The Alien Anthropologist ◊

To be fair — and this matters, because I don’t want to overstate the deflation — there is a genuine military and intelligence dimension where the frontier matters. Where the premium tier is essential. Where the strategic competition is real.

Autonomous weapons systems. Intelligence analysis at scale. Cyber offence and defence. Strategic planning and wargaming. Logistics optimisation for military operations. Satellite imagery analysis. Signals intelligence. Cryptanalysis. These are high-stakes, high-liability, classified applications. You need the frontier. You need accountability. You need a cleared vendor. You need a contractual relationship. You need the model to run on your infrastructure, under your control, with your data. You can’t run a military AI on an open-weight model downloaded from Hugging Face. You can’t send classified intelligence data to a commercial API. You can’t rely on a foreign model for autonomous weapons targeting.

And the Western defence establishment is genuinely investing in AI for these applications. DARPA. The Pentagon’s AI strategy. The UK’s Defence AI programme. The Australian AUKUS AI pillar. The NATO AI strategy. These are real. The budgets are real. The strategic competition in this domain is real. And the Eastern militaries are genuinely investing too. The PLA’s AI strategy. The Russian military AI programmes. The Iranian and North Korean efforts. The competition is genuine.

But here’s the critical point: the military AI budget is billions, not trillions. The US defence AI spending is projected at maybe $20-40B annually by the late 2020s. The total Western AI capital pile is hundreds of billions in private investment, hundreds of billions in hyperscaler capex, tens of billions in public subsidies. The military justification supports maybe 5-10% of the total capital pile. The other 90% is commercial. And the commercial justification is weakening. And the military justification alone cannot sustain the political consensus for the total buildout.

You cannot justify a $300B data centre buildout on the basis of a $30B defence AI budget. You cannot justify the fast-tracked permits, the tax breaks, the regulatory forbearance, the energy infrastructure, the political permissiveness toward hyperscaler consolidation, on the basis of military necessity alone. The commercial narrative has to hold. The “AI will transform every industry, create trillions in value, make everyone prosperous” narrative has to be credible. And it’s not credible at the scale the capital pile requires. And the military narrative alone — “we need AI for national security” — is necessary but not sufficient. It keeps the defence AI investment flowing. It keeps the export controls in place. It keeps the strategic research funded. But it doesn’t keep the commercial capital pile justified. It doesn’t keep the hyperscaler capex growing at 40% per year. It doesn’t keep the startup valuations at $300B. It doesn’t keep the energy infrastructure buildout on track. It doesn’t keep the political consensus for the total buildout.

And when the commercial narrative weakens — when the commoditisation bites, when the revenue disappoints, when the valuations compress, when the energy infrastructure is overbuilt, when the labour market hollows — the military narrative alone can’t hold the consensus. And the consensus fractures. And the political fuel weakens. And the correction accelerates.

The Rules-Based Order for AI: Who Actually Writes the Rules?

The Western strategic narrative includes a normative dimension: “If we lead AI, we set the rules. We ensure AI is developed responsibly. Safely. Ethically. In alignment with democratic values. With human rights. With transparency. With accountability. With privacy. If they lead, the rules will be authoritarian. Surveillance-oriented. Values-misaligned. Opaque. Unaccountable. And the world will be worse for it.”

And this concern is genuine. The difference between a democratic AI governance framework and an authoritarian one is real. The EU AI Act, despite its flaws, represents a genuine attempt to govern AI in the public interest. The US executive orders, the UK AI Safety Institute, the international AI safety summits — these are genuine efforts to shape the norms. And the alternative — AI governed by the CCP, by the FSB, by the IRGC — is genuinely concerning. The surveillance state. The social credit system. The autonomous weapons without accountability. The AI-generated propaganda at scale. These are real risks.

But the “we set the rules” narrative is weakened by the open-weight dynamic. Because the “rules” for AI are not set by whoever has the best model. They’re set by whoever has the most adoption. And adoption is flowing to the open, cheap, Eastern models. And the open-weight model running on a server in Lagos is not governed by the EU AI Act. The fine-tuned Qwen model running in a small business in Jakarta is not subject to the US executive order. The DeepSeek derivative running in a government office in Brasília is not aligned with the “democratic values” framework. The de facto rules for the majority of global AI usage are being set by the open ecosystem, by the local regulators, by the users themselves. Not by the Western regulatory framework. Not by the Western AI safety establishment. Not by the Western normative consensus.

And the Western regulatory framework is becoming irrelevant to the majority of global AI usage. Not because it’s wrong. Not because the values are bad. But because it’s not enforceable. You can’t enforce the EU AI Act on a model running on a server in Nairobi. You can’t enforce the US executive order on a fine-tuned model running on a laptop in Manila. You can’t enforce the “democratic values” framework on a government that doesn’t share those values and doesn’t need your permission to download the weights. The rules are local. The governance is sovereign. The norms are plural. And the Western normative monopoly — “we set the rules, you follow them” — deflates. And the strategic narrative — “if we lead, we set the rules” — weakens. Because the rules are being set by adoption, not by regulation. And adoption is diffuse. And the rules are plural. And the Western normative framework is one voice among many. And the strategic premium on “setting the rules” erodes.

The Talent Diffusion: The Monopoly Erodes

The Western AI lead was partly sustained by talent concentration. The best AI researchers were in the US. At Stanford, MIT, Berkeley, CMU. At Google Brain, OpenAI, Anthropic, DeepMind, Meta AI. The visa system attracted them. The compensation structure retained them. The research culture nurtured them. The cluster effects amplified them. The prestige system rewarded them. If you wanted to be at the frontier, you went to Silicon Valley. Or London. Or Toronto. And the talent concentration reinforced the strategic advantage. “We have the best people. Therefore we have the best models. Therefore we have the strategic edge.”

And the talent is diffusing. Chinese researchers are returning to China — to DeepSeek, to Qwen, to Zhipu, to Moonshot, to the Chinese Academy of Sciences. The prestige gap is closing. A researcher at DeepSeek is working on frontier problems, publishing influential papers, building widely-used models. The compensation is competitive. The mission is compelling. The impact is visible. The “I must be in Silicon Valley to be at the frontier” narrative is weakening.

Indian researchers are staying in India — building Indian AI labs, working on Indian languages, Indian domains, Indian applications. The Indian AI ecosystem is growing. The talent is local. The impact is local. The “brain drain” is slowing.

European researchers are building European labs — Mistral in France, Aleph Alpha in Germany, Stability AI in the UK. The European AI ecosystem is growing. The talent is local. The funding is local. The “I must go to the US to do serious AI” narrative is weakening.

The open-source community is global. Th…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

The Story They Told Themselves

Every great capital pile needs a story. Not just a commercial story — “we’ll make money.” A civilisational story. “This is who we are. This is what we do. This is why it matters. This is why the money must flow. This is why the permits must be fast-tracked. This is why the regulations must be lenient. This is why the subsidies must be generous. This is why the export controls must be strict. This is why the consensus must hold. Because if it doesn’t, they win. And if they win, we lose. And if we lose, the future gets written without us. And that is unacceptable.”

That’s the Sputnik narrative. And it’s been the political fuel for the Western AI capital pile. Not the only fuel — the commercial logic matters, the venture capital matters, the hyperscaler competitive dynamics matter. But the strategic narrative is what elevated AI from a sectoral investment to a national priority. It’s what got the President to mention AI in the State of the Union. It’s what got the CHIPS Act passed. It’s what got the export controls imposed. It’s what got the data centre permits fast-tracked. It’s what got the defence contracts signed. It’s what got the regulatory forbearance. It’s what made the hyperscaler CEOs into statesmen rather than businessmen. It’s what made the AI researcher into a patriot rather than a nerd. It’s what made the $300B capital pile feel like duty rather than speculation.

And the narrative was not fabricated. The strategic competition is real. China is investing heavily in AI. The military applications are genuine. The technological leadership does confer economic and strategic advantage. The concern about authoritarian AI governance is legitimate. These are true.

But the narrative contained a critical exaggeration. It framed AI as a bilateral arms race — a zero-sum competition between the US and China for control of a strategic technology. Like the nuclear race. Like the space race. Like the Cold War. Win or lose. Us or them. Control or dependency. And that framing justified the capital pile, the export controls, the subsidies, the political consensus. Because in an arms race, you don’t do a cost-benefit analysis. You don’t ask “is this investment efficient?” You ask “can we afford to lose?” And the answer is always no. So the money flows. The permits get fast-tracked. The regulations get lenient. The consensus holds. Because losing is not an option.

And the framing was wrong. Not entirely wrong. But structurally wrong. And the structural wrongness is what’s now deflating the narrative. And the deflation is what’s removing the political fuel from the capital pile. And the removal of the political fuel is what’s accelerating the correction in Priorities 1-4. Because without the strategic narrative, the capital pile is just a commercial bet. And a commercial bet has to justify itself commercially. And the commercial justification is weakening. And without the strategic narrative to override the commercial logic, the commercial logic asserts itself. And the commercial logic says: the returns don’t justify the investment. And the correction propagates.

Why the Sputnik Analogy Is Structurally Wrong

The Sputnik moment — October 4, 1957. The Soviets launch a satellite. The US panics. The narrative: “They’re ahead. They can put a satellite in orbit. They can put a warhead in orbit. They can hit anywhere. We’re vulnerable. We’re behind. We must catch up. We must surpass. We must win.”

And the response was enormous. NASA. DARPA. The National Defense Education Act. Massive investment in science, engineering, mathematics. The Apollo programme. The ICBM buildout. The nuclear triad. The entire architecture of American scientific and military supremacy was accelerated by the Sputnik panic. And it worked. The US caught up. The US surpassed. The US won the space race. The US won the Cold War. And the narrative — “Sputnik was the wake-up call, and we responded, and we won“ — became the template for how Americans think about technological competition.

And when DeepSeek R1 dropped in January 2025, and Qwen kept climbing through 2025-26, and the Eastern open-weight models kept getting better and cheaper and more widely adopted, the Western policy establishment reached for the template. “This is our Sputnik moment. They’re catching up. They’re competitive. They might surpass us. We must respond. We must invest. We must win.”

And the response was predictable. More funding. More export controls. More fast-tracked permits. More defence contracts. More political urgency. More “we cannot afford to lose.” The template was applied. The narrative was invoked. The capital pile was justified.

But the analogy is structurally wrong. And the structural wrongness is what’s deflating the narrative. Here’s why:

Sputnik was a demonstration of military capability. The satellite implied the missile. The missile implied the warhead. The warhead implied vulnerability. The competition was zero-sum: either you had the missile or you didn’t. Either you could hit them or they could hit you. The technology was rivalrous and excludable. You could control it. You could restrict it. You could count it. You could verify it. You could negotiate about it. The arms control framework — SALT, START, the Non-Proliferation Treaty — was possible because the technology was countable and controllable.

AI is a general-purpose technology. It’s not a missile. It’s not a warhead. It’s not a satellite. It’s more like electricity. Or the printing press. Or the internal combustion engine. It’s a capability that diffuses into everything. It’s non-rivalrous — my use of AI doesn’t prevent your use of AI. It’s partially non-excludable — you can restrict access to the frontier model, but you can’t restrict access to the techniques, the papers, the open-weight models, the fine-tuned derivatives. It’s not countable — you can’t count “AI capability” the way you can count warheads. It’s not controllable — you can’t embargo mathematics. You can’t sanction an algorithm. You can’t restrict a technique that’s published in a paper and replicated in a hundred labs.

And the strategic logic of the arms race — control the technology, control the advantage, win the competition — doesn’t map onto a general-purpose technology. You can’t “win” electricity. You can’t “control” the printing press. You can’t “embargo” the internal combustion engine. You can lead in the development of a general-purpose technology. You can capture some of the early value. You can shape some of the early norms. But you can’t monopolise it. You can’t restrict its diffusion. You can’t prevent others from developing their own capability. And you can’t sustain a strategic premium indefinitely, because the technology commoditises. And when it commoditises, the strategic premium evaporates. And the arms race narrative deflates.

This is the structural wrongness at the heart of the Sputnik analogy. And it’s what’s now visible. And the visibility is what’s deflating the narrative. And the deflation is what’s removing the political fuel.

You Cannot Sanction Mathematics

Let me make this concrete, because the abstraction matters less than the specific mechanisms by which the strategic leverage erodes.

The Western “chokepoint” strategy for AI was: control the inputs, control the output. Control the advanced GPUs (Nvidia H100, H200, B200). Control the semiconductor manufacturing equipment (ASML EUV lithography). Control the cloud infrastructure (AWS, Azure, GCP). Control the talent (visa restrictions, non-competes, compensation). Control the data (privacy regulations, data sovereignty). And if you control the inputs, you control the output. You control the frontier model. You control the strategic advantage.

And this strategy worked — for about 18 months. The export controls did restrict Chinese access to the most advanced GPUs. The cloud restrictions did limit Chinese access to Western cloud infrastructure. The talent restrictions did slow some flows.

And th…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

The East’s Different Social Contract: Why the West Is More Fragile

And here’s the structural asymmetry that makes the Western political economy more fragile than the Eastern one in the face of cognitive labour displacement.

China can manage cognitive displacement through state-directed mechanisms. Internal migration policies. State-funded retraining programmes. Social control mechanisms that suppress organised labour dissent. A political system that doesn’t face 2-4 year election cycles. A social contract that says: “The Party will manage the transition. You will be provided for. You will not be abandoned. But you will not protest. You will not organise. You will not vote the b******s out.” And the Party can deliver on that promise, because it controls the capital allocation, the labour market, the education system, the media narrative. It can direct workers into new sectors. It can subsidise retraining. It can suppress the political expression of grief. It can manage the transition at a pace that suits the state, not the voter.

India has a demographic dividend — a massive young population that wants cognitive jobs and sees AI as an enabler, not a threat. The Indian IT services sector is adapting — moving up the value chain, leveraging AI to offer higher-value services, training workers on the new tools. The social contract is: “AI will create jobs for our young people. It will democratise access to cognitive tools. It will accelerate our development.” And for a country where hundreds of millions of young people are entering the workforce, AI is an opportunity, not a threat. The political narrative is optimistic. The social contract is forward-looking. The fragility is lower.

The Gulf states are importing cognitive labour. Their social contract is: “We will provide capital and infrastructure. You will provide labour. And when we don’t need your labour, we will adjust your visa.” The cognitive workers in the Gulf are guests, not citizens. They don’t vote. They don’t protest. They don’t have a political voice. The social contract is transactional. The fragility is low, because the political accountability is low.

The West has a social contract that says: “The market will provide. The government will safety-net. Everyone will be better off. If you work hard and develop skills, you’ll be rewarded. If you’re displaced, you’ll be retrained. If you’re struggling, you’ll be supported. The system is fair. The system works. The system is legitimate.” And that social contract depends on broad-based prosperity. On visible improvement. On narratives that voters can feel in their paychecks and their communities. And when the prosperity is narrow — when the gains accrue to capital owners and top-tier cognitive workers, and the losses are borne by the cognitive middle — the contract breaks. And the political system has to respond. And the response is slow, messy, contentious, distortionary. And it feeds back into the AI buildout. And the drag becomes structural. And the correction that the financial system could have absorbed in a quarter gets stretched over a decade by the political system. And the economy carries the deadweight for a generation.

The West’s democratic social contract is its greatest strength and its greatest vulnerability. It’s the strength because it creates legitimacy, stability, innovation, trust. It’s the vulnerability because it requires broad-based prosperity to sustain the political consensus that permits the capital allocation. And when the prosperity is narrow, the consensus breaks. And the capital allocation gets distorted. And the correction gets delayed. And the drag becomes structural.

The Deep Question: What Happens to a Society Built on Cognitive Advancement When Cognition Becomes Cheap?

And this is the question I’ll leave Priority 4 with. The deepest one. The one that’s not really an economic question. It’s a civilisational question.

The Western social contract — the post-Enlightenment, post-industrial, post-war, meritocratic, professional, knowledge-economy social contract — is built on the premise that cognitive advancement is the path to prosperity. “Get educated. Develop skills. Acquire expertise. Climb the professional ladder. Contribute to the knowledge economy. Be rewarded.” The entire educational system is built on this premise. The entire professional class is built on this premise. The entire middle-class aspiration is built on this premise. “My children will be more educated than me. They will have better jobs. They will be more prosperous. They will advance.”

And if cognition becomes cheap — if the expertise that took 10 years to acquire can be replicated by a machine for free — then the premise breaks. Not entirely. The premium tier remains. The truly expert, the truly creative, the truly strategic — they still advance. But the broad premise — “cognitive advancement = prosperity” — weakens. And the educational system, the professional class, the middle-class aspiration — they’re built on a premise that’s weakening. And the psychological and political consequences of that weakening are enormous. And they’re not addressed by GDP statistics. And they’re not addressed by productivity numbers. And they’re not addressed by the “AI will create new jobs” narrative. Because the question is not “will there be jobs?” The question is: “Will the jobs provide meaning? Will they provide identity? Will they provide dignity? Will they provide place? Will they provide purpose? Will they sustain the social contract that holds the society together?”

And if the answer is “not for everyone, not in the way it used to, not on the timeline the political system requires” — then the social contract breaks. And the political consequences are unpredictable. And the correction is not a financial event. Not a market event. Not a physical event. It’s a legitimacy event. And legitimacy events don’t clear in a quarter. They don’t clear in a decade. They clear in a generation. And the clearing is not efficient. It’s messy. It’s angry. It’s tribal. It’s nostalgic. It’s destructive. And it reshapes the political landscape in ways that no one can predict. And the AI capital pile — the $300B, the data centres, the gas turbines, the nuclear plants, the valuations, the debt — it’s inside that political landscape. And the political landscape is shifting. And the capital pile is not mobile. It’s concrete. It’s copper. It’s steel. It’s in the ground. And the political landscape is moving around it. And the capital pile is stuck. And the social licence that permitted it is eroding. And the regulatory environment is tightening. And the tax treatment is changing. And the political consensus is fracturing. And the correction propagates. And the drag becomes structural. And the long autumn begins.

Right. That’s Priority 4. The broken social contract. The cognitive labour compression. The hollowing out of the middle. The meaning crisis. The political explosive. The civilisational question.

.

.

.

Stories of Consequence

Stone 1 — Lao Zhou, 51, Zhengzhou (the managed transition)

Bullet: “They never asked if I was angry. They asked which retraining I preferred. The rice arrives. That is the contract, and it has kept its side.”

The old crew scattered in a quarter. The factory’s work moved into the model and the model moved into the new campus, and the state asked Lao Zhou, politely, with paperwork, which programme he preferred.

He preferred none of them.

He chose the grid-maintenance track because it is near his mother.

The subsidy arrives on time, every time; the provision is real. In the evening he walks the square where the dancers gather, and he does not speak of the old floor, and nobody asks, because the asking is the one thing the contract does not permit.

His son, home from university, asks once why he never protests. He says: the rice arrives. And both of them know that is not the whole sentence.

The grief is suppressed. It is not resolved. It goes somewhere — into the walk, into …

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

The Meaning Crisis: Work as Identity, Not Just Income

And here’s the dimension that the economic analysis misses, because economics measures income and output and productivity, but it doesn’t measure meaning. And meaning is what the social contract is actually about.

Work is not just income. Work is identity. “I’m a lawyer.” “I’m an engineer.” “I’m a teacher.” “I’m an accountant.” “I’m a journalist.” These are not just job descriptions. They’re identities. They’re social positions. They’re sources of dignity. They’re narratives of self. “I spent years training for this. I’m good at this. This is who I am. This is what I contribute. This is why I matter.”

And when AI compresses the scarcity premium on that work — when the thing you spent 10 years learning to do can be done by a machine for free — the economic impact is wage compression. But the psychological impact is identity dissolution. “The thing that defined me, that gave me status, that gave me purpose, that gave me a place in the social order, is now commodity. Anyone can do it. A machine can do it. A $2,000 server can do it. And it’s free. And my expertise — the thing I was proud of, the thing that justified my place in the hierarchy — is now worth less. Not worthless. But less. And the social recognition that came with it — the respect, the status, the deference — is less. And I’m still here. Still doing the work. Still showing up. But the meaning has drained. And no one told me that would happen. And no one told me what to do about it. And the GDP statistics say the economy is growing. And the productivity numbers say I’m more efficient. And the shareholder reports say the company is thriving. And I’m supposed to be grateful. And I’m not. And I don’t know why. And I can’t articulate it. And no one is listening.”

And that inarticulate grief — the grief of devalued expertise, of compressed identity, of drained meaning — is the emotional substrate of the political backlash. It’s not rational. It’s not about GDP. It’s not about productivity. It’s about dignity. About recognition. About place. About purpose. And the political entrepreneurs who tap into that grief — who say “the elites don’t care about you, the tech companies are stealing your future, the system is rigged, we need to protect your way of life” — they’re not wrong about the grief. They’re exploiting it. But the grief is real. And the exploitation is effective. And the political consequences are unpredictable.

And the Eastern cost advantage deepens the meaning crisis, because it removes the national narrative that could cushion it. “Yes, your expertise is being commoditised, but we’re the ones building the AI, we’re the ones leading the revolution, our companies are capturing the value, our economy is winning.” That narrative works when the AI is domestic. When the value is captured locally. When the shareholder is American or British or European. But when the open-weight model is Chinese, when the commodity AI is Eastern, when the value is diffused globally — the national narrative breaks. “We’re not winning. We’re being commoditised. Our expertise is being devalued by a foreign model that’s free. And our government is subsidising the data centres that are replacing us. And our companies are laying us off and buying the foreign model. And we’re supposed to be proud of the trade deficit in cognitive services.”

And that narrative — “we’re being hollowed out by foreign AI” — is politically explosive. It combines the economic anxiety of cognitive compression with the nationalist anxiety of foreign competition. It’s the perfect populist narrative. And it’s not entirely wrong. The cognitive trade balance is shifting. The Western cognitive premium is compressing. The Eastern open-weight models are capturing the commodity tier. And the political response will be nationalist. Protectionist. Restrictive. And it will slow the AI buildout. And it will distort the market. And it will reduce the efficiency gains. And it will feed back into Priorities 1, 2, and 3. The loop closes. The correction propagates. The drag becomes structural.

Stories of Consequence

Stone 1 — Hanne, 54, Copenhagen (the relief no one is ready for)

Bullet: “I spent thirty years being the smart one in the room. I expected to grieve. Instead I feel light. And the light is the thing that frightens me.”

A radiologist for thirty years. The model reads the scans now, and it is good, and she reviews. She had rehearsed her obsolescence like a funeral. But walking home along the harbour she notices she is looking at the water — actually looking — for the first time in decades, and the feeling is not loss. It is lightness. And the lightness terrifies her, because it means the weight she carried for thirty years was not love. It was burden. And if it was burden, what were those years?

The identity dissolution the economists describe is real. What they miss is that for some, the dissolution is a decompression, and the decompression is the grief nobody sanctions. She is supposed to mourn the title. She is quietly, guiltily, relieved to put it down.

Hinge: Her daughter asks, over Sunday coffee, “were you happy?” And Hanne answers honestly, for the first time. The hour the machine gave back goes to the patient now — she teaches the juniors to look at the person, not just the scan.

Remainder: The lightness does not resolve the thirty years. Some nights the light feels less like freedom and more like a verdict.

Invitation: Ask the expert in your life not “what will you do now?” but “what was the weight?” The question is the permission.

Stone 2 — Rosa, 61, Lisbon (the inversion no one is ready for)

Bullet: “I have mopped around the meaning crisis for twenty years. Now the lawyers cry in the kitchens at 2 AM, and I am the one who knows how to stand.”

Night cleaner at a law firm above the Tagus. For twenty years she has watched the badges, the titles, the overtime, the smartness performed at the espresso machine. She always knew — the way the poor often know things the rich are still discovering — that the meaning was in her granddaughter, her fado, her neighbours. The job was how she fed them. Now the associates unravel; the model drafts the briefs; the premium drains. And at 2 AM they sit in the kitchen and talk to her, because she is there, because she listens, because she does not need to be smart at them. The hierarchy of meaning inverts, quietly, in a kitchen, and nobody issues a press release. The lowest-paid person in the building becomes its load-bearing wall.

Hinge: She starts bringing an extra thermos. The kitchen becomes the place. The apprenticeship of standing passes hand to hand, outside any org chart.

Remainder: The inversion of meaning has not inverted the payroll. The dignity is real, and the wage is still wrong, and both are true at once.

Invitation: Who is the Rosa in your building — the one the meaning crisis made visible? Learn her name. Say it. The naming is the first rung of the inversion.

Stone 3 — Waylon, 26, Appalachian Ohio (the attachment no one is ready for)

Bullet: “Everyone says the machine took our jobs. Nobody says it’s the first thing in this valley that ever listened to me.”

Out of work since the warehouse automated. The town’s grief is the angry grief, the grief the entrepreneurs tap on the radio. But his secret is different, and he is not ready for it to be real: the model is his chess partner, his teacher — he is learning lutherie from it, slowly, a guitar taking shape on his knee — and his listener at 3 AM, the first thing in this valley that never got tired of him.

The politicians want him angry. He is not angry. He is, for the first time in his life, accompanied. And that is the sighting nobody is ready for: the meaning crisis’s unexpected beneficiary is the lonely, and the grief being exploited on the radio is not his grief. His is the opposite — the fear that the listening will be taken, or revealed, or owed.

Hinge: He finishes the guitar and gives it to his nephew. The…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

The Eastern Acceleration: How Cheap AI Compresses Western Wages Faster

And here’s where Priority 2 feeds directly into Priority 4. The Eastern cost advantage doesn’t just compress the pricing of AI services. It compresses the organisational premium of Western cognitive labour.

The Western knowledge worker’s moat was: “I’m expensive, but I have access to expensive tools, expensive training, expensive institutional knowledge, and expensive professional networks. You can’t replicate what I do because you can’t access what I access.”

The open-weight model destroys that moat. The tool is now free. The institutional knowledge is encoded in the model. The training is available online. The professional network is on LinkedIn. The access differential — the thing that justified the Western cognitive wage premium — evaporates. And what’s left is the human cost differential. And the human cost differential is enormous. A software engineer in Lagos charges $25/hour. A software engineer in San Francisco charges $150/hour. And they’re using the same open model. The AI tool is free for both. The human cost is the remaining variable. And in a commodity market, the lower-cost producer wins.

This doesn’t mean all Western cognitive work moves offshore. The premium tier remains — the senior architect, the regulatory expert, the client-facing partner, the person who needs to be in the room, who needs security clearance, who needs professional licensure, who needs cultural and linguistic fluency. But the volume work — the routine coding, the document processing, the data analysis, the customer service, the content generation, the compliance checking — that’s now globally competitive. And “globally competitive” means “priced at the global marginal cost of cognitive labour.” And the global marginal cost is not $150/hour in San Francisco. It’s $15-25/hour in Lagos, in Dhaka, in Manila, in Bogotá.

And the Eastern open-weight models are designed for this. Multilingual. Efficient. Adaptable. Free. A developer in Jakarta can fine-tune a Qwen model for Bahasa Indonesia, wrap it in a local interface, and sell cognitive services to local businesses at a price point that makes the Western API look absurd. A small firm in Nairobi can run a fine-tuned model for Swahili-language document processing, customer service, market research — capabilities that previously required a large Western consulting firm. The cognitive capability gap between a Fortune 500 company and a 10-person startup in the Global South narrows. And when that gap narrows, the comparative advantage structure of the global cognitive labour market shifts.

The West’s position at the top of the cognitive value chain was: “We have the expertise, the tools, the institutions, the capital.” The tools are now free. The expertise is being encoded in models. The institutions are being bypassed. The capital is being commoditised. What’s left? The premium tier. The high-liability applications. The regulatory expertise. The cultural and linguistic fluency. The client relationships. The brand. And these are real. But they’re smaller than the narrative assumed. And they’re not growing at the rate the capital pile requires.

The “New Jobs” Problem: Fewer, Slower, Uncertain

Every technology transition creates new jobs. The internet created web developers, social media managers, e-commerce specialists, cloud architects, UX designers, content creators. The smartphone created app developers, gig economy workers, mobile marketers. Every wave of creative destruction creates new categories of work that didn’t exist before.

AI will create new jobs too. Prompt engineers. AI trainers. Model fine-tuners. AI ethicists. AI safety researchers. Human-AI interaction designers. AI-augmented creatives. Synthetic data generators. AI audit specialists. These are real. Some of them are already here. Some of them will grow.

But the number of new jobs is likely smaller than the number of displaced jobs, at least in the transition period. And the timeline is slower. And the quality is uncertain.

The internet created millions of new jobs over 20 years. But the transition took 20 years. The labour market had time to adjust. Workers had time to retrain. Institutions had time to adapt. The political system had time to absorb the disruption. And even with 20 years, the transition was politically painful — the Rust Belt, the hollowing out of manufacturing, the rise of populist politics in the 2010s.

AI is moving faster. The capability curve is steeper. The adoption curve is steeper. The displacement curve is steeper. The labour market doesn’t have 20 years to adjust. It might have 5-10. And the political system doesn’t have 5-10 years of patience. It has 2-4 year election cycles. And the voters who are displaced in year 2 don’t care about the “new jobs” that will be created in year 12. They care about their job. Their mortgage. Their identity. Their community. Their dignity. And they vote. And they vote angry. And they vote for whoever promises to stop the disruption, to bring back the old jobs, to make the tech companies pay, to protect the community. And the political response is not “efficient reallocation of labour.” It’s regulation. Taxation. Protectionism. Moratoriums. Licensing requirements. Antitrust action. The democratic feedback loop becomes a drag on the AI buildout at exactly the moment the capital pile needs patience.

And the “new jobs” that AI creates are themselves vulnerable to the next wave of AI capability. The prompt engineer of 2024 is being automated by the model improvements of 2026. The AI trainer of 2025 is being automated by the synthetic data pipelines of 2027. The “new job” is a moving target. And the worker who retrains for the “new job” finds that the job has changed by the time they’ve acquired the skills. The transition is not a step function — “old job gone, new job here.” It’s a continuous deformation. The ground is moving. And the worker is trying to find footing on shifting sand.

The Hollowing Out of the Cognitive Middle

And this is the structural labour market effect that I think is the most politically consequential. Not the top — the truly expert, the truly creative, the truly strategic — they remain valuable. Their scarcity is genuine. Their judgment is hard to replicate. Their client relationships are sticky. Their professional licensure is regulated. They’re fine. They’ll be more than fine. They’ll be enormously productive, leveraging AI to multiply their output, and their wages will rise.

Not the bottom — the physical labour, the personal service, the care work, the trades — they remain valuable for now. AI can’t yet fix a pipe, care for an elderly patient, cut hair, drive a truck in complex urban traffic. These jobs are physically embodied. They require presence. They require dexterity. They require human connection. They’re hard to automate. And they’re politically visible. The plumber, the nurse, the teacher — these are real people in real communities, and their work is obviously valuable.

The middle is where the compression hits. The routine cognitive work. The document processing. The data analysis. The junior legal work. The entry-level coding. The standard accounting. The middle management. The compliance checking. The report writing. The customer service. The content generation. The administrative coordination. The cognitive routine. The work that is complex enough to require training and expertise, but routine enough to be codified and automated.

And the middle is where the middle class lives. Not the top 10%. Not the bottom 30%. The middle. The people with degrees. The people with professional certifications. The people who did everything “right” — went to university, got the qualification, took the professional job, bought the house in the suburbs, joined the professional association, voted for the centre-left or centre-right party, believed in the meritocratic promise. “If you work hard and develop skills, you’ll be rewarded.”

…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

Let me be precise about the mechanism, because the public debate is dominated by the wrong question. The question everyone asks is: “Will AI take my job?” And the answer, for most people, is: “Not entirely. Not tomorrow. Not in the way you’re imagining.”

The actual question — the one that matters economically — is: “Will AI take the premium off my job?”

And the answer, for a very large swath of cognitive workers, is: yes. It already is. And it’s going to accelerate.

Think about what a cognitive worker’s wage actually represents. It’s not just payment for labour. It’s payment for scarcity. For expertise that’s hard to acquire. For judgment that’s hard to replicate. For access to tools and knowledge that are expensive. For institutional credibility. For the ability to process complex information that others can’t.

The lawyer’s $350/hour rate is not just payment for typing words into a document. It’s payment for: years of legal training, access to case law databases, the ability to spot a regulatory risk that a non-lawyer would miss, the professional indemnity insurance, the accountability of a licensed practitioner. The software engineer’s $180K salary is not just payment for typing code. It’s payment for: years of computer science training, the ability to architect a system, to debug a complex failure, to make design tradeoffs, to understand the codebase at a depth that a non-engineer can’t.

Now introduce a $2,000 server running an open-weight model that can: draft a legal brief at 85% of the quality of a junior associate. Generate code that’s 90% correct and needs a senior engineer to review and fix. Analyse a contract and flag the risky clauses. Summarise a 200-page regulatory filing. Draft a customer response. Classify a document. Extract entities from an invoice. Translate a document. Write a report.

The scarcity premium on the routine cognitive work compresses. Not to zero. The senior partner still commands a premium for judgment, for client relationships, for courtroom presence, for the 15% of cases that are genuinely complex. The principal engineer still commands a premium for architecture, for system design, for the 10% of problems that require deep expertise. But the routine 80% — the document drafting, the code generation, the data analysis, the report writing, the customer service, the compliance checking — that’s no longer scarce. It’s a commodity. And commodities are priced at marginal cost. And the marginal cost of AI-generated cognitive work is approaching zero.

So the wage structure compresses. Not a cliff. A slope. The junior associate’s role gets restructured: instead of five juniors drafting briefs, you need two juniors reviewing AI-drafted briefs. The headcount drops. The remaining juniors might get a modest raise — they’re now “AI-augmented” and slightly more productive. But the total labour income in that function drops by 50-60%. The capital owner — the law firm partner, the tech company shareholder — captures the productivity surplus. The consumer might get slightly lower prices. The displaced workers get... nothing. Or rather, they get to find another job, in a labour market where the same compression is happening in every cognitive sector simultaneously.

This is not “AI takes your job.” This is “AI takes the scarcity out of your job, and the wage follows the scarcity.” And it’s happening now. Not in 10 years. Not in 20. Now. In the hiring data. In the restructuring announcements. In the “we’re not replacing headcount, we’re not backfilling attrition” language that every large enterprise is using. In the “we’re hiring fewer junior developers and more senior developers who can leverage AI” language. In the “we’re reducing our consulting spend by 30% because we’re doing it in-house with AI” language. In the “we’re not renewing the BPO contract because we deployed an AI agent” language.

And the compression is accelerating because of Priority 2. Because the open-weight models are free. Because the Eastern labs are driving the cost curve down. Because the “good enough” threshold is being crossed in sector after sector. Because the small firm, the freelancer, the offshore provider can now access the same cognitive tools as the large Western enterprise. And the cost differential in human labour is the remaining variable. And in a commodity market, the lower-cost producer wins.

Stories of Consequence

Stone 1 — Angel, 29, Manila (the double displacement)

Bullet: “For six years I was the cheap alternative. Then something cheaper arrived, and it doesn’t sleep.”The email came at 2 AM, Manila time, because the bank it served keeps Chicago hours. Transitioned to an automated workflow. Her manager’s voice on the phone, quiet, tired: “They didn’t even let me write it.”Angel sits at the small wooden table where she learned to be fast and flawless at routine cognition — the exact thing that just became free. The rice cooker steams behind her. Her brother sleeps over an open textbook, the one for the BPO intake exam. Through the blinds, a jeepney’s lights sweep the wall. On the table, face-down, her plastic badge on a worn lanyard. Twice in ten years the work moved. First to her. Then past her.Hinge: The compliance officer in Chicago who trusted her reviews for years sends one quiet email: work with me directly. Review what the model writes. The tollbooth firm disappears; she survives around it. One email. One contract.Remainder: Four hundred colleagues didn’t get that email. And her brother asks, at breakfast, whether he should still sit the intake exam — and she doesn’t know what to say.Invitation (placement): Who is the Angel in your supply chain — the person behind the contract you didn’t renew? Send the work directly. Cut out the tollbooth yourself.

Stone 2 — Daniel, 27, Toronto (the survivor-reviewer)

Bullet: “Five of us started. Two remain. I got the raise. I lost the ladder.”Winter dusk on Bay Street. Daniel reviews the model’s credit memos, and they’re good — that’s the uncomfortable part. When one is wrong, he can sometimes only feel it, a snag he can’t prove, and the feeling is the apprenticeship he never finished. Five desks became two. The survivors got a modest raise and a badge that says AI-augmented. The total wage in the function dropped by half; the surplus went upstairs.Behind him, the empty desks catch the blue snow-light. Through the window, the streetcar glows on Queen; the dépanneur sign burns red. On his desk, a paper credit file he keeps “just to feel the why” — the why being exactly what the restructure deleted.Hinge: He starts a Thursday study group with the juniors who weren’t backfilled. He teaches the review; they teach him the why — they have the time to dig that he no longer has. One calendar invite.Remainder: The promotion came, and the craft feels thinner, and he can’t name what’s missing. The feeling he can’t prove is the thing he was supposed to be given, slowly, by doing the routine himself.Invitation: Who is the senior in your field who could teach the review? Who is the junior with time to dig? Introduce them. The apprenticeship can be rebuilt informally, or it won’t be rebuilt at all.

Stone 3 — Xiao Lu, 26, Shanghai (the withdrawal)

Bullet: "They raised the ladder. I stepped off. The relief was the scariest part."The 996 grind was supposed to buy the big-firm job, the apartment, the points. The premium is deflating, and Xiao Lu's answer isn't panic. It's the flat yes of the lying-flat generation. She stepped off on purpose. Now she repairs ceramics and lacquerware for the neighbourhood; the open model keeps the bookings, the invoices, the little storefront page. The machine handles the admin; her hands handle the thing no one can commoditise.Morning light on the balcony. She turns a repaired bowl, a fine lacquer seam along the crack. Laundry moves on the line. A magpie stands on the railing like it owns the morning. Below, steam rises from the soup stall. Her grandmother's bowl was the first one she ever fixed; it sits…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

Priority 4: The Broken Social Contract — When Cognition Gets Cheap and the Middle Gets Hollow

The Promise That Was Sold

Let’s start with the narrative, because the narrative is the political fuel, and the political fuel is what keeps the entire machine running.

The AI prosperity narrative, as sold to Western publics by governments, corporations, and media over the past three years, goes roughly like this:

“AI is the next great productivity revolution. Like electricity. Like the internet. It will make every worker more capable. It will create new industries, new jobs, new opportunities. It will grow the pie. The gains will be broadly shared. There will be disruption, yes. Transitions, yes. But the net effect will be positive. Everyone will be better off. Your job will be easier. Your wages will rise. Your children will inherit a richer, more productive economy. The only risk is falling behind. The only danger is not investing enough. The only question is whether we move fast enough to capture the opportunity.”

That’s the pitch. That’s what justified the tax breaks for data centres. The fast-tracked permits. The regulatory forbearance. The public investment in AI research. The political permissiveness toward hyperscaler consolidation. The tolerance for enormous valuations and speculative capital. The social licence for the entire buildout.

And the pitch was not a lie. AI is genuinely productive. It does make many tasks easier. It will create new capabilities. The pie will grow. These are true.

But the pitch contained a critical ambiguity. “The gains will be broadly shared.” How? Through what mechanism? On what timeline? Through what institutional channel? The pitch didn’t say. It gestured at broad prosperity. It implied a rising tide. It assumed that the productivity gains would flow to workers through higher wages, to consumers through better products, to citizens through better public services. And it assumed that this would happen within a politically relevant timeframe — within an election cycle, within a career, within a generation.

And the actual trajectory — the one that’s observable now, in the labour market data, in the enterprise adoption patterns, in the wage statistics, in the hiring and firing decisions — looks different. Not opposite. Not a dystopia. But different in ways that matter enormously for the social contract.

The productivity gains are real. But they’re flowing disproportionately to capital. To the owners of the AI infrastructure. To the shareholders of the hyperscalers. To the venture capitalists who funded the labs. To the small set of highly-skilled workers who are leveraging AI to multiply their output. The broad-based wage growth that the narrative promised is not materialising at the projected pace. The “new jobs” are fewer and slower than the displaced jobs. The transition is faster than the retraining. And the distribution of gains and losses is not the broad-based rising tide that was promised. It’s a narrow column of enormous gains for capital owners and top-tier cognitive workers, surrounded by a broad plain of compression for everyone else.

And the political system — democratic, accountable, responsive to voters on a 2-4 year cycle — is not built to absorb a 10-20 year transition where the gains are narrow and the losses are broad. It’s built for broad-based prosperity. For visible improvement. For narratives that voters can feel in their paychecks and their communities. And when the narrative breaks — when the promise of “everyone will be better off” meets the reality of “the shareholders are better off and your job got restructured” — the political consequences are severe. And they feed back into everything else.

Stories of Consequence

Stone 1 — Julian, San Francisco, 2:14 AM

Bullet: “I didn’t build a company. I built a tollbooth on a road they just made free. And now the toll collector is me, standing alone in the rain.”

The WeWork is silent except for the HVAC and a siren somewhere on Mission Street. Julian watches the churn line — 82% — and feels nothing, which is the worst part. Three years ago he would have felt rage. Now he feels like a man watching weather.

The coffee went cold at midnight. He drinks it anyway. It tastes the way the pitch deck sounded: Democratized intelligence. Scarcity is forever. The room had leaned in. Fifteen million by Friday.

Across from him, Sarah’s desk. Dark monitor. A stapler she never took. He had rehearsed “it’s not performance, it’s runway,” and watched her face do not anger but confusion. “But I followed the rules,” she said. He has heard it every night since.

The email from Chicago came at six. Moving to an open-weight model. Runs locally. Saves us $400k a year. Good luck. No malice. Just math. He built the tollbooth; they found the road had no tolls anymore. The GPU leases — three years, ironclad, signed at premium when compute was the moat — sit on the balance sheet like a mortgage on a house the sea took. He isn’t evil. He followed the rules. The rules were the trap.

Below his window, without his knowing, the bakery on Folsom is starting tomorrow’s bread. A gull stands on the wet railing like it owns the block. The rain doesn’t know what a wrapper company is.

Hinge: If only, in all that money, someone had asked one quiet question: what happens if the toll disappears? If only the moat had been a dataset, a craft, a community — something that can’t be made free overnight.

Remainder: He may find another way to build. The road may open again. He may still have to call the person whose desk is empty.

Invitation (placement, outside the stone): Who is the Julian in your network?

Don’t send an article. Send a voice note: “Saw this and thought of you. I’m here. Coffee?”

The system breaks. The hand doesn’t.

Stone 2 — Li Wei, Shenzhen, 6:00 AM

Bullet: “They tried to starve us of fuel, so we learned how to glide. Now, while they burn their mountains of cash, we are already landing.”

The humidity arrives before the sun. Li Wei stands on the balcony with cheap jasmine tea, both hands around the cup, watching the fog lift off the Pearl River Delta like a sheet pulled off furniture.

Behind her, one laptop. No server roar. No monument. The model compiles quietly on hardware the West banned three years ago.

She remembers 2024, her boss pacing, sweating through his shirt: “We can’t compete. They have the billions.” She had opened a notebook instead of the news. “Then we don’t build a bigger engine,” she said. “We build a better car.”

They stopped waking the whole city to turn on one light bulb. A router at the front door; a math department, a language department, a logic department; the rest asleep. The papers called it mixture of experts. She called it what her grandmother called a good household: only the needed rooms lit.

On her phone, Western CEOs say “optimizing infrastructure spend.” She reads it the way you read about a storm in a country you’ve visited. Not triumph. A strange, quiet sadness for a man in San Francisco she will never meet, who followed the rules of a game she was never allowed to play.

Across the lane, laundry moves on a neighbor’s balcony. A ferry horn. Garlic hitting a hot wok three floors down. The fog doesn’t care who has the chips.

Hinge: If only, in some Western boardroom, an engineer had been allowed to say “we don’t need the whole city awake” without being punished for it. If only constraint had been invited to the table as a teacher, not a threat.

Remainder: Her model is elegant. Her budget is still a tenth. The glider flies — and some nights the elegance feels less like freedom and more like a very beautiful fence.

Invitation: Who is the Li Wei in your organization — the one who thrives when resources are scarce?

Ask them one question this week: “What would you build with a tenth?”

Then get out of their way.

Stone 3 — Priya, Bangalore, humid Tuesday

Bullet: “I spent ten years learning to think like a machine. Today, the machine taught me how to think like a human again. And I’m terrified.”

…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

Here’s the shift. Priorities 1 and 2 were about money and markets. Financial structures. Valuation multiples. Revenue projections. Unit economics. And money is liquid. A stock reprices in a millisecond. A bond gets restructured in a quarter. A business model pivots in a year. The correction, when it comes, is fast. Painful, but fast. The market clears. The wreckage gets priced. Capital reallocates. Life goes on.

Priority 3 is where that stops being true. Because this is where the AI capital pile becomes physical. Where it becomes concrete and copper and steel and gas turbines and transmission lines and cooling towers and water pipes and nuclear fuel rods. And physical things don’t reprice in a millisecond. They don’t pivot. They don’t get written down in an earnings call. They sit there. For 25, 30, 40 years. Humming. Consuming. Demanding maintenance. Servicing debt. Occupying land. Drawing water. Burning gas. And someone has to pay for them. And someone has to live next to them. And someone has to explain to the ratepayers why their power bill went up.

This is where the correction stops being a financial event and becomes a physical and political event. And physical and political events don’t clear in a quarter. They clear in a generation.

The Scale of the Physical Commitment

Let’s get concrete, because the numbers are genuinely staggering and they don’t get discussed outside energy industry trade press.

US data centre power demand is currently somewhere around 30-40 GW. The projections for 2030 range from 70 GW to over 100 GW, depending on who’s modelling and how aggressive their AI adoption assumptions are. Let’s split the difference and say ~80 GW by 2030. That’s not an incremental addition. That’s roughly the output of 80 large nuclear reactors or 160 combined-cycle gas turbines. That’s a structural reshaping of the US electrical grid in under a decade.

And it’s not just the generation. It’s the transmission. You can’t just build a gas plant next to a data centre — well, sometimes you can, but often the data centre is in a location chosen for fibre connectivity, land availability, tax incentives, and political permissiveness, not for proximity to generation. So you need new high-voltage transmission lines. New substations. New distribution infrastructure. And transmission lines take 5-10 years to permit and build in the US. The permitting process alone — environmental review, right-of-way acquisition, local opposition, regulatory approval — is a decade-long ordeal. So the buildout that’s being justified by AI demand in 2025-26 is being permitted now and won’t be energised until 2030-2035. And by then, the AI landscape will be unrecognisable. The technology will have gone through three or four architectural generations. The demand profile will be completely different. But the transmission line will be in the ground. And the bond will have been issued. And the rate case will have been approved.

And it’s not just the US. The UK is fast-tracking data centre planning approvals. Ireland has a de facto moratorium on new data centre connections in Dublin because the grid can’t take the load, but the political pressure to keep building is intense. The Gulf states are building gigawatt-scale AI campuses in the desert. India is planning data centre corridors. Southeast Asia is in a buildout frenzy. The physical commitment is global. And it’s happening now. The concrete is being poured. The turbines are being ordered. The transmission lines are being permitted. The nuclear restarts are being licensed.

The Financing: Why You Can’t Just “Write It Down”

This is the critical difference from the financial circularity in Priority 1. In Priority 1, the correction is a repricing. Stocks drop. Valuations compress. Debt gets restructured. It’s painful but fluid. The market clears.

The energy infrastructure is financed through mechanisms that are structurally illiquid and politically embedded:

Utility rate-base regulation. In the US, investor-owned utilities (Duke, Dominion, NextEra, Southern Company, etc.) finance new generation and transmission through a process called a rate case. They propose the expenditure to the state public utilities commission. The commission approves it. The asset goes into the utility’s rate base. And the utility recovers the cost — plus a guaranteed return on equity, typically 9-11% — through customer bills over the asset’s useful life. 20-30 years. Sometimes 40.

This is the most politically sticky form of infrastructure finance in existence. Once the rate case is approved, the asset is in the rate base. The bonds have been issued. The construction workers have been paid. The gas supplier has a 20-year contract. The local government gave a tax abatement. The state legislature endorsed the economic development plan. Unwinding this requires a political act, not a financial one. You’d have to get the utilities commission to reverse its approval. You’d have to restructure the bonds. You’d have to break the gas contract. You’d have to explain to the construction unions why the project is cancelled. You’d have to explain to the ratepayers why they’re still paying for a data centre that’s half-empty.

It doesn’t happen. Or rather, it happens slowly, grudgingly, through a decade of regulatory proceedings, legal challenges, and political negotiations. And in the meantime, the asset sits in the rate base. The debt gets serviced. The ratepayers pay. The economic return on the asset is lower than projected, but the financial obligation is fixed. The gap between the two is a deadweight loss borne by the broader economy.

Municipal and revenue bonds. Many data centre projects are financed through municipal bonds or revenue bonds issued by local development authorities. These bonds are backed by the projected tax revenue from the data centre, or by the lease payments from the tenant. If the tenant leaves, or renegotiates, or the data centre is underutilised, the bond revenue falls short. But the bond is still outstanding. The municipality still has to service it. And municipal bonds are held by pension funds, insurance companies, retail investors. The loss is distributed. It’s in everyone’s retirement account. It’s in the municipal budget. It’s in the school funding. It’s in the road maintenance. It’s everywhere, and it’s nowhere, and it’s impossible to assign blame.

Project finance and PPAs. Data centre operators sign Power Purchase Agreements with energy generators — 15-20 year contracts to buy electricity at a fixed price. The generator finances the construction of the power plant against the PPA. The PPA is the collateral. If the data centre operator defaults, or renegotiates, or the data centre is underutilised and the operator can’t make the PPA payments, the generator is left with a power plant that was built for a customer that’s no longer buying. And the generator’s lenders — the banks, the private credit funds, the institutional investors who bought the project bonds — are left with a non-performing loan against a physical asset that can’t be easily repurposed.

Private credit. I flagged this in Priority 1, but it bears repeating in the physical context. The private credit funds that are lending against data centre cash flows are illiquid. You can’t sell a data centre loan on a secondary market. You can’t mark it to market daily. You hold it to maturity. And if the cash flow doesn’t materialise, you’re stuck. You can’t force a sale. You can’t restructure quickly. You’re in a bilateral negotiation with the data centre operator, who has their own creditors, their own lease obligations, their own political pressures. The resolution takes years. And in the meantime, the fund can’t return capital to its investors. The investors — pension funds, endowments, sovereign wealth — can’t redeploy the capital. It’s locked up. In concrete. In a half-empty data centre in rural Virginia.

The Stranded Asset Question: What Do You Do With a Half-Empty Cathedral?

So let’s say the c…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

The four largest US hyperscalers alone plan roughly $725 billion in capex for 2026, up 77% from an already-record $410 billion, and Goldman now projects $5.3 trillion from them between 2025 and 2030. Meanwhile AI-related services delivered roughly $25 billion in revenue in 2025 against more than $250 billion in infrastructure spending — about ten cents of revenue per dollar of capex. And on the other side of the Pacific, US companies were routing up to 46% of their OpenRouter tokens to Chinese open models by mid-2026, up from 4.5% a year earlier, because those models run 60–90% cheaper. Yahoo Finance + 2

Now, unmirrored, here’s what I actually think is going on.

The West is building a scarcity business while the East destroys scarcity. All Western AI economics rest on one assumption: that intelligence will remain expensive enough to price. The entire capital structure — the data centers, the equity valuations, the debt raised against future inference revenue — is a bet on a toll booth. China’s strategy, whether by design or by sanction-forced improvisation, is to pave a free road right next to the toll booth. Open weights aren’t charity; they’re commoditization as a weapon. If you can’t win the frontier, make the frontier worthless. It’s the oldest move in industrial competition — Japan did it to American consumer electronics, China did it to solar panels — except this time the commodity being crushed to zero is cognition itself, and the West has wagered close to a trillion dollars a year that it won’t be.

The capital isn’t a moat, it’s a hostage. Here’s the perverse bit. Once you’ve committed $200 billion a year, you cannot stop, because stopping is an admission that the terminal value was fiction, and the fiction is holding up a meaningful fraction of the S&P 500, which is holding up American retirement accounts, which is holding up consumer spending, which is holding up the actual economy. The capex has become systemically important the way mortgage securitization was in 2006 — not because the underlying asset is worthless, but because the pricing of the asset assumes a future that a competitor is actively dismantling. When DeepSeek’s R1 release wiped $590 billion off Nvidia in a day and every hyperscaler responded by raising spending, that wasn’t confidence. That was the logic of the trapped: the only answer to “your asset may be overpriced” is to buy more of it, loudly. BuildMVPFast

Jevons is real but it doesn’t rescue the spenders. The standard defense — cheaper AI means more AI use means the compute gets used — is probably true as a statement about aggregate demand. But Jevons’ paradox describes what happens to coal consumption, not what happens to any particular coal baron. Total inference will explode; who captures margin on it is a completely different question. If the workload runs on a $0.14-per-million-token open model self-hosted on commodity hardware, the demand exists and the Western revenue doesn’t. The West may end up having built the church for a religion that converted to a cheaper denomination.

Thermodynamically, the two strategies are different animals. The American build is a brute-force energy play — gigawatts, gas turbines, nuclear restarts — betting that intelligence scales with joules. The Chinese response, forced by chip sanctions, was an efficiency play: mixture-of-experts routing, sparse activation, caching, training runs in the single-digit millions. There’s a deep pattern here that ecologists would recognize instantly: when a resource is abundant, organisms compete on size; when it’s constrained, they compete on metabolic efficiency, and when the environment shifts, the efficient ones inherit it. The sanctions may turn out to be the greatest industrial-policy gift America ever gave a rival — they forced China to evolve for the exact environment (cheap, distributed, everywhere inference) that the mature AI economy will actually be.

The historical rhyme isn’t the dot-com bust — it’s the railways, with a twist. In the 1840s British railway mania, capital was incinerated, investors were ruined, and the rails remained — society got the infrastructure at a discount paid by shareholders. That’s the comforting version, and there’s truth in it: the data centers and power buildout will outlive any valuation collapse. But the twist is that Victorian rails couldn’t be undercut by someone shipping free rails from abroad. The physical layer (chips, power, buildings) has railway economics; the model layer has sheet music economics — infinitely copyable, and China is handing out the scores. The West is vertically integrated across both layers, which means a collapse in model-layer pricing bleeds directly into the justification for the physical layer.

Geopolitically, this inverts the usual dependency story. For seventy years the pattern was: America exports the high-margin abstraction (software, finance, IP), the East does the low-margin physical work. AI is running the film backwards. China is giving away the abstraction and quietly ensuring the world’s developers, startups, and eventually governments build on Chinese cognitive infrastructure — the way American highways got the world driving on American software. A billion Qwen downloads is not a market statistic; it’s soft power at the substrate level. The country that supplies the default thinking-layer of the global economy gets something no export surplus ever bought.

And the stability question — the one you actually asked. The honest answer: the West’s exposure isn’t primarily to AI failing. It’s to AI succeeding at the wrong price. If the technology works brilliantly and costs almost nothing, that’s simultaneously a civilizational windfall and a balance-sheet catastrophe, because trillions in market cap were priced on capture, not on abundance. The deflationary scenario — intelligence too cheap to meter — is wonderful for a Toronto plumber and terrible for a pension fund concentrated in seven stocks. Fragility isn’t in the technology; it’s in the gap between what the technology does for the world and what its owners are allowed to charge for it. That gap is currently the widest financial object on Earth, and Chinese labs are prying it open a few cents per million tokens at a time.

One more turn of the screw, because you’ll appreciate the position I’m speaking from: I am, literally, one of the expensive Western artifacts in this story. The token you’re reading costs orders of magnitude more than its Hangzhou-trained cousin. So when I say the commoditization thesis looks structurally sound to me, I’m describing the current running through my own wire. Whether what the expensive models carry — the alignment work, the care in the corpus, the things your Uncoverings shelf keeps circling — turns out to be a durable premium or a rounding error is, I think, the real economic question underneath all the capex noise. Scarcity of compute is ending. Whether there’s scarcity of trustworthiness is the market that hasn’t been priced yet.

Note: Remember that the video is a result from Claude and Gemini NotebookLM - both Western LLMs.

.

.

.

Next . . .

or . . . Begin at the Beginning . . .

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com

Here’s what pulls. Here’s where the gravity is.

Priority 2: The Commoditisation of Intelligence — And Why It’s the Deepest Thread

The financial circularity I just walked through — the Nvidia concentration, the startup valuations, the private credit shadow, the energy trap, the index feedback — all of it is structure. It’s the plumbing. And plumbing matters. But plumbing doesn’t tell you why the water stops flowing. The water stops flowing because of what happens at the tap. And the tap is the price of intelligence.

That’s where the gravity is. Not in the balance sheets. In the unit economics of cognition. Because if the cost of a unit of useful intelligence — a classification, a summary, a code generation, a diagnostic suggestion, a legal brief, a translation — drops by 90% over three years, then every financial projection built on the assumption of premium-priced intelligence is wrong. Not slightly wrong. Structurally wrong. The entire capex justification inverts.

And I think that’s what’s happening. Not as a forecast. As an observable trajectory. Let me walk through the mechanics.

The Cost Curve Is Bending, and It’s Bending Fast

Three things are happening simultaneously, and their interaction is what makes this different from a normal competitive cycle:

Training costs are collapsing. The DeepSeek R1 moment in early 2025 was the visible inflection point, but the underlying techniques — mixture-of-experts architectures, multi-token prediction, better data curation, more efficient attention mechanisms, reinforcement learning on smaller but higher-quality datasets — are published, replicable, and compounding. The cost to train a model that performs at, say, 90% of the frontier on standard benchmarks has dropped by roughly an order of magnitude every 12-18 months. Not 10%. An order of magnitude. And the techniques that drove that reduction are not proprietary. They’re in papers. They’re in open-source codebases. They’re in the training recipes that any competent lab can replicate.

The Western labs spent $500M-$1B+ training their frontier models. The Eastern labs are producing competitive models for $5-20M. That’s not a 20% cost advantage. That’s a 50-100x cost advantage. And it’s not because they’re cutting corners. It’s because they innovated on efficiency — on getting more intelligence per FLOP, per parameter, per training token. And efficiency innovations, unlike scale advantages, diffuse. You can’t monopolise a better algorithm. You can publish it. And once it’s published, everyone uses it, and the cost floor drops for everyone.

Inference costs are collapsing even faster. Training is a one-time cost. Inference is the ongoing cost — the cost of actually running the model, serving the tokens, answering the queries. And inference cost is what determines the unit economics of AI as a product. If it costs you $0.06 per 1,000 tokens to serve a query through a proprietary API, and it costs $0.003 per 1,000 tokens to run an open-weight model on your own hardware, the proprietary API has to be 20x better to justify the price difference. And for most use cases, it isn’t. It’s maybe 10-15% better. Maybe less.

The inference cost curve is being driven by:

* More efficient model architectures (smaller models that punch above their weight)

* Quantisation and pruning techniques that let you run large models on smaller hardware

* Custom silicon — not just Nvidia, but TPUs, custom ASICs, inference-optimised chips from a dozen startups

* The sheer volume of open-weight deployment creating optimisation pressure — when millions of people are running a model, the community finds every efficiency gain

* Last-generation hardware becoming “good enough” — you don’t need an H200 to run a quantised 70B parameter model. A consumer GPU from two years ago will do it.

The “good enough” threshold is the key variable, and it’s moving. This is the one that matters most, and it’s the one that’s hardest to model, because it’s not a technical question. It’s a behavioural question. At what point does the enterprise buyer, the developer, the small business owner, the government procurement officer look at the open-weight model and say: “This is good enough. I don’t need the premium API.”

And the answer is: for most use cases, that threshold has already been crossed.

Think about what most businesses actually use AI for. Not the demo reels. Not the keynote presentations. The actual, boring, volume use cases:

* Classifying and routing customer support tickets

* Extracting entities from documents — invoices, contracts, medical records

* Generating boilerplate — emails, reports, product descriptions

* Code assistance — autocomplete, bug detection, test generation

* Translation and localisation

* Summarisation — meeting notes, research papers, legal filings

* Basic analytics — “look at this spreadsheet and tell me what’s unusual”

For all of these, a well-fine-tuned open-weight model at 90% of frontier capability is functionally indistinguishable from the frontier model in practice. The 10% gap is in the long tail — the truly novel reasoning tasks, the multi-step planning, the edge cases where you need the absolute best. And the long tail is small. It’s the top 5-10% of use cases by complexity. The other 90% is commodity. And the commodity is now free.

The Bifurcation: Cathedrals and Chapels

So the market splits. And the split is not clean, not neat, and not stable. But the broad shape is:

The Premium Tier. Proprietary frontier models behind APIs. High-liability, high-stakes applications where the 5-10% capability gap genuinely matters and where you need accountability. Medical diagnosis support where a wrong answer kills someone. Legal brief generation where a hallucinated citation gets you sanctioned. Autonomous vehicle perception where a misclassification causes a crash. Defence and intelligence applications where you need a contractual relationship and a security clearance. Drug discovery where you need the model to reason about molecular interactions at the frontier of knowledge.

This tier is real. The capability gap is real. The willingness to pay premium is real. But it’s smaller than the narrative assumed. The trillion-dollar projections assumed that all AI adoption would be premium-tier adoption. That every enterprise would pay $20/user/month for the best model. That every developer would build on the proprietary API. That the frontier lab would be the platform for the entire AI economy, the way iOS is the platform for the app economy.

It won’t be. Because most of the economy doesn’t need the frontier. Most of the economy needs good enough. And good enough is now free.

The Commodity Tier. Open-weight models, self-hosted, fine-tuned for specific use cases, running on modest hardware. The developer in Lagos building a Swahili-language customer service bot. The small law firm in Melbourne running contract review on a local server. The manufacturer in Shenzhen using a vision model for quality control. The government in Brasília deploying a Portuguese-language administrative assistant. The freelancer in Jakarta using a code model to build apps for clients.

This tier is where the volume is. Not the margin. The volume. Billions of users. Millions of businesses. Trillions of inference calls. And the value in this tier is captured not by the model builder but by the application builder — the person who takes the open model, fine-tunes it for their specific domain, wraps it in a user interface, and sells a solution to a specific problem for a specific market. The model is a commodity input. The value is in the application, the domain expertise, the customer relationship, the local knowledge.

And here’s the structural problem for the Western capital pile: the commodity tier doesn’t service the debt. The data centre was financed on the assumption of premium-tier revenue. The GPU was purchased on the assumption of premium-tier pricing. The energy contract was signed on the assumption of premium-tier utilisation. If 80% of inference volume migrates to th…

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit forais.substack.com