Elon Musk Imagines A Terawatt Of Compute, Or About 1.43 Billion Gpus And 2.1x The Power Output Of The Us

By now, we’re each beautiful utilized to Elon Musk’s large dreams crossed his various companies. Tesla is transitioning nan world to sustainable energy, and real-world AI successful nan shape of humanoid and driverless robots, SpaceX is trying to make humanity multi-planetary (Mars) and Neuralink is trying to revolutionlise healthcare pinch a generalised encephalon interface to sanction a few.

The latest posts from Musk are connected a full different level – Musk posts that he’s been considering what nan fastest measurement to bring a terawatt of compute online would be.

To understand conscionable really ample that connection is, nan world compute capacity today, is apt successful nan scope of 1 to 10 zettaFLOPS (10^21 to 10^22 FLOPS), dominated by information centers, unreality providers, and AI infrastructure successful nan U.S., China, and Europe.

A terawatt of compute, assuming 10^11 to 10^12 FLOPS/W, translates to 10^23 to 10^24 FLOPS (100 zettaFLOPS to 1 yottaFLOPS). This is 10 to 1000 times nan estimated world compute capacity of 1 to 10 zettaFLOPS (10^21 to 10^22 FLOPS) successful 2025.

This would require a monolithic summation successful powerfulness infrastructure, balanced to 20x existent information halfway powerfulness aliases 1% of world electricity, making it a awesome theoretical exercise, alternatively than a applicable reality today.

Musk shared an estimate that suggests nan powerfulness request would beryllium nan balanced of each powerfulness produced successful America today. With immoderate thief from Grok, it looks for illustration that whitethorn beryllium a small connected nan debased side.

A terawatt of compute requires 1 TW of power, producing 10^23 to 10^24 FLOPS (100 zettaFLOPS to 1 yottaFLOPS). The U.S. generates ~4,200 TWh/year successful 2025, balanced to an mean powerfulness of 0.48 TW.

Thus, 1 TW of compute powerfulness is astir 2.1 times nan full U.S. mean electrical powerfulness output and ~77% of its installed capacity (1.3 TW).

Musk’s existent pinned station connected X is simply a reply to a station by Jesse Peltan, which appears to person started Musk down this statement of thinking. In nan reply, Musk suggests nan world would request a LOT much star to execute this.

Then power harnessed will summation possibly a billionfold if we make it to Kardashev II, pinch abstraction star power, and different billionfold if we harness nan power of our galaxy.

The Kardashev Scale, introduced by Nikolai Kardashev successful 1964, measures a civilization’s technological advancement by its power consumption, pinch Type I harnessing planetary power (4 × 10^12 W), Type II stellar power (4 × 10^26 W), and Type III galactic power (4 × 10^37 W).

The station envisions humanity progressing from Type I to Type III, perchance expanding power usage by a trillionfold done star and space-based power.

Musk provides different reference to thief our mini brains get astir this ridiculously ample figure, suggesting that nan power produced by astir 10 Starships would beryllium akin to power level required to present 1 Terawatt of computational power.

One point we’re not considering present is what you’d do pinch that compute capacity. AI is surely nan hungriest of quiet beasts and arsenic yet, it’s expertise to amended does not look to beryllium impacted by scaling laws. This yet intends nan sophistication and capacity of AI is constricted by your wallet (and power).

Imagine a web of large-scale information centres strategically located crossed nan globe, possibly successful areas pinch abundant renewable power sources for illustration star aliases geothermal.

What would it cost?

Let’s for a infinitesimal entertain what achieving 1TW of power whitethorn cost.

The costs of operating 1 terawatt of compute (10^23 to 10^24 FLOPS) is astir $7.3 trillion to $12.9 trillion per year, pinch a midpoint of $10.1 trillion/year. This includes:

• Electricity: $700 billion-$1.14 trillion/year (midpoint $911 billion, assuming PUE 1.3, $0.08/kWh).
• Capital (hardware, information centers, maintenance): $6.62-$11.74 trillion/year (midpoint $9.18 trillion, assuming mixed hardware and 4-year lifespan).

This costs is ~10% of world GDP, balanced to 25-30x existent world information halfway spending aliases 2-3x U.S. energy consumption, making it economically and energetically impractical today.

As we see compute and what’s possible, we besides request to see what this really looks for illustration successful position of hardware.

Let’s presume nan compute is GPU-based (e.g., NVIDIA H100 aliases balanced successful 2025). A H100 consumes ~700 W and costs ~$30,000 (2023 price; presume $25,000 successful 2025 owed to marketplace scaling).

The number of GPUs you would request to create 1 TW of compute, is astir 1.43 cardinal GPUs (10^12 W ÷ 700 W). Given nan largest buys from Nvidia astir measured successful nan hundreds of thousands today, this is much than 1000x nan amount, again nosy to deliberation about, but acold from practical.

It’s nosy to dream sometimes.