INSUFFICIENT DATA: Why 70 Years of Computing Advances Still Can't Reverse Entropy

TL;DR From UNIVAC-scale machines in 1956 to today’s 1.102-exaflop Frontier supercomputer, every system asked how to reduce the universe’s net entropy has returned some version of “insufficient data,” a verdict reinforced by DESI’s October 2024 cosmology results showing accelerating expansion toward heat death.

• U.S. electricity was 68% coal in 1956 with finite uranium reserves; solar reached 1,419 GW globally by 2023 but supplies just 5.4% of electricity
• Single AI model training runs now consume 10–100 GWh, mirroring the story’s geometric energy-growth problem
• Landauer’s principle proves information erasure itself increases entropy; no verified mechanism exists to reverse it at cosmic scale
• Space-based solar demos have transmitted only 200 mW over 1 km so far

In the summer of 1956, UNIVAC I had been crunching census data for five years and engineers were already imagining self-correcting computers that could optimize an entire planet’s energy supply. The United States generated 68% of its electricity from coal that year, uranium reserves were known to be limited, and the second law of thermodynamics loomed as the ultimate constraint on all physical processes. Those early machines were promptly tasked with humanity’s biggest question: could we ever reverse entropy on a universal scale and escape the heat death of the cosmos? Seventy years later, with AI training runs consuming as much electricity as small cities and DESI’s latest results strengthening evidence for permanent cosmic expansion, the same question is being asked again—this time by both researchers and casual users prompting frontier models. The answer has not improved.

When Room-Sized Computers First Tackled Planetary Energy

The 1956 computing hardware filled 29,000-pound cabinets and drew 125 kW, yet it was already plotting rocket trajectories that would lead to the Moon and Mars missions within fifteen years. Engineers recognized that coal and uranium could not scale to interstellar travel, so the focus shifted to harvesting solar energy directly at planetary scale. By the early 1960s conceptual designs existed for mile-wide orbiting stations beaming power to Earth, exactly the infrastructure needed to retire fission and fossil plants. Fast-forward to 2023: installed solar photovoltaic capacity hit 1,419 GW worldwide and generated 5.4% of global electricity. This is precisely why the 1950s optimism feels both prophetic and sobering—progress is real but still orders of magnitude short of total primary-energy replacement. Meanwhile data-center demand is doubling every few years, driven by models that each require 10–100 GWh to train, recreating the exponential consumption curve imagined decades earlier. [1][2][3]

Landauer’s Limit Turns Computation Into an Entropy Tax

The technical barrier is more fundamental than fuel supply. Since Rolf Landauer’s 1961 proof, we have known that erasing one bit of information must dissipate at least kT ln(2) energy as heat, making every computation an irreversible entropy increase. The monolithic, planet-spanning computer concept of the 1950s has been replaced by millions of distributed servers; Frontier now delivers 1.102 exaflops yet still converts most of its input electricity into waste heat. Asking these systems to devise a net-negative-entropy process for the observable universe collides with the same thermodynamic wall—current quantum thermodynamics offers no workaround. This differs sharply from efficiency gains such as better solar inverters or liquid-cooled racks; those merely slow the approach to equilibrium. The real question is whether any future architecture, even one that merges biological and silicon substrates, can escape a law that applies to all information-bearing systems. [4][5]

Big Tech’s Solar PPAs and the 24/7 Data-Center Gap

Google, Microsoft, and Meta have begun signing long-term solar power purchase agreements and exploring small modular reactors precisely because their AI clusters cannot run on intermittent generation. Caltech’s MAPLE experiment in 2023 successfully beamed 200 milliwatts across one kilometer in orbit—a proof of concept for the invisible power beams envisioned in the 1950s, yet still nine orders of magnitude below utility scale. Bridging day/night cycles for always-on compute requires either enormous battery farms or constant baseload, both of which raise costs and material demands. DESI’s Year 3 data released in October 2024 further cements the cosmological constant’s role, pointing toward a Big Freeze with no natural mechanism for entropy reversal. The hype that “AI will solve climate” therefore misses the deeper constraint: without new physics, more powerful models simply accelerate the approach to the universal energy limit they were supposed to transcend. Real-world adoption will hinge on whether space-based solar can reach gigawatt scale by the 2030s or whether we must accept strict efficiency caps on model size. [6][7]

The pattern is clear: each leap in computing power lets us pose the entropy question more elegantly, yet the physical universe has so far refused to upgrade its answer. As hyperscale data centers begin to rival national electricity budgets, the practical stakes have moved from philosophical curiosity to engineering reality. The lingering puzzle is whether we will eventually discover a loophole in the second law—or whether the only genuine solution is the one that requires us to step outside the universe itself.

References

[1] U.S. Energy Information Administration, Monthly Energy Review historical tables (1950–2024) - https://www.eia.gov/totalenergy/data/monthly/

[2] IRENA, Renewable Capacity Statistics 2024 - https://www.irena.org/Publications/2024/Mar/Renewable-capacity-statistics-2024

[3] HPC Top500, June 2024 - https://top500.org/lists/top500/2024/06/

[4] DESI Collaboration papers, arXiv (October 2024) - https://arxiv.org/abs/2411.12022

[5] Isaac Asimov, I. Asimov: A Memoir, Doubleday, 1994 - https://www.goodreads.com/book/show/209024.I_Asimov

[6] Caltech Space Solar Power Project, MAPLE experiment results (2023) - https://www.caltech.edu/about/news/in-a-first-caltechs-space-solar-power-demonstrator-wirelessly-transmits-power-in-space

[7] Isaac Asimov: The Last Question (1956) - https://hex.ooo/library/last_question.html