Suleyman's Big Claim
Microsoft AI chief Mustafa Suleyman argued publicly that continued data growth, compute improvements outpacing Moore’s Law and clean-energy pathways could lead to a future of 'cognitive abundance'—a strong endorsement of scaling as the way forward for AI. His framing pushes back on skeptics who warn scaling has diminishing returns and suggests Microsoft remains committed to investment at the systems level. That perspective matters because company posture shapes where partners and customers place bets on infrastructure and product strategy. (x.com)
Mustafa Suleyman, the executive running Microsoft AI, went public on April 8 with a blunt claim: artificial intelligence is not running out of road, and the real story is an “exponential compute ramp” that is still accelerating. (technologyreview.com) He is arguing against a popular idea in artificial intelligence that bigger models are hitting diminishing returns, meaning each extra dollar buys less progress than the last. In his essay, he says the opposite: more data, more computing power, and better systems design are still compounding together. (mustafa-suleyman.ai) His core measure is floating-point operations, which are the tiny math steps a machine does while training a model, like counting how many turns an engine can make in a day. Suleyman says frontier training runs grew from about 10^14 floating-point operations in 2010 to more than 10^26 by 2025, which he calls a trillion-fold jump. (mustafa-suleyman.ai) He says the reason this kept growing even as Moore’s Law slowed is that the gains are no longer coming from one place. Faster chips, denser memory, better networking, and smarter software are now working together like widening every lane on a highway at once. (technologyreview.com) One example in his argument is Nvidia’s jump in raw chip performance, which he says went from 312 teraflops in 2020 to 2,250 teraflops in 2026. Another is high bandwidth memory, which stacks memory vertically and feeds processors fast enough to keep expensive chips from sitting idle. (technologyreview.com) He also points to Microsoft’s own Maia 200 chip, launched in January 2026, and says it delivers 30% better performance per dollar than any other hardware in Microsoft’s fleet. In his longer slide deck, he says the chip is built on Taiwan Semiconductor Manufacturing Company’s 3 nanometer process and delivers more than 10 petaFLOPS per second. (technologyreview.com) (mustafa-suleyman.ai) The clean-energy part of his case is just as important as the chip part. Skeptics often say artificial intelligence will hit an electricity wall, but Suleyman argues energy supply can expand alongside computing supply instead of capping it. (mustafa-suleyman.ai) This is not a detached essay from a commentator. On April 2, Bloomberg reported that Microsoft wants to build large cutting-edge models by 2027, and Suleyman said the company’s goal is to reach “state-of-the-art” across text, image, and audio systems. (bloomberg.com) That same week, Microsoft released three in-house foundation models through its Microsoft AI unit: a transcription model, a voice model, and an image model. TechCrunch reported that the models came from the MAI Superintelligence team, which Microsoft formed in November 2025 under Suleyman. (techcrunch.com) The company is still balancing two bets at once. Suleyman told Bloomberg that Microsoft will keep hosting outside models while building its own long-term “AI self-sufficiency” over the next three to five years. (bloomberg.com) That makes his scaling argument more than philosophy. If Microsoft is telling customers, chip partners, and data-center builders that the bottleneck is not model ideas but usable compute, the money keeps flowing toward servers, networking, power, and custom silicon rather than away from them. (technologyreview.com) (semafor.com) The phrase he uses for the destination is “cognitive abundance,” meaning machine intelligence becomes cheap and plentiful enough to spread through everyday software the way electricity spread through factories. Whether that future arrives on his timeline or not, Microsoft is clearly spending as if the scaling era is still in full swing. (technologyreview.com)
