Lawyer Lin Shang-lun's article: When you ask AI what to eat for lunch today, the world is reconfiguring the energy landscape in response to this question

Lawyer Lin Shanglun dissects the economic essence of AI Tokens: Tokens are not investable digital assets, but rather units of usage measurement similar to "degrees." Every question asked behind the scenes involves a complete energy chain from power plants, grids, data centers to graphics cards, with data centers converting electricity into billable Token services. Huang Renxun repeatedly warns that energy will be insufficient, reflecting the real gap seen at the supply chain frontlines.
(Background recap: Serenity, the white-haired stock god, calls out SIVE again! Secures an $8.2 million military order, Ka-band chips enter the U.S. military satellite supply chain)
(Additional background: Amazon opens up AI-related merchandise: giving Alexa a phrase to print custom T-shirts + Prime delivery to your door)

Recently, one of the most frequently appearing yet easily misunderstood terms in the global tech scene is AI Token. Every time you open a chat box and input a question, behind the scenes is a whole energy supply chain—from power plants, grids, data centers to graphics cards—operating for you, with Tokens serving as the unit to measure this entire process.

Breaking it down, the most discussed and often misunderstood issue is whether Tokens are a form of money, whether the $17 ChatGPT subscription fee truly covers all costs, why some data centers are located in Silicon Valley and science parks while others are in deserts or fjords, and why Huang Renxun repeatedly emphasizes that energy will inevitably be insufficient. These four questions are key to understanding the AI economy.

Is AI Token really a form of money?

Market discussions are increasingly describing Tokens as a new type of asset, even circulating phrases like "AI Token investment." Many first hear the word Token and naturally associate it with cryptocurrencies, thinking it might be a digital asset that can be accumulated, appreciated, or exchanged for something. But fundamentally, AI Tokens are more like a usage metric than a store of value or tradable digital asset.

Starting from the most basic unit concept: the conversion of Tokens at the output (generation) end varies significantly with language. In English, one Token is roughly equivalent to 0.75 words or about four letters. Common short words like "apple" are usually one Token; longer words like "hamburger" are broken down into "ham," "bur," "ger," costing 2-3 Tokens. In Chinese, it consumes more—since Traditional Chinese characters occupy more space in encoding, one Token roughly equals 0.5 to 1 Chinese character. Common characters like "我" or "的" typically take one Token; complex or rare characters may be split into 2-3 Tokens.

On the input (reading) side, one Token is so small that it barely constitutes a complete sentence. To give a concrete sense of "reading files," industry standards often use 1,000 Tokens as a benchmark: reading an English document of 1,000 Tokens is about 750 words, roughly 1.5 pages of A4 Word document; reading a Chinese document of 1,000 Tokens is about 500-800 characters, roughly half a page of A4, like a short article or a typical news report. Uploading a ten-plus-page contract to AI consumes tens of thousands of Tokens just to read, not including the Tokens used for generating responses.

More importantly, Tokens are divided into Input Tokens and Output Tokens, each with precise and separate pricing. Input Tokens measure the amount of data (files, videos, audio) the AI reads to understand; Output Tokens measure the amount of text, images, or code the AI generates. These are not abstract concepts but concrete billing units used by major providers (OpenAI, Google, Anthropic, etc.), priced per million Tokens. Currently, the market rate is about a few dollars per 1 million Tokens.

From this perspective, Tokens are more like "degrees": a unit of usage measurement for AI services. You wouldn’t say you used 30 degrees of electricity and thus own an "asset" of 30 degrees; nor would you think of your electricity consumption as a commodity or currency. The logic of AI Tokens closely resembles this.

Of course, in the future, Tokens could evolve into financial products like futures, prepaid quotas, or quota trading—similar paths taken by oil, electricity, and carbon markets. But fundamentally, understanding Tokens as "a unit to measure AI usage" is closer to their true role than treating them as an independent store of value. For ordinary users, this conceptual gap directly influences how you view the next question: AI service subscriptions.

Does the $17 monthly subscription cover all costs?

This is probably the most straightforward question for ordinary users. The subscription prices for AI services, from ChatGPT Plus at $20 to enterprise plans, seem affordable. Many naturally assume that the actual cost of AI should be roughly equal to this subscription fee. But a closer look at industry financial reports shows it’s not that simple.

Most leading AI companies are still operating at a loss; OpenAI, Google, Anthropic’s AI divisions are heavily investing, with a significant portion of their operating capital coming from funding rather than service profits. This means current subscription fees do not fully reflect the true costs of the service. When heavy users ask AI to help with image editing, long conversations, or generate large amounts of content, their actual Token consumption may already exceed what the monthly fee covers, with the difference mainly borne by the companies and investors. This is a typical "user education" pricing strategy—to cultivate usage habits and expand the user base. Whether prices will be adjusted in the coming years remains an ongoing observation.

This pricing model also creates an interesting phenomenon: the usage barrier is almost entirely eliminated. In the past, using oil as an industrial energy required owning vehicles, machinery, or factories—an economic filter. But in the AI era, the entry barrier is nearly nonexistent. The same amount of Token consumption can produce a due diligence report and M&A contract, or a medical research summary, or just chat about lunch. The computational power and electricity consumed are comparable, but the social value created varies greatly.

Why do data centers choose such different locations?

This is another common confusion. Some data centers are located in Silicon Valley or science parks, appearing as cutting-edge tech facilities. Yet, news reports also show data centers built in deserts in Dubai, fjords in Norway, rural Ireland, or even next to hydroelectric plants. This raises the question: are data centers high-tech products?

A reasonable understanding is that both views are correct. Data centers are indeed highly specialized infrastructure involving cooling engineering, power management, networking, cybersecurity, and advanced chip integration—building and operating them is not trivial. Many Taiwanese companies in this field have strong engineering capabilities, which is why some choose locations near tech hubs for talent, clients, and supply chain support.

However, from a global perspective, another critical factor influences data center placement: "it heavily depends on stable and relatively cheap electricity." When computing scale reaches a certain level, electricity costs directly impact operational expenses. That’s why Dubai, Abu Dhabi, leveraging desert solar energy; Norway, utilizing hydropower and low temperatures; Ireland, with favorable energy policies—become hotspots for data centers. In other words, the key advantage is that data centers are energy-dependent nodes: "where there is stable power, there is potential for new data centers."

From this angle, data centers are a crucial node in the AI era: they convert electricity—an energy that’s hard to store and transport across borders—into Tokens, a billable, tradable service. All user questions, corporate files, and research data are processed here, with fees based on Token consumption. This naturally leads to the fourth and most critical question.

Why does Huang Renxun keep saying energy will be insufficient?

In recent years, NVIDIA founder Huang Renxun has repeatedly emphasized in speeches and interviews: energy will inevitably be insufficient, and the bottleneck for AI development will ultimately be electricity. Many hear this and think it’s just marketing hype or GPU manufacturer rhetoric. But when connecting the previous three questions, it becomes clear that his statement is backed by solid industry logic.

From the Token perspective, each question involves a computation, which consumes electricity. From a subscription model view, current subsidized pricing encourages unlimited use, leading to exponential growth in overall consumption. From the data center perspective, regions with energy resources are rapidly expanding computing capacity; each new data center significantly impacts local power demand. When these three trends combine, the energy pressure escalates rapidly.

That’s why Huang Renxun emphasizes energy issues more than GPUs themselves—because no matter how powerful GPUs are, without enough electricity, they can’t run. NVIDIA knows well the scale of their clients’ compute orders and the corresponding electricity gap. His statement that "energy will be insufficient" reflects the real situation seen at the supply chain frontlines, not just a slogan.

This also indicates that the energy industry will be a key long-term focus in the next era. Regions with conditions—deserts, abundant hydro, suitable coastlines for wind farms, countries with nuclear development potential—are actively expanding power generation. China, with its massive green energy system, holds a clear energy advantage in this race; Taiwan’s policies on nuclear and green energy will also be reevaluated in the context of AI’s growth. It’s foreseeable that the overall electricity consumption of global AI infrastructure will surpass many existing industrial sectors in the coming years. Currently, the global AI user base is still in early growth, and demand curves are expected to continue upward.

How should users and investors respond?

In summary, two long-term focus areas emerge. First, the overall supply chain of AI infrastructure—including power generation, grid upgrades, data center construction, cooling, advanced packaging, and core compute supply—has relatively clear order visibility. As AI demand continues to grow, energy and infrastructure gaps will be structural issues, not short-term fluctuations.

Second, the efficiency of Token usage itself will become an important topic. As Token pricing gradually becomes transparent, users will start to recognize cost differences across tasks, and companies and individuals will develop more rational usage habits.

For the AI industry, three basic principles are key:

1. Understand that Token is fundamentally a billing unit, which helps you judge various financial narratives and extensions.
2. Recognize the role of data centers—they are highly specialized infrastructure but also location-dependent energy consumers.
3. Pay attention to long-term energy trends, as they will become the most critical bottleneck and opportunity for AI—this is the core message repeatedly emphasized by AI leaders.

The rapid development of AI over the past year has outpaced many’s understanding of its underlying cost structure. This doesn’t mean AI is a bubble or that the industry will collapse; rather, as a new industry matures, pricing, resource allocation, and infrastructure will gradually find a new balance. The demand curve for AI is expected to keep rising, but energy and cost issues will become increasingly important for users, investors, and policymakers to consider.

This tiny unit, Token, links back to power plants, grids, data centers, chips, models, applications, and extends into every user’s daily choices. Understanding its essence helps us better see where opportunities and costs lie in the AI era.