The AI Industry Is No Longer Competing on Intelligence, But on Cost

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

In 2025, Silicon Valley changed the axis by which it measures AI. For the past few years, all discussions revolved around Scaling Law: whether models would continue to get smarter, when AGI would arrive. That year, the mainstream narrative shifted to token consumption: consumption kept growing, up more than 20% month-over-month in July, with a curve that looked exactly like traffic and retention during the peak of mobile internet.

The strange part is that the same year GPT-5 launched to mixed reviews—the model didn’t get noticeably smarter, yet usage kept climbing. Put these two facts together, and only one explanation holds: the industry changed playing fields. People stopped betting on the next intelligence leap and started managing the intelligence they already have.

Changing Axes Isn’t Decline, It’s Standard Industrial Practice

Why did expectations shift? Because demand was unlocked by existing capabilities. B2B wants cost reduction and efficiency gains; C2C wants search replacement and work assistance. These scenarios have long had sufficient intelligence requirements. What’s holding users back was never that models weren’t smart enough, but that products weren’t smooth enough to use. As users switched from “waiting for a smarter next model” to “actually using what’s available,” the measurement naturally shifted from capability to consumption.

GPT-5’s true identity becomes clearer in this context: it didn’t try to prove it was smarter, but split capabilities into separately priced modes like Instant and Thinking, and integrated scattered models, information, and interfaces into a better product. This is textbook industrialization. Historically, electricity adoption didn’t rely on brighter lightbulbs, but on continuously falling electricity prices. When performance converges, cost determines penetration speed. The “intelligence” metric itself has also failed. When models win gold and silver at the International Mathematical Olympiad, humans no longer have an appropriate yardstick to measure who’s smarter. Among remaining competitive dimensions, cheapness is the only one still measurable.

Companies at each position therefore take their places: model companies make each token more valuable, infrastructure companies make tokens faster and cheaper, application companies figure out how to exchange consumed tokens for more data feedback. Agents particularly resemble apps from the mobile internet era: where every product used to need an app, now every scenario needs an agent. And agents’ token utilization efficiency is still very low. Whoever gets efficiency up first eats the biggest chunk in the middle.

Be Careful—Token Consumption May Be a New Vanity Metric

Here I need to throw cold water, and this is the article’s most important independent judgment: consumption is a cost-side number, not a value-side one. When a hundred million tokens burn through, what comes back? Metrics no one questions will sooner or later become vanity metrics.

The mobile internet era went through an identical disease course. First traffic worship, with portals and apps competing on install numbers; later they learned to look at retention and ARPU, only then realizing half the traffic was false prosperity. The equivalent metric for the token era should be “revenue or retention returned per unit consumption”—the recovery rate per token. Seeing consumption curves rise beautifully and judging the industry healthy is like judging mobile internet maturity by download numbers in 2012.

Even more worrying is a term that became popular in Silicon Valley: Vibe Revenue. It means users know AI is useful but can’t articulate what it’s actually useful for, so much product revenue comes from inflated expectations and following trends rather than real usage. It’s a deeper problem than ARR quality—ARR is at least real money, while with Vibe Revenue, users themselves may not even know what they bought.

The detection method is ready-made: don’t look at new acquisitions, look at renewals. Whether the same cohort’s actual task volume has grown over consecutive quarters, what percentage of paid users repurchase—these two curves can’t lie. Products with fast revenue growth but flat usage depth are likely in Vibe Revenue territory.

Company Boundaries Are Disappearing, the Wrapper Layer Gets Squeezed from Both Sides

The axis shift brings a structural consequence: the division of labor between model companies and application companies has blurred. OpenAI recruited a batch of startup founders to build products; Cursor turned around and started training its own model; Manus builds both product and technology while proposing engineering methodologies like Context Engineering. Ambitious companies are all trying to integrate model, product, and infrastructure end-to-end, no longer accepting the old framework of “defining yourself in one sentence.” ByteDance is an app factory, Google is talent reserves—these sentence patterns have failed in the AI era.

For the wrapper layer in between, this isn’t good news. Going up, model companies casually build out scenarios with validated PMF, with each new model release compressing the wrapper’s value; going down, inference optimization is largely open-sourced with questionable margins, leaving independent middle layers unstable. What can hold are the two ends: those with models and those with scenarios and data. Companies stuck in the middle with neither are seeing their survival space structurally narrow.

Primary market capital flows corroborate this. At the time in the US market, the valuation gap between first-tier and second-tier AI companies reached levels unseen in over a decade, with some star companies starting at multi-billion dollar valuations. Heat multiplied tenfold, deal count didn’t double, but invested companies received ten times the money. Capital is concentrating at the top, and there’s still no reason to see this trend reverse.

The Boundaries of This Judgment

I must be clear: the above views took shape in mid-2025, a mid-cycle judgment, not an endgame. Two clearly defined validation points were worth checking against at the time. First, whether Meta’s heavily assembled model team could deliver a significantly stronger model within 6 to 12 months; second, whether stable token-consuming application scenarios could emerge in the market. If both gave negative answers, the judgment that “AI valuations are generally inflated” would start to materialize. The bubble here refers to inflated valuations, not that the technology has no future.

For readers wanting to track this continuously, watch three numbers:

One final sentence to close. The axis shift itself isn’t bad news—when a technology moves from proving itself to deploying itself, metrics naturally should change. But the moment of changing metrics is easiest for muddying waters: in the intelligence era you tell stories about capability, in the industrialization era you tell stories about consumption, and consumption happens to be easiest to inflate with subsidies and expectations. When judging an AI company, far more important than “how many tokens it burned” is what each token returned. The former is an admission ticket; the latter is the answer to whether this business can actually work.