Commentator Phil Trubey Claims One Smartphone's Computing Power Equals a 1980s Data Center
Commentator Phil Trubey estimates that if 1980s technology were used to achieve the current iPhone's approximately 2.5 TFLOPS of FP32 graphics computing power, about 15,000 Cray-1 supercomputers would be needed, consuming around 1.7 gigawatts of power, with a size comparable to a large modern data center. He states that the concept of a "supercomputer in your pocket" underestimates the gap.
Data can be cross-verified for magnitude. The Cray-1 was delivered from 1976 to 1977, with a peak performance of about 160 MFLOPS, a total power consumption of about 115 kilowatts, and a weight of about 5.5 tons including cooling, priced around $8 million. Dividing 2.5 TFLOPS by 160 MFLOPS gives approximately 15,625 times, and multiplying by 115 kilowatts results in about 1.7 gigawatts. The IBM personal computer began shipping in 1981, almost concurrently with the microprocessor revolution and this cabinet-sized supercomputer.
The current iPhone's graphics peak performance is around 1.7 to 2.3 TFLOPS FP32, while the neural network engine typically claims around 35 TOPS, which is not the same metric as floating-point peak performance. The estimate only considers the GPU and does not account for the CPU and neural network engine, nor the additional amplification from memory bandwidth, interconnect, and cooling on vector machines. The Cray-1 is a vector register machine, while the smartphone is a system on a chip; dividing the peaks is a thought experiment, not a directly comparable workload.
The implication lies in energy efficiency rather than nostalgia. The same magnitude of floating-point throughput has been compressed from data center motor units and Freon cooling into a few watts in a pocket battery. Supercomputers have transitioned from national-level orders to standard consumer electronics, and the narrative of "big data centers" for training and inference contrasts with the already reduced endpoint peak performance by more than four orders of magnitude.
Mechanically, there is no capital turnover, which serves as a price anchor in the history of computing power. The buyers are users who still regard smartphones as terminals, while the sellers are foundries that sell nanotechnology into consumer electronics. The 1.7 gigawatts serves as a reminder: today, the electricity cost of a training cluster is equivalent to what an entire city would have needed in 1980 to build a chip small enough to fit in a pocket. Beneficiaries are endpoint model deployments; those under pressure are still planning computing power based on cabinet counts.
Source: Public Information
ABAB AI Insight
The 115 kilowatts of the Cray-1 was necessary to push 160 million floating-point operations through a 12.5 nanosecond clock, with power supplied by motor units and cooling by Freon. The iPhone compresses a similar magnitude of peak performance into a few square millimeters, relying on forty years of transistor density, dedicated matrix units, and packaging cooling. Trubey only multiplied the GPU, already achieving gigawatts; if the neural network engine's trillion operations were stacked in the same crude manner, the data center would need to be enlarged further.
The capital pathway is that semiconductors have transformed national-level capital expenditures into consumer electronics bills. In 1980, purchasing a Cray required a national laboratory budget; by 2026, equivalent peak performance will be spread across the average price of a smartphone. Foundries and packaging plants have taken over the rents that once belonged to supercomputer manufacturers. Cloud training clusters may appear as "new Crays," but they are actually re-stacking already miniaturized computing power back into data centers, aimed at model scale rather than single floating-point operations.
The analogy is not a simple Moore's Law poster, but rather a comparison of cars to horse-drawn carriages: after the collapse of energy and volume per unit of transport capacity, society changes its ownership structure. The computing industry is in a split period where endpoints have already "outgrown supercomputers" while the training side is re-establishing "supercomputers."
Structural judgment belongs to technological substitution. What is being substituted are cabinets, cooling towers, and dedicated operation and maintenance teams. The mechanism is: as the joules per unit of computing power collapse, computing power transforms from a public good into a pocket accessory; people are rebuilding data centers not because smartphones are not fast enough, but because models need to aggregate billions of pockets into a single training step.
ABAB News · Cognitive Laws
The peak performance in your pocket, measured by old technology, is like a power plant.
After supercomputers shrink, large models are stacked back into data centers.
People marvel at smartphones resembling supercomputers, often forgetting that supercomputers once resembled power plants.