Mustafa Suleyman has a message for anyone predicting an AI ceiling: look at the math. In a detailed essay published by MIT Tech Review, Microsoft’s AI chief lays out a compelling case that the forces driving AI forward aren’t just holding steady, they’re compounding faster than most people realize.
The core argument is straightforward. Three hardware advances are converging simultaneously, and their combined effect dwarfs anything Moore’s Law would predict.
The 50x surprise
Suleyman breaks the acceleration into three layers:
- Faster chips: Nvidia’s GPUs jumped from 312 teraflops in 2020 to 2,250 teraflops today, a sevenfold leap in six years. Microsoft’s own Maia 200 chip, launched in January, delivers 30% better performance per dollar than anything else in its fleet.
- Faster memory: High bandwidth memory (HBM3) stacks chips vertically, tripling data throughput so processors stay busy instead of sitting idle.
- Massive interconnection: Technologies like NVLink and InfiniBand now link hundreds of thousands of GPUs into warehouse-scale supercomputers that operate as single systems.
The result? A training job that took 167 minutes on eight GPUs in 2020 now finishes in under four minutes. Moore’s Law would have predicted a 5x gain. The actual number: 50x.
What stands out here is how Suleyman frames the software side. Research from Epoch AI shows the compute needed to hit a fixed performance level halves every eight months, roughly three times faster than traditional Moore’s Law cycles. Serving costs for some recent models have collapsed by a factor of 900 on an annualized basis.
The 1,000x horizon
The forward-looking numbers are where things get genuinely wild. According to MIT Tech Review, leading labs are growing capacity at nearly 4x per year. Global AI-relevant compute is forecast to reach 100 million H100-equivalents by 2027. Suleyman projects roughly another 1,000x in effective compute by the end of 2028.
His prediction for what this enables: the shift from chatbots to near-human-level agents. Not assistants that answer questions, but “teams of AI workers that deliberate, collaborate, and execute”. These are systems that write code for days, manage logistics, and negotiate contracts over weeks-long projects.
This is significant because it comes from someone who controls the infrastructure budget. Suleyman isn’t a pundit speculating from the sidelines. He runs Microsoft AI and oversees the company’s chip strategy, data center buildout, and product roadmap. When he says scaling won’t stall, he’s also signaling where Microsoft is placing its bets.
The energy question
Suleyman acknowledges the obvious constraint head-on. A single AI rack now consumes 120 kilowatts, enough to power 100 homes. His answer: solar costs have dropped nearly 100x over 50 years, and battery prices fell 97% in three decades. He sees a “pathway to clean scaling coming into view.”
That’s the weakest part of the argument. Solar and battery cost curves are real, but building out 200 gigawatts of new compute capacity annually by 2030, equivalent to the peak energy use of the UK, France, Germany, and Italy combined, requires permitting, grid infrastructure, and political will that don’t follow exponential curves.
What practitioners should take away
If Suleyman is even directionally right, the practical implications are clear. Companies building AI products should plan for capabilities that are radically cheaper and more powerful within 18–24 months. Architectures designed around current model limitations will age fast. And the competitive moat increasingly shifts toward whoever can deploy agents, not just models, at scale.
The skeptics may eventually be proven right about some wall, somewhere. But the data so far keeps pushing that wall further out. More details on Suleyman’s full analysis are available at MIT Tech Review.