Why Time Will Become the Core Pricing Factor for Tokens Strategic Research Report • Compute Economics Series

I. Why Time Is Currently Excluded — Because Pricing Power Lies with the “Production Side,” Not the “Exchange” OpenAI charges per million tokens. Anthropic charges per million tokens. DeepSeek, Zhipu, Tongyi Qianwen — all charge per million tokens. Input $X/MTok, Output $Y/MTok. That is it. Time? Non- existent. Why? Not because they are foolish. It is because the current pricing model is a legacy of the SaaS era, not a product of the commodity era. The logic of SaaS is: I am a factory (the model), I provide you with a product (inference), and you pay based on usage. As for whether my factory’s production speed is fast or slow — that is an internal efficiency issue for me, not a matter of choice for you. You can only choose “whether to use my factory,” not “whether you want it fast or slow.” But the logic of the commodity market is completely the opposite. The spot price of West Texas Intermediate (WTI) and the three-month forward contract price are never the same. A barrel of oil delivered to your refinery today and a barrel arriving at the port three months from now — the former is always more expensive. This price difference is called the convenience yield — because you can use it today, you are willing to pay a premium. AI inference has a convenience yield that has been completely overlooked by everyone. And this yield will skyrocket in the Agent era.

II. 0.15ms vs. 15 Seconds — Not a “Difference in Speed,” but a “Difference in Viability” You are right: delivering a barrel of oil in 0.15ms versus 15 days cannot possibly cost the same. But AI inference is even more extreme than oil. Oil has inventory to act as a buffer. If you lack a barrel today, it might still be in the warehouse. AI inference has no inventory. The exact moment an Agent initiates an inference request — if it takes 15 seconds to return, the Agent’s entire decision chain breaks down. This is not a case of “poor user experience.” This makes the product completely unusable. • An autonomous driving Agent needs to detect obstacles within 100ms — 15 seconds? The car has already crashed. • A high-frequency trading Agent needs to make an arbitrage decision within 5ms — 1 second? The spread has already vanished. • A real-time dialogue Agent needs to generate the next sentence within 200ms — 3 seconds? The conversation is already dead. For these Agents, “15-second Tokens” and “0.15ms Tokens” are not the same commodity. Just as “crude oil arriving today” and “crude oil arriving in three months” are entirely distinct products. Yet, the current pricing framework sells both at the exact same price — $0.15 per million tokens. This is not price discovery. This is price distortion.

III. Why This Distortion Has Persisted Until Now — Because Historically, There Has Only Been One Customer: Humans In the GPT-3.5 era, who was OpenAI’s customer? It was humans. A human types in a chat box and waits 3 seconds for a reply — 3 seconds is perfectly fine. Human tolerance for latency is measured in seconds. Therefore, OpenAI did not need to distinguish between “fast inference” and “slow inference” — all customers

had a similar sensitivity to time. The time dimension was collapsed into a single variable: “model quality.” But Agents are not humans. Agents tolerate latency in milliseconds. Furthermore, different Agents have widely varying latency requirements — just as different refineries have distinct schedule requirements for crude oil arrivals. When you swap the customer base from “ten million chatting humans” to “one hundred million autonomously deciding Agents” — time shifts from an “ignorable user experience variable” into a “core pricing factor.”

IV. KAI’s Opportunity — The First Time Someone Can Turn “Inference Speed” into a “Pricing Dimension” Let me push this logic to its natural conclusion. When there are three hundred model vendors on a KAI clearing network — SiliconFlow might achieve a 50ms time-to-first-token (TTFT) latency; DeepSeek’s official API might be 200ms; and a smaller provider’s inference might take 2 seconds. In today’s world, the token prices of these three are likely identical — because no one factors latency into the price. However, KAI’s pricing engine can incorporate latency. Spot Inference: How quickly can you return the first token to me? The lower the latency, the higher the price. The market clears in real-time — all vendors compete on a single order book. Latency-Guaranteed Inference: I require the P99 latency to be under 100ms. Can you guarantee that? Providers who can promise this earn a premium. Those who cannot can take “batch inference” orders. Batch Inference: I do not care about time. You can finish computing whenever you want, but you must give me the absolute lowest price. Three hours? Fine. Running it late at night when idle compute capacity is high? Even better.

These three categories are not “different prices for the same commodity.” They are three distinct commodities. Just as the crude oil market has spot, forwards, and options — these are not different prices for a single barrel of oil, but separate products under different delivery conditions, timelines, and risk profiles. KAI will be the world’s first token exchange to treat and price “inference latency” as a “delivery timeline.”

V. The True Meaning of Token Futures What I previously termed “Token Futures,” many understood simply as “pre-booking future compute capacity.” That is only the shallowest layer. The deeper meaning of Token Futures is: formally incorporating the dimension of “time” — which has been overlooked by the tech industry for two decades — into the pricing function of AI inference.

1. Today’s World: Token Price = f(Model Quality)
2. KAI’s World: Token Price = f(Model Quality, Inference Latency, Supply, Demand, Delivery Time) This is not merely adding another variable. This is a paradigm shift from “one-dimensional pricing” to “multi- dimensional pricing.” And multi-dimensional pricing is the fundamental raison d’être for any exchange. No one goes to a local wet market to trade crude oil futures, because a wet market deals in only one dimension: “How much for this item?” An exchange exists because commodity pricing must manage multiple dimensions — time, quality, location, and delivery conditions. AI inference is currently evolving from a wet market into a sophisticated exchange. KAI is the infrastructure carrying this evolution forward.

VI. Why Now, and Why KAI Because the paradox you raised — “why 0.15ms and 15-second tokens cost the same” — is only logical under two preconditions:

1. All customers share the identical sensitivity to latency (The human chat era — which has ended).
2. There is insufficient supply diversity to create competition along the latency dimension (The oligopoly era — which is ending). Right now, both conditions are collapsing simultaneously. The Agent economy has driven a massive divergence in consumer latency requirements. Concurrently, three hundred model vendors have introduced significant latency variance on the supply side. When differentiated demand converges with differentiated supply, an exchange inevitably emerges. It is not that KAI actively ambitions to be this exchange; the physics and gravity of the market mandate its emergence. Your “strange question” is actually the fundamental blueprint for the KAI pricing engine’s design. A 0.15ms token should inherently command a premium over a 15-second token. How large should that premium be? It is not determined by arbitrary human judgment. It is the exact number that crystallizes from the real-time game played out between all Agents and all model vendors on KAI’s order book. This is the ultimate significance of a clearing network.