Trong kỷ nguyên AI, cuộc chiến cung-cầu Token là màn diễn biến cuối cùng

Video Title: The Supply and Demand of AI Tokens | Dylan Patel Interview

Video Author: Invest Like The Best

Compiled by: Peggy, BlockBeats

Editor’s Note: As AI model capabilities continue to leap forward, and tools like Claude Code and Cursor are adopted on a large scale by enterprises, the industry conversation is shifting from "how powerful is the model" to "how to integrate the model into production." However, when AI programming, automated analysis, and data modeling become the new consensus, a more fundamental question arises: when execution costs are rapidly compressed, what truly becomes scarce – manpower, capital, or the right to access frontier models and tokens?

Left: Host Patrick O'Shaughnessy, Right: Dylan Patel

This article is compiled from a conversation between Patrick O'Shaughnessy and SemiAnalysis founder Dylan Patel. Dylan has long focused on AI infrastructure, the semiconductor supply chain, and model economics. In this dialogue, starting from the surge in his own company's Claude Code expenses, he discusses how AI is changing enterprise organization, information services, token demand, the computing supply chain, and social sentiment.

The most noteworthy aspect of this conversation isn't that a particular model has set a new benchmark again, but that it provides a way to understand the AI economy – viewing AI as a production system that is reallocating execution capability, organizational efficiency, and industrial profits, rather than just a software tool upgrade.

This conversation can be broadly understood from five perspectives.

First, execution costs are being broken.  In the past, ideas were not scarce; the real difficulty was turning an idea into a product, a system, and a deliverable service. Now, Claude Code allows non-technical people to write code, build applications, and perform data analysis. Tasks that once required a team's long-term maintenance are now accomplished by a few people using models. SemiAnalysis's annualized spending on Claude Code has reached $7 million, exceeding a quarter of its payroll. This indicates that AI is no longer just a productivity tool but is becoming a new form of production capital for enterprises.

Second, the information services industry is the first to be rewritten.  Dylan's business essentially sells analysis, consulting, and datasets – an area most susceptible to commodification by AI. Chip reverse engineering, energy grid modeling, and building macroeconomic indicators once required a team's long-term investment, but can now be prototyped by a few people in weeks. This means the pressure on information service companies isn't "whether AI will replace people," but "who can rebuild their competitor's product faster." Companies that don't adopt AI will be commodified by faster ones, while those that do adopt must constantly raise the bar to avoid being replaced by the next wave of more efficient competitors.

Deeper still, tokens are becoming a new means of production.  In the past, companies bought software subscriptions, and the core question was whether the tool was useful. Now, access to frontier models, rate limits, enterprise contracts, and token budgets directly determine production capacity. A more powerful model doesn't necessarily mean higher cost because smarter tokens might complete higher-value tasks in fewer steps. The real competition is shifting from "who uses AI" to "who can get the most powerful models and deploy the most expensive tokens on the highest-value scenarios."

This demand will continue to propagate through the entire supply chain.  Surging token usage ultimately translates into sustained pressure on GPU, CPU, memory, FPGA, PCB, copper foil, semiconductor equipment, and wafer fab capital expenditure. The "bullwhip effect" mentioned in the article perfectly illustrates this logic: what appears downstream as merely increased model call demand can manifest upstream as massively amplified orders, capacity expansion, and price hikes. Therefore, profit distribution in the AI industry won't stop at model companies and NVIDIA; it will spill over along the semiconductor and data center supply chain.

Finally, social backlash against AI may arrive sooner.  As AI genuinely enters workflows, public concerns about job displacement, energy consumption, data center expansion, and power concentration will rise in tandem. Dylan even predicts large-scale protests against AI within three months. For model companies, continuing to emphasize that "AI will change the world" may not alleviate anxiety but could instead reinforce the public's sense of loss of control. The AI industry's next task is not just to prove technological capability, but how it can create concrete, tangible public value right now.

Today, the core question of AI is shifting from "what can models do" to "who can get the models, how to use them, and who can capture the value they create." In this sense, this article's subject is no longer just Claude Code, Anthropic, or a single AI company, but a structural reorganization centered around productivity, capital expenditure, organizational efficiency, and social acceptance.

The following is the original content (edited for readability):

TL;DR

·The core variable of AI is shifting from "can it be done" to "is it worth doing." As execution costs plummet, what truly becomes scarce are high-value ideas that can be amplified by models.

·Claude Code spending at 25% of payroll is just the beginning; AI is transforming from a software tool into a new form of enterprise production capital.

·Competition for frontier models is no longer just about capability, but about access to tokens. Those who can access the most powerful models earliest and most stably can build new business moats.

·The information services industry will be restructured by AI first, because the production cost of data, analysis, and research is rapidly declining. Slow companies will be commodified by faster ones.

·Token demand will not slow down just because older models become cheaper. Each model improvement unlocks new high-value use cases, pushing users towards more expensive frontier models.

·The biggest change AI brings isn't less work for people, but a few people achieving multiple times the output in the same time. Those who cannot create and capture value from tokens risk being locked into a "permanent underclass."

·The computing shortage is spreading through the entire semiconductor supply chain, from GPUs, CPUs, and memory to PCBs, copper foil, and equipment manufacturers. AI demand has become a price driver for the entire industry chain.

·The economic value of AI is difficult to capture with traditional GDP metrics. The real question isn't just how much money model companies make, but how much "phantom GDP" is created by the decisions, efficiency, and cascading effects generated by tokens.

Interview Transcript:

Claude Code Becomes the New Workforce

Patrick O'Shaughnessy (Host):

You once told me a fascinating story about the massive change in your team's token usage this year. Could you tell it again? What does it tell you about what's happening in the world?

Dylan Patel (Founder, SemiAnalysis):

Last year, we thought we were heavy AI users. Everyone was using ChatGPT, everyone was using Claude. I provided the team with any subscriptions they wanted. Back then, the company's spending was in the tens of thousands of dollars.

But this year, spending started to skyrocket. The real inflection point was around the end of last December with the arrival of Opus. This also includes Doug, our president, Douglas Lawler. He essentially took the lead in pushing non-technical people to use AI for coding. He sort of dragged the whole company into it, little by little. Of course, engineers were already using it, but starting this January, our spending clearly turned a corner and then exploded rapidly.

We later signed an enterprise contract with Anthropic. Last time we spoke, our annualized spending was about $5 million; now it's $7 million.

Patrick O'Shaughnessy:

And that was last week's number.

Dylan Patel:

Right, a large part of it is just usage volume. What's really interesting is that people who have never coded before are now using Claude Code, and some can spend thousands of dollars a day single-handedly. But from a company-wide perspective, our annual spending on Claude Code is now $7 million, while our payroll is around $25 million. That means Claude Code spending already exceeds 25% of our payroll.

If this trend continues, by the end of the year, it could even surpass 100% of payroll. That's a bit scary. Fortunately, I don't have to choose between "people" and "AI" right now because the company is growing fast. It's more like: I don't need to hire people as quickly, but I can spend more on AI, and it works, enabling the company to grow even faster.

But I think other companies will eventually face this question: If one person using Claude Code can do the work of 5, 10, or even 15 people, what happens next? First, maybe you should indeed lay people off; second, the use cases are currently very broad.

For example, we have a reverse engineering lab in Oregon that's been under construction for a year and a half. It's full of high-end equipment like microscopes and scanning electron microscopes. The lab's core purpose is to reverse-engineer chips, extract their architecture, and analyze the materials used in their manufacturing. This is one of the datasets we sell.

But analyzing this data used to be a very slow process. Now, one person on our team spent just a few thousand dollars on Claude tokens and built an application. It uses GPU acceleration and runs on our servers at CoreWeave. We just feed it a chip image, and it automatically labels the position of every material on the image: here's copper, here's tantalum, here's germanium, here's cobalt. Then you can very quickly perform finite element analysis on the entire chip stack structure, all visualized with a full GUI and dashboard.

This person used to work at Intel, and he said that in the past, this would have been a full team's task to build and maintain. Having similar things happening across the entire company is simply incredible.

Another example I find particularly interesting is Malcolm. He used to be an economist at a large bank. That bank's economics department probably had 100 to 200 people. What he's built now is quite astonishing.

He integrated various data sources, including FRED data, employment reports, and other datasets from different APIs. We also signed contracts with some data vendors and got API access. Then he pulled all the data in and started running regressions to analyze the inflationary or deflationary impacts of different economic changes.

The U.S. Bureau of Labor Statistics has a whole classification of tasks, about 2,000 items. Malcolm used AI to assess which tasks could now be completed by AI and which could not, scoring them according to a rubric. The results showed that about 3% of tasks could now be completed by AI.

So he created a metric to measure which things could be done by AI, and the deflationary effect when they are. Output might increase, but because costs fall so drastically, GDP could theoretically contract. He calls this "Phantom GDP."

Based on this concept, he built a comprehensive analysis and also established a new language model benchmark containing about 2,000 evals.

Patrick O'Shaughnessy:

All this was done by him alone?

Dylan Patel:

Yes, all by himself. He told me, "Dude, this would have been something a 200-person economics team would take a year to do." He's completely immersed in Claude now, saying everything has changed.

Patrick O'Shaughnessy:

As a business operator, how do you make sense of this? You went from almost zero spending on this to it being nearly 25% of payroll and still rising. At what point do you think, "Wait, should I hit the brakes? Should I control spending? Maybe we don't always need the absolute latest frontier model released today, like Opus 4.7, and could switch to a cheaper one?"

Dylan Patel:

Ultimately, I'm in the information business. We sell analysis, do consulting, and create datasets. I see no reason why these things won't be completely commoditized at a pretty rapid pace.

If I don't continuously improve, my first data product sold is now seeing more people doing similar things. We can still sell it because we keep making it better and more detailed. But the way we were doing it in 2023 isn't that different from how others are doing it now. If I don't keep raising the bar, I get commoditized. If I'm not fast enough, I lose my edge.

So the question is: Yes, AI will commoditize many things, just as it is commoditizing software. But those who act fast enough, maintain customer relationships, consistently deliver excellent service, and continuously improve their service won't shrink; they'll grow faster. The incompetent, those who do nothing, will lose.

So it's a bit of a survival problem: If I don't adopt AI, someone else will, and they'll beat me.

Another simple example is the energy sector. We've had a few energy analysts for the past year or so, trying to build an energy model. It's very complex, and the energy data services market is about $900 million, so it's clearly a large market I want to enter. But despite our team having someone working on it for a year, we hadn't really cracked the energy data services business.

Then, the "Claude Code psychosis" hit. One of our people responsible for data center energy and industrial business, Jeremy, started using Claude Code, and things changed dramatically. In three weeks, he spent a lot of money – about $6,000 a day, which is indeed crazy. But he scraped every power plant in the US, every transmission line above a certain voltage level, built a map of the entire US grid from various public data sources, and also integrated a lot of demand-side data.

We turned it into a dashboard showing and analyzing power shortages and surpluses in different micro-regions across the US, along with many details. This thing was built in a few weeks.

Later, we showed it to some clients who already buy our data center datasets, including energy traders. They saw it and said, "Wow, how long did this take? This is pretty good, better than Company X." Then we investigated further and found that "Company X" has 100 people who have been working on this for ten years.

Of course, our product isn't as complete or robust as theirs yet, but in some aspects, it's already better. So now I am commoditizing these energy data service companies. But conversely, if I don't run faster, who will come to commoditize me?

So, from a business owner's perspective, the question isn't "am I spending a lot of money." Yes, I am. The question is, what is this money getting me? Is it generating more revenue? If the answer is yes, the money is worth it.

Patrick O'Shaughnessy:

Do you worry that, ultimately, those who control capital and decide where to invest it – the people who often hire you for what you do – might say, "We have analysts too, and they're smart, why don't we just do it ourselves?" If this becomes so easy, at what point does it all flow back in-house to investment firms? They are, after all, the ones most likely to get the biggest leverage from this data and insight.

Dylan Patel:

First, fundamentally, any information services business works like this: The value I get from a piece of information is obviously less than the value the client gets from it.

If I sell you information for $1, you're willing to pay that $1 because you know that information helps you make a decision that earns you more than $1. You have an arbitrage opportunity. You make more money from me than I make selling you the information.

Investment funds certainly have their own information service capabilities. Firms like Jane Street and Citadel are very granular and deep in their data usage. Yet these institutions still buy our data, continue to buy it, and our relationship with them is growing.

I think there's an "it factor" here. We move faster, are more agile, have a smaller team, and focus on a very specific niche: AI infrastructure and the massive changes it's causing, including AI, the token economy, and the whole package. We can see directions earlier and build things faster.

So, investment professionals will certainly try to do some of what we do in-house. But more often, they buy our data and then build on top of it. For them, buying our data and building on it is usually cheaper than building from scratch. Someday, someone will inevitably try to do it all themselves.

Tokens Become the New Means of Production

Patrick O'Shaughnessy:

I feel like every time we talk, I end up back at the same question: the supply and demand of tokens. It's the most interesting thing in the world to me right now. Have your own experiences given you a new understanding of the demand side? After feeling this so viscerally yourself, has your judgment on token demand changed?

Dylan Patel:

Stepping back and looking at the macro picture, Anthropic's ARR might have grown from $9 billion to around $35-40 billion. By the time this episode airs, it might be $40-45 billion.

But their compute hasn't grown at the same pace. If you do the math and assume they haven't reduced R&D compute – which they clearly haven't, as they keep releasing new models like Metis, Opus 4, Opus 4.7 – it means one thing: even if all the new compute went solely to inference, their gross margin floor would be around 72%.

In reality, some of that new compute likely went to R&D, so their actual gross margin is probably higher than 72%. Remember, earlier this year someone leaked part of their fundraising documents,