LeCun Says LLMs Can't Get Robots Off the Ground
Hyperscaler Capex Is About to Get Even Bigger.
LeCun Says LLMs Can't Get Robots Off the Ground
Yann LeCun, now running AMI Labs after leaving Meta with roughly $1B raised, is done being polite about LLM limitations. At VivaTech and AI Forum Europe this year, he called LLMs "largely hopeless for robotics" and said they're "not a path towards human level or even animal-like intelligence" — the core problem being no world model, no way to handle real, messy physical data [1][2].
His alternative is JEPA — Joint Embedding Predictive Architecture — built around abstraction and prediction in physical environments rather than next-token statistics. He's not hedging on timing either: he expects the paradigm shift away from LLM-based robotics to be "obvious" by early 2027, and he's drawing a direct line to the early autonomous-driving hype cycle — impressive demos, years from the real thing [3].
Worth taking seriously given who's saying it. This is a useful check on any org assuming that scaling the current chat-and-agent stack automatically gets you to embodied AI. It doesn't, according to one of the people who helped build the current stack.
Hyperscaler Capex Is About to Get Even Bigger
Goldman Sachs bumped its 2026 US hyperscaler AI capex estimate to ~$800B (up from ~$750B) and, more strikingly, revised 2027 up to $1.4T from a prior consensus of $1.2T [1][2]. Cumulative spend 2026–2031 could hit $7.6T. Morgan Stanley separately projects a 57% jump in 2027 capex from the big four alone [3].
Microsoft, Amazon, Alphabet, Meta, and Oracle are the names driving this, and the justification has shifted from "build it and see" to supply constraints and actual revenue backlogs — commercial AI deployment is now the bottleneck, not speculative bet-hedging [1]. Some of these companies are heading into negative free cash flow territory and tapping debt and equity markets to fund it [2].
This is now a macro story, not just a tech story — capex at this scale is being framed as a meaningful contributor to US GDP growth. If you're a buyer of AI infrastructure or inference, expect pricing and availability dynamics to keep shifting under you as this spending lands.
What This Means For Your Business
Three stories, one thread: the value is moving from writing code to directing systems that write and execute it themselves. Agentic frameworks are maturing fast because the market has already decided single-model, single-prompt workflows don't cut it for real work — you need decomposition, verification, and orchestration across multiple agents. That's an organizational skill now, not just an engineering one. The teams winning here aren't the ones with the best prompt — they're the ones who've figured out how to structure work so agents can be checked, contained, and trusted.
LeCun's pushback on LLMs-for-robotics is a useful reality check on hype cycles generally: don't assume the current architecture generalizes to every problem just because it's dominant. If your roadmap depends on LLMs eventually "figuring out" physical-world reasoning, read the fine print — the people building alternatives don't think that's coming from this architecture. Meanwhile the hyperscaler capex numbers tell you the industry is betting hundreds of billions that inference and agent workloads are the durable demand driver, not general intelligence breakthroughs. Those two things — massive infrastructure spend on agentic workloads, and skepticism about LLMs as a universal architecture — aren't in tension. They're the same signal: the winners are building infrastructure and judgment layers around known-good models, not waiting for a bigger model to solve everything.
For any company making AI decisions right now, the practical read is this: invest in your orchestration and verification layer, not just your model access. The code is increasingly commodity. The judgment — how you split tasks, verify outputs, and decide what not to automate — is where the durable advantage sits.
Key takeaway: The frontier has moved from "which model" to "how do you orchestrate, verify, and contain many models" — and that shift is where the next competitive edge gets built.
Sources
- https://www.kdnuggets.com/10-agentic-ai-frameworks-you-should-know-in-2026
- https://toolradar.com/guides/best-ai-agent-frameworks
- https://theairankings.com/best-ai-agent-frameworks/
- https://www.bbc.co.uk/news/articles/cj6gr0xkyr3o
- https://knowledge.insead.edu/technology-innovation/what-next-ai-revolution-may-look
- https://www.humanoidsdaily.com/news/yann-lecun-predicts-paradigm-shift-in-robotics-by-2027-dismisses-llms-as-intrinsically-unsafe
- https://cryptobriefing.com/goldman-sachs-ai-capex-1-2-trillion-2027/
- https://batreports.online/reports/goldman-sachs-us-equity-views-ai-capex-required-revenues-2026-09-24/report.pdf
- https://www.tipranks.com/news/hyperscaler-ai-capex-set-to-jump-57-in-2027-says-morgan-stanley
Stay ahead of AI
No spam. Unsubscribe anytime.
Want to go deeper?
Reading the news is one thing. Exploring the frontier is another. See what we're building.