There is a moment in every technology cycle when the underlying resource stops being the advantage. It happened with electricity. It happened with cloud compute. And according to Chamath Palihapitiya, it is happening right now with AI inference.

His “barrel of intelligence” analogy — one million useful tokens as the unit of measure — puts the cost drop in stark terms: from roughly $26 a barrel a couple of years ago to under $0.50 today. That is a 98 percent reduction in the cost of generating intelligence at scale.

He is mostly right. But the analogy has limits worth understanding.

Is Inference Actually Becoming a Commodity?

Yes. That part is directionally correct.

Over the last 18 to 24 months, better chips, quantization, distillation, sparse models, caching, and software optimization have all driven token costs down faster than most people expected. Open-weight Chinese models like DeepSeek, Qwen, and GLM accelerated the trend further.

Raw inference is moving toward commodity pricing. That is no longer a prediction. It is already happening.

Are the Dollar Figures Right?

Roughly, yes — with context.

The exact numbers shift constantly and depend on input versus output tokens, cached context, reasoning mode, the model chosen, enterprise discounts, and batch APIs. The figures are not an accounting statement. They are an illustration of the order of magnitude difference that now exists in the market.

The direction is more important than the precision.

Does Cheaper Inference Mean Equal Intelligence?

No. This is the biggest omission in the commodity framing.

A million tokens from one model are not interchangeable with a million tokens from another. Enterprises making purchasing decisions care about reasoning quality, hallucination rate, coding and agent reliability, tool use, multimodal capability, latency, uptime, security, and compliance.

If a $30 model saves a lawyer one hour, it is still dramatically cheaper than paying the lawyer. Premium pricing can absolutely be justified — it just has to be demonstrated, not assumed.

What Does This Mean for the Big Labs?

They can no longer rely on brand name alone.

OpenAI, Anthropic, and others have to prove measurable value through better reasoning, enterprise reliability, agent performance, ecosystem integration, and developer experience. The labs that win long-term will be the ones that build the most trusted platform, not the ones holding the most expensive weights.

The Bigger Idea Chamath May Be Missing

He is treating AI like oil.

I think it will look more like cloud computing.

AWS did not win because compute stayed expensive. Compute got incredibly cheap. AWS won because of infrastructure, management, APIs, ecosystem, reliability, enterprise sales, and security. The commodity layer became the foundation, not the finish line.

Similarly, the long-term value in AI is migrating toward agents, workflows, memory, orchestration, vertical expertise, integrations, trust, and proprietary data — not the base model itself.

What This Means for Reflekta

This trend actually helps builders who are creating on top of AI rather than competing on inference costs.

Reflekta’s value is not “we have the smartest LLM.” It is captured family memories, identity, trust, emotional design, long-term continuity, and proprietary legacy content. If the underlying intelligence becomes 100 times cheaper, gross margins improve while the moat — accumulated life stories and human relationships — stays intact.

The commodity layer is becoming the floor. The products built on top of it are where the durable value lives.

FAQ

What is the “barrel of intelligence” analogy?
Coined by investor Chamath Palihapitiya, it compares one million useful AI tokens to a barrel of oil — a standardized unit for measuring the falling cost of AI inference.

Is AI inference really becoming a commodity?
Yes. Token costs have dropped roughly 98 percent in two years due to better hardware, open-weight models, and software optimization. Raw inference is commoditizing quickly.

Does cheaper AI mean all AI is equal?
No. Price reflects more than compute. Quality, reliability, reasoning depth, security, and ecosystem integrations still vary significantly between providers.

What is the best comparison for how AI markets will evolve?
Cloud computing. AWS did not lose when compute got cheap — it won by building the infrastructure layer on top of cheap compute. AI is likely to follow the same path.

How does AI commoditization affect AI-native companies?
For companies building products on top of AI — rather than selling raw inference — falling token costs improve margins without eroding the real moat: proprietary data, user trust, and unique workflows.

About the Author

Miles Spencer is an entrepreneur, author, and explorer. He is the Co-Founder and CEO of Reflekta.ai, a platform for intergenerational storytelling and legacy preservation. He is also the author of A Line in the Sand and Havana Famiglia. He writes about technology, human stories, and the places the two meet at Miles to Go.

Follow along: milesspencer.com | Reflekta.ai | Book a Zoom