A New Wild Cat, and Why the Chip Supercycle Might Be Wrong

Scientists have formally identified and described the first new wild cat species discovered in a century. Found in Bolivia — where it had apparently been present in museum specimen collections but misclassified as a known species until genetic and morphological analysis confirmed it was distinct — the animal's formal identification is the first addition to the roughly 37 known wild cat species since the 1920s. Formal description is not merely taxonomic: species that are not officially recognized are effectively invisible to legal protection frameworks, including CITES trade regulations and endangered species habitat protections. For an animal in Bolivia, a country with significant habitat under pressure from agriculture and development, that recognition matters immediately.

Against the backdrop of AMD's trillion-dollar valuation, the dominant market narrative holds that agentic AI workloads — systems running continuously and requiring constant computation — will drive a prolonged supercycle in semiconductor demand. The bull case is straightforward: more AI agents mean more chips, and chipmakers are structurally positioned to benefit for years. But several scenarios could make this consensus wrong.

First, efficiency gains could outpace demand growth. Every generation of AI hardware is dramatically more compute-efficient than the last, and algorithmic improvements have repeatedly allowed newer models to match older models' performance with far less compute. A historical parallel exists in storage: late-1990s investors correctly identified the internet's storage needs but dramatically underestimated how quickly storage efficiency would improve, collapsing per-unit hardware spending even as total deployed storage grew. Semiconductor bulls might be right about the direction of demand and wrong about the magnitude.

Second, the agent workload thesis may be premature. Most enterprise AI deployment right now is inference — running queries against trained models — not continuous autonomous agents. If the transition to agentic workflows takes longer than expected, or if most businesses find that productivity gains do not justify the infrastructure costs, the demand surge could be smaller and later than valuations currently price in. Third, geopolitical disruption: AMD manufactures at TSMC, the Taiwanese foundry. Tightening U.S. restrictions on semiconductor exports to China affect AMD's China revenue directly; escalating tensions over Taiwan affect TSMC's ability to operate at all. A trillion-dollar valuation assumes a relatively stable supply chain in an environment that is anything but.

The signals worth watching: if major hyperscalers — Google, Microsoft, Amazon — begin reporting that AI infrastructure spending is growing more slowly than AI product revenue, efficiency is winning over demand, and chip valuations will need to recalibrate. If agent workloads are slower to materialize, the evidence will appear in enterprise software adoption figures — specifically whether AI agent platforms move from pilot programs to production deployments at scale. The difference in compute demand between a pilot and full production is, by one estimate, roughly tenfold. Q3 earnings reports, beginning in approximately three weeks, will provide the first hard data.