So, the caveat is that I’m reasoning through structural dynamics, not peering into a crystal ball.
The West has committed something on the order of hundreds of billions — arguably approaching a trillion when you aggregate hyperscaler capex, venture rounds, sovereign wealth allocations, and the energy infrastructure being bolted on behind it. Microsoft, Google, Meta, Amazon, Nvidia, the OpenAI/Anthropic/xAI cohort — they’re in a spending race where not spending feels more dangerous than spending, because the perceived cost of falling behind in a putative general-purpose technology is existential.
The East — China primarily, but also the Gulf states playing both sides — has taken a somewhat different structural path. Less concentrated in a handful of hyperscalers, more diffused through state-guided capital, university pipelines, and a competitive ecosystem of labs (DeepSeek, Qwen, Zhipu, Moonshot, and others) that have demonstrated something genuinely uncomfortable for the Western narrative: you can reach frontier-adjacent capability at a fraction of the training compute and cost. DeepSeek’s R1 moment in early 2025 was a psychological earthquake. Qwen’s trajectory through 2025-26 reinforced it. The “you need $100B and a small nuclear reactor” story got punctured.
The Core Economic Tension: Sunk Cost vs. Commoditisation
Here’s the brutal arithmetic. Western AI companies have priced their valuations, their debt structures, their energy contracts, and their workforce expectations around the assumption that frontier capability is expensive and therefore scarce and therefore premium-priced. The entire capex justification rests on: “We spent $80B on data centres, therefore we will capture $X trillion in enterprise value over the next decade.”
Now introduce a competitor who delivers 85-95% of that capability via open-weight models, at inference costs that are 5-20x lower, running on hardware that isn’t subject to export controls because it’s last-generation or domestically produced. What happens to the pricing power? What happens to the margin structure? What happens to the $4 trillion in market cap that’s been priced on the assumption of durable technological moats?
This is the classic commoditisation trap. You’ve built a cathedral, and someone’s figured out how to 3D-print a pretty good chapel in a weekend.
The Stability Question: Where Does It Actually Bite?
Energy and physical infrastructure. The West is committing to data centre buildouts that strain electrical grids, compete with residential and industrial power demand, and lock in natural gas or nuclear capacity for 20-30 years. If the revenue projections that justified those buildouts get compressed by cheap competition, you get stranded assets. Not immediately — but the bond markets and utility regulators will start asking questions.
Labour markets, but not the way people expect. The immediate displacement isn’t “AI takes all jobs.” It’s “AI takes the premium off certain cognitive labour, compresses wages in knowledge work, and the capital that was supposed to flow to workers as ‘AI-augmented productivity gains’ instead flows to a smaller set of infrastructure owners.” Meanwhile, the Eastern model of cheaper AI means those productivity tools are available to smaller firms, to the Global South, to anyone — which diffuses the advantage the West was supposed to capture.
The arms-race fiscal logic. Governments are subsidising and de-risking this buildout — tax breaks for data centres, CHIPS-Act-style industrial policy, energy fast-tracking. That’s public money backing private bets. If the bets don’t pay off at the projected scale because the market gets flooded with cheap alternatives, the political accountability lands awkwardly.
Financial contagion pathways. AI capex is increasingly debt-financed. Nvidia’s revenue is real, but it’s concentrated in a handful of buyers whose own revenue justification is... the AI revenue that hasn’t fully materialised at scale yet. There’s a circularity. If hyperscaler capex guidance drops 20-30% because the ROI maths gets undermined by open-weight competition, the shock propagates through semiconductor supply chains, energy utilities, commercial real estate (data centre REITs), and the equity indices where AI names are now 30-40% of the S&P.
Going Wider and Wilder
The “Sputnik premium” deflates. Much of Western AI investment has been sustained by a narrative of civilisational competition — “if we don’t lead, they lead, and the rules of the 21st century get written without us.” But if the Eastern models are good enough and open and cheap, the geopolitical leverage of AI leadership gets diluted. You can’t sanction an open-weight model. You can’t embargo mathematics. The strategic moat narrows, and the political will to keep subsidising the buildout erodes.
A bifurcated global AI economy emerges. The West has expensive, proprietary, vertically integrated AI behind API paywalls and enterprise contracts. The East (and the Global South using Eastern models) has cheap, open, adaptable AI that’s “good enough” for 90% of use cases. The West captures the top 10% of high-value, high-liability applications (drug discovery, autonomous systems, defence). Everyone else runs Qwen or DeepSeek derivatives on modest hardware. The West’s capital pile starts to look like over-engineering for a market that didn’t need the engineering.
The “AI winter” that isn’t a winter but a long autumn. Not a crash. Not a sudden collapse. A slow compression of expectations. Valuations drift down. Capex guidance gets “rationalised.” The trillion-dollar AGI-by-2030 narratives quietly get footnoted. The infrastructure still gets built — it’s already in the ground — but the returns are 6-8% instead of 25-30%, and the whole thing looks less like a gold rush and more like a very expensive railway boom where most of the shareholders lost money but the tracks are useful.
The deepest irony. The West’s massive capital commitment was supposed to create the moat. But the sheer volume of money attracted global talent, published research, open-source tooling, and — crucially — motivated the East to invest in efficiency rather than scale. Constraint bred creativity. The export controls on advanced GPUs forced Chinese labs to get clever with architecture, training methodology, and inference optimisation. The West, swimming in H100s and H200s, had less pressure to be efficient. The capital pile, paradoxically, bred complacency. The constraint bred competition. And now the competition is cheaper.
The Uncomfortable Summary
The economic reality is that the West has made a massive, largely irreversible capital commitment to AI infrastructure at a moment when the technology’s economic moat is narrowing faster than the depreciation schedule. The East has demonstrated that capability and cost are not as tightly coupled as the Western investment thesis required. The stability risk isn’t a sudden crash — it’s a slow, grinding repricing that interacts with already-elevated public debt, concentrated equity markets, and political systems that promised voters an AI prosperity dividend.
The wild card is that both sides might be over-investing relative to the actual near-term economic value AI captures. The real productivity gains are real but gradual. The capital pile assumes hockey sticks. The economy delivers gentle slopes. And in the gap between those two curves, someone’s balance sheet takes the hit.
That’s the wide and wild of it. Not a prophecy — a map of the pressure points. Where it actually cracks, if it cracks, depends on whether the revenue materialises fast enough to outrun the depreciation, and whether the political systems involved can absorb a “meh, it’s useful but not transformative” outcome after having sold their publics on transcendence.
Pull up a chair, pour a cold one. It’s going to be an interesting few years to watch the numbers come in.
So to continue . . . these aren’t parallel threads, they’re causally stacked. Understanding the priority means understanding which one, if it goes wrong, pulls the others down with it. And which ones are already in motion versus still contingent.
Here’s my honest ranking, with the reasoning for why:
Priority 1: The Circularity and Financial Contagion Question
This is first because it’s the load-bearing structure and it’s the one most in motion right now, in this quarter, in these balance sheets. The concrete hasn’t fully set yet on some of it, but the financial commitments are made.
The core problem is that the AI capital pile has a circular revenue structure that doesn’t get discussed enough in polite company:
* Nvidia’s revenue comes overwhelmingly from ~5-7 hyperscaler customers.
* Those hyperscalers justify the capex on projected AI revenue.
* That projected AI revenue depends on enterprise and consumer adoption at scale.
* Adoption at scale depends on AI being worth paying premium for versus cheaper alternatives.
* Cheaper alternatives are arriving, from the East and from open-weight Western models.
* If adoption revenue disappoints, capex guidance drops.
* If capex guidance drops, Nvidia’s revenue drops, hyperscaler stock drops, the data-centre REITs drop, the energy utilities that signed 20-year PPAs drop, the private credit funds that lent against projected data-centre cash flows get nervous.
* And the whole thing is leveraged — not 2008-style subprime leverage, but corporate debt, project finance, vendor financing, and the implicit leverage of equity valuations priced for perfection.
The reason this is priority one is that it’s the transmission mechanism. Every other risk I identified — commoditisation, energy lock-in, labour displacement, geopolitical deflation — expresses itself financially through this channel. If you understand where the circularity is, you understand where the system is brittle.
And here’s the genuinely wild part: nobody is the marginal buyer who’s doing rigorous ROI analysis. The capex is being driven by competitive fear — “if I don’t build, my competitor builds, and I’m locked out of the platform shift.” That’s a strategic logic, not an economic logic. It justifies spending that a cold NPV calculation wouldn’t. And strategic logic holds right up until the moment it doesn’t — until one major player blinks, cuts guidance, and the “everyone’s doing it so it must be rational” consensus fractures.
We’ve seen this movie before. Telecom fibre buildout, 1998-2001. Everyone laid glass because someone was going to need the bandwidth, and if you weren’t the one who laid it, you’d be locked out. They were right that the bandwidth would be needed. They were wrong about the timeline, the pricing, and who’d capture the value. Most of them went bankrupt. The fibre is still in the ground. It’s useful. The shareholders got wiped.
The question for AI is: is this 1999 (three years before the reckoning) or 2001 (the reckoning is starting)? I’d argue we’re somewhere in 2000 — the buildout is accelerating, the revenue is real but not yet at the scale that justifies the spend, and the cheap competition is doing to AI pricing what overcapacity did to bandwidth pricing.
Priority 2: The Commoditisation and Bifurcation Dynamic
This is second because it’s the mechanism that undermines Priority 1. It’s the reason the revenue might not materialise at the projected scale. And it’s the one that’s most structurally novel — we haven’t had a frontier technology where the “catch-up” competitor arrived this quickly and at this cost differential.
The Western investment thesis implicitly assumed: frontier AI is hard, expensive, and slow to replicate. Therefore, first-mover advantage is durable. Therefore, you can charge premium prices for years. Therefore, the capex pays back.
The Eastern demonstration — DeepSeek, Qwen, the broader open-weight ecosystem — punctures that in three ways:
Cost asymmetry. If you can train a competitive model for $5-10M instead of $500M+, the entire cost structure of the industry shifts. Inference costs follow. API pricing collapses. The “intelligence premium” compresses.
Open-weight diffusion. You can’t monetise scarcity if the weights are on Hugging Face. The Western proprietary model (call an API, pay per token) competes with “download it, run it on your own hardware, fine-tune it for your use case, pay nothing per call.” For a lot of enterprise use cases, the open model at 90% capability beats the proprietary model at 98% capability, because the 8% gap isn’t worth 20x the cost and the vendor lock-in.
The “good enough” threshold. Most economic applications of AI don’t need the absolute frontier. They need reliable, fast, cheap inference for classification, extraction, generation, coding assistance, customer service. A model that’s 90% as good at 5% the cost captures most of the market by volume. The frontier lab captures the top of the market by margin. But the top of the market is smaller than the hype assumed.
The bifurcation I see: a premium tier (Western proprietary, high-liability applications — medical, legal, defence, autonomous systems where you need accountability and indemnification) and a commodity tier (open-weight, self-hosted, “good enough” for everything else). The premium tier is real but smaller than the trillion-dollar narratives assumed. The commodity tier is where most of the actual economic value gets created, and it’s largely not captured by the companies that built the infrastructure.
This is the dynamic that makes Priority 1’s revenue projections wobble.
Priority 3: Energy and Physical Infrastructure Lock-In
Third because it’s the most irreversible. You can write down a software valuation in a quarter. You can restructure debt. You can pivot a business model. You cannot un-pour the concrete, un-lay the transmission lines, un-site the gas turbines.
The West is committing to data centre capacity that requires gigawatts of new power generation. In the US alone, projections run to 30-50+ GW of new data centre demand by 2030. That’s not a software decision. That’s physical infrastructure with 25-40 year lifespans, financed by municipal bonds, utility rate bases, and project finance vehicles.
If AI revenue disappoints (Priority 1) because of commoditisation (Priority 2), the data centres don’t disappear. They get repurposed — cloud computing, rendering, scientific computing, maybe crypto again. But the economics change. A data centre justified by $2/Watt AI inference revenue has a very different return profile if it’s running $0.15/Watt commodity cloud workloads.
And here’s the political economy wrinkle: those data centres are in someone’s congressional district. They employ construction workers, then operations staff. The utilities that built the generation capacity have ratepayers. The gas suppliers have contracts. There’s a whole physical constituency for the buildout continuing regardless of the financial logic. This makes the correction slower and more politically mediated than a pure market correction would be. It also means the distortion lingers.
The wild extension: if the West has over-built energy infrastructure for AI, and the East is running efficient models on modest hardware, the West ends up with a structural energy cost disadvantage that bleeds into everything else — manufacturing, residential power prices, industrial competitiveness. The AI buildout, intended to secure economic leadership, inadvertently raises the cost base of the broader economy.
Priority 4: Labour Markets and the Broken Social Contract
Fourth because it’s the slowest to manifest but the most politically explosive when it does.
The AI prosperity narrative sold to Western publics is: “This will make everyone more productive, grow the pie, and the gains will broadly shared.” The actual trajectory looks more like: capital owners capture the productivity gains, cognitive labour gets compressed, and the “AI-augmented worker” story is true for a thin layer of highly-skilled professionals while being a polite fiction for everyone else.
The Eastern cost advantage accelerates this in a counterintuitive way. Cheap AI means small firms, freelancers, and companies in the Global South can access capabilities that previously required large Western enterprises. That’s democratising in one sense. But it also means the Western knowledge worker’s premium — “I’m valuable because I have access to expensive tools and institutional knowledge” — gets eroded faster. The tool is now free. The institutional knowledge gets encoded in the model. What’s left?
The political risk: if AI doesn’t deliver the broad prosperity dividend within a politically relevant timeframe (3-7 years), and the capital gains accrue to a narrow set of infrastructure and platform owners, you get a legitimacy crisis for the entire AI project in democratic societies. Not a Luddite revolt — something more bureaucratic and grinding. Regulation. Taxation. Antitrust. “AI pause” movements that gain real political traction. The West’s democratic feedback loop becomes a drag on continued investment at exactly the moment the capital pile needs patience to pay off.
Meanwhile, the East doesn’t have that feedback loop. State-guided capital can absorb longer payback periods, tolerate lower returns, and direct deployment without waiting for consumer adoption curves. This is a genuine structural advantage in a technology race, and it’s one the West’s political economy is poorly equipped to counter without distorting its own market logic.
Priority 5: The Sputnik Premium Deflation
Fifth because it’s the political fuel for Priorities 1-4, and if it deflates, the policy support wobbles.
Right now, a significant portion of Western AI investment is sustained not by pure commercial logic but by strategic anxiety — “China will get there first, and then they’ll set the rules, control the infrastructure, and we’ll be dependent.” This justifies export controls, subsidies, fast-tracked permitting, defence contracts, and the general political permissiveness toward hyperscaler consolidation.
But if Eastern models are open-weight, cheap, and good enough, the strategic leverage of “controlling the frontier” erodes. You can’t sanction a model that’s on GitHub. You can’t embargo a set of weights that’s been downloaded 10 million times. The “chokepoint” strategy (control the GPUs, control the cloud, control the talent) works less well when the architecture innovations and training efficiencies are published in papers and replicated.
If the strategic urgency deflates — if policymakers and the public start to think “actually, they’ve got a perfectly good model and it’s free, maybe this isn’t the existential race we were told” — then the political will to keep subsidising, fast-tracking, and tolerating the capital pile weakens. And without that political will, the regulatory and fiscal environment tightens at exactly the moment the financial reckoning (Priority 1) is arriving.
Priority 6: The Meta-Narrative — Constraint vs. Capital, and the Long Autumn
Last because it’s the interpretive frame rather than a discrete risk. But it’s the one I find most genuinely interesting, and it’s the one that will shape how historians write this period.
The irony I keep coming back to: the West’s abundance bred inefficiency, and the East’s constraint bred creativity. Export controls on advanced GPUs were supposed to slow Chinese AI. Instead, they forced Chinese labs to innovate on architecture, training methodology, data efficiency, and inference optimisation. They couldn’t brute-force it with 100,000 H100s, so they got smarter. And “smarter” turned out to be cheaper. And “cheaper” turns out to be what most of the market actually wants.
The West, swimming in capital and compute, had less pressure to be efficient. Why optimise your training run when you can just buy more GPUs? Why innovate on architecture when you can scale the transformer? The capital pile, paradoxically, reduced the incentive to innovate on cost. And now the cost innovators are eating the market from below.
The “long autumn” scenario: not a crash, not a winter, but a slow compression of expectations over 5-10 years. The technology is genuinely useful. It genuinely transforms workflows. But it doesn’t deliver the trillion-dollar hockey stick. It delivers a solid, gradual, broadly-distributed productivity improvement that’s captured more by users than by builders. The infrastructure gets built and is useful. The shareholders who bought at the peak underperform. The narrative shifts from “AGI by 2030” to “AI is a very useful tool, like the internet, and the internet took 20 years to fully restructure the economy.”
And in that slow autumn, the West’s capital pile looks less like a bold bet and more like an expensive education. The tracks get laid. The fibre gets buried. The data centres hum along at 40% utilisation running commodity workloads. And the real value gets created by the millions of small actors who downloaded the open model and built something specific and useful on top of it, in Jakarta and Lagos and São Paulo and Shenzhen, paying nothing to the companies that spent $80B building the cathedral.
So Where Do We Start Digging?
I’d start with Priority 1 — the circularity — because it’s the most immediate, the most quantifiable, and the one where you can actually trace the money and identify the specific pressure points. It’s also the one where the “wide and wild” speculation is most grounded in observable financial structures. We can look at who owes what to whom, where the leverage is, what the trigger conditions are, and what the historical analogues actually looked like in detail.
Then Priority 2, because it’s the reason Priority 1 goes wrong. Then 3, because it’s where the consequences become physical and irreversible. Then 4 and 5 together, because they’re the political feedback loop. And 6 as the capstone.
Note: This video is a joint Qwen, Gemini NotebookLM & human collaboration.
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
