Alibaba Slashes AI Costs While Kling AI Hits $18B

Here is the number that should make every AI developer stop scrolling: 99 percent. That is how much Alibaba claims its new framework can reduce the token consumption of AI agents — which, if it holds up in the real world, is not an incremental improvement. It is a structural rethink of how agents work.

To understand why this matters, you need to know what AI agents actually spend their time doing. Most agent systems work by loading every available tool into the model's context window at the start of each task. Think of it like handing a contractor the entire Home Depot catalog every time they need to fix a leaky faucet. The model has to read and process all of it, even when 95 percent of those tools are irrelevant to the job at hand. Tokens get burned. Costs pile up. Latency creeps in.

Alibaba's framework takes a different approach. Instead of front-loading every tool into the context, it selectively retrieves only the tools relevant to a given task at any specific moment. The agent figures out what it needs, fetches just that, and moves on. It sounds almost obvious in hindsight, which is usually the hallmark of a genuinely clever idea.

This matters well beyond Alibaba's own products. Token costs are one of the biggest practical barriers to deploying AI agents at scale. Enterprises experimenting with multi-step autonomous workflows are often shocked by how quickly the bills accumulate, especially when agents are running dozens or hundreds of tasks simultaneously. A 99 percent reduction in token usage would transform the economics of those deployments almost overnight.

The skeptic's question is whether that 99 percent figure survives contact with messy, real-world use cases. Benchmark numbers from the lab and production performance have a complicated relationship in AI. Alibaba has every incentive to present its framework in the best possible light, and independent verification will matter a lot here.

That said, the underlying logic is sound, and Alibaba is not exactly a newcomer to large-scale AI infrastructure. The company runs some of the most demanding cloud and e-commerce workloads on the planet, so the engineering instinct to aggressively optimize resource consumption is baked into its culture.

For the broader AI industry, this is a reminder that the frontier is not always about building bigger models. Sometimes the most valuable work is making the systems we already have dramatically cheaper and faster to run. Efficiency gains like this one could do as much to accelerate AI adoption as any new model release — maybe more, because they lower the floor for who can afford to build with this technology in the first place.

An $18 billion valuation for an AI video tool that has existed for barely a year. That is the number being floated around Kling AI, the video generation platform spun out of Chinese short-video giant Kuaishou, as it reportedly closes in on a $3 billion funding round with Tencent now in the mix.

To put that in perspective, Kling AI launched in June 2024. It does what a growing crowd of AI video tools do — turn text prompts and static images into video clips — but it has moved fast enough and impressed enough people that Kuaishou is already talking about a standalone IPO as early as the first quarter of 2027. That is an aggressive timeline for a product that is still a toddler by startup standards.

The Tencent angle is worth paying attention to. These are two of China's biggest internet companies, and their relationship has historically been more rivalry than partnership. Tencent's decision to write a check here signals something: it believes AI video generation is a category worth owning a piece of, and it would rather invest in a credible competitor than risk being left out entirely. In Chinese tech, that kind of strategic hedging is common, but it still carries weight.

The valuation trajectory here is also eyebrow-raising in its own right. Kuaishou initially went into investor conversations seeking a $20 billion valuation for Kling AI. The reported $18 billion figure suggests some pushback — investors liked the asset but not quite that price — which is actually a healthy sign. Frothy AI valuations with zero negotiation would be the more alarming outcome.

For context, the global AI video generation space is getting crowded fast. OpenAI's Sora, Runway, and a handful of other well-funded players are all competing for the same creative and enterprise customers. Kling AI's edge has been performance on certain benchmarks and competitive pricing, but sustaining that advantage while simultaneously preparing for a public offering is a significant operational challenge.

The planned restructuring — bringing in external investors through a formal reorganization before the IPO — is a fairly standard playbook for Chinese tech companies looking to establish an independent valuation for a business unit. It creates a cleaner story for public market investors and lets the parent company, Kuaishou, potentially realize value from an asset it built internally.

The bigger question hovering over all of this is whether AI video generation becomes a genuine platform business or remains a feature that gets absorbed into larger products. If it is the former, an $18 billion bet could look conservative in a few years. If it is the latter, this round will look like peak hype in the rearview mirror.