Greg Jensen's compute-concentration forecast is specific enough to be wrong, and that is what makes it worth watching

Greg Jensen has put a number on something most observers have left vague. Within a couple of years, Jensen argues, OpenAI and Anthropic will control 35 percent of the world’s compute capacity. He adds that other estimates he takes seriously push that figure as high as 50 percent. The call is falsifiable, the horizon is short, and the structural implications of it being anywhere near correct are worth sitting with.

The concentration Jensen describes does not have a close precedent in computing infrastructure. Control of a third to a half of global AI compute sitting with two private companies would represent a degree of resource concentration that most regulated industries are specifically designed to prevent. Industries built on network infrastructure, on systemic financial exposure, on natural resource extraction have all generated governance frameworks precisely because unchecked concentration in those sectors produced outcomes that markets alone did not correct. Compute, on Jensen’s numbers, would be entering that category faster than any of those precedents took shape.

The regulatory gap this implies is not trivial. Jensen’s projected timeline leaves little room for a multi-year legislative or regulatory process. If the forecast is close to correct, any governance question becomes urgent well before a framework could be drafted, debated, and enacted. That mismatch between the pace of infrastructure accumulation and the pace of institutional response is where the real pressure builds.

Jensen’s framing treats the current moment as a narrow window, one in which the structural conditions for concentration are assembling faster than the institutions designed to manage systemic risk can respond. That framing deserves scrutiny. Jensen is a financial professional offering a structural forecast, not a technical researcher mapping compute deployments against audited figures. The 35 percent figure is his projection. The 50 percent figure he cites is attributed to unspecified others whose models he does not detail in his remarks. Neither number comes with a public methodology attached, and neither has been independently verified against actual infrastructure data. A sophisticated reader should hold both figures as estimates, not measurements.

What makes the call worth tracking is not its precision but its specificity. Jensen is not saying AI concentration is growing in a general sense. He is saying two named companies will hold a named share of a measurable resource within a named timeframe. That is a checkable bet. If OpenAI and Anthropic’s combined compute holdings remain well below 35 percent in two years, the structural concern Jensen is raising loses its empirical foundation. If they approach or exceed that range, the absence of any governance architecture becomes considerably harder to explain away.

The argument Jensen appears to be making is not that either company is behaving improperly now. It is that concentration at the scale he is projecting would raise questions that current institutions are not structured to answer. Both companies are private. Neither is subject to disclosure requirements that would make compute holdings publicly auditable today. Those factors together mean that by the time the data needed to evaluate his forecast is available, the window for a measured policy response may already have closed.

Whether Jensen’s argument gains traction depends less on his credibility than on whether concentration data, when it eventually surfaces, tracks anywhere near his projections. The forecast will be testable before most governance processes could complete a first draft. That gap between the pace of infrastructure buildout and the pace of institutional response is the core of what Jensen is pointing at. The specific numbers, 35 percent or 50 percent, matter less than whether any institution is positioned to act on them if they turn out to be in the right range.