The Era of the “Token Factory”: Has AI Computing Power Reached a Strategic Singularity?

AI computing power is advancing at warp speed, and China’s building out a national computing network like there’s no tomorrow.

The National Development and Reform Commission (Guojia Fazhan He Gaige Weiyuanhui 国家发展和改革委员会) estimates that total investment in the “Six Major Networks” — which includes water infrastructure, new electric grids, computing power, next-generation communications, urban underground pipes, and logistics networks — is expected to exceed ¥7 trillion RMB ($968.3 billion USD) this year alone.

Here’s what’s wild: China’s three major telecom operators just entered the token business, and listed companies are now building out what’s being called “Token Factories.”

This shift signals something bigger — a fundamental rethinking of how AI computing power gets priced, distributed, and monetized across the entire economy.

—

When Telecom Giants Discovered the AI Gold Mine

For decades, telecommunications operators have been trapped in a brutal squeeze.

They own massive infrastructure but operate as what insiders call “dumb pipes” — essentially commoditized distribution channels with razor-thin margins.

Token-based AI services represent their escape route.

On May 16, China Telecom (Zhongguo Dianxin 中国电信) Shanghai Branch officially launched its token computing service.

Here’s the pricing: ¥1 RMB ($0.14 USD) gets you 250,000 credit points.

Using the Kimi-K2.5 model as a benchmark, this translates to roughly 250,000 input tokens — enough to power meaningful AI inference workloads for businesses and developers.

The same day, China Unicom (Zhongguo Liantong 中国联通) Shanghai announced token services specifically targeting its One Person Company (OPC) clients.

The hook?

A free initial 30 million tokens for testing through the end of June.

After that trial period ends, Shanghai OPC customers can renew at rates as low as ¥1 RMB ($0.14 USD) per million tokens.

Twenty-four hours later, on May 17, China Mobile (Zhongguo Yidong 中国移动) Shanghai branch joined the party with a general token service aimed at small businesses and individual users.

Their pitch: ¥1 RMB ($0.14 USD) for 400,000 tokens.

Notice the pattern?

All three operators launched within 48 hours of each other with aggressive pricing that signals intense competition and a race to capture market share.

—

The Infrastructure Play: “Token Factories” Are Rising

• Super-computing Cluster: Huawei Ascend 384-node
• Initial Deployment: 4 Super-node Servers
• Total Capacity: 1,536 compute cards across 4 systems
• Location: Wuxi High-tech Zone, Jiangsu Province

It’s not just the telcos moving fast.

Local governments and publicly traded companies are betting big on specialized computing infrastructure.

On May 15, computing power leasing provider Kinwong Electronic (Hongxin Dianzi 弘信电子) partnered with the Wuxi High-tech Zone to establish what they’re calling Jiangsu Province’s first Huawei Ascend (Huawei Shengten 华为昇腾) 384-node super-computing cluster.

Let’s break down what that means:

• The “Token Factory” will initially deploy four Huawei Ascend 384 super-node servers
• Each system houses 384 cards (meaning GPU/compute units)
• Together they form a massive centralized computing cluster designed to process AI workloads at scale
• This infrastructure can serve thousands of downstream users and businesses

The economics here are crystal clear: build once, monetize many times over through token sales.

—

The Token Economics Model: Reinventing the Internet Playbook?

Here’s the question everyone’s asking: will the token economy just replicate the business models from the traditional internet era?

An industry source close to the operators offered some candid insights:

“Operators see an opportunity to stop being just ‘dumb pipes.’ Given the pressure on their profitability, they are actively testing these waters.”

The key difference from traditional internet data packages?

Tokens directly monetize computing assets.

Unlike paying for bandwidth (measured in gigabytes), token pricing reflects the actual computational work being done — the inference, the model inference, the GPU cycles consumed.

Think of it like this:

• Old model: You bought a data package (10GB for ¥50 RMB). Whether you used it or not, same price.
• Token model: You buy tokens directly. Each AI inference call consumes X tokens. You only pay for what you actually use.

According to Soochow Securities (Dongwu Zhengquan 东吴证券), the business model for computing power leasers is quietly undergoing a massive shift.

It’s moving from renting “bare metal” hardware (just selling you server access) to providing Model-as-a-Service (MaaS) or token-sharing models.

The strategic implication?

Shifting from “selling computing power” to “selling tokens” is expected to significantly improve profitability and optimize valuation systems for these companies.

From a business perspective, this makes total sense:

• Bare metal hosting = low margin commodity business
• Token services with built-in models = high-margin SaaS business
• Better margins = better valuations from investors

—

The Big Question: Is AI Computing Hitting a “Universal Access” Singularity?

The real question beneath all of this: does the entry of major operators mean AI computing demand has reached a singularity where costs are becoming truly inclusive for the masses?

Not necessarily.

An industry expert offered a cautionary tale through the story of “OpenClaw” (nicknamed ‘Lobster’), which exploded in popularity earlier this year.

Here’s what happened:

• OpenClaw let users operate their computers via AI instead of just chatting
• Users found a loophole: they could use it to access Anthropic’s Claude subscription for free
• This caused massive losses for Anthropic
• Once Anthropic cut off third-party access and forced users into pay-as-you-go APIs, costs skyrocketed 10 to 50 times higher
• The user base vanished instantly

The lesson?

Cheap access doesn’t mean sustainable access.

According to the same industry source: “The current market reality is that token producers, including operators, are targeting the rigid demand of professional users to run high-margin businesses.”

Translation: we’re still in the trial phase.

Right now, the focus is on encouraging adoption among power users and professionals who’ll pay for reliability and consistency.

Mass market pricing comes later — if it comes at all.

—

Efficiency, Stability, and the Path to True Cost Reduction

For AI to achieve large-scale, truly inclusive application, efficiency and stability are paramount, according to Li Aijun (Li Aijun 李爱军), CTO of Intellifusion (Yuntian Lifei 云天励飞).

Here’s why that matters:

Every single interaction in the inference stage consumes computing power and electricity.

As applications become more frequent across millions or billions of users, lowering costs becomes not just nice-to-have — it becomes mission-critical.

To actively drive down costs, Intellifusion recently joined several partners to launch the “1001 Plan” initiative.

The goal?

Ambitious and moonshot-level: 10 billion tokens for just ¥0.01 RMB ($0.0014 USD).

That’s roughly 1 million tokens per ¥0.000001 RMB — orders of magnitude cheaper than current pricing.

If they pull it off, it would represent a fundamental shift in AI economics.

—

What This Means for Investors, Builders, and the Market

The token factory era signals several things happening simultaneously:

• Margin compression is coming. As more players enter the market, pricing pressure will intensify.
• Infrastructure becomes the new battleground. Companies with assets (like telcos and data center operators) have an advantage over software-only plays.
• The SaaS model is eating computing power. Direct hardware rental margins are too thin. Operators are racing upstack to capture higher-margin AI services revenue.
• Profitability is still driven by efficiency, not volume. Even with cheaper tokens, only companies that can operate at massive scale with razor-thin margins will win long-term.
• User adoption is the real variable. Cheap tokens don’t matter if developers and businesses don’t build on them. The winner will be whoever captures developer mindshare first.

The “token factory” phenomenon isn’t just a pricing gimmick or temporary promotional war.

It’s a structural reorganization of China’s computing infrastructure stack — one where tokens become the universal unit of currency for AI services, much like how bandwidth became the unit of currency for the internet era.

Whether AI computing power truly reaches universal access depends on three things: sustained cost reduction, reliable infrastructure, and killer applications that drive adoption.

We’re watching the token factory era unfold in real time.

—

References

• “token工厂”来了 AI算力需求迎奇点？ – Shanghai Securities News
• Official Website – China Unicom
• Official Website – China Telecom
• Corporate Profile – Kinwong Electronic (Hongxin Electronics)