Agents over hype: the real economics of AI infrastructure

There's a shift happening in how we talk about AI, and it's quietly significant. For months now, the debate has circled around whether we're in a bubble-whether all this spending on infrastructure is justified. But agents are changing that conversation entirely. Not chatbots that answer questions. Not even reasoning models that think before they respond. Agents that autonomously break down problems, call tools, verify results, and iterate without human involvement. This is the third fundamental shift in large language models, and unlike the previous two, it's fundamentally changing what compute we actually need.

Ben Thompson makes a compelling case that we're no longer in a speculative bubble-we're in the early stages of something genuinely transformative. The reason is economics. Agentic AI demands exponentially more compute than chatbots or even reasoning models, but it also enables something that chatbots never could: it removes the need for widespread human adoption. One person with agency can control multiple agents. That means you don't need millions of users adopting AI for the infrastructure investment to pay off. A handful of knowledge workers deploying agents across their organisations creates genuine economic value at scale. This reframes the entire question around Anthropic and OpenAI's valuations-they're not betting on consumers; they're betting on enterprises willing to pay for measurable productivity gains.

The architecture question no one wants to face

But there's a wrinkle in this narrative, and some prominent voices are starting to acknowledge it openly. Sam Altman recently conceded something he would have dismissed two years ago: we need architectural breakthroughs beyond mere scaling. Gary Marcus has been saying this for years, and he's now documenting a quiet shift among the industry's biggest names. Altman, Musk, Zuckerberg, LeCun, Sutskever-they're all starting to publicly question whether throwing trillions at data centers is actually the path forward. That's a significant change. The scaling narrative-the idea that bigger models trained on more data automatically get better-is running into real limitations. The conversation is shifting from "how big can we go" to "what are we missing architecturally."

Physical AI getting real-and expensive

Meanwhile, humanoid robotics is moving from labs into the real world, and the challenges are more practical than theoretical. Grippers are expensive, safety is complex, and most humanoids today move slowly and need careful supervision. The dream of a $20,000 general-purpose robot is still just that-a dream. What's actually shipping are specialised systems: cobots in factories, wheeled robots in warehouses, tele-operated units requiring remote operators. The economic case exists, but it's narrower than the marketing suggests. Tesla's Optimus, Boston Dynamics' Spot, Unitree's humanoids-they're all impressive engineering, but they're solving specific problems, not replacing household staff. The real shift will come when manipulation becomes reliable, safety becomes guaranteeable, and cost drops closer to competitive with human labour. We're not there yet, but the trajectory is clear.

For developers and builders, the practical tools are getting sharper. New ROS2 utilities are making it easier to debug multi-node robotics systems and integrate AI agents into robotics workflows. There's real momentum here, even if the consumer narrative around humanoids still runs ahead of the engineering reality. The infrastructure for agentic robotics is being built right now, quietly, in open-source projects and enterprise deployments. That's where the next wave of value creation will happen-not in flashy product launches, but in tools that let teams actually ship.