https://youtu.be/XQnY2WONwqE?si=97haCFis3qrHUXWz
Yes. Superconductive (or superconducting) computing is one of the most promising technologies for making AI infrastructure dramatically more energy-efficient. However, it is unlikely to make today’s giant AI data centers disappear overnight. Instead, it may allow far more computation to fit into a much smaller power envelope.
Here’s why.
What is superconductive computing?
Ordinary computer chips are made from silicon. Every time electricity flows through the transistors and metal wires, some energy is lost as heat because of electrical resistance.
A superconductor is a material that, when cooled below a critical temperature, has essentially zero electrical resistance.
That means:
- almost no heat generation
- almost no energy lost in wiring
- signals travel extremely fast
- much higher efficiency
Instead of conventional CMOS transistors, superconducting computers often use devices called Josephson junctions.
Why AI data centers are becoming gigantic
Today’s AI data centers are not huge primarily because computers are physically large.
They are huge because of heat.
A modern GPU may consume:
- 700–1000 W
- nearly all of which becomes heat
A cluster of
- 100,000 GPUs
may consume
- 100–150 MW
The computers themselves occupy only part of the building.
The rest is devoted to
- cooling
- power delivery
- transformers
- backup systems
- water pipes
- fans
- air handling
In many AI facilities, cooling infrastructure occupies almost as much space as the computing hardware.
What superconducting computing changes
If electrical resistance nearly disappears,
then
Heat ↓↓↓
Power consumption ↓↓↓
Cooling ↓↓↓
Power supplies ↓↓↓
The result is that much more computing can be packed into the same footprint.
Some estimates suggest superconducting logic could reduce energy per operation by 100× to even 1,000× for certain workloads, though achieving this in practical AI systems remains an active research challenge.
But there is a catch…
Superconductors must stay extremely cold.
Typically:
- 4 K (liquid helium)
- or 20–77 K depending on the material
So instead of cooling GPUs with chilled water,
you cool the computer inside a cryostat.
This seems wasteful, but surprisingly:
Cooling a superconducting computer can consume less total energy than cooling thousands of conventional GPUs.
Think of it this way:
Current AI
produce enormous heat, then spend enormous energy removing it.
Superconducting AI
avoid producing most of the heat in the first place.
Can it shrink AI data centers?
Probably yes—but not because the computers themselves are magically tiny.
Instead:
Current AI data center
100 MW
70 MW computing
30 MW cooling
Future superconducting center (illustrative)
15 MW
12 MW computing
3 MW cryogenics
The exact numbers are uncertain and depend on the technology, but the direction is clear: much higher computational density with much lower energy use.
Could AI eventually fit into a shipping container?
Possibly.
Many engineers envision something like this:
Today
One GPT-scale model
= multiple warehouse-sized buildings
Future
Same computation
= one room
later
= one cabinet
later
= perhaps a refrigerator-sized system
This progression depends not only on superconducting electronics but also on advances in memory, networking, packaging, and algorithms.
Other technologies that will likely combine with superconducting computing
Superconductors alone are unlikely to define the future. More likely, they will be combined with several other technologies:
- Photonic computing — uses light instead of electrons for extremely fast, low-energy communication.
- Neuromorphic computing — hardware inspired by biological neural systems for greater efficiency on certain AI tasks.
- 3D chip stacking — places memory directly above compute to reduce data movement.
- Cryogenic memory — memory designed to operate efficiently at superconducting temperatures.
The winning architecture may integrate all of these.
My longer-term prediction (2035–2050)
Your previous questions about space computing, neuromorphic computing, embodied AI, and agentic AI all point toward a broader trend.
Today’s AI infrastructure resembles the vacuum-tube computers of the 1950s:
- enormous
- power hungry
- centralized
- expensive
Over the next two decades, I expect AI hardware to become progressively:
- cooler
- denser
- more distributed
- dramatically more energy efficient
Superconducting computing could play a major role in that transition, especially for AI inference and specialized accelerators, though silicon will likely remain dominant for many applications for years.
That also connects to one of the themes we’ve discussed before: the “物的反撲” (the return of things). As computation becomes cheaper and more efficient, intelligence is likely to diffuse out of hyperscale data centers into robots, vehicles, factories, scientific instruments, and everyday objects. Rather than AI living only in a few giant warehouses, computation may become embedded throughout the material world—moving from centralized “AI factories” toward a pervasive computational environment.
