Two sentences, one of them seemingly satirical, are enough to tear the veil off one of the most contentious dynamics in enterprise AI. On Reddit, a user quotes an unnamed CEO stating that “token efficiency needs to drop to as much as 20% over the next 12 months, and 90% by the following year.” He then adds: “just write \no_think before you ‘summarize this email’ prompts.”

Behind the sarcasm lies more than a joke about prompting. The remark attributed to the CEO, regardless of the poster's intent, captures a profound disconnect between efficiency as an engineering virtue and efficiency as an economic variable for the parties that sell those tokens. Taken at face value, the statement describes a future where models become vastly more verbose – or far less capable of distilling a task into a few tokens – forcing every interaction to consume ten times today's resources.

For cloud-hosted workloads, an explosion in tokens per query translates into a linear increase in operational cost, but also into multiplied revenue for the platform provider. The incentive is aligned: the more tokens flow, the higher the margin for whoever rents GPUs by the minute. Self-hosting flips this equation entirely. In a corporate server room or on an edge node, TCO is anchored to hardware: VRAM, effective throughput, electricity. Every extra token erodes remaining capacity, saturates inference queues, and shortens the useful life of the upfront investment without generating any upside.

Organizations that defend data sovereignty – banks, public administrations, defense, regulated sectors – have built their business cases on the assumption that an on-premise LLM can serve a given request volume with a known, predictable GPU footprint. If token efficiency collapses, that math falls apart. A cluster sized for 500 queries per minute would suddenly manage 50, forcing the purchase of additional accelerators in a market already strained by bottlenecks and high prices. This is not a software problem; it is a physical constraint tied to memory bandwidth and compute capability of the cards.

There is also a less visible technical implication. Token efficiency is not just a matter of prompt engineering: it is baked into model architecture, quantization, and the training process itself. Deliberately pushing toward inefficiency would mean favoring more “wasteful” architectures – likely larger, less optimized models – that fit poorly with inference on consumer hardware or enterprise servers with strict VRAM limits. The result would be a further widening of the gap between the trajectory of foundation models, ever more resource-hungry, and the realistic possibilities of local deployment.

In this light, the CEO's provocation – even if delivered with a smile – has the merit of making explicit a conflict of interest that usually stays beneath the surface. On one side stands the rhetoric of AI democratization; on the other, a structural push to make models less token-efficient, because the token is the smallest unit on which cloud pricing plans are built. On-premise, by definition, plays no part in this volume game: it pays for electricity, not per token. And it needs exactly the opposite of what that controversial sentence calls for.