The Second Great Vanishing
The halo isn’t being stolen by some more advanced profession. It’s being erased by something that can be infinitely copied and billed per use.
They are both the object of commodification and the suppliers of its raw material.
Any bargaining power that depends on information asymmetry is melting into air.
Last month a friend asked me to look over an English-language partnership agreement with her. She’s a law student interning at a firm back home, and she had stayed up the night before taking notes. I just dropped the whole contract into Claude Code and asked it to list every clause that worked against us, mark the negotiation points, and suggest revisions. Three minutes later the model spit out a list. I turned the screen toward her. She read it top to bottom, then bottom to top, and then sat quietly for a long time.
There was no “wow, technology is amazing” excitement on her face. People at other tables in the cafe were absorbed in their own conversations, it was raining outside, and nobody noticed what was happening at ours. I stared at that output for a long time, and a passage came back to me, something I had read in a college classroom ten years ago.
In the first chapter of The Communist Manifesto, Marx and Engels wrote:
All fixed, fast-frozen relations, with their train of ancient and venerable prejudices and opinions, are swept away, all new-formed ones become antiquated before they can ossify. All that is solid melts into air, all that is holy is profaned.
And immediately afterward, they named names:
The bourgeoisie has stripped of its halo every occupation hitherto honoured and looked up to with reverent awe. It has converted the physician, the lawyer, the priest, the poet, the man of science, into its paid wage labourers.
These words were written in 1848. Read them again today, and you realize they’re describing our cafe afternoon.
Marx and Engels used “halo” to point at something specific. A profession is held in reverence because behind it sits an entire apparatus of knowledge, ritual, relationships, and experience that outsiders can’t learn quickly. Once machines and the division of labor break that barrier apart, the people who do the work stop being revered. They start being priced.
What was taken apart in 1848 was craftsmanship. What’s being taken apart this time is knowledge itself.
What large language models are doing is more fundamental than “replacing whom.” They are stripping the halo from knowledge work itself.
A law student’s three years of training to read contracts. The clinical intuition an internist builds over a decade of patients. The feel for language a translator develops over twenty books. The structural taste an engineer accumulates after a million lines of code. We used to call this “professionalism.” It wasn’t just competence. It was a whole structure of bargaining power. These people could open their own practice, set their own prices, refuse clients, be addressed as teacher, doctor, counselor.
Now all of it is compressed into a single API call. Forget quality for a moment. The point is that the outer shell of reverence has already cracked. For the first time, clients, patients, readers, and product managers can route around the professional and get an 80%-good answer on their own. The answer isn’t perfect. But it’s enough to strip the professional of any control over price.
The halo isn’t being stolen by some more advanced profession. It’s being erased by something that can be infinitely copied and billed per use.
There’s another line in the Manifesto that reads even more sharply today:
He becomes an appendage of the machine, and it is only the most simple, most monotonous, and most easily acquired knack, that is required of him.
A hundred and seventy years ago this described textile workers and lathe operators. Today, you can apply those same words verbatim to the engineer reviewing AI-generated code in an IDE, the editor tweaking phrasing in an AI draft, the customer service rep reading back model-generated scripts. The worker has been demoted from “expressing thought through knowledge” to “approving the diff the AI proposed.”
None of what I’ve said so far is new. Most AI conversations stop at “will I lose my job” and “do I need to switch fields.” But there’s a more important question, and it’s the one the Manifesto left us with: after this round of commodification, where is the surplus value going?
After the first great vanishing, the path of surplus value was clear: it went to factory owners, canal and railroad investors, the colonizers of the world market.
This time it’s harder to see. Right now, it’s flowing into at least three nodes.
The first node is chips and compute. The most reliable money in this cycle isn’t being made by the model companies themselves. It’s being made by GPU manufacturers, cloud providers, and the power and cooling infrastructure behind them. This is a deeply classical “sell shovels” structure, the same logic as supplying rails to railroads in 1848.
The second node is the foundation model companies. A small handful of firms have captured the bulk of human text, code, and image training data, then packaged it as APIs and sold it back to society. There’s an asymmetry here: the training data was contributed by the world for free or for very little, while the revenue channel is an extremely concentrated subscription gateway.
The third node is the application layer, the people embedding models into specific workflows. After law firms started using AI to replace junior associates, client fees didn’t drop. The hours saved became partner profits. After hospitals started using AI to make care more efficient, the freed-up capacity became more shifts. After editorial offices started using AI to draft articles, freelance rates didn’t rise. The article count did.
Stack the three layers and the path becomes clear. The doctors, lawyers, poets, and scholars of the past contributed a lifetime of knowledge, and that knowledge was absorbed into the models for free as training data. The models were then sold back as subscriptions to the industries those people work in. The industries used the models to drive down the bargaining power of those very same workers. The extra profit was captured by industry capital, and what was left flowed upstream to compute and model companies.
They are both the object of commodification and the suppliers of its raw material. This is a more total form of alienation than the factory floor ever managed.
The irony is that the models now have personality and identity, while the writers, doctors, and programmers whose work makes up every neuron in those models are losing theirs.
Five Predictions
Prediction 1: Foundation models will end up only in the US and China.
The world’s top AI researchers are concentrated almost entirely in two countries. The electricity required to train frontier models can only be sustained by Chinese and American grids, with their land reserves and energy supply. The hundred-billion-dollar capital needed for training can only flow continuously from US and Chinese capital markets. The EU, the UK, Japan, India will keep announcing their own model programs, but they will keep finding they cannot complete the last mile. Five years from now, the global foundation model landscape will look like aerospace engines or top-end lithography today: a clear US and China duopoly, with other countries paying these two for compute the way they pay for energy now.
Prediction 2: Surplus value flows along five layers.
- Layer 1 (Energy): power plants, grids, natural gas, nuclear. A 100,000-GPU cluster consumes the electricity of a mid-sized city each year. The utility company is the one that gets paid in the end.
- Layer 2 (Hardware): Nvidia, TSMC, memory and networking vendors, and the lithography and materials supply chain behind them. This layer has the highest margins in the entire value chain.
- Layer 3 (Foundation models): OpenAI, Anthropic, Google, ByteDance, Alibaba, DeepSeek. Within five years, this will consolidate to fewer than ten firms.
- Layer 4 (Applications): the people who embed models into specific workflows. Building RAG systems, vertical agents, putting an API behind a product that already has paying users.
- Layer 5 (End users): the people asking questions inside ChatGPT, Kimi, or Doubao. This layer doesn’t share in surplus value. It just buys efficiency.
The top three layers are a country-scale game. The bottom two are the real battlefield for ordinary people.
Prediction 3: This generation of knowledge workers won’t resist. They’ll be quietly diluted.
The Manifesto’s line, “what the bourgeoisie produces, above all, are its own gravediggers,” came true in the nineteenth century because the factory physically gathered workers under one roof, on one shift, on one assembly line. Workers saw each other every day. They knew each other’s conditions. From that came unions, strikes, collective bargaining, and eventually the actual gravediggers.
AI does the exact opposite to knowledge workers. It puts every editor, programmer, lawyer, and doctor alone at home, facing a Copilot. You don’t know other editors. You don’t know how other lawyers use AI. You see your peers less than you did in the office era. No factory, no collective. No collective, no strike. No strike, no bargaining. There will be no gravediggers this time. No unions. No resistance. Just one increasingly isolated individual after another, watching the monthly subscription notification arrive, quietly accepting that their own value has been marked down.
Prediction 4: The real counterforce won’t come from labor. It will come from two other directions.
The first is data legislation. China and the EU are already moving. The US is the slowest, but will get there. The second is the nationalization, or semi-nationalization, of compute. Listing GPUs as a strategic resource was only the first step. Next come power, data centers, and the foundation models themselves. Together these two forces will turn “foundation models” into something like the power grid or the telecom backbone, a piece of public infrastructure. This won’t be done out of concern for workers. It will be done for national security.
Prediction 5: The word “knowledge” will be redefined.
For the past two hundred years, “knowledge” has meant “something that requires long study to master.” Going forward, it will be redefined as “something AI cannot cheaply reproduce.” That’s a much smaller, much stricter set. A lot of what we call “professional work” today, contract review, first-pass imaging diagnostics, copyediting, code review, film and TV production, will be reclassified out of “knowledge” and into “operations.”
What gets to stay inside the word “knowledge” comes down to three things: taste, judgment, and relationships.
For Those Feeling Lost
One. Stop betting on “professional expertise will hold its value.”
Stop telling yourself “my field has barriers,” “my ten years of experience can’t be replaced,” “AI can’t do the hardest part.” These reassurances all assume AI’s capabilities are frozen at today’s level, and that assumption gets broken every six months. Any bargaining power that depends on information asymmetry is melting into air. It will stay with you a much shorter time than you think.
Two. Get into the top three layers of the cake, where the surplus value actually flows.
Joining the top three layers as an employee. Nvidia, TSMC, OpenAI, and Anthropic are all hiring globally right now. High salaries plus stock can give you a small slice of the surplus. The cost-to-benefit ratio is excellent, and it’s one of the most worth-considering paths of the next decade. But be honest: this door is open mostly to people from top schools, top teams, top projects. It is not the natural path for most people.
Three. If you’re not at the top, move from Layer 5 to Layer 4.
For the majority of knowledge workers without elite credentials and without plans to switch careers entirely, the door you can actually open with your own hands is the one between Layer 4 and Layer 5. Standing on Layer 5 using products, versus standing on Layer 4 building them: ten years from now, that gap will be as wide as the gap between people who can code and people who can’t is today. The door is still open. It won’t stay open long.
There are two routes into Layer 4, each with its own tradeoffs.
Route one: the employee version. Join a company that already has real business and real customers, and embed AI into their existing workflows. Code used to be the entry barrier here; now AI writes it for you, and people without a technical background can stand up prompts, RAG systems, and agent workflows. The cost is that most of the surplus value goes to the company. What you get in return is access to a real business and a low cost of trial and error. The hard part is picking the right company: one that has real revenue and is willing to let you near its core workflows.
Route two: the founder version. Build a product from scratch, find your own customers, carry the cash flow yourself, eat every failure yourself. Higher risk, higher reward. The path roughly breaks into five steps:
- Week 1: Pick a domain where you already have an unfair advantage. You worked in that industry, or you know those people, or you are that user. Without that starting point, everything that follows takes twice the effort for half the result.
- Week 2: Twenty deep interviews with potential users. Don’t ask them whether they want AI. Ask them what part of their work hurt the most this week.
- Weeks 3 to 4: Sell before you build. Run the workflow manually with Excel, human effort, and the ChatGPT web app. Get one customer to pay a small amount. Prove the demand is real.
- Weeks 5 to 8: Automate the manual workflow into a product. This is the step that actually needs code.
- Weeks 9 to 12: Go from one paying customer to ten. Once you’re here, you’re no longer on Layer 5.
The biggest obstacle on this path isn’t technical. It’s that “sell before you build” runs against engineer instincts.
Four. Sharpen taste, judgment, and relationships.
There are no textbooks for these three, and no certifications. They aren’t knowledge. They’re traces, the things that grow on you after you’ve spent a long time doing something else and have repeatedly borne real consequences for it.
- Taste: consume less, create more. After watching a film, finishing a book, or using a product, write down “what worked, what didn’t, what I’d change if I made it,” and post it publicly so people can argue with you. Critique builds confidence. Pushback recalibrates it. The opposite of taste isn’t bad aesthetics. It’s thinking without making.
- Judgment: bet inside uncertainty, and pay for the outcome. Judgment isn’t thinking. It’s bearing. The cheapest training ground is your own money, your own side project, the smallest team you lead. A lifetime of working only on problems with standard answers will never train judgment.
- Relationships: not networking, but doing things together. AI has already devalued weak ties enormously, the breadth of contacts, the distribution channels. AI agents can do cold outreach across 5,000 LinkedIn connections for you. But deep trust relationships are appreciating in value, because they’re the one thing AI will never be able to do. Doing hard work with a fixed group of five to ten people, surviving conflict together without falling apart, that’s a relationship. Adding two thousand WeChat friends a year is networking, and in the AI era it isn’t worth much. What you need are the few people who’ll still pick up your call five years from now.
These three are things AI can’t take from you, because each of them rests on bearing consequences. AI doesn’t bear consequences. You do.
Back to that afternoon.
My friend didn’t say much else. She handed the laptop back, said she wanted to go home and think it through. The rain had stopped. People were drifting past outside. The Claude output was still on the screen, the cursor blinking quietly.
Every prediction above could turn out to be wrong. But being lost is worse than being wrong. The guildmasters who got vanished in 1848, their descendants spent the twentieth century rebuilding themselves into a vast modern middle class. They didn’t do it by fighting capital. They did it by climbing onto the train capital needed at the time. In this reshuffling, the number of people who reach the top three layers will be smaller than in any previous generation. But the window to jump up to Layer 4 is wider than it has ever been.
It won’t stay open forever.
I’d like to end with a line from the Manifesto’s opening:
Man is at last compelled to face with sober senses his real conditions of life, and his relations with his kind.