● AI Token Wars
AI era where it’s hundreds of millions of won per month becomes real… a real polarization scenario that happens when it shifts from subscription to usage-based pricing
Core point to look at first right now: “Even AI gets used faster and more by people who have money.”
The conclusion of this issue is simple. As AI services quickly move from a “subscription” model to “usage (pay-per-use) billing,” the AI gap is likely to harden into a gap in available computation, tokens, and speed.
In particular, from Anthropic’s coding agent tool (Claude Code), talk about raising the pricing/cost estimates (in effect, designing it to be “more expensive”) has spread a mood where “the era of AI costing hundreds of millions of won per month is coming soon” might not just be hype.
There are exactly three things this article covers today.
1) Why subscription cannot hold and usage-based billing will inevitably grow (token/agent costs) 2) Why IPO and big-tech performance pressure pushes pricing policy toward “usage-based billing” 3) Warning points for how AI could move not toward democratization, but toward “utility (infrastructure)ification + class entrenchment”
1) Why “subscription” is shaking: AI agents consume tokens like crazy
① The structural cost problem faced by Claude Code
The core of the article/talk track is this. Claude Code–type coding agents aren’t just simple chat; they carry out APIs called repeatedly in a development environment, multi-step tasks, and even complex interactions like screens and input.
So it’s not just a matter of “generating the answers” like before— token consumption explodes while automating the work.
② Why “usage-based billing,” not monthly billing, becomes natural
If the service provider has to cover the costs, it only makes sense that “as usage increases, costs should increase too.”
That’s where pay-per-use comes in. With a structure that charges by what you use, costs get passed on more precisely to users (or companies) doing lots of high-end work.
③ The section where users’ “plan dissatisfaction” explodes in a way that’s felt
From the user’s perspective, these changes are felt like this.
- With a low-tier plan, “you hit the limit quickly while working”
- Better performance/fast processing/higher capacity move into a more expensive tier
- Complaints grow like “Why do I reach the limit so fast?”
In other words, on the surface it’s a “pricing policy,” but underneath it becomes a difference in “compute access rights.”
2) Why a price increase isn’t just an “individual service” issue: cost pressure spreads across big tech
① The common dilemma for AI companies preparing for an IPO
The point I find especially scary in this piece is the IPO timing. If AI companies like OpenAI and Anthropic are approaching listing (an IPO), the market looks not only at “revenue growth speed,” but also whether there’s a structure where profits can be made.
But with AI, even if revenue grows, training/inference/infrastructure costs grow alongside it. So the question of “are the margins sufficient relative to revenue?” keeps coming up.
② Big-tech earnings season: AI spending affects the stock price
The trend revealed in recent big-tech earnings announcements is this.
- The more a company spends on AI, the more its investment efficiency can be questioned
- As a result, layoffs/cost reductions proceed at the same time
So changing pricing (billing) policy may be closer to a “survival strategy” than “marketing.”
③ Cost-cutting competition: workforce restructuring and AI-native organizational formation
Following the flow of the main text, it also mentions that, starting with Meta, multiple big tech companies carry out personnel reductions and move to reorganize their organizations into AI-native ones.
This is not only about making AI better— it strongly implies “reducing the cost of running AI within the overall cost structure.”
3) The end of the subscription economy, the expansion of usage-based pricing: “tokens can become the new currency”
① The age of “cost per seat” → the age of “cost per agent”
Now a major shift is coming. Conventional software typically billed based on users (seats).
But AI agents aren’t a single person— in a company/one workflow, multiple agents are run.
Then the cost basis changes too.
- Subscription: costs scale “with the number of seats”
- Usage-based: costs scale “with the amount of work/tokens/number of calls”
So if you run 100 agents, you end up with a structure where the 100x cost risk grows.
② Who becomes more advantaged: a period where capital becomes “compute access rights”
This is where polarization can widen a lot.
- Those with capital: secure better models/faster processing/more tokens
- Those without capital: productivity drops due to slower models/limited quotas/reduced scope of work
Ultimately, rather than “who can do what better,” it can shift into a period where “who can run more and faster” becomes the competitive advantage.
4) AI literacy can become the “last line of defense”
① The ability to use tools accumulates like “cultural capital”
The message in the main text is along these lines. Even when using AI the same way, results differ.
Some people ask good questions, some design agents well and automate them, and some turn them into products.
So going forward, it’s not just “whether you can use AI,” but rather how you operate it (workflow/prompt/agent design) that may create the gap.
② AI literacy = the ability to get more outcomes with less capital
The less capital you have, the more important it is to “reduce waste and increase efficiency.”
- Design that doesn’t waste tokens meaninglessly
- Optimize the number of calls through automation of repetitive tasks
- Quality control through verification steps (human in the loop)
With these capabilities, you can create a “relative advantage” even in a usage-based environment.
5) Refinement of “ads and targeting”: AI can entrench class like infrastructure
① AI can become more than search/recommendations—it can become a persuasion device
The scariest expansion scenario in the main text is advertising. When AI receives questions and provides answers, those answers can become increasingly sophisticated not as mere information, but as targeted recommendations/persuasion.
In other words, a moment comes when the boundary of “is it an ad or not” becomes blurry. Users may feel “I chose it,” but in reality, it’s the model that steers the direction.
② A possibility of infiltrating not just viewpoints but “psychology/behavior” like political leaning and purchase tendencies
It’s extreme wording, but the core is this. As targeting becomes more sophisticated, it can be read as a warning that the possibility of manipulation based on cognition—like gaslighting—could also grow.
③ Typical power-creation pattern: fully open models vs partially restricted models
Another point is the scope of model disclosure. The stronger the model, the more likely you may see a flow where, instead of “fully open,” it gets released only to certain partners/companies in a limited way— and that itself can lead to an access gap.
In the end, contrary to the slogan “democratization of intelligence,” there’s a view that in reality, hierarchy can form due to costs and control.
6) Investment/strategy perspective: now is the time to look at it as an “AI infrastructure” issue
① Survival strategy in the usage-based era: “equity” in good tech may be the answer
The direction proposed in the main text is simple. If the gap in AI utilization widens, the side that provides AI (platform/chips/infrastructure/models/operating solutions) is likely to benefit.
So instead of just “people who use AI tech,” there’s a perspective that “people who own part of the AI ecosystem” could become the ones who benefit.
② What individuals/teams should do: literacy + cost (unit cost) management
Individuals or teams can approach it like this.
- Identify where “token/call costs” blow up for each type of work
- Automate agents, but design verification loops to reduce failure costs
- Run only the necessary tasks quickly, and route the rest through lower-cost paths
This is a practical response for the usage-based era.
The most important additional takeaway that this article says “elsewhere” doesn’t talk about enough
The real core of this news isn’t “Claude’s prices went up.” Once the unit of work for AI agents shifts from “conversation” to “automated execution,” the cost structure solidifies into usage-based billing, and as a result, the gap in access rights (tokens/compute/speed) can become entrenched like a class structure.
And when you combine corporate financial pressure—like IPOs, earnings, and layoffs— there’s a strong possibility that pricing policy becomes more conservative (and stricter).
Finally, as advertising/targeting becomes more refined, a warning follows that AI can expand beyond a simple tool into “infrastructure that guides choice and behavior.”
Main points to convey (one-line conclusion)
Within the flow of subscription AI ending and shifting to usage-based AI, AI literacy and cost optimization can become the last line of defense that reduces the gap in skill. And from a corporate/investment perspective, it becomes important to view AI infrastructure and cost structure together.
SEO keywords naturally incorporated (viewpoints covered in the article)
Today’s article was organized around the flow summarized by keywords like AI gap, usage-based billing, AI agents, big-tech cost reduction, and AI infrastructure.
< Summary >
– AI agents don’t just do simple conversations; they carry out automated execution, so token consumption explodes, making it likely that usage-based billing will spread rather than subscription – IPO and performance pressure (cost burden, doubts about margins) push pricing policy toward “usage-based billing” – The more agents you run, the more directly costs are linked, and AI access rights (speed/compute/tokens) can connect to capital inequality – As advertising/targeting becomes more sophisticated, AI turns into a persuasion device, and there’s a warning that the boundary of what’s an ad can blur – Individuals/teams can mitigate the gap with key responses: AI literacy and cost optimization (reducing token waste)
[Related articles…]
AI polarization created by usage-based billing
In the era of AI agents, token costs are the competitive advantage
*Source: [ 월텍남 – 월스트리트 테크남 ]
– 월 수백만원 AI시대 곧 온다..이제 AI도 양극화 ㄷㄷ..(feat. 클로드 종량제)
