Mustafa Suleyman: AI Development Won't Hit a Wall Anytime Soon—Here's Why
Mustafa Suleyman opened his MIT Technology Review essay with a crisp diagnosis of the problem: we evolved for a linear world, and that makes us catastrophically bad at perceiving exponential change. The argument flows cleanly from there.
Suleyman, who co-founded DeepMind and now runs Microsoft AI, has been inside the compute curve since 2010. By his count, the amount of training data going into frontier models has grown by a trillion times over that period—from roughly 10¹⁴ floating-point operations to numbers that require scientific notation to state without embarrassment. The skeptics who keep predicting a wall keep being wrong, he argues, not because they misunderstand the individual constraints (Moore’s Law deceleration, data exhaustion, energy limits) but because they underestimate how many parallel levers the industry is pulling simultaneously.
The essay is partly a rebuttal to a strain of AI pessimism that has gained credibility as benchmark saturation makes headlines and lab valuations draw scrutiny. The argument is familiar from the scaling hypothesis crowd: every plateau has been engineered around, and the engineering surface area keeps expanding. New chip architectures, inference-time compute, synthetic data generation, and multimodal training pipelines represent distinct axes of improvement, each capable of extending the trajectory even when pre-training on internet text approaches its ceiling.
There’s a valid counter: knowing that exponential growth has continued doesn’t tell you when it stops. The trillion-times figure is impressive on its face, but the marginal capability gain per FLOP has not been constant, and the practical use-case gap between benchmark performance and deployed utility remains real. Suleyman is gesturing at inevitability, which is a different claim from demonstrating a specific mechanism.
Still, the core observational point holds. AI development has consistently surprised its skeptics on the upside, and the people closest to the infrastructure—those building the chips, the data pipelines, the training runs—are not currently acting like men who think they’re near the end. The capital flowing into data centers is not a bet on diminishing returns. Suleyman is making explicit what that capital is already expressing implicitly.
The compute explosion, as he frames it, is the technological story of our time. Whether that framing survives contact with the next two years of model releases is the only test that matters.