Mustafa Suleyman's insider argument asserts that exponential AI training compute growth—from 10^14 to over 10^26 flops since 2010—faces no near-term wall despite skepticism on Moore's Law slowdowns, data limits, and energy (MIT Technology Review, 2026).

Suleyman details three converging advances: Nvidia GPU performance rising eightfold from 312 to 2,500 teraflops in six years, HBM3 tripling memory bandwidth, and NVLink/InfiniBand linking 100,000+ GPUs into single cognitive entities, yielding 50x efficiency gains from 2020-2026 versus Moore's Law's 5x; this synthesizes with Epoch AI's observation that compute for fixed performance halves every eight months and aligns with Kaplan et al.'s 2020 scaling laws paper showing predictable returns from scale (arXiv:2001.08361; Epoch AI, 2024). Original coverage missed the full software stack efficiencies that have reduced model serving costs by up to 900x annualized and underplayed how these trends consistently falsify prior wall predictions seen in coverage of AlexNet-era systems.

The piece connects to patterns in related events such as the shift from two-GPU AlexNet training in 2012 to contemporary 100,000-GPU clusters and labs growing capacity 4x annually, with global AI compute forecast to reach 100 million H100-equivalents by 2027; what coverage got wrong was insufficient emphasis on unified hardware-software scaling outpacing isolated bottlenecks, a synthesis that shapes expectations for another 1,000x effective compute by 2028 and continued capability gains by 2030.