks to raise $5B at a $50B valuation collapsed in January 2026 as senior researchers (Zoph, Metz, Schoenholz) defected back to OpenAI. The lab sits at its original $12B with ~150 staff, two live products, billions in Nvidia and Google infrastructure commitments — and its own frontier models still promised.

World LabsDISSENT

SAN FRANCISCO · ~$1B RAISED (REPORTED) · MARBLE

Building: "large world models" — spatial intelligence instead of language.

Fei-Fei Li's argument: language is a lossy, low-bandwidth encoding of reality, and intelligence requires reasoning about 3D space, objects, and physical interaction. World Labs generates persistent, navigable 3D environments with Marble — a direct architectural alternative to the text-token frontier.

NdeaDISSENT

SAN FRANCISCO · $43M (REPORTED) · ARC-AGI LINEAGE

Building: program synthesis guided by deep learning — abstraction over memorization.

François Chollet (creator of Keras and the ARC-AGI benchmark) holds that LLMs interpolate over memorized patterns and fail at genuinely novel abstraction. Ndea's bet, with Zapier co-founder Mike Knoop, is that searching over programs — with neural networks as the guide — is the missing mechanism for general intelligence.

Ai2 (Allen Institute)COMMODITY

SEATTLE · NONPROFIT · OLMO 3.1 · NSF OMAI

Building: the fully open counter-example — data, code, and weights, all published.

Olmo 3.1 (Think and Instruct, 32B) is the strongest "truly open" model line: end-to-end transparency over training data and decisions, not just released weights. Ai2 anchors the US public-interest position while leadership transitions (Farhadi stepped down as CEO in March 2026).

Verses AIDISSENT

LOS ANGELES · PUBLIC CO · GENIUS ENGINE

Building: active-inference agents from neuroscience, not next-token prediction.

With Karl Friston as chief scientist, Verses builds Bayesian reasoning agents on the free-energy principle: intelligent systems act to minimize uncertainty about the world. The claim is that LLMs are passive mimicry, while active inference yields agents that model causes, quantify what they don't know, and act accordingly.

Europe & Canada

Sovereignty · enterprise trust · dissent
Mistral AICOMMODITY

PARIS · €11.7B → TALKS AT €20B · ARR $400M → $1B TARGET

Building: Europe's sovereign AI stack — open weights, on-prem agents, own data centers.

The flagship of European strategic autonomy: air-gapped, self-hosted deployments that sidestep US Cloud Act exposure, ARR that grew twentyfold to $400M by February 2026 (target: $1B by year-end), and a reported €3B raise in the works at a €20B valuation. Its compute is becoming sovereign too: an $830M debt-financed data center near Paris came online in Q2 2026, 200MW of EU capacity is targeted by end-2027, and a 1.4GW AI campus with MGX, Nvidia, and Bpifrance is slated for 2028.

AMI LabsDISSENT

PARIS · $1.03B SEED · $3.5B PRE — LARGEST EU SEED EVER

Building: JEPA world models — the largest single bet that LLMs are a dead end.

Yann LeCun left Meta after the Superintelligence Labs reorg and launched AMI in Paris in December 2025 with CEO Alexandre LeBrun; by March 2026 it held Europe's largest-ever seed round, backed by Bezos Expeditions, Eric Schmidt, Xavier Niel, and the Berners-Lees. The thesis he's argued for years: autoregressive token prediction cannot plan, lacks persistent memory, and won't reach general intelligence at any scale. AMI trains joint-embedding predictive architectures aimed at industry, robotics, and healthcare — prediction in representation space, not in words.

CohereCOMMODITY

TORONTO · ~$7B VALUATION · $240M ARR · IPO WATCH

Building: security-first enterprise agents, explicitly not chasing AGI.

North — GA since August 2025, deployed by RBC, Dell, and LG CNS — runs agents inside a customer's own infrastructure, with a HiddenLayer partnership targeting prompt injection and data leakage. Cohere closed 2025 at $240M ARR with >50% quarter-over-quarter growth, hired Uber's IPO-era finance chief as CFO, and is the clearest "AGI is not the business" position among Western labs — sovereignty and data control are. No S-1 has been filed as of mid-2026.

China — the open-weights offensive

Permissive licenses · brutal pricing · distribution
DeepSeekCOMMODITY

HANGZHOU · V4 PRO 1.6T / V4 FLASH 284B · MIT WEIGHTS

Building: frontier-class reasoning at a fraction of frontier price.

The V4 preview (April 24, 2026) — a 1.6T-parameter Pro and 284B Flash, tuned for agentic tasks and tightly integrated with Huawei silicon — priced output at $3.48 per million tokens against OpenAI's $30 and Anthropic's $25, then cut further with promotional and cache discounts within days. DeepSeek alone now serves 16.3% of all OpenRouter tokens — more than any other single provider — and its efficiency innovations forced repricing across every Western lab's API sheet.

Alibaba (Qwen)COMMODITY

HANGZHOU · APACHE 2.0 · 1B+ HUGGING FACE DOWNLOADS

Building: the default base model of the open ecosystem — to sell cloud.

Qwen owns the ecosystem: past one billion cumulative Hugging Face downloads, it overtook Llama as the most-downloaded, most-derived open model family in the world, released at relentless cadence. The models are demand generation for Alibaba Cloud and consumer surfaces — the clearest statement that models themselves are a commodity.

Moonshot AI (Kimi)COMMODITY

BEIJING · K2.6 / K2.7-CODE · 1T MOE, 32B ACTIVE · MODIFIED MIT

Building: open agentic models tuned for long tool-use loops.

Kimi K2.6 (April 2026) is a 1T-parameter mixture-of-experts (32B active, 256K context) whose Agent Swarm coordinates up to 300 domain-specialized sub-agents across 4,000 steps in one autonomous run; K2.7-Code followed in June, cutting reasoning-token use 30%. Five major releases in under a year — the open-weights answer to Western agentic flagships, at a cadence Western labs don't match.

Zhipu AI (GLM)COMMODITY

BEIJING · GLM 5.X · MIT WEIGHTS · CODING-FIRST FLAGSHIP

Building: open coding models that trade blows with the Western frontier.

GLM 5.1/5.2 is the coding-first open flagship, breaking through on planning quality and long-horizon coding; on coding benchmarks it competes directly with closed Western models at a fraction of the cost, making it a staple of the agentic-coding stack outside the US.

ByteDance (Seed / Doubao)COMMODITY

BEIJING · SEED 2.1 · 200M DAU POST-LUNAR-NEW-YEAR

Building: the most-used consumer AI in the world's largest market.

Doubao dominates Chinese consumer AI — 155M weekly users in early 2026, past 200M daily actives after Lunar New Year — while overseas twin Dola targets 30M DAU this year. Seed 2.1 (June 2026) is pitched as an agent judged by task completion, priced far below Western tiers, with strong media/video capability and hardware tie-ins. Value is captured in the app and ad surfaces, not the model.

Baidu (ERNIE)COMMODITY

BEIJING · ERNIE 5.0 / 5.1 · 2.4T OMNIMODAL MOE · KUNLUN CHIPS

Building: an open, home-grown full stack — from silicon to search.

The mirror image of Meta's pivot: from zero Hugging Face releases in 2024 to 100+ in 2025. ERNIE 5.0 — a 2.4T-parameter omnimodal MoE activating under 3% of parameters per inference, trained on Baidu's own Kunlun silicon — serves an assistant with 200M monthly users; ERNIE 5.1 cracked the global top five in 2026, and the new Kunlun M100/M300 chips extend sovereignty to the silicon level.

Tencent (Hunyuan)COMMODITY

SHENZHEN · WECHAT-FIRST · 3D/MEDIA OPEN MODELS

Building: AI woven into WeChat, gaming, and cloud rather than sold as a model.

Hunyuan lives inside Tencent's ecosystem — WeChat mini-programs, games, enterprise SaaS — with selective open-sourcing of 3D and media models. A 2026 "self-correction" pivot refocused spending on surfaces where Tencent already owns the user.

The price war, in one chart

Frontier-tier API price, $ per 1M output tokens · April 2026, at DeepSeek V4's launch

OpenAI frontierclosed $30.00
Anthropic frontierclosed $25.00
DeepSeek V4-Proopen weights, MIT $3.48

Two days after launch DeepSeek added a 75% promotional discount on V4-Pro, then cut cache-hit input prices to a tenth of their previous level — the gap above is the conservative version.

§3

Fault lines: where the labs contradict each other

Five questions on which the industry gives incompatible answers. Positions are inferred from what labs ship, license, and say — hover any dot for the full name.

SCALE HEDGED COMMODITY DISSENT

Fault line 1 — Should frontier weights be open?

The open end is now overwhelmingly Chinese plus Mistral and Ai2 — and it's winning on volume. The closed end justifies itself with safety and with the one market that still pays a premium: frontier coding agents. The line divides business models, not values: open labs monetize distribution, cloud, and sovereignty; closed labs monetize the model itself.

Who actually serves the tokens

Composition of token volume on OpenRouter, the largest neutral model router

Chinese open-weight models US models Everything else
Late 2024
US ~70% other
May 2026
Chinese open weights 61% US ~30%

DeepSeek alone serves 16.3% of all OpenRouter tokens — the largest single provider, ahead of Google, Anthropic (13.3%), and OpenAI. Four of the five most-used models are Chinese; Llama, the open-weight leader two years ago, has fallen off the rankings entirely.

Fault line 2 — Does the current recipe (transformers + RL + compute) reach general intelligence?

This is the deepest schism in the field. The left end is spending tens of billions on the assumption the recipe holds; the right end has now raised billions on the assumption it doesn't. DeepMind is the pivotal middle case: it ships the world's most-distributed LLM while its CEO argues LLMs alone can't understand reality — hedging both outcomes.

Fault line 3 — When do systems broadly better than humans arrive?

Anthropic's formal submission to the US government expects powerful AI by late 2026 or early 2027 — among the most aggressive claims from any CEO. LeCun says decades and a different substrate. Nothing in 2026 narrowed this spread; if anything the poles both hardened, and both got funded.

Fault line 4 — Commercialize now, or straight shot to superintelligence?

SSI's "straight shot" — $3B raised, nothing shipped — is the purest rejection of the commercialize-as-you-go consensus. Everyone else has concluded that products fund the mission (and, per OpenAI and Anthropic, that deployment is how you learn to make systems safe). Sutskever's counter is that products create gravity that pulls research off course — and Thinking Machines' January defections are his exhibit A.

Fault line 5 — Is the model the product, or a loss-leader for something else?

The pure-play labs (left) must win on model quality because the model is all they sell. The platforms (right) can give the model away forever — which is exactly what China's ecosystem players and Google are doing, and why the open-weights flood is a rational strategy rather than altruism.

Three ironies worth naming

The open-source flip-flop

Meta, the company that made open weights a movement, went proprietary with Muse Spark in the same twelve months that Baidu — long a closed shop — released 100+ open models and Qwen overtook Llama as the world's most-downloaded open family. The largest reversals in the debate happened in opposite directions, for the same reason: each side concluded its old strategy wasn't winning.

The landlord's tenant

Colossus 1's mixed GPU architecture proved unusable for training Grok — so it was leased to Anthropic, a direct competitor, for inference: 220,000 GPUs and 300MW. Musk now makes money when Claude answers a question, and Colossus 2 capacity is being marketed to outside tenants too. Infrastructure economics has quietly overridden the model war.

The hedged champion

The CEO shipping the world's most-distributed LLM — Demis Hassabis, with Gemini on 3B+ devices — is also the most senior voice arguing that language models "can't understand reality" and that world models are required. DeepMind is simultaneously the strongest evidence for and against the scaling thesis.

§4

Common ground: what everyone now agrees on

Strip away the doctrine and the labs' actual 2026 behavior converges on six points — several of which were contested as recently as 2024.

Agents are the product; chat is over

Every flagship release of 2026 is framed as an agent that plans, uses tools, and completes multi-step work — GPT-5.6's multi-agent mode, Opus 4.8's dynamic workflows, Kimi's Agent Swarm, Seed 2.1's task-completion framing. The API business itself is being reshaped around delegation rather than completion.

OpenAI · Anthropic · DeepMind · Microsoft · Moonshot · Cohere · ByteDance

Post-training and reasoning are the active scaling axis

Naive pretraining scaling has quietly plateaued as a strategy everywhere; the gains of 2025–26 came from reinforcement learning, verifier-guided reasoning, and test-time compute. Even Sutskever's "the age of scaling is over" is only a sharper version of what every lab's training pipeline already concedes.

Universal — including the scaling believers

Coding is the killer app that pays

Agentic coding is the one market where customers demonstrably pay a large premium for frontier intelligence — which is why every lab ships a dedicated coding model (MAI-Code-1-Flash, Kimi K2.7-Code, GLM's coding-first line, Claude's coding agents, GPT-5.6's developer positioning) and why coding benchmarks became the de facto scoreboard.

Universal

Compute gets built regardless of doctrine

Colossus 2 at 2GW, Stargate, Fairwater, a 1.4GW Paris campus — and even the architecture dissenters raised ten-figure rounds that go substantially to GPUs. Whatever intelligence turns out to be, every faction is behaving as if it will be expensive to compute.

Universal — even the LLM skeptics

Science is the next frontier claim

The stated endgame has shifted from "assistant" to "discovery": Isomorphic's AI-designed drugs approaching trials, GPT-5.6's scientific-research positioning, Anthropic's research institute agenda, Wang's "biggest discoveries in history." Accelerating science is now the shared justification for the entire buildout.

DeepMind · OpenAI · Anthropic · Meta · Microsoft

Everyone hedges with world models

The dissenters' core idea is being absorbed by the mainstream: DeepMind's world-model convergence, video-generation-as-simulation efforts, ByteDance naming world models a 2026 priority, and embodied bets across the frontier labs. The disagreement is no longer whether world models matter — it's whether they emerge from scaled LLMs or must replace them.

DeepMind · OpenAI · Meta · ByteDance · AMI · World Labs

The buildout, in gigawatts

Announced single-site or single-program AI compute capacity · press-reported, mid-2026

Stargate programOpenAI+partners · announced target ~10 GW target
Colossus 2, MemphisSpaceXAI · 555K GPUs · $18B 2.0 GW
MGX Paris campusMGX · Nvidia · Bpifrance · ops 2028 1.4 GW planned
Anthropic's Colossus 1 lease220K GPUs · rented from SpaceXAI 0.3 GW
Mistral EU capacityown target by end-2027 0.2 GW target

Lighter bars are announced targets, not operating capacity. For scale: 1 GW is roughly the output of a large nuclear reactor, running continuously.

§5

The dissenters: who is against LLMs, and why

"Against LLMs" here means something precise: these researchers accept that LLMs are useful products but argue they are the wrong substrate for general intelligence — and each has a specific technical objection and a funded alternative. What changed in 2025–26 is that this camp stopped writing essays and started raising billions.

The counter-bet, capitalized

Disclosed funding of labs betting against (or beyond) pure LLM scaling · $B

SSI — Sutskeverhalf-dissent: new recipes, still deep learning $3.00B
AMI Labs — LeCunJEPA world models $1.03B
World Labs — Lispatial intelligence ~$1.0B reported
Ndea — Cholletprogram synthesis $0.04B

Verses AI (Friston, active inference) is publicly traded and omitted. The hatched bar marks SSI's ambiguous position: not anti-LLM, but anti-"just add GPUs."

WhoVehicle & backingCore objection to LLMsThe alternative bet
Yann LeCunex-Meta chief AI scientist AMI Labs, Paris — $1.03B seed at $3.5B pre-money, Europe's largest ever; backers include Bezos Expeditions and Eric Schmidt Autoregressive next-token prediction cannot plan, has no persistent memory, and no model of how the world evolves. "Scaling them up will not allow us to reach AGI." JEPA — joint-embedding predictive architectures that learn world dynamics from video and act by planning in representation space, not in words.
Fei-Fei LiStanford, ImageNet creator World Labs, San Francisco — ~$1B raised (press-reported) Language is a lossy, low-bandwidth compression of reality. Intelligence is grounded in 3D perception and action; text-only systems inherit the blind spot. Spatial intelligence: "large world models" that generate and reason over persistent, navigable 3D environments (Marble).
François CholletKeras & ARC-AGI creator Ndea — $43M (reported) with Zapier co-founder Mike Knoop LLMs interpolate over memorized patterns and fail novel abstraction — which is why ARC-style tasks that are trivial for humans stayed hard. Benchmark gains ≠ generality. Program synthesis guided by deep learning: search over discrete programs, measuring intelligence as skill-acquisition efficiency.
Karl Fristonneuroscientist, UCL Verses AI, Los Angeles — public company, Friston as chief scientist LLMs are passive statistical mimicry: no agency, no intrinsic drive to resolve uncertainty, no generative model of causes behind observations. Active inference on the free-energy principle: Bayesian agents (Genius) that act on the world to minimize uncertainty, with calibrated beliefs.
Richard SuttonRL pioneer, "Bitter Lesson" author No lab — University of Alberta / Keen collaboration The sharpest irony: the author of the pro-scaling "Bitter Lesson" argues LLMs learn from human imitation rather than experience, and imitation caps out at human knowledge. The "era of experience": agents that learn continually from their own interaction and reward, in the classic RL sense.
Gary MarcusNYU emeritus, critic-at-large No lab — books, testimony, and a widely read newsletter Hallucination and brittle reasoning are structural properties of pure neural approaches, not bugs to be scaled away; the field mistakes fluency for understanding. Neurosymbolic hybrids — explicit knowledge and symbol manipulation fused with learning.

The half-dissenters inside the citadel

Demis Hassabis runs the strongest LLM franchise on Earth and still argues language models lack physics, causality, and spatial understanding — his position is convergence: LLMs plus world models plus AlphaGo-style planning, not LLMs alone. Ilya Sutskever — arguably the person most responsible for the scaling era — now says that era is over and that progress returns to algorithmic research; SSI is not anti-LLM so much as anti-"just add GPUs." That the two most decorated researchers of the deep-learning age are both hedging the pure-LLM thesis is, quietly, the strongest signal in this entire report.

Why does this camp exist at all, given LLMs' commercial success? Three recurring reasons: (1) the generalization critique — models excel on distributions they were trained on and degrade on genuine novelty (Chollet's ARC results); (2) the grounding critique — text is a shadow of the world, so systems trained only on it can't acquire causal, spatial, physical understanding (LeCun, Li, Hassabis); and (3) the learning-mechanism critique — imitating human output is not the same as learning from experience, and can't exceed its source (Sutton, Friston). The scaling camp's rebuttal is empirical: every year the "wall" moves, agents keep absorbing capabilities that critics said required new architectures, and revenue compounds. 2026 has not settled it.

§6

Method, scope & sources

Compiled 13 July 2026 from public reporting, lab publications, and industry analysis, then fact-checked the same day against primary and news sources (the corrections pass revised SSI's total raised from $6B to $3B, updated Thinking Machines' valuation to reflect the collapsed $50B round, dated Cohere's North GA to August 2025, and refreshed all model names and figures to their July 2026 state). "Top 20" was interpreted as the 22 most consequential model-building labs across the US, UK/Europe, Canada, and China; hyperscaler-adjacent efforts that primarily serve internal platforms (Amazon's Nova/AGI group, NVIDIA's Nemotron line) and smaller research shops (AI21, Reka, Nous, Liquid AI, MiniMax) were noted but not profiled. Doctrine placements on the fault-line spectrums are this report's editorial judgment, inferred from behavior — licensing, pricing, shipping cadence, and leaders' own words — not from self-description.

Valuations, GPU counts, and model specifications are drawn from press reports and have not been independently verified; figures marked "reported" (World Labs' total raise, Ndea's funding, Grok 5's specifications, the Stargate target) rest on thinner sourcing than the rest. The landscape described here moves monthly — treat this as a dated snapshot, not a standing reference.

Principal sources