Against Whom Is 'Physical AI' Arming?

Three Researchers, Three Billion Dollars: Against Whom Is "Physical AI" Arming?

The price of a company with no product is not the price of the product; it is the price of the target.

17 September 2026. According to Bloomberg and the Financial Times, London-based Emulate is in advanced talks for a "seed" round of up to 700 million dollars, co-led by Index Ventures and Lightspeed Venture Partners; the valuation is about 3.7 billion dollars. The company was founded in August 2026; its founders are Jack Parker-Holder (lead researcher on DeepMind's Genie 2 and Genie 3 world models), Matthew McGill and Philip Ball — all three from DeepMind's world-model team, none of the three with prior experience of running a company. The company has no announced product, no website, no disclosed team size. A few weeks earlier The Information had written that the same team had raised 100 million dollars; in a month the target grew sevenfold. The focus: "world models" — not language, but artificial-intelligence systems that simulate a physical environment and predict how the environment will change after an action. The umbrella word the press uses: "physical AI."

Dear Young Comrades,

We are building this piece not on a piece of news but on a word. The word is "physical." When the adjective "physical" is added to an artificial-intelligence company, the sentence says this: what we have done until now was not physical. Correct; language models were producing text. Against whom does a machine that produces text work? Against those who work with text — that is, against white-collar labour. And against whom does a "physical" machine work? Against those who work with objects — that is, against the warehouse worker, the assembly line, the driver. The companies' own word draws capital's map of fronts: the first front has been opened and its balance sheet is in view; now the second front is opening. This piece sets those two fronts side by side and tries to show that the first front does not end with the second; on the contrary, the second deepens the first.

Earlier, in Reading GPT-6 Astra Through a Class Lens, we had set a reading template: what is being sold, from whom has it been taken, to whom is it being sold, who pays the bill? In this piece we ask the same four questions, this time when there is as yet nothing to sell. Because putting 3.7 billion dollars on a company with no product does not void the four questions; on the contrary, it leaves the answers bare.

First let us set the fact down correctly

Class analysis begins not with exaggeration but with the figure. Emulate's news is not on its own; it is the last row of 2026's world-model investment table.

Company	Founder / origin	Raised in 2026	Valuation	Product status
Emulate (London)	Parker-Holder, McGill, Ball — DeepMind Genie team	700 million $ (in talks)	~3.7 billion $	No product, no site
Skild AI	Carnegie Mellon origin	1.4 billion $	14 billion $	Robot "brain"
World Labs	Fei-Fei Li	1 billion $+	~5.4 billion $	Marble (3D world generation, on sale)
AMI Labs (Paris)	Yann LeCun, left Meta	1.03 billion $	3.5 billion $	No product
Ineffable Intelligence (London)	David Silver, left DeepMind	1.1 billion $	—	No product
Runway	"Pivot" from video generation	315 million $	5.3 billion $	Gen-4.5
Decart	—	300 million $	4 billion $	Oasis 3
General Intuition	—	320 million $	2.3 billion $	—
Odyssey	—	310 million $	1.45 billion $	—

Sources: Finance Magnates, Forbes, TNW, Introl. In the first six months of 2026 alone more than 2.3 billion dollars went into this field; Forbes's figure is 3 billion. If Emulate's round closes, the year-end total will pass 4 billion.

Note the shared investor names in the table as well: SoftBank, Andreessen Horowitz, Lightspeed, Index, Sequoia, Khosla, General Catalyst, Bezos Expeditions; and as corporate investors Nvidia, AMD, Amazon, Google, Toyota, Autodesk, Adobe. This list is almost the same as the investor list of the language-model wave. Not new capital; the same capital's new target.

And a context: after Demis Hassabis stepped back from the CEO role at DeepMind to the chair, senior-researcher exits sped up; in the last 18 months 112 DeepMind alumni are reported to have founded, or to be founding, companies, more than a quarter of them in Britain. In City AM's phrase DeepMind has become a "founder factory" for venture capital. Keep that phrase in mind; we will come back to it in the fifth section of the piece.

What is a world model, and how does it differ from a language model?

We cannot discuss the class content without getting the technical content right; so let us go short but carefully.

A language model predicts the next piece of a text. A world model predicts the next state of the environment given an action: if I push the glass will it fall, if the robot arm turns at this angle will it hit the shelf, if the lorry takes the corner at this speed will the load shift. By Fei-Fei Li's definition a world model has three criteria: to produce worlds with perceptual, geometric and physical consistency; to be multimodal by design; and to give as output the next states of the world according to input actions.

DeepMind's Genie 3 produces, from a text prompt, interactive 3D environments that stay consistent for several minutes at 24 frames a second; the physics is not hand-coded, it is learned from data. Nvidia's Cosmos is trained on 20 million hours of real video and is used by Figure, Agility, 1X, XPENG and Uber in robot and autonomous-vehicle training. Meta's V-JEPA is the foundation of LeCun's AMI Labs.

The compute cost is above that of language models: 8–32 GPUs per request for inference, thousands to tens of thousands of GPUs for training, video data at petabyte scale. This figure will be decisive when we discuss the monopoly question in the sixth section of the piece.

The promised uses are written out plainly as well: training warehouse robots "in thousands of simulations before putting a real robot on the floor"; generating rare accident scenarios for autonomous vehicles; factory, construction-site, infrastructure inspection; and games, visual effects, virtual-reality content. The first half of the list is the working field of blue-collar labour, the second half of white-collar creative labour. Do not forget this; the fourth section of the piece is built exactly on this duality.

"Physical AI": the word itself is a programme

Capital most often explains its own targets, even before its rivals do, in its own marketing language; we need only know how to read the language.

The phrase "Physical AI" spread in Nvidia's launches in late 2025 and became an investment category in 2026. Market reports write that this field will grow from 1.5 billion dollars in 2026 to 15.2 billion in 2032. The word has three functions.

First, it draws a boundary. To say "physical" is to admit that the previous wave was not physical. This admission is the confession that language-model capital has hit a limit: a machine that produces text cannot enter work that is not done with text. Warehouse, factory, road, building site — the labour-intensive body of the world economy — stayed outside the language model's reach. "Physical AI" is the ticket into that body.

Second, it opens a new investment cycle. The language-model market concentrated in a few monopolies' hands in three years; the cartelisation we discussed in The Class of a Resignation: What an Anthropic Researcher's Farewell Says, and What It Cannot Say is the official form of this concentration. In a concentrated market there is no room for a newcomer; and venture capital makes its money by investing in the newcomer. So a new market must be invented. Runway's "pivot" from video generation to world models, taking its valuation to 5.3 billion, is the clearest example of this invention: the same company, the same team, the same technology; a new label, a new valuation.

Third, and most important, it defines a target. Every technology's value for capital is measured by which living labour it can replace. What Marx said about the machine in Capital is simple: the capitalist buys the machine not to lighten labour but to cheapen the product — that is, to cut the share of waged labour inside the product. The target list of "physical AI" is written in its own use-cases: warehouse picking, packing, lorry driving, assembly, the building site. These are the sectors in which unionisation is still possible in the world, in which strike power is still felt. In 2026 a strike in an Amazon warehouse or a port terminal can still stop the movement of goods. "Physical AI" promises to undermine this power to stop.

Let us say one thing plainly: this promise has not yet been realised; robotics is still expensive, brittle and slow. But capital's target can be read independently of capital's success. And the target is written in the investment table.

White-collar work: the first front's balance sheet and the second front's hidden target

Now we come to the heart of the piece. The discourse of "physical AI" says two things to white-collar labour: first, we are done with you; second, you are no longer the target. Both are false. Let us take them in order.

A. The first front has not closed; its balance sheet is only now being published

The effect of the language-model wave on white-collar labour is no longer a matter of forecast but of measurement. We take the figures from Fortune's April 2026 compilation:

Indicator	Source	Figure
Early-career employment in occupations exposed to artificial intelligence	Stanford Digital Economy Lab (Brynjolfsson), November 2025	a 16 percent fall since the end of 2022
Employment of software developers aged 22–25	The same study	a fall of nearly 20 percent from the peak
Software-development job postings (US)	Indeed	a 53 percent fall from the late-2022 floor
Unemployment among new computer-science graduates	New York Fed	7.0 percent (computer engineering 7.8 percent)
Monthly employment effect of artificial intelligence (US)	Goldman Sachs, April 2026	~16,000 jobs a month
Of entry-level white-collar jobs within five years	Dario Amodei (Anthropic CEO)	half may be gone

The class reading of this table is the same as the frame we set in TÜİK: Two Figures, One Country: 8.1% and 30.6%: artificial intelligence is not putting existing workers out the door in a mass; it is closing the entrance. Junior posts are not being opened, internship programmes are being cut, "let one person do a three-person job" is being said. The youngest detachment of the industrial reserve army is being enrolled before it has even entered its first job. We named this, in The Tearing of the Illusion, the "mental proletariat": a layer whose only thing to sell on the market is mental labour-power, and which still thinks itself "professional" while the price of that labour-power falls.

Now the discourse of "physical AI" says to this layer, indirectly, this: the part that concerns you is complete. That is, white-collar wage and employment loss is no longer "news" but "accepted data." A front's leaving the news is not the closing of the front; it is the routinisation of the war.

B. The world model targets white-collar work as well — and says so in its own publicity copy

The most important claim of this piece is this: "physical AI" is not directed only at physical labour. What a world model produces is not robot motion but the design of robot motion. And design is white-collar work. Let us make it concrete.

What does it mean to "train a warehouse robot in thousands of simulations"? Today this work is the work of simulation engineers, robotics software workers, test and verification specialists, process planners. A factory's production line is modelled by industrial engineers; a logistics network is optimised by operations researchers; accident scenarios for an autonomous vehicle are written by safety engineers. The world model is handing to the machine exactly the work these people do — the work of modelling the world. What is "physical" is the robot's arm; the labour being targeted is the head of the engineer who thinks the robot's arm.

The creative-labour side is even clearer. Keywords Studios, one of the large service companies of the games sector, in its piece on world models, writes that these models will "speed up the prototyping of mechanics," will make it possible to "test mechanics before full implementation," will "shorten iteration cycles in pre-production." It then offers a reassurance: "The creative direction stays human-led. Programmers still define the constraints. Designers still shape the intent." And the proposal: "team training for collaboration with AI, not replacement by AI."

Let us learn to read this text, because in the years ahead you will read hundreds of this kind. What does "shortening the iteration cycle in pre-production" mean? In a games studio pre-production is one of the most crowded stages: level designers, technical artists, prototype programmers. If the cycle shortens, a portion of these people is no longer needed. The sentence "designers still shape the intent" says nothing about the number of designers; three people doing the work of ten designers by "shaping intent" is fully compatible with the sentence. "Not replacement but collaboration" is the same formula we unpacked in A Class Analysis of the World of Artificial Intelligence in Light of the Anthropic Survey: work that is not replaced is not work in which the number of people doing it stays fixed; it is work that, when done with the machine, is done with fewer people. When capital speaks of protecting "the work" it is not speaking of the worker.

So let us draw the table correctly:

Front	Target in the discourse	Real target	Capital's word
Language models (2023–)	"Knowledge work," "productivity"	Entry-level white-collar: software, law, finance, content, customer service	"Assistant," "copilot"
World models / "physical AI" (2026–)	Warehouse, factory, road, building site	Blue-collar plus the white-collar that designs their work: simulation, robotics software, industrial engineering, test, games and VFX pre-production	"Physical AI," "simulation," "collaboration"

The two fronts are not two separate wars. The second front, while carrying the first into the physical world, also takes in the white-collar layer the first did not target — the engineer, the designer, the planner. Capital's target is not "knowledge work" or "physical work"; it is waged labour in its entirety. The words only say whose turn it is.

C. The white-collar error and the blue-collar opening

This table carries a particular lesson for the computing worker in Turkey. In The Computing Worker's Handbook we had described the courier and the warehouse worker as a "ring of alliance." "Physical AI" turns this description from an abstract wish for solidarity into a concrete unity of interest: the simulation engineer who trains the warehouse robot and the warehouse worker the robot is to replace are two targets of the same investment round. The engineer may think themselves among "those who do"; the worker among "those who are done to"; capital does not make this distinction. The sentence in the handbook is even truer today: the computing worker is not limited to the one who writes code; the one who writes code is of the same class as the worker the code is to replace.

3.7 billion in a month: what is capital buying?

That a company with no product is worth 3.7 billion dollars is not a scandal; it is a mechanism, and the mechanism is instructive.

Charlie Dai of Forrester's statement to City AM is clear: the valuation is pricing "future capability production, not current revenue or commercial traction." FourWeekMBA's analysis is clearer still: a formation-stage valuation is "the output of an option-pricing bargain, not discounted cash flow"; what is being priced is "not an asset but an option." And the decisive finding: this is "a thin specialist market"; in frontier fields "a departure from an institution strengthens the rival at the same time."

Let us translate these three sentences into Marxist language.

What is being priced is the knowledge in three people's heads. Where did that knowledge accumulate? At DeepMind; in a laboratory on which Google spent tens of billions of dollars over ten years, in which thousands of researchers worked in common, fed by public papers and open-source tools. Genie 3 is the product not of three people but of a collective; and that collective is built on the research commons of the previous decade. This is what Marx in the Grundrisse called general intellect: socially produced knowledge becomes the productive force itself. In The Revolt of Crystallized Labor: A Call to the 20th Karaburun Science Congress we described the expropriation of this knowledge by locking it inside the models. The Emulate affair shows the second layer of the expropriation: expropriated knowledge is now being turned into private property a second time, in three people's names.

The finding of a "thin specialist market" is a finding of a scarcity of labour-power. There are perhaps a few dozen people in the world who have trained a world model at Genie scale. This scarcity puts an extraordinary rent on those people's labour-power. How does capital pay this rent? Not as a wage; as equity. The three researchers, however high it was, were taking a wage at DeepMind; at Emulate they are taking a share of property. The wage relation is being converted into an equity relation. The class consequence of this matters: a waged researcher can object to their company's decisions, can resign like Coxon, can write a letter with 1,386 signatures. A founder-shareholder is identical with the company's decisions. The "founder factory" is a class-conversion machine that carries the most capable research workers from a position in which they can object to a position in which they cannot. The one who becomes a founder ceases to be a worker; but they do not become a capitalist either — they turn into a manager appointed by the funds. After the round, Index and Lightspeed will sit on Emulate's board.

The option price is the price of uncertainty; and uncertainty is a thing capital likes. Putting 700 million dollars in on 3.7 billion means the funds are buying about 19 percent. These funds work on the assumption that nine of ten investments will fail and one will return a hundredfold. That is, Emulate's failure is not a problem for the system; it is the expected result. The creation of three billion dollars of "value" in a month is not a production of value; it is a wager; we will ask who will pay the cost of the wager in the sixth section.

At this point it is necessary to recall the concept of finance capital that Lenin described in Imperialism: capital detached from production, playing wagers on the future of production without investing in production. The price put on a product Emulate has not yet produced is exactly the measure of this detachment.

Capital's circulation through itself: who puts the money in, to whom does it return?

Look at the investor list again: Nvidia, AMD, Amazon, Google. These four companies are putting money into world-model startups. And what do those startups buy with that money? GPUs and cloud. Who sells the GPU? Nvidia and AMD. Who sells the cloud? Amazon and Google.

Think of this not as a conspiracy but as a circuit. If Emulate raises 700 million dollars, a large part of this money — by sector averages more than half — will return to Nvidia and the cloud providers as the rent and purchase of tens of thousands of GPUs. That is, Google's losing its researchers turns into Google's cloud revenue; Nvidia's investing in rival startups turns into Nvidia's chip sales. As Rudolf Hilferding showed in Finance Capital, monopoly does not abolish competition; it takes competition inside its own circuit. The nine startups that look independent are in practice the four monopolies' external research laboratories: if they fail, the cost stays with the funds and the pension investors; if they succeed, they are bought. DeepMind itself was bought exactly this way in 2014.

And let us see the compute reality behind the discourse of "entrepreneurship." The tens of thousands of GPUs required to train a world model are today in the hands of only five or six companies. This means the world-model field has monopolised before it is born: whichever startup wins, the owner of the compute wins. In Learning from Everyone Is Permitted, Learning from the Monopoly Is a Crime we described this as "openness that serves whoever has the chip." World models are a "new front" that serves whoever has the chip.

Who pays the bill of the new front? Three groups. First, the pension funds, university endowments and sovereign wealth funds that are the source of the money that turns into GPU rent — that is, indirectly, the savings of the world's workers. Second, the communities that share the electricity and water of the data centres in which those GPUs run. Third, and principally, the labour that is targeted to be replaced by the simulation that will train the robot and by the engineer who designs the simulation. If the wager is won the bill is dumped on labour; if it is lost, on the fund.

What does science say? Capital is giving the answer before science does

Young comrades, read this section carefully; because against the discourse that "technology is inevitable," your strongest instrument is technology's own scientific argument.

There is a workshop in the 2026 programme of CoRL, one of the most important conferences in robot learning; its name is "Do Robots Need World Models?" The workshop's framing text lists three doubts: end-to-end trained policies (visuomotor models, diffusion policies) may already contain the required predictive structure implicitly; a separate world model, having to model a much wider scope than is needed for action, runs into problems of abstraction; and explicit models can bring "hallucinated predictions, modelling errors, slow inference or unnecessary complexity." The concluding sentence is this: "Existing benchmarks cannot tell us whether world models are necessary."

That is, while scientists are only now asking the question, capital has priced the answer at 3.7 billion dollars. This is not an investment contrary to science; it is an investment indifferent to science. What venture capital is interested in is not whether the world model is necessary; it is whether the world model is saleable. And saleability springs not from scientific correctness but from what the buyer — the warehouse operator, the car company, the defence contractor — hopes for.

Let us apply the principle "the absence of data is also a datum" here as well. Emulate's compute agreement, infrastructure partner, cash-burn projection, even team size have not been disclosed. The company has no website. The press has not reached the company. That a three-person team reaches 3.7 billion dollars in a month shows that it is not knowledge that is being priced but the absence of knowledge; because an option gains in value as uncertainty rises.

This does not mean the risk is fake. Simulation may truly be necessary for robots to work safely in the real world; producing rare accident scenarios synthetically may truly save lives. Our problem is not with the technology; it is with the necessity of the technology being decided not by science but by the wager.

From Britain to Turkey: tenant countries

Emulate, Ineffable, Wayve: London is being presented as a "world-model capital"; the British press writes with joy that "more than a quarter of DeepMind alumni are in Britain." But look at the money: Index (London–San Francisco), Lightspeed (Menlo Park), Creandum (Stockholm). Look at the chip: Nvidia (Santa Clara). Look at the cloud: Amazon, Google, Microsoft. The "British artificial-intelligence ecosystem" is an arrangement in which US capital rents British labour with a US chip. Britain is exporting its most capable researchers as "founders"; the return is a few hundred high-waged jobs in London and share value flowing to US funds.

Turkey's position is two steps below even this. We produce neither founders nor chips; we are, as we wrote in The Class of a Resignation: What an Anthropic Researcher's Farewell Says, and What It Cannot Say, a market in which a dollar-based machine is used with lira-based labour. How will "physical AI" come to us? As a robotic picking system in an e-commerce warehouse, a simulation licence in a car plant, autonomous crane software in a port terminal; that is, as an imported product. And the labour that imported product will replace is already, in Turkey, the most insecure, the lowest-waged, the least organised labour: the courier and warehouse worker we wrote of in the Gig Economy Dossier, the shop worker we wrote of in Sandalyeyi Kim Kaldırdı?.

The question for the computing worker in Turkey is the same: when these products arrive, who will adapt them, who will maintain them, who will train them? Today this work is done in Turkey; tomorrow "simulation" may pull these too into the centre. The white-collar of the tenant country shares the same fate as the blue-collar of the tenant country; only with a delay of a few years.

A comparative table: the discourse and its class counterpart

Discourse	Class counterpart
"Physical AI"	The confession that language-model capital has hit a limit; the ticket into the labour-intensive body; a programme aimed at blue-collar strike power
"World models are different from language models, a less crowded field"	A new investment cycle invented because there is no room for a newcomer in the concentrated language-model market; Runway's "pivot" is the proof
"Train robots in thousands of simulations before putting them in the real world"	Unpaid apprenticeship for the robot; the handing to the machine of the work of the engineer who today does the simulation — white-collar work is in the target as well
"Designers still shape the intent; not replacement but collaboration"	The formula that protects the work, not the worker; ten people's work to three
"Future capability production is being priced"	Not the product but the target is being priced; option = wager; the cost of the wager, if won, to labour, if lost, to the fund
"DeepMind is a founder factory"	Socially produced general intellect turned into private property a second time; the conversion of the waged researcher who can object into the founder-manager who cannot
"Nvidia, AMD, Amazon, Google as investors"	Capital's circulation through itself; startups as the monopoly's external laboratory; the field monopolised before it is born
"Britain, world-model capital"	US capital + US chip + British labour; a tenant economy; Turkey two steps below
"Do robots need a world model?" (CoRL 2026)	While science asks the question, capital has priced the answer; investment indifferent to science
Undisclosed: team, compute agreement, cash burn, website	The absence of data is a datum: as uncertainty rises the option gains in value

If we reduce the table to a sentence: the word "physical" describes the machine, not the target; the target, physical or mental, is waged labour in its entirety.

Three common errors

First: saying "white-collar work is finished, now it is blue-collar's turn." This is to take over capital's word as it stands. We showed above: the world model produces not the robot but the design of the robot; the engineer and the designer are in the target as well. The fronts are not successive; they are simultaneous. The political consequence of this error is heavy: a left that counts white-collar work "saved" and blue-collar work "lost" breaks the alliance between the two layers before it is built.

Second: saying "a bubble, they will all burst." Emulate may fail; seven of the nine companies may fail. But when the bubble bursts the GPUs stay at Nvidia, the data at Google, the trained models with the monopoly that buys them; what bursts is only the fund's money. Bubble talk ignores that monopolisation proceeds independently of the bubble. Amazon and Google came out of the 2000 dot-com collapse; the collapse did not destroy the monopoly, it selected it.

Third: saying "technology is neutral, the question is use." A world model's training data, architecture and cost structure make it a tool that only those with tens of thousands of GPUs can use. This tool, before it is used "badly" or "well," is a tool whose who-can-use-it has already been determined. To say, without a public compute infrastructure, "let workers' cooperatives use the world model too" is not a technical demand but a political one — and it should be formulated as such.

Concrete proposals

1. Every time you hear the word "physical," ask "whose work." Who today designs the simulation in which a robot is trained, who today does the work the robot is to replace? The two answers are not two separate class positions; they are a single target list. To ask this question is the first sentence of the alliance between the engineer and the warehouse worker.

2. Put "simulation labour" on the computing union's agenda. Simulation engineers, robotics software workers, technical artists, test specialists working in Turkey in automotive, the defence industry, logistics and games: this layer today thinks itself "unaffected by artificial intelligence." The world model targets exactly their work. The principle in The Computing Worker's Handbook that "the computing worker is not limited to the one who writes code" should be made concrete so as to cover this layer.

3. Build the alliance with warehouse and logistics workers first at the technical level. When a robotic picking system arrives in a warehouse, the person who knows in which simulation that system was trained, at what error rate it works, which of the worker's tasks it has taken on, is the engineer. This knowledge is the worker's strongest card at the collective-bargaining table. The circulation of knowledge inside the class is the precondition of organising.

4. Demand an "automation clause" in collective agreements. In the 2022–2024 agreements of the international dockers' unions (ILWU, ILA) there are provisions for union approval, a retraining fund and a job guarantee for the bringing in of autonomous cranes and vehicles. In Turkey there is as yet no such clause in any collective agreement. The clause should be written before "physical AI" arrives.

5. Demand public compute infrastructure as a budget line. World-model training wants tens of thousands of GPUs; this hands the field to five companies. Public GPU capacity that universities, public research institutions and workers' cooperatives can reach is not a utopia; it is a heading for a parliamentary budget debate. What we said for the science budget in the Space Dossier holds here as well.

6. Do not count the researcher who becomes a "founder" as cut off from their class; but ask of whom the board is composed. The three researchers may, six months later, be working under a CEO appointed by Index and Lightspeed. The founder myth isolates the research worker. In international computing unions' contact with workers at frontier laboratories, the equity, confidentiality and non-compete contracts in spin-off processes should be an agenda item.

7. Pull the scientific argument onto the class's side. The CoRL workshop's question — "do robots need a world model?" — looks like an engineering question but it is a public question: who is paying the cost of this uncertainty? Popular-science channels will relay this news as "artificial intelligence now understands physics"; let your commentary be "which work, for whom, with what evidence."

8. Watch its arrival in Turkey and record it. In which warehouse, which factory, which port is the first robotic/simulation system coming in; how many workers are being put out, how many engineers "repositioned"? TÜİK will not collect this data. The class must collect its own data itself; in TÜİK: Two Figures, One Country: 8.1% and 30.6% we said "the absence of data is also a datum" — filling the absence is our work as well.

Dear Young Comrades,

In August three researchers left a laboratory; in September 3.7 billion dollars was put on the company they founded, when it had as yet produced nothing. This figure is not a measure of success; it is a measure of a target: capital is wagering on which labour the knowledge in three people's heads — and in fact produced by a collective over ten years — can replace. The name of the wager is "physical AI"; the word says where the machine will enter: the warehouse, the factory, the road, the building site. But the simulation that teaches the machine that path is done today by engineers, that game is built today by designers. While the word says "physical," the target covers mental labour as well.

We now know in figures what the language-model wave did to white-collar work: a 16 percent fall in early-career employment, 53 percent in software postings, 7 percent in new-graduate unemployment. This front has not closed; it has left the news. Now the same funds, the same chips, the same cloud are opening the second front, and the second front contains the first. The line they want drawn between white-collar and blue-collar work is not capital's; it is a line in the minds of those who must organise against capital. To erase that line is this piece's only proposal.

Young comrade, when you read the next "physical AI" story the question you will ask is clear: where is the machine entering, who works there today, who designs that machine today — and why are these two not at the same table?

Comradely.

Knowledge belongs to everyone.

Sources

The Emulate news

Bloomberg, "DeepMind Offshoot Emulate Closes In on $700 Million Seed Round", 17 September 2026
Tech Funding News, "DeepMind's Genie creators launch Emulate, seeking $700M…", 17 September 2026 — founders, investors, 112 DeepMind alumni, "physical AI" market projection
City AM, "Month-old DeepMind spinout nears $4bn on 'future capability' bet", 17 September 2026 — Forrester/Charlie Dai assessment, "founder factory"
The Next Web, "Former DeepMind world-model researchers' startup Emulate in talks to raise $700M", 17 September 2026 — Creandum, Genie 3, DeepMind management change, Ineffable and AMI rounds
The Information, "Ex-Google DeepMind Researchers Raising $100 Million to Build 'World Models'", August 2026
FourWeekMBA, "Emulate and the $700 Million Question: What Investors Are Actually Pricing at Formation", September 2026 — option pricing, thin specialist market
PYMNTS, "Weeks-Old AI Startup Emulate Approaches $4 Billion Valuation", 17 September 2026

The world-models market

Finance Magnates, "Big Tech and VCs Pour Billions into World Models to Make Physical AI a Reality", 2026 — company/round/valuation table, investor list, Fei-Fei Li definition
Forbes, "World Model Startups Raise Billions As VCs Bet Past LLMs", 30 June 2026
Introl, "World Models Race 2026" — Genie 3, Cosmos, Marble, AMI Labs; GPU requirements
TechCrunch, "AI video startup Runway raises $315M at $5.3B valuation, eyes more capable world models", 10 February 2026
BusinessWire, "Odyssey Raises $310 Million to Accelerate World Simulation", 17 June 2026
StartupFundraising, "AI Physical Intelligence & World Models Fundraising (2026)" — physical-AI rounds 10–30 times larger than software
Google DeepMind, Genie 3
Fei-Fei Li / World Labs, "A Functional Taxonomy of World Models"

The white-collar balance sheet

Fortune, "AI won't kill your job — it will kill the path to your first one", 29 April 2026 — Stanford Digital Economy Lab (16 percent / 20 percent), Indeed (53 percent), New York Fed (7.0 percent / 7.8 percent), Goldman Sachs (16,000 a month), Amodei ("half of entry-level jobs")
Keywords Studios, "From Pixels to Physics: Why World Models Deserve A Serious Look", 2026 — world models in games pre-production; "not replacement but collaboration" discourse

The scientific argument