AI's Spending Spree Meets a Capability Reality Check

A $3.4B GPU deal, Alibaba's bid to be "China's AI factory," and Yann LeCun arguing the models still aren't that smart — all in one week. What the capex-vs-capability gap means for what you build.

The capital and the candor pointed in opposite directions this week. While IREN locked in a $3.4 billion deal to deploy NVIDIA Blackwell GPUs in Texas, Yann LeCun spent the day on X reminding everyone that today's AIs are "nowhere near human intelligence." For anyone building on these models, the gap between how much money is chasing compute and how modest the capability claims have become is the signal worth reading.

The capital

IREN's $3.4 billion agreement, reported by Edge Infrastructure Review, is the latest escalation in what the outlet calls a "neocloud" battle — specialized providers racing to stand up GPU capacity outside the hyperscalers. The same week, Alibaba moved to vertically integrate, unveiling its Qwen3.7-Max model alongside custom silicon in a bid to become what the South China Morning Post called "China's AI factory." The Max-Preview ranks 13th for text performance on LM Arena's leaderboard as of May 2026, according to Value The Markets, with a stated focus on logic and problem-solving over creative writing.

The candor

Against that backdrop, LeCun's framing matters. His argument, posted to X, isn't that the models are useless — it's that they compensate for weak reasoning, planning, and common sense through sheer accumulated knowledge. He's made the point consistently; in a follow-up he lined up translation systems, AlphaGo, Watson, and Deep Blue as past milestones that each triggered premature "AI surpassed humans" claims. The working mental model: today's systems are broad retrieval-and-recombination engines, not reasoners, and they improve fastest where breadth of knowledge carries the task rather than multi-step planning.

What it means for builders

This reframes where to place your bets. swyx argued on X that one-shot "deep research" reports have been roughly dead since o3, and that interactive loops — read, think, ask, repeat — beat dropping a thoughtless prompt into a long report nobody reads. If LeCun is right about why these models are useful, swyx is right about how to use them: design for tight human-in-the-loop exchanges that lean on the model's recall, not its long-horizon planning. For buyers, the neocloud price war and Alibaba's vertical integration point the same direction — inference costs should keep falling and model choice should keep widening, especially as challengers like Qwen compete on price as much as on benchmarks.

The bottom line: Build for what these models actually do well today — fast, broad recall inside interactive loops — and let the capex race lower your inference bill while the reasoning catches up.

Sources