Hook
Sam Altman recently declared that intelligence will become a utility, priced in tokens, and that token consumption will grow exponentially. The statement is seductive—a clean narrative for a world where AI is as ubiquitous as electricity. But as someone who has spent years auditing smart contracts and dissecting protocol assumptions, I’ve learned that the most elegant narratives often hide the most brittle code. The problem is not the vision; it’s the unstated theorem: exponential token consumption requires exponential improvements in inference efficiency, cost reduction, and value creation. None of these are given.
Context
Altman’s framing is not new. OpenAI’s API has always charged by the token, a unit derived from the Transformer architecture’s autoregressive generation. Each token corresponds to a fixed amount of compute—roughly 1–2 petaFLOPs for a forward pass of GPT-4, depending on context. The “utility” label is a natural extension: token as the kilowatt-hour of intelligence. But the analogy breaks under scrutiny. Electricity’s unit cost fell by orders of magnitude over a century, driven by physics and engineering. AI’s token cost is still dominated by hardware, energy, and model architecture, and the rate of improvement is already showing diminishing returns. The industry’s current focus on inference optimization—quantization, speculative decoding, pruning—is real, but it’s incremental, not transformative. I’ve seen this pattern before in blockchain scaling: the belief that L2 throughput can grow exponentially while maintaining security and decentralization. The code always reveals the trade-offs.
Core: The Exponential Assumption’s Hidden Constraints
Let’s trace the logic. Token consumption grows exponentially because: (1) more users, (2) more tasks per user, (3) more complex tasks consuming more tokens per task. This is the classic “Jevons paradox” applied to AI—cheaper token costs lead to more usage, not less. But the cost side must cooperate. If the marginal cost per token does not fall at least as fast as usage grows, total spending on AI becomes a drag, not a utility. In my 2022 audit of a modular blockchain’s data availability sampling, I found a similar tension: the protocol assumed exponential growth in data throughput without accounting for the escalating cost of validator bandwidth. The result was a theoretical model that collapsed under realistic load. Altman’s claim is the same—a hypothesis waiting to break.
Tracing the gas leak in the untested edge case: The unstated variable is the price elasticity of token demand. If token price falls 10% and usage grows 20%, revenue grows 8%. But if price falls 50% and usage grows 100%, revenue grows 0%. OpenAI’s revenue growth depends on the balance. Altman’s “exponential” narrative is carefully vague: it doesn’t specify the base, the time horizon, or the price trajectory. This is not a technical forecast; it’s a capital market pitch. The underlying assumption is that unit cost will drop fast enough to keep revenue growing faster than costs. But the engineering reality is that inference cost improvements are hitting a wall. The next generation of hardware (custom ASICs, optical interconnects) may offer a 2–3x improvement, not a 10x. And the energy cost of inference at scale is already a constraint. In 2024, I optimized a ZK-rollup prover for a Layer2 project, and I saw firsthand how a 15% reduction in proof generation time required weeks of circuit refactoring. The gains are real, but they are hard-won and diminishing.
Modularity isn’t an entropy constraint—it’s an architectural choice. In blockchain, modularity allows L1 and L2 to specialize, but it introduces latency and complexity. In AI, the “utility” model assumes that the token is a fungible unit of intelligence, but intelligence is not fungible. A token used for a simple classification is not the same as a token used for a complex reasoning chain. The market will price them differently, breaking the “one token, one intelligence” analogy. The real economic structure is more like a multi-tiered market: high-value tokens for complex queries, low-value tokens for simple ones. This is not a utility; it’s a differentiated commodity.
Optimizing the prover until the math screams is what I did in 2024. The same obsessive focus is needed on token cost. But the industry is still in the “gas fee” mentality: treat token cost as a necessary friction, not a business model. Altman’s narrative is an attempt to naturalize that friction, to make us believe that paying for every token is as natural as paying for electricity. But electricity’s cost is transparent, regulated, and predictable. Token costs are opaque, volatile, and subject to API pricing changes. The comparison is premature.
Contrarian: The Blind Spots in the Utility Narrative
First, the utility framing ignores the regulatory dimension. If intelligence is a utility, governments will regulate it—price caps, universal access, liability for errors. OpenAI’s current business model (proprietary API, closed weights) is incompatible with public utility status. The moment an AI system causes a large-scale failure (e.g., a financial meltdown triggered by a hallucinated token stream), the regulatory response will be swift. The “exponential growth” narrative will then be a liability, not an asset.
Second, the token-as-utility model creates a cost governance problem for enterprises. In my 2025 cross-chain bridge audit, I saw how hidden reentrancy vulnerabilities could drain liquidity. Similarly, hidden token consumption spikes from AI agents could drain budgets. The need for “AI FinOps” is real, but it’s a symptom of the model’s inefficiency, not a feature. Altman’s statement implicitly acknowledges this—he admits “new consumption and cost management strategies” are needed. That’s not a sign of a mature utility; it’s a sign of a broken pricing model.
Third, the competitive landscape. If token becomes a commodity, the lowest-cost producer wins. Today, OpenAI is not the lowest-cost—Meta’s Llama models, Google’s Gemini, and open-source alternatives are cheaper per token. The gap is closing. Altman’s “utility” narrative is a strategic move to position OpenAI as the standard, but the market is moving toward fragmentation, not monopoly. The code is a hypothesis waiting to break: the assumption that OpenAI will maintain its cost advantage is not supported by current trends.
Takeaway
Altman’s exponential token claim is a powerful narrative, but it’s built on unverified technical assumptions. The real value in the AI stack may not be in the model layer at all, but in the infrastructure that enables token cost reduction—compute, energy, networking. As a Layer2 researcher, I see parallels: the L1 (model layer) is congested, but the value is in the data availability and execution layers below. The winners of the AI utility era will be those who solve the gas leak in the untested edge case—the hidden cost of token consumption. The rest is just marketing.