The Next Phase of Artificial Intelligence

Yann LeCun and JP Vert discussed how artificial intelligence and LLMs may be translated into the physical world, including the new techniques and infrastructure required. The conversation highlighted where physical AI components could be built, but the article contains no specific financial figures, company guidance, or near-term market catalyst. Overall impact appears limited and primarily informational.

Analysis

The investable shift here is not “better models,” it is the migration of AI spend from software-only inference to a full-stack industrial bill of materials. That broadens the winner set away from obvious model vendors toward semis, networking, power, cooling, sensors, and robotics integration — the parts of the stack where bottlenecks, not hype, determine deployment pace. In the near term, the market will likely over-assign value to frontier-model optionality while underpricing the capex intensity required to make physical AI work at scale.

Second-order effects matter more than the headline: if AI begins to control real-world systems, latency, reliability, and safety certification become gating factors, which favors incumbents with industrial distribution and punishes pure-play software vendors with no edge in hardware integration. That creates a medium-term competitive moat for companies sitting at the intersection of compute and the physical world, especially those with installed bases in automation, defense, logistics, and factory control. The supply chain beneficiaries may emerge in waves: first compute/power infra, then edge devices and industrial sensors, then systems integrators.

The biggest contrarian point is that this transition is likely slower than the market narrative implies. Physical deployment cycles are measured in quarters to years because every failure mode gets stress-tested in the real world, and that slows adoption relative to consumer AI. If regulatory scrutiny, labor pushback, or safety incidents rise, sentiment can compress quickly even if the long-term thesis remains intact.

For trading, the cleanest expression is to own the infrastructure picks-and-shovels while fading excess valuation in speculative AI software that depends on near-term monetization. The near-term catalyst window is 3-6 months around capex guidance, datacenter buildouts, and industrial pilot announcements; the longer tail is 12-24 months as edge AI moves from demos to procurement budgets.

Key Decisions for Investors

• Long SMH vs short an equal-weight basket of high-multiple AI software names for a 3-6 month relative-value trade; risk/reward favors the hardware side as physical AI adoption shifts spend toward semis, networking, and power infrastructure.
• Add to infrastructure beneficiaries such as VRT and ETN on pullbacks; 6-12 month view with upside tied to datacenter power/cooling demand and downside limited by secular capex visibility.
• Initiate a basket long on industrial automation leaders like ABB and ROK versus short low-quality software-only AI names with no hardware tie-in; thesis is that physical deployment rewards distribution and integration over narrative.
• Buy 6-12 month call spreads on defense/industrial tech exposure like LMT or HON if you expect early autonomous/physical AI procurement to surface in regulated environments; asymmetric payoff if pilot programs convert to budgeted contracts.
• Avoid chasing front-end LLM beta after rallies; if the market starts pricing physical AI revenue too early, fade with put spreads on the most optically exposed names that have no evidence of edge deployment.