> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Sam Altman Says AI Now Uses 100,000x Less Energy — So Why Is OpenAI Building Nuclear Reactors?
- URL: https://wire.fourthweb.ai/sam-altman-says-ai-now-uses-100-000x-less-energy-so-why-is-openai-building-nuclear-reactors/
- Published: 2026-11-03T05:30:57.000Z
- Updated: 2026-11-03T05:31:00.000Z
- Description: The price per query crashed 99.999%, but OpenAI is still building nuclear reactors — turns out efficiency is only half the equation. Sam Altman told the ADIPEC energy conference that AI queries now use up to 100,000 times less energy than they did a few years ago, a dramatic efficiency gain
- Author: Travis Wright
- Tags: Real World Assets, AI Agents, AI Infrastructure, OpenAI, Anthropic

**The price per query crashed 99.999%, but** [**OpenAI**](https://wire.fourthweb.ai/tag/openai/) **is still building nuclear reactors — turns out efficiency is only half the equation.**

### The Summary

- [Sam Altman told the ADIPEC energy conference that AI queries now use up to 100,000 times less energy](https://cryptobriefing.com/sam-altman-ai-energy-efficiency-adipec/?ref=wire.fourthweb.ai) than they did a few years ago, a dramatic efficiency gain
- [Despite efficiency improvements, soaring AI demand means energy remains a critical bottleneck](https://cryptobriefing.com/sam-altman-ai-energy-efficiency-adipec/?ref=wire.fourthweb.ai) — more efficient models don't matter if everyone runs 10,000 more queries
- [Altman also declared AI is "now mostly smarter than top scientists"](https://cryptobriefing.com/sam-altman-ai-smarter-than-top-scientists/?ref=wire.fourthweb.ai) at white-collar work, signaling a shift from automation of rote tasks to knowledge work
- [He dismissed AI tokens as "a kind of fake unit"](https://cryptobriefing.com/sam-altman-ai-tokens-fake-unit/?ref=wire.fourthweb.ai) for measuring business value, suggesting the industry needs better metrics than consumption

### The Signal

OpenAI has made [AI queries 100,000 times more energy-efficient](https://cryptobriefing.com/sam-altman-ai-energy-efficiency-adipec/?ref=wire.fourthweb.ai) than they were just a few years ago. That's not incremental progress. That's the kind of algorithmic leap that changes cost structures overnight. A query that once required a small power plant now runs on the equivalent of a nightlight. For context, that's roughly the difference between driving across the country and walking across your living room.

But here's the part Altman didn't sugarcoat: [demand is growing so fast that total energy consumption is still a critical challenge](https://cryptobriefing.com/sam-altman-ai-energy-efficiency-adipec/?ref=wire.fourthweb.ai). When something becomes 100,000 times cheaper, people don't use it 100,000 times more — they use it 1,000,000 times more. This is Jevons paradox playing out in real time. More efficient models don't solve the energy crisis if every knowledge worker suddenly has 47 [AI agents](https://wire.fourthweb.ai/tag/ai-agents/) running in the background.

> "Efficiency gains get eaten by exponential demand growth."

Which explains why OpenAI and other AI labs are still [cutting deals with nuclear power companies](https://cryptobriefing.com/sam-altman-ai-energy-efficiency-adipec/?ref=wire.fourthweb.ai) and building out massive infrastructure. Altman's message at the energy conference was clear: we've solved the per-query problem, but not the aggregate problem. The grid can't keep up with what's coming.

Meanwhile, [Altman is now claiming AI is "mostly smarter than top scientists" at white-collar work](https://cryptobriefing.com/sam-altman-ai-smarter-than-top-scientists/?ref=wire.fourthweb.ai). Not "approaching" or "getting close." Smarter. Present tense. This isn't about autocomplete anymore. It's about agents that can do knowledge work at a level that previously required a PhD and a decade of experience. The efficiency story and the capability story are the same story: the economics of intelligence just changed.

The token critique adds another layer. [Altman called AI tokens "a kind of fake unit"](https://cryptobriefing.com/sam-altman-ai-tokens-fake-unit/?ref=wire.fourthweb.ai) for measuring business value. He's right. Charging by token consumption is like charging for electricity by the kilowatt-hour when what customers actually want is light, heat, or [compute](https://wire.fourthweb.ai/tag/ai-infrastructure/). It's an input metric masquerading as an output metric. If AI is really doing white-collar work at PhD level, the value is in the work product, not the number of tokens it took to generate it.

**Key tensions:**

- Per-query energy cost: down 100,000x
- Total infrastructure demand: up and accelerating
- Value metric: still stuck on consumption, not outcomes

### The Implication

If you're building on top of OpenAI or [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) or any other foundation model, your cost basis just dropped through the floor. But your competitors' did too. The efficiency gains don't give you an edge — they just reset the baseline. The real question is what you build now that intelligence is this cheap.

For infrastructure investors and energy companies, Altman's comments are a roadmap. The AI labs are going to keep buying power, even as models get more efficient, because query volume will outpace efficiency gains for years. Nuclear, geothermal, and grid-scale storage are the new picks and shovels. And if AI really is smarter than top scientists at knowledge work, the next wave of infrastructure won't be designed by humans alone. It'll be designed by agents with PhDs we didn't have to pay for.

### Sources

[Crypto Briefing](https://cryptobriefing.com/sam-altman-ai-energy-efficiency-adipec/?ref=wire.fourthweb.ai) | [Crypto Briefing](https://cryptobriefing.com/sam-altman-ai-smarter-than-top-scientists/?ref=wire.fourthweb.ai) | [Crypto Briefing](https://cryptobriefing.com/sam-altman-ai-tokens-fake-unit/?ref=wire.fourthweb.ai)