Six months. One billion yuan.

On July 13, Approaching.AI (趋境科技) officially announced its Series A round. Cumulative funding over the past six months exceeds 1 billion yuan, led by Henan Investment Group's Huirong Fund, with all existing shareholders over-subscribing.

A 1 billion yuan raise isn't the biggest in the current AI infrastructure race — CoreWeave went public at a market cap of tens of billions; Nebius has raised billions too. But what makes Approaching.AI's story interesting is that it's not just another cloud provider "selling GPU hours." It's proposing a new concept: Token as a Service (TaaS).

Put simply: before, you bought "GPU hours." Now, you buy "high-quality tokens." It sounds like just rebranding, but the underlying business logic is entirely different.

From MaaS to TaaS: Selling Results, Not Compute

For the past two years, the dominant AI infrastructure model has been MaaS (Model as a Service) — I set up the platform, integrate a bunch of models, and you pay by token volume. OpenAI, Anthropic, every cloud provider — all follow this model.

But MaaS has a problem: what customers actually need isn't "how many tokens they've called" — it's "how many business outcomes they've achieved." You spend a million on API calls, but the generated content quality is all over the place, errors are frequent, speed fluctuates — was that million well spent? Nobody really knows.

Approaching.AI's TaaS model flips this: I don't sell you compute, I don't sell you models — I sell you "high-quality AI tokens." What does "high-quality" mean? Under high concurrency, it still maintains: low first-token latency, stable high output speed, consistent output quality, reliable structured output and function calling, and controllable unit cost.

None of these metrics are particularly hard on their own. But achieving all of them simultaneously under real production loads? The difficulty rises exponentially. Different combinations of capabilities can produce efficiency gaps of several times — even tens of times. Approaching.AI's core competitive edge is making all these metrics work together in real production environments, and doing it at scale.

The numbers speak for themselves: since the 2026 Spring Festival, Approaching.AI has increased per-GPU token production efficiency by more than 3x, and total high-quality token capacity has grown over 30x. For one leading trillion-parameter model, daily high-quality output has steadily broken through the trillion level. Most importantly: some mature businesses have already crossed the profitability line.

"Crossed the profitability line" — those five words carry more weight in the AI infrastructure industry than almost anything else. There are tons of companies doing inference services today, but very few are actually profitable — everyone's burning cash to grab market share, racing for scale, undercutting each other on price. When Approaching.AI says parts of its business are already profitable — that says a lot.

"Few Models, Deep Optimization": A Counter-Consensus Technical Route

Approaching.AI's technical approach is also interesting. They call it "few models, deep optimization" — exactly the opposite of the industry mainstream "many models, one-stop-shop" approach.

Today's MaaS platforms compete on how many models they've integrated — OpenAI, Anthropic, Google, open-source ones, the more the better, ideally with one-click switching. But Approaching.AI doesn't play that game. They focus on just a handful of models with real production demand, and optimize them to death: model sharding, VRAM management, heterogeneous coordination, cache reuse, failover, elastic scaling — everything that can be optimized, they optimize.

Why? Because enterprise customers ultimately pay for business outcomes, not for the number of compatible models. You can integrate 100 models, but if each one is slow, expensive, and unreliable — what's the point? Better to optimize two or three models people actually use to the absolute maximum.

This logic sounds straightforward, but it's actually a minority position in the industry. Why? Because "many models" makes for good slides, good stories, and good fundraising. "Deep optimization" is hard, tedious, and doesn't show quick results — short-term payoff is invisible. But when you get to production environments and customers vote with their feet, the value of "deep optimization" becomes apparent.

Approaching.AI's funding story is essentially about one thing: AI infrastructure competition has moved from "who has the biggest plate" to "who has the highest efficiency" phase. The past two years everyone was grabbing territory, deploying compute, piling up scale. Now the tide is going out, and we'll see who's been swimming naked. The companies that maximize token output per unit of compute, that break through the cost line, that customers actually use — those are the ones that survive.

The era of the Token Factory is just beginning.

See you tomorrow.