Compute Exponential Stack

Masayoshi Son's specific arithmetic for why the current compute investment wave is not an overspend, first stated in full public form at FII Miami 2026. The argument multiplies three independently compounding factors, chip volume, per-chip performance, and model algorithmic efficiency, to produce a combined improvement rate that makes traditional cost-benefit critiques category-incorrect.

The three factors

Factor one, chip volume. The Stargate infrastructure buildout provides at least 10x the number of chips per year. This is a capital-allocation claim rather than a physics claim: more money deployed into chip procurement means more total compute available.

Factor two, compute per chip. Each new chip generation delivers roughly 10x more compute capability per chip, approximately true across successive Nvidia GPU generations and the hardware-engineering parallel to Moore's Law. It requires no algorithmic improvement, just continued semiconductor research and development.

Factor three, model capability per unit compute. Each new OpenAI model generation delivers roughly 10x more capability per unit of compute: the algorithmic efficiency dimension, meaning better training procedures, architectures, and reasoning techniques extract more intelligence from the same hardware budget.

Combined: 10 times 10 times 10 equals 1,000x effective intelligence per cycle. Each cycle runs 12 to 18 months. Three cycles from today's starting point yields roughly a billion x cumulative improvement, in Son's words, "three cycles is coming in just a several years."¹

The investment rationale

Son uses the billion-x endpoint to build a direct return-on-investment argument that rebuts what he calls the DeepSeek efficiency critique, the view that Stargate-scale spending is overspending because more efficient training reduces the required compute investment. The rebuttal runs as follows: a billion-x superintelligence will replace 5 to 10 percent of global GDP, estimated at roughly 180 trillion dollars, within ten years. Five percent of 180 trillion dollars is 9 trillion dollars a year; ten percent is 18 trillion dollars a year. Against 9 to 18 trillion dollars a year in perpetuity, spending 500 billion dollars, or even 2 to 3 trillion dollars, on infrastructure costs a fraction of a single year's return. And in a winner-take-most market, with OpenAI holding roughly 80 percent market share in Son's framing, a small quality difference captures a hugely disproportionate share of that return.¹

Son's summary: "If return is $9 to $18 trillion per year, why should you save, why should you try to be efficient? For what? I don't get it."¹

Relationship to scaling laws

The Compute Exponential Stack is a capital-markets extension of the standard scaling laws observation that model capability scales as a power law with compute. The key addition Son makes is separating the three axes rather than treating compute as a single variable: compute itself is growing 100x per cycle, from 10x volume times 10x per-chip efficiency, even before any algorithmic improvement, and then a third 10x multiplier is stacked on top from model architecture improvement.¹

The benchmark evidence

Son anchors the starting point of the stack with a concrete benchmark: OpenAI's O3 model has already passed PhD-level exams in mathematics, physics, medicine, and law simultaneously. His implicit claim is that if the starting point of the billion-x journey is already PhD-level across major knowledge domains, the destination is so far beyond current human comprehension that precision in the exact multiplier is beside the point. "What is the definition of AGI? Already it is so smart. But just imagine that intelligence, becoming a billion x smarter."¹

Tensions and open questions

The claim that model capability improves a sustained 10x per generation is empirically contested; capability gains have been uneven across benchmarks, with some showing dramatic improvement and others plateauing. The 5 to 10 percent GDP replacement figure and ten-year window are asserted without a bottoms-up model, eliding labor market transitions, political resistance, regulatory friction, and deployment timelines. And the efficiency critique Son dismisses may still matter for the distribution of who captures the AI value even if the total market is large: efficient training lowers the compute moat, which threatens the market-share durability of the current leader even if the value pool itself remains enormous.