Hook
Sam Altman just declared that intelligence will become a utility, with token consumption growing exponentially. The statement landed on Crypto Briefing, a venue that converts AI jargon into crypto-native speculation. No data. No time horizon. No cost curve. Just a promise that the units of intelligence—tokens—will multiply like electrons in a grid. As someone who spent three weeks reconciling FTX's ledger to find a $1.8 billion hole, I know a narrative without a balance sheet when I see one. Let's run the numbers that weren't provided.
Context
Altman's quote is not new. OpenAI has been charging per token since its API launch in 2020. The utility framing is a marketing upgrade: it repositions OpenAI from a software vendor to an infrastructure provider. The comparison to electricity or water is seductive—both are essential, metered, and subject to natural monopolies. But tokens are not electrons. An electron's cost per unit has historically declined with scale (Swanson's law). Tokens? Their cost is tied to inference compute, which is currently dropping but not at a rate that guarantees exponential usage without exponential spending. The article's author tacked on a single line: this growth will require new consumption and cost management strategies. That's a platitude that hides the real tension: if usage grows exponentially and unit cost drops linearly, total spending still grows. Who pays?
Core
Let's isolate the variables. Altman's claim rests on three assumptions: (1) token generation cost will continue to drop, (2) model capability will create new use cases that absorb more tokens, and (3) the price elasticity of demand is high enough that cheaper tokens unlock orders of magnitude more consumption. None of these are proven. From my audit experience, I've seen how human-in-the-loop security protocols catch what automated scanners miss. The same principle applies here: a narrative that skips the proof-of-concept is a variable I refuse to define.
First, the cost curve. OpenAI's token prices have fallen roughly 10-20x since GPT-3 era, but that's a single data point. The marginal cost of inference is dominated by hardware and energy. Nvidia's H100s cost $30,000 each, and a single inference query can use hundreds of GPU-seconds. If token consumption grows 100x, the compute required grows 100x unless efficiency improves by 100x. There's no evidence that inference efficiency is improving at that rate. The industry is still in the early phase of moving from training to inference optimization. Expecting a 10x improvement in cost per token over the next three years is optimistic; expecting a 100x is fantasy.
Second, the demand side. Token consumption is not a pure function of price. It's constrained by the economic value of the output. A company using AI to generate SEO spam will consume tokens regardless of value, but that's a bubble, not a utility. The real utility demand comes from replacing human labor with automated reasoning. Each such task has a budget. If a call center agent costs $20/hour, and an AI agent costs $10/hour in tokens, the firm switches. But if the AI agent consumes $19/hour in tokens, the switch doesn't happen. The break-even point is narrow. Altman's exponential growth assumes that the value created per token far exceeds its cost, but that's a hypothesis, not a fact. In my 2024 audit of AI-generated contract exploits, I found that the cost of generating malicious code was trivial, but the value of the exploit was massive. That asymmetry is a feature, not a bug—it means token consumption can explode without proportional value creation.
Third, the infrastructure reality. Token usage at exponential scale means exponential energy demand. The IEA estimates that data centers could consume 1,000 TWh by 2026, up from 460 TWh in 2022. That's a doubling in four years. If AI token consumption grows at 50% CAGR, we'd need to double energy every 1.7 years. That's not sustainable without a breakthrough in energy generation or chip efficiency. Volatility is just liquidity leaving the room; exponential token demand is just energy conversion leaving the grid.
Contrarian
Now, what do the bulls get right? The utility framing itself is structurally sound. Electric utilities did experience exponential growth for decades, and they became the backbone of the modern economy. The same could happen for AI. The bulls are correct that the market for AI services is nowhere near saturation. Companies like Microsoft, Google, and Meta are already embedding AI into every product. The token-as-a-unit-of-intelligence metaphor is useful for pricing and interoperability. If OpenAI can lock in the standard, it becomes the meter reader for the entire economy. That's a powerful position.
But the bulls ignore the regulatory and competitive risks. If intelligence is a utility, governments will regulate it. Altman's own Worldcoin project has already faced privacy crackdowns. The assumption that OpenAI can charge what it wants while maintaining a natural monopoly is naive. Trust is a variable I refuse to define, but regulation is a constant I refuse to ignore. Moreover, open-source models are eating the cost curve. Llama 3 can run on a single consumer GPU for free. The unit cost of inference for open-source models is effectively zero for the user (ignoring hardware). If open-source models become good enough, the utility becomes a commodity, and the margin disappears. Altman's exponential token growth thesis depends on OpenAI being the only provider, which is a bet against the open-source ecosystem.
Takeaway
Altman's statement is a fundraising pitch dressed as a futurology lecture. It contains no testable hypothesis, no data, and no time frame. The crypto audience that lapped it up from Crypto Briefing should remember that tokens in AI are not tokens in blockchain—they are units of cost, not units of value. The real question is not whether token consumption will grow, but whether it will grow faster than the cost of producing them. If the answer is no, the utility narrative collapses into a cost spiral. If the answer is yes, the winners won't be the model providers—they'll be the energy companies and chip manufacturers who own the real infrastructure. Code doesn't lie. People do. And the code of Altman's prediction is missing a line: the cost function.
