How Much AI Inference Does a 1-Gigawatt Data Center Really Deliver?

A growing narrative claims that AI is running into hard limits: power, energy, and infrastructure. That claim does not survive contact with basic arithmetic.

Using today’s hardware and today’s models, a single 1-gigawatt (GW) AI data center already delivers more inference capacity than an entire population can realistically consume. This article walks through the numbers and explains what that means for AI infrastructure, energy economics, and the so-called “AI bubble.”

No speculation. Just current-state math.

What 1 Gigawatt of AI Compute Actually Means

A 1 GW facility is large, but it is not hypothetical.

Metric	Value
Utility power	1,000 MW
Typical PUE	~1.1
Power available to compute	~909 MW

Facilities at this scale are already being planned and permitted.

Hardware Assumptions (Conservative)

This analysis assumes current-generation inference hardware, not roadmap claims.

Component	Assumption
GPU class	H200-class
GPUs per node	8
Sustained draw per node	~8.5 kW
Total nodes	~107,000
Total GPUs	~855,000

This fits cleanly inside a 1 GW envelope with standard cooling and redundancy.

From Power to Inference Capacity

Raw inference throughput

Benchmarks on production-grade language models show approximately: ~3,000 output tokens/sec per GPU

Metric	Value
Total output tokens/sec (raw)	~2.5B
Tokens/day (raw)	~2.2 × 10¹⁴

The reasoning tax (the part most analyses omit)

High-quality answers require internal reasoning. A conservative assumption of 100 reasoning tokens per output token yields:

Metric	Value
Net output tokens/day	~2.2T
Tokens/person/day (330M people)	~13,000
Approx. words/person/day	~10,000

Power cost per person

Metric	Value
Power per capita	~3 watts
Equivalent	Small LED bulb

Conclusion: Expert-level reasoning is already cheap at scale.

Image Generation (Quality-Adjusted)

Raw image benchmarks overstate real output. Applying a ~5× quality penalty to match modern, high-fidelity generation still leaves massive surplus capacity.

At this scale, limiting image output is a product decision, not a hardware constraint.

Energy Economics: Power Is Not the Bottleneck

Metric	Approximate Share
Share of U.S. electricity	~0.2%
Capital cost (order of magnitude)	~$50B
Dominant ongoing cost	Hardware refresh, not power

Electricity matters—but hardware lifecycle economics dominate. Claims that AI will imminently stall due to energy limits are overstated.

What This Actually Says About the “AI Bubble”

A surface reading of these numbers raises a fair question: If 1 GW already delivers ample daily inference for an entire population, why are hyperscalers targeting 50 GW+ buildouts?

The answer is not “waste.” It is workload mix.

• Inference for text reasoning is already abundant
• Training frontier models is not
• Video, real-time simulation, and interactive environments are far more compute-intensive

So the real issue is not whether the buildout is excessive—but whether additional capacity is being converted into durable capability, or just raw output.

We’ll address that distinction directly in a future post.

Why This Matters to HitWit

At HitWit, we focus on using AI to deliver cognitive depth to everyone, not bespoke intelligence for a few.

That requires:

• Multimodal content tailored to an individual’s current cognitive state
• Grounding in the information they need to act
• Economics that work at population scale, not consulting scale

Understanding real inference capacity and energy costs is foundational to building systems that meet those criteria.

Read the Full Technical Analysis

For full assumptions, benchmarks, and sensitivity analysis, see the white paper:

→ The National Hive Mind: A Blueprint for AI Abundance