Huang Renxiong and Jack Ma Are Both Competing for It: See Clearly the Business Model Behind Tokens

Last night, NVIDIA CEO Jensen Huang delivered the keynote speech at GTC 2026. Just the day before, Alibaba established the Token Hub Business Group, with CEO Wu Yongming personally leading the team.

Two giants, on the same day, focusing on the same thing: Token. How exactly does Token make money? Today, we’ll clarify that.

What is a Token?

Huang Huang used an analogy: Tokens are the oil of the digital world.

Ask ChatGPT a question, generate a video, or make a robot move—all of these consume Tokens. It’s the basic unit for AI to understand and generate information, and also the measurement unit for all future intelligent services.

How is Token charged? How much do companies need to pay?

Huang Huang publicly revealed the tiered pricing system for Tokens for the first time. This isn’t just a concept; it’s a real charging standard:

• Free Tier: High throughput, low speed, used to attract users
• Basic Tier: $3 per million Tokens, for ordinary users
• Advanced Tier: $6 per million Tokens, for larger models and faster speeds
• High-Speed Tier: $45 per million Tokens, supporting long context and deep reasoning
• Top Tier: $150 per million Tokens, for ultra-long research tasks and real-time critical path responses

Huang Huang did some quick math: “A researcher using 50 million Tokens daily, at $150 per million, is completely acceptable for a research team.”

What does this mean? Tokens are not a one-time purchase but a continuous consumption. You burn through them today, and you’ll burn through them tomorrow. As long as AI is running, Tokens are being used.

CUDA’s Business Model: Why Huang Huang Dares to Say “Compute Power Gets Cheaper the More You Use It”

So how does NVIDIA make money from Tokens? The core is CUDA. You can think of it as the “operating system” of the AI world. All AI applications run on CUDA, just like all mobile apps run on Android or iOS.

CUDA’s moat is its hundreds of millions of installed bases.

NVIDIA spent 20 years enabling hundreds of millions of GPUs to run CUDA, penetrating every cloud, every computer company, and nearly every industry. This installed base creates a flywheel effect: large installations attract global developers; developers create new algorithms, spawning new markets; new markets demand more computing power, further expanding the installed base.

The result of this flywheel is a continuous decrease in computing costs.

Huang Huang gave an counterintuitive example: GPUs based on the Ampere architecture shipped six years ago now have rising cloud pricing. Why? Because more applications can run on these old chips, and NVIDIA’s software optimizations allow all existing users to get free speed boosts.

What does this mean? The computing power you buy today may not become outdated in a few years—in fact, it could become more valuable. This is almost a nightmarish scenario for hardware industries.

Huang Huang’s words: “Once you install NVIDIA GPUs, you gain long-term performance upgrades. It’s not just initial acceleration, but ongoing cost reduction.”

This is CUDA’s business model: Build an installed base with hardware, continuously create value with software, and lock in users through the ecosystem. Hundreds of millions of GPUs worldwide run CUDA, and every Token consumed must pass through NVIDIA’s platform.

This changes the entire business logic. Previously, companies bought AI cards as a one-time capital expenditure. Once bought, it was done; no further relationship with the vendor. Now, AI cards become production lines for Tokens—companies spend money on “fuel” rather than just equipment. Today’s burn is tomorrow’s need—shifting from a one-time expense to ongoing operational costs.

This is called rent collection. You pay today, and you’ll pay again tomorrow.

Why Did Alibaba Rush In? An Unavoidable Hard Battle

Looking at Alibaba’s actions: on March 16, Alibaba established the Alibaba Token Hub Business Group, with CEO Wu Yongming personally leading. The goal is divided into three parts: Create Tokens, Transport Tokens, and Apply Tokens.

Why is Alibaba so urgent?

The core pain point is: AI capabilities are too scattered. Over the past two years, Alibaba’s AI has been spread across Tongyi Laboratory, Alibaba Cloud, DingTalk, Taotian, and other systems. Different teams work independently, resources are duplicated, and collaboration costs are rising. User awareness is also fuzzy—Tongyi Qianwen, Qianwen App, Quark AI—all use similar technology but produce multiple products, confusing users.

Just 12 days ago, “Qianwen’s soul figure” Lin Junyang suddenly resigned, behind which was a clash between “scientists wanting to publish papers” and “management wanting revenue.” Such conflicts are common in the AI industry.

Wu Yongming’s solution is straightforward: Turn Tokens into the only command baton. Future Alibaba’s evaluation system will no longer debate which technical route is correct but will directly measure Token usage and revenue. Departments that generate Token growth will stay; those that don’t will be marginalized.

Deeper logic: Alibaba believes that future digital work will be supported by hundreds of billions of AI Agents. Wu Yongming stated in an internal letter, “These AI Agents will be driven by Tokens generated by models.” And Agents don’t sleep—they run 24/7, meaning Token demand is continuous and high-frequency.

If we compare Tokens to “electricity” in the AI era, the roles of the two giants are clear: NVIDIA is the “generator”—selling generators and building power plants, locking in customers through the CUDA ecosystem, making all AI applications pass through its platform.

Alibaba is the “power grid + major consumer”—with Tongyi Laboratory responsible for creating Tokens, MaaS platforms for transporting Tokens, and Qianwen, Wukong, etc., for consuming Tokens. From consumer AI assistants to enterprise workflows, Alibaba aims to connect the entire “generation–transmission–consumption” loop.

Who has the bigger advantage? NVIDIA has technical barriers and ecosystem lock-in, with a dominant cost advantage per Token. Alibaba has scenarios and data—e-commerce, cloud, DingTalk, Amap, Ele.me—producing massive data daily (603138), and consuming huge amounts of Tokens. If the loop can be completed, Tokens can circulate and appreciate within Alibaba’s ecosystem.

The giants are already racing in real money.

The Endgame of the Token Business

The essence of the Token business is transforming computing power from “asset” into “service”. Moving from one-time sales to continuous charges. From selling shovels to selling oil.

Whoever masters the lowest Token production cost will hold the pricing power in the AI era.

NVIDIA’s 20-year CUDA ecosystem, hundreds of millions of installed bases, and the scale advantage of software optimization benefit all users with each software update. Alibaba has complete business scenarios and the determination to treat Tokens as strategic resources.

Both giants share the same judgment: Tokens are the hard currency of the AI era. As for when NVIDIA’s stock price will recover? Wait until the market stops viewing it as just a hardware company and starts valuing it as an AI infrastructure platform.