$700B Hyperscaler Splurge Can't Outrun Physics, Review Argues

By Tom Whitfield

Six hyperscalers are on track to spend more than $700 billion this year on computing infrastructure, roughly double the record set in 2025. The capex buys performance by pouring concrete, copper, and silicon into new sites faster than the electrical grid can feed them. It does not patch a quieter problem. Five trend lines that defined 20th-century computing have bent or broken.

Which scaling laws have stalled?

• Moore's Law (transistor density): slowed.
• Dennard's Law (power per chip area stays flat as transistors shrink): ended around 2004, when clock speeds stalled near 3 GHz and Intel canceled its planned 4 GHz chip.
• Koomey's Law (computation per joule): flagging.
• Kryder's Law (magnetic storage density): crawling.
• Keck's Law (fiber bandwidth): flattening.

None is a physical law. All are observations about engineered systems.

How far has the descent actually gone?

In the 78 years since the first transistor, density has climbed 10 orders of magnitude, to more than 100 million transistors per square millimeter. Speed has risen 12 orders. Energy per operation has dropped 17 orders, from about 100 joules for an early vacuum-tube machine to roughly a femtojoule per logic-gate switch today. That is the ceiling side. The floor sits far lower.

What does physics actually allow?

In a 2000 Nature paper, MIT physicist Seth Lloyd asked how much computing a kilogram of matter, squeezed into one liter, could perform if driven to absolute limits. He called the imaginary machine the ultimate laptop. By E=mc², that kilogram holds about 9×10^16 joules, near what the U.S. draws from the grid every two days. Spend it all, and the machine runs at roughly 5×10^50 operations per second, about 35 orders of magnitude beyond a leading chip today. Concentrated, that energy becomes a plasma of roughly a billion degrees, closer to the universe's first minutes than to any laptop.

Three other limits govern energy, memory, and communication.

• Energy. In 1961, IBM's Rolf Landauer proved that erasing one bit at room temperature must release at least 2.9×10^-21 joules. Current chips spend several hundred thousand times that per switch; thermal noise pins any reliable transistor a hundred to a few thousand times above the Landauer floor, no matter how small it gets.

• Memory. In 1981, Jacob Bekenstein, working out black hole thermodynamics, showed that any bounded region's information is set by its size and energy. Lloyd's kilogram could in principle hold about 10^42 bits, or roughly 10^36 bits per cubic millimeter, about 30 orders of magnitude beyond today's storage density.

• Communication. In 1973, Soviet mathematician Alexander Holevo fixed the cap on what a stream of quanta can convey: about 10^17 bits per second per watt of light. Today's optical interconnects deliver roughly one-millionth of that rate for the same power.

Where is the gap narrowest?

Energy. Less than six orders separate today's chips from the Landauer floor. Conventional CMOS cannot fully close it. Reversible logic can. Charles Bennett at IBM showed in 1973 that any computation can be arranged to throw nothing away and pay no erasure toll. Vaire Computing, a startup commercializing that approach, tested its first chip in 2025; its resonant circuits recover part of the switching energy instead of venting it as heat.

Communication is next, six orders from the Holevo cap. Nvidia and Broadcom now ship optical interconnects built directly against the silicon, with signals leaving the chip package already on photons. As classical interconnects descend, they meet quantum interconnects that already operate on single photons. Both groups bump the same ceiling at the bottom.

Memory sits about 30 orders away. DNA holds a bit in roughly 30 atoms; Feynman, playing conservative in 1959, guessed 125. Density is not the barrier. Writing cost and speed are, both still orders of magnitude from anything data-center practical.

The compute-speed gap, 35 orders, is the largest and least bridgeable. Closing it would require a machine that computes on every degree of freedom in a lump of matter. Quantum computers go part of the way by using atoms and spins as computing elements. Today's machines do the opposite, spending trillions of atoms per fragile qubit; thousands of them combine into one reliable logical qubit.

How do you shrink a 35-order gap?

Build from the physics up. The transistor was a physics experiment before it was an industry. The laser was a laboratory curiosity before it carried transoceanic calls. Giant magnetoresistance went from physics discovery to the read head of every hard drive within a decade. In each case, the lab demonstration showed the floor was reachable; engineers then scaled it up.

The 2026 ceiling problem is real. Hyperscaler capex has doubled because compute demand grows faster than transistor gains. Training compute for AI models has doubled roughly every six months since 2010, well past Moore's pace. Power grids, land, and cooling water cap how much scale-out any region can absorb. The room above is closing.

The room below, by the four physical floors, remains mostly empty. As Volkan Gurses, a postdoctoral researcher at MIT and a visiting researcher at Nvidia, wrote in MIT Technology Review: "These gaps are invitations. We can progress by building more or by approaching the floor, which means making each device do more with less." For R&D planners, that turns the next decade's road map into a portfolio question: how much capital chases the ceiling, and how much funds the lab experiments that descend toward the floor. The next machines, Gurses argues, will not be reached by descent. They will be built from the bottom up.