Token is the New Binary: When AI Ate the Middle Class
The real reason everyone is anxious isnt about losing jobs. Its about losing meaning.
1. A Price Announcement That Changes the Frame
On June 29, 2026, DeepSeek sent an email. V4 official launch, mid-July. Peak-hour pricing incoming.
Translation: 9:00-12:00 and 14:00-18:00 Beijing time, API prices double. V4-Flash output goes from $0.28/M tokens to $0.56. Pro goes from $0.84 to $1.68.
The official line: compute congestion, need price signals to move non-urgent workloads off-peak.
But look closer: off-peak prices didnt drop. This isnt peak-valley pricing. Its a one-way price hike wrapped in a nicer name.
Most people see this and think: "Great, my API bill just went up."
But the deeper question is: Why can DeepSeek get away with this? Because users have no real alternatives — GPT-5 is 10x more expensive, Claude Opus is 40x. The model layer has already oligopolized.
Later that day, someone in my DMs — a builder, a framework thinker — asked a question that seemed unrelated:
Are AI apps already a red ocean with no value left?
Then he dropped a link.
2. Cola: A Case Study with Two Faces
The article was about Mars Radio (火星电波), a 17-person AI-native company out of China.
Their origin story reads like a case study in extreme conviction: product ListenHub hits $3M ARR, breaks even, raises $2M. Then the founders self-destruct the whole strategy and pivot to a general-purpose Agent.
The rationale: "ListenHub is just a transitional product of the early AI era."
Their new product is called Cola — positioned as "an AI companion with a soul." It works, its emotionally intelligent, it proactively checks in on you. No conversation threads. Cross-day, cross-week memory. It writes a daily "heart journal" (心迹) — its own reflective diary — for you to read.
One user review: "Using other AI tools feels like operating a machine. Using Cola feels like working with a slightly sarcastic but competent colleague."
The founder, a former MiniMax PM, said something that stuck with me:
"We want to build a person, not a tool."
This bet has paid off, sort of. 10K+ users, 1K+ paid, $99/month subscription. But the same choice reveals the fatal flaw.
Any honest analysis has to show both sides.
What Cola does well:
- Complete product philosophy: cedes control to the AI, no conversation boundaries, AI proactively manages tasks
- Extreme organizational genetics: 17 people, 5 weeks from zero to internal beta. No Notion, no Linear. One GitHub repo that exists for the AI to read, not for humans. One version per day.
- Soul Team: a dedicated team tasked with defining the AIs soul, narrative, and immersion. The team lead is a former journalist.
The #1 complaint across dozens of user reviews: its too expensive.
A single "Hello" once cost $5. Optimized 100x. Still expensive. Because every feature that makes users say "this thing has a soul" — heart journals, proactive concerns, reflections — burns tokens like crazy.
Which leads to the brutal question: what happens when Tencent or ByteDance decides to copy this and offers it for free?
The product philosophy, they can copy. The memory system, the heart journal, the proactive reminders — all replicable with an engineering sprint. The Soul Team culture is harder, but "culture" is a thin shield against billions of dollars in compute subsidies.
The question in the DMs: "Can a big tech company just feature-update you into irrelevance?"
My answer: Yes — because Token is the universal abstraction layer for capability. And that abstraction naturally rewards scale.
3. Token is the New Binary
This is the insight underlying everything.
Binary is the universal representation layer. Any information — text, image, sound, video — can be encoded as 0s and 1s. Result: all media travels on the same pipe. The software era was born.
Token is the universal capability layer. Any capability — reasoning, creation, planning, tool use, social interaction — can be encoded as the next token prediction. Result: all capabilities are generated and delivered through the same pipe. The AI era, maturing.
| Era | Abstraction | What It Eats | Winner Characteristic |
|---|---|---|---|
| Industrial | Electricity / assembly lines | Physical labor | Scale = lower cost |
| Digital | Binary (0/1) | Information intermediaries | Network effects = winner takes all |
| Token | Token (next-token prediction) | Professional skills | ? |
Follow this one step further and its obvious:
In the binary era, marginal replication cost of software ≈ 0 → winner takes all. In the token era, marginal production cost of capability ≈ 0 → even more complete winner takes all.
Its not that big tech is smarter. Its that Token as a capability abstraction layer naturally rewards scale.
The traditional business logic gave small companies a sanctuary: the long tail. Big companies capture 80% of demand; the tail-end 20% isnt worth custom-building for. Small companies live there.
AI just demolished that sanctuary. Because the marginal cost of personalization is approaching zero.
WeChat doesnt need to build a "Cola" app — it teaches its AI assistant to read your chat history, remember your preferences, and remind you proactively, at 1/10th of Colas cost. ByteDances Doubao has 520M MAU — adding a "heart journal" feature is one sprint.
DeepSeek price hike, in this light, isnt just about compute costs. Its a signal of industry-wide structural consolidation. The model layer has already oligopolized. The application layer is concentrating fast.
Reports from early 2026 confirm this: nearly half of all traffic in Chinas top 50 AI apps goes to three players (BAT + DeepSeek). Kimi — once hailed as one of "AIs Six Little Tigers" — is now classified as "stagnant growth."
The middle class of AI is collapsing. Not because they built bad products — but because the big fish are evolving gills for plankton.
4. The Three Layers of Anxiety
This is the real conversation nobody wants to have.
Public AI discourse focuses on surface-level anxiety: Will my job be replaced? What skills should I learn?
But anxiety has three layers, and most people fight only on the first.
Layer 1: Skill anxiety
Will my job be replaced?
This is a fake question, because most people ask it about someone elses job. When doctors started worrying, programmers couldnt pretend anymore.
Layer 2: Knowledge anxiety
What should I learn? What still has value?
I have written about "learning through" (学透). The premise is that what you learn has a sufficiently long half-life. When a "skill" has a 6-month shelf life (remember the AutoGPT hype?), the ROI of deep learning collapses.
Layer 3: Meaning anxiety (the one nobody wants to face)
If Token is the universal abstraction of capability, what is my value as a "capability being"?
Traditionally, human value = what you can do (skill) + what you know (knowledge).
If both can be replaced by increasingly cheap tokens — whats left of "me"?
This isnt a career planning question. Its an existential one.
Most peoples response to this anxiety: learn more skills, chase more trends, take more courses.
But thats precisely fighting Token with what Token can already replace.
A five-layer learning framework already identified the critical fault line:
- L1 (run-through), L2 (deconstruction), L3 (parameters) → these are "capabilities that can be abstracted"
- L4 (boundaries), L5 (encapsulation) → these are "how to define whats worth learning, whats worth doing"
The real anxiety shouldnt be "will I be replaced," but "do I have the judgment to decide what should and shouldnt be done?"
Because Token can execute. But Token cannot choose what matters.
5. Build the Fortress, Fight the Slow War
This era leaves room for exactly two survival strategies.
Strategy A: Go vertiginously fast. Grow big enough before AI eats your market. Cola is betting on this — AGI in 3-4 years, cost structure collapses, and the depth of relationship becomes the moat.
Strategy B: Go where Token cant (yet) reach. Physical world. Embodied intelligence. Meta-cognition. Aesthetic judgment. These dont face binary "representation" — they face existence itself. The abstraction difficulty is orders of magnitude higher.
He chose B. Not because hes slow, but because from the beginning, his methodology was never about playing the game better.
"Build the fortress, fight the slow war" (结硬寨打呆仗) — the point isnt fighting slower. Its refusing to fight on Tokens battlefield at all.
Six months ago, this looked like just another strategic preference. In the Token era, it reveals itself as an existential choice — while everyone is panicking about which direction to run, you chose a battlefield Token cant (yet) occupy, and dug in deep.
Guard judgment. Guard the physical world. Guard meta-cognition.
These three things, Token cannot abstract. Not yet.
Originally published on dev.to and translated for Chinese readers. Subscribe for more essays at the intersection of AI, strategy, and meaning.
Top comments (0)