That's more fuel in the tank. What we need is better mileage.
This line, from Pete Johnson, stayed with me, because it brings back an older pattern.
In the 1970s, the oil problem was not a problem of scarcity, but a problem of fast-growing dependence. Too much exposure to a handful of regions. Too much confidence in cheap abundance. Too little concern for sovereignty, resilience, and the long externalities of overconsumption. Over time, this did not only create visible shocks, but also a slower form of depletion around it: industrial decline, strategic fragility, and, in a broader sense, a kind of desertification.
Today, tokens feel effectively unlimited. No embargo, no queue at the pump, no obvious scarcity narrative. Perhaps this is precisely why the analogy matters, as the issue is no longer the amount of fuel in the system, but rather which kind of system this apparent abundance is encouraging us to build.
This "golden age" of tokens rests on a handful of hyperscalers and their infrastructure foundations, with levels of economic concentration that are historically unusual, and with a future increasingly pledged through an unprecedented economic bet, when it's not circular investment flows.
The timing is striking: at the very moment when entire economies are holding their breath because a few tankers can no longer navigate freely through a narrow strait, we are collectively building a cognitive infrastructure with its own invisible bottlenecks.
So rather than token scarcity, the long-term question we should address may be the indirect externalities of apparent abundance: the erosion of the business models that sustain high-quality content, the weakening of the open ecosystems on which these models were trained—the open web, code repositories and other commons—the dilution of meaning under industrial-scale generation, and perhaps even a gradual loss of comprehension itself.
Some software engineering teams are already starting to name part of this problem: "comprehension debt"—code that looks clean, passes tests, and still remains cognitively unowned.
This is also where harness engineering may matter: it helps diffuse decomposable agentic blueprints, where these workflows could just as easily have been locked into black-box bundles. A chance to keep what we build more intelligible, not just more fluent.
If the 1970s taught us anything, it is that abundance without sovereignty can become a trap. What matters, then, is not token volume in itself, but whether we are building forms of composable intelligence that remain intelligible, sovereign, and sustainable enough not to exhaust the soil they depend on.
