The US leads China in frontier AI by roughly six to eight months, and that gap is not growing

Four people with different vantage points on the AI competition between the US and China have landed on numbers that are remarkably close to each other. The gap at the frontier sits somewhere between six and nine months, measured by model capability. The agreement is worth examining, because the estimates arrive independently and because they point toward a specific mechanism that export controls alone cannot address.

Marc Andreessen, the venture capitalist and co-founder of Andreessen Horowitz, puts the US lead at approximately six months. Sebastian Mallaby, an economic historian and journalist, puts it at eight. Patrick O’Shaughnessy, an investor who tracks AI development closely, offers a range of six to nine months and adds a detail that gives the figure its edge: China’s ability to stay within that window is partly a function of distillation, the process of training smaller models on the outputs of larger ones. American frontier models, in other words, are themselves a resource that Chinese developers draw on to close the gap.

David Sacks, a venture capitalist, is the most direct of the four in pushing back on the narrative that China has caught up. “I don’t believe that China has caught up,” he has said. “I still think we are 6 months ahead.” That framing treats the six-month figure not as a ceiling but as a floor that, in his assessment, still holds.

I don't believe that China has caught up. I still think we are 6 months ahead. David Sacks

What makes the tight clustering significant is that these figures come from people working in different domains and with different interests. A venture investor, an economic historian and journalist, and two capital allocators do not typically converge on the same number by accident. The range of six to nine months also reflects a shared underlying read on the state of the competition: that chip export controls have not been decisive, and that Chinese developers have found routes to remain competitive despite constrained access to the most advanced hardware.

The distillation point O’Shaughnessy raises deserves more attention than it usually gets in discussions framed around compute. Export controls restrict hardware. They do not restrict the public model outputs that enable distillation. A Chinese lab that cannot access the most advanced chips can still, in principle, train on the behavior of models that were built with those chips. The strategy is efficiency-focused rather than scale-focused, and the six-to-nine-month gap suggests it is working well enough to keep Chinese developers within striking distance.

None of the four speakers claim that the gap is widening. None suggest that export controls have been decisive. The implicit read across all four assessments is that the US holds a real but bounded advantage: enough to matter in capability terms, not so large that it insulates American developers from meaningful competition. Four observers arriving at the same narrow range without coordination is a data point that deserves weight.

The practical question the six-to-eight-month figure raises is what that gap means in product and deployment terms. A six-month lead in model capability translates differently depending on how fast the capability curve is moving. If the frontier is advancing quickly, six months represents a significant head start. If the pace slows, the same gap compresses into something easier to close. None of the four speakers address the rate of change directly, which is the limiting factor in using their estimates to forecast where the race stands two or three years out. What the estimates do establish is that the current moment is one of genuine, sustained competition rather than a runaway American lead, and that the tools China is using to maintain proximity include ones that export policy cannot easily reach.