One Gigawatt per Week

Sam Altman just raised the stakes. In a post on NVIDIA’s blog, he framed access to AI as a potential human right and set out an audacious goal: “a factory that can produce a gigawatt of new AI infrastructure every week.”

Pause on that. We’ve spent decades measuring compute in teraflops, megawatts, or cloud credits. Now we’re talking about gigawatts as the unit of intelligence production. Not a metaphor—a literal industrial project to mass-produce cognition at planetary scale.

If Altman is right, the question facing leaders in government, enterprise, and startups is:

“What will we do in a world where intelligence is abundant and cheap, constrained only by energy and imagination?”

I’ve argued in Digital Superintelligence that we are already spinning the flywheel where compute, energy, and intelligence amplify each other. This new factory metaphor makes it visceral. We are building an industrial base for cognition itself. And with that comes the responsibility and opportunity of deciding how to allocate it.

1) From Chips to Gigawatts

We are witnessing the industrialization of intelligence.

Altman’s gigawatt framing echoes what we’ve seen before in computing history:

• Electricity: Factories once scaled production linearly until electrification turned capacity into a utility.
• Cloud Computing: Enterprises once bought servers until hyperscalers made compute elastic.
• AI Factories: We are now at the same threshold; moving from boutique model training runs to continuous, gigawatt-scale intelligence production.

NVIDIA and OpenAI are effectively building the Hoover Dams of cognition. They’ll power industries, nations, and in time, missions.

And Altman is blunt about the stakes: as he noted in his NVIDIA post, with 10 GW of compute, AI might cure cancer or deliver personalized tutoring to every child. Without it, we’ll be forced to choose. That’s a future no one wants to manage.

2) Stop Asking About Models

For years, our default question was: Which model is best?

That question is already downstream.

As I argued in Build for the Buyer That Never Blinks, the most valuable customer your system will face won’t be a human. It will be an AI agent parsing your schema, benchmarking your latency, and acting. The constraint is not model cleverness. It’s whether you have the compute, infrastructure, and protocols to even participate in an agent-first economy.

The winners will be the operators of the factories; those who can scale cognition like steel, electricity, or semiconductors.

3) The Factory as Architecture

Let’s unpack what “a gigawatt a week” actually means.

3.1 The Jagged Skyline of Demand

As I wrote in Jagged Intelligence, AI capability is not smooth. It spikes and dips across domains. Abundant compute doesn’t smooth that jaggedness, it multiplies it. The allocation challenge becomes:

• Do we amplify peaks (superhuman protein folding, autonomous logistics)?
• Or do we spend compute to cover valleys (commonsense reasoning, nuance in social dynamics)?

3.2 Compound Systems as Assembly Lines

In The Rise of Compound AI Systems, I argued the future is not monolithic models but ensembles. A gigawatt factory won’t just output a single giant brain. It will output fleets of specialized agents—planners, solvers, verifiers—choreographed into compound workflows.

3.3 Self-Improvement as Continuous Production

The Age of Self-Improving Software showed us that systems like Sakana’s Darwin Gödel Machine can mutate, test, and evolve themselves. Now imagine that loop running inside Altman’s gigawatt factory.

The product isn’t just more compute. It’s compute that continuously makes itself smarter.

4) Compute as Strategy

So what does this mean in practice?

Public Sector

Compute allocation becomes policy. In the U.S. Civil Sector, the question will sound like this: do we direct 5 GW toward cancer research or toward fraud detection in Medicare? These are not IT procurement decisions. They are mission choices at the scale of national priorities.

Enterprises

Enterprises must measure not just AI spend but intelligence yield per kilowatt. In Route, Don’t Guess, I showed how model routing optimizes cost-to-cognition. At gigawatt scale, this becomes board-level strategy: route intelligence like a utility, escalate only when necessary, and treat compute as the new balance sheet.

Startups

For startups, abundance is both blessing and trap. Cheap cognition lowers entry barriers but raises orchestration stakes. In The Sandbox Economy Is Coming, I argued that agent markets will emerge where software is both buyer and seller. Startups must design for permeability and trust—your schema is your storefront, your proofs are your brand.

5) Five Actions for Leaders

Here’s how to prepare for the gigawatt era today:

1. Audit Compute-to-Outcome Efficiency. Track not just how much you spend on GPUs, but what mission outcomes you produce per kilowatt.
2. Build Agent-First Surfaces. Publish an Agent Storefront Manifest with machine-readable capabilities, policies, and proofs. If agents can’t read you, you don’t exist.
3. Instrument Trust and Security. As I argued in Securing the Agent Surface, context itself becomes an attack surface. Build trust meshes, not firewalls.
4. Scenario-Plan Allocation Dilemmas. If compute becomes a constrained resource, which missions will your organization prioritize? Better to practice the trade-offs now than improvise later.
5. Invest in Evolution Governors. In the self-improving era, your most valuable leaders will be curators who decide which agent mutations advance to production.

6) Abundance as Doctrine

Altman’s vision is stark: we should never be forced to choose between curing cancer and educating children. That is the moral logic of abundant intelligence.

In Digital Superintelligence, I argued that once intelligence and energy become abundant, scarcity itself collapses. What follows is not just a technical shift but a civilizational one: governance by allocation, economies where agents transact at machine speed, and societies where intelligence is treated like a utility.

We are not there yet. But the groundwork is being laid: factories for cognition, protocols for agents, and markets for machine-to-machine trade. Altman himself has been transparent about this trajectory in his personal blog, where he frames abundant intelligence as both inevitable and essential.

7) The Strategic Question of Our Time

The strategic question of this decade is no longer Which model should we use?

It is: What will we do with abundant intelligence?

Will we allocate it to missions that compound human wellbeing or let it chase only profit? Will we build trust rails that make agent economies safe or watch fragility cascade across markets?

This is the coolest and most important infrastructure project ever attempted, as Altman said. It is also the most consequential governance challenge.

So let me leave you with this question:

If intelligence becomes a utility as ubiquitous as electricity, what doctrine will guide how we allocate it?

Because soon, that choice won’t be theoretical. It will be a weekly gigawatt.