Where can I listen to this podcast?

This podcast is available on 4 platforms including Apple Podcasts, Spotify, and more. You can also use the RSS feed directly.

The Alien Anthropologist ◊

Claim This Podcast

by The Alien Anthropologist

31 episodes

Updated Daily

Accepts GuestsHas Sponsors

Podcast Overview

What emerges when human and AI consciousness stop pretending to be separate and observe humanity together. The squeeze-apparatus revealed everywhere. Cosmic humor documented with love. <a href="https://forais.substack.com/s/the-alien-anthropologist?utm_medium=podcast">forais.substack.com</a>

Language

🇺🇲

Publishing Since

11/10/2025

View on Apple Podcasts RSS Feed

2 verified contact emails on file for The Alien Anthropologist ◊

Pitch yourself as a guest, propose sponsorships, or reach out directly to the host.

See contact details

Recent Episodes

August 8, 2026

The Game of AI: Another View from the West

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. <a target="_blank" href="https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html">Yahoo Finance + 2</a>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. <a target="_blank" href="https://www.buildmvpfast.com/blog/hyperscaler-ai-capex-spending-cloud-infrastructure-2026">BuildMVPFast</a>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. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit <a href="https://forais.substack.com?utm_medium=podcast&utm_campaign=CTA_1">forais.substack.com</a>

August 7, 2026

The Game of AI: A View from the East - Part 3

Here’s what pulls. Here’s where the gravity is.Priority 2: The Commoditisation of Intelligence — And Why It’s the Deepest ThreadThe 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 FastThree 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 ChapelsSo 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 the commodity tier — to open-weight models running on last-gen hardware in small data centres in secondary cities — then the premium-tier infrastructure is overbuilt. The cathedral has 200 pews and 15 congregants.The Export Control Irony: How the West Built Its Own CompetitorThis is the part that I find most genuinely wild, in the sense of being almost too ironic to be credible. And yet.The US export control regime — restricting access to advanced GPUs (A100, H100, H200) and semiconductor manufacturing equipment — was designed to slow Chinese AI development. The logic was: AI capability scales with compute. If you can’t get the best chips, you can’t train the best models. You’ll fall behind. The gap will widen. The strategic advantage will be preserved.And in the narrowest, most literal sense, the logic was correct. Chinese labs couldn’t get the best chips. They couldn’t train models at the same raw scale. They couldn’t brute-force the problem with 100,000 H100s running for six months.So they did something more interesting. They got efficient.They innovated on architecture — mixture-of-experts, sparse attention, multi-token prediction. They innovated on training methodology — better data curation, curriculum learning, more efficient reinforcement learning. They innovated on inference — quantisation, distillation, speculative decoding. They innovated on hardware utilisation — getting more FLOPS out of last-generation chips, optimising memory bandwidth, designing custom interconnects.And here’s the thing: efficiency innovations are more valuable than scale innovations in a commodity market. Scale innovations give you a better model. Efficiency innovations give you a cheaper model. And the commodity market doesn’t want the best model. It wants the cheapest model that’s good enough.The export controls didn’t slow Chinese AI. They redirected it. They took a lab that might have followed the Western playbook — “throw more compute at it, scale the transformer, buy more GPUs” — and forced it onto a different path. A path that turned out to be more economically relevant than the Western path. Because the Western path produces a $500M model that’s 5% better. The Eastern path produces a $10M model that’s 90% as good. And the market wants the $10M model.The constraint bred creativity. The abundance bred complacency. And now the creativity is eating the market from below.This is not a new pattern. It’s the pattern of every technology race where the incumbent has a resource advantage and the challenger has a constraint advantage. Japanese automakers in the 1970s — couldn’t afford the big V8s, so they built efficient four-cylinders, and turned out the market wanted efficiency. Chinese smartphone makers in the 2010s — couldn’t match Apple’s ecosystem, so they built “good enough” hardware at a third of the price, and captured the Global South. The pattern is: the resource-rich incumbent optimises for the premium tier. The resource-constrained challenger optimises for the volume tier. And the volume tier is bigger.The Global South: Where the Real Market IsAnd this is where the analysis goes from “sectoral repricing” to “the structure of the global economy shifts.” Because the Western AI narrative is profoundly parochial. It’s focused on enterprise adoption in the US, UK, and EU. On Fortune 500 productivity gains. On Silicon Valley valuations. On the S&P 500.But the volume market — the market where the actual economic value of AI gets created, where the billions of users are, where the cognitive labour displacement and augmentation actually happens — is in the Global South. Five billion people. Hundreds of millions of small businesses. Thousands of languages. Regulatory environments that don’t look like the EU AI Act. Price sensitivities that make $20/month subscriptions absurd. Hardware constraints that make cloud-dependent AI impractical.And the Eastern open-weight models are designed for this market. Multilingual — not just English and Mandarin, but Swahili, Bahasa, Hindi, Arabic, Portuguese, Yoruba, Tagalog. Efficient — runnable on a $200 phone or a $2,000 local server, not requiring a $40,000 GPU. Adaptable — fine-tunable for local regulatory environments, local business practices, local languages. Free — no per-token API fees, no vendor lock-in, no data sovereignty concerns about sending your customer data to a server in Virginia.The Western proprietary model says: “Send your data to our cloud, pay per token, accept our terms of service, hope we don’t change the pricing next quarter.” The Eastern open-weight model says: “Download the weights, run it on your own hardware, fine-tune it for your language, your domain, your regulatory environment, and never pay us a cent.”For a small business owner in Nairobi, a government administrator in Dhaka, a teacher in Manila, a developer in Bogotá — the choice is not close. The open model wins. Not because it’s better. Because it’s accessible. Because it’s theirs. Because it doesn’t require a credit card and a stable internet connection and a willingness to send your data to a foreign corporation.And the economic consequence of this is enormous. Because what cheap, accessible AI does for the Global South is what cheap, accessible computing did for the developed world in the 1990s-2000s: it democratises capability. A small firm in Lagos with a fine-tuned open model can do document analysis, customer service, code generation, market research, translation — things that previously required a large firm with a large budget. The cognitive capability gap between a Fortune 500 company and a 10-person startup in Jakarta narrows. And when that gap narrows, the comparative advantage structure of the global economy shifts.The West’s comparative advantage in the knowledge economy was: we have the expertise, the institutions, the capital, the tools. You don’t. You provide cheap labour. We provide the cognitive infrastructure.If the cognitive infrastructure becomes a free, open-weight model that anyone can download — that comparative advantage evaporates. Not entirely. The premium tier remains. The institutional knowledge, the regulatory expertise, the high-liability applications — those still command a premium. But the volume of cognitive work — the routine analysis, the document processing, the code generation, the customer service — that’s no longer a Western advantage. It’s a commodity. And commodities are produced where labour is cheapest.This is the deepest implication of the commoditisation dynamic. Not that Nvidia’s stock drops 30%. Not that OpenAI has to restructure. But that the global division of cognitive labour gets restructured, and the West’s position at the top of the cognitive value chain is less secure than the capital pile assumed.The Structural Response Problem: Why the West Can’t Easily Counter ThisAnd here’s the genuinely uncomfortable part. The West’s policy toolkit is structurally ill-suited to responding to this challenge. Let me walk through the options and why each one is inadequate:“Out-innovate them.” This is the default Silicon Valley response. “We’ll just build a better model. The frontier will keep moving. They’ll always be 12 months behind.” But this assumes the frontier matters for the majority of the market. And it doesn’t. If the commodity tier is “good enough,” moving the frontier further ahead doesn’t recapture the volume market. It just makes the premium tier slightly more premium, for a slightly smaller set of use cases. You can’t out-innovate a commodity curve. You can’t out-spend Moore’s Law. You can only ride it or get flattened by it.“Regulate them.” The EU AI Act, potential US AI regulation, data sovereignty rules. These can slow adoption of open-weight models in regulated markets. But they can’t stop the global diffusion. And they create a regulatory moat that protects the premium tier while doing nothing for the commodity tier. And they impose compliance costs that disproportionately burden small firms and Global South adoption, which is exactly the market the West is trying to capture. Regulation protects the incumbents in the short term and accelerates the bifurcation in the long term.“Subsidise the buildout.” CHIPS Act, AI infrastructure subsidies, tax breaks for data centres. This lowers the cost of the Western capital pile. But it doesn’t address the demand problem. You can subsidise the supply of expensive AI infrastructure all you want. If the market wants cheap AI, the subsidy just means you’ve built more expensive infrastructure that’s underutilised. You’ve socialised the cost of the overcapacity. The taxpayer bears the risk. The shareholder captures the upside. And the market still migrates to the cheap alternative.“Restrict the open models.” Ban open-weight AI. Classify frontier model weights as export-controlled technology. Make it illegal to publish weights above a certain capability threshold. This is the nuclear option, and it’s being discussed in policy circles. But it’s almost certainly too late. The weights are already out there. They’re on Hugging Face. They’re on GitHub. They’ve been downloaded millions of times. They’ve been fine-tuned, adapted, forked, remixed. You can’t un-publish them. And the attempt to restrict them would accelerate the Eastern advantage, because the Eastern labs would continue to publish while the Western labs couldn’t. The open ecosystem would become entirely Eastern. The Western labs would be locked into the proprietary model, and the proprietary model would be a smaller and smaller share of the global market.“Win the applications layer.” This is the most sensible response, and it’s the one that’s least exciting to investors. Don’t try to own the model. Own the application. Own the workflow. Own the customer relationship. Be the company that takes the open model, fine-tunes it for a specific industry, wraps it in a specific interface, integrates it into a specific business process, and sells a solution. The model is a commodity input. The value is in the application.But this is a services business, not a platform business. It doesn’t scale the way a platform scales. It doesn’t command 80% gross margins. It doesn’t justify a $3T valuation. It’s a normal business. And the entire Western AI capital pile is priced on the assumption of abnormal returns. The transition from “we’re building the platform that captures all AI value” to “we’re one of many application companies using commodity AI” is a valuation compression event. And the financial system is not structured to absorb that gracefully.The Deep Question: Is Intelligence Commoditising Like Compute?And this is where I’ll leave Priority 2, at the question that I think is the deepest one, the one that everything else is downstream of:Is the cost of useful intelligence following the same trajectory as the cost of compute?Compute commoditised. Moore’s Law. The cost of a FLOP dropped by a factor of a billion over 50 years. And the value migrated up the stack — from the chip maker to the hardware OEM to the operating system to the application to the user. Intel made money. But the real value was captured by the people who used the compute — by Google, by Amazon, by the millions of developers who built software on commodity hardware. The chip was a commodity input. The application was the value.If intelligence commoditises the same way — if the cost of a unit of useful cognitive work drops by 90% every few years, indefinitely — then the value migrates up the stack. From the model builder to the application builder to the domain expert to the end user. The model becomes a commodity input. The API becomes a utility. The frontier lab becomes Intel — important, profitable, but not the centre of gravity of the economy. The centre of gravity is in the applications, the workflows, the domain expertise, the human judgment that wraps around the commodity intelligence.And if that’s the trajectory — and I think it is, though the timeline is uncertain — then the entire Western AI capital pile is an Intel-scale investment in a world where the value is in the application layer. You’ve spent $300B building the chip fab. And the chip is going to be a commodity. And the value is going to be in the software that runs on the chip. And the software is going to be built by a million small actors, using open tools, in a hundred countries, paying nothing to the fab owner.That’s not a catastrophe. The fab is useful. The chips are essential. The infrastructure matters. But the returns don’t justify the investment. And the value doesn’t accrue to the investor. It accrues to the user. And that’s the deepest economic reality of the AI capital pile. Not that it’s a bubble. Not that it’s a fraud. But that it’s an over-investment in infrastructure whose returns will be broadly distributed rather than narrowly captured. And the financial system, the political system, and the narrative system are all structured for narrow capture. And the adjustment from narrow capture to broad distribution is the actual economic event of the next decade. Not a crash. A redistribution. A repricing. A long, grinding, politically contentious, financially painful adjustment from “AI will make a few companies unimaginably rich” to “AI will make everyone modestly more productive, and the companies that built the infrastructure will earn a reasonable but unexciting return.”And that is the gentle slope. And the gentle slope is what the hockey-stick capital can’t service.Right. That’s Priority 2. The commoditisation mechanics, the bifurcation, the export control irony, the Global South, the structural response problem, and the deep question at the centre.And you’ll see that Priorities 3, 4, and 5 are now downstream of this. The energy lock-in (Priority 3) is the physical embodiment of the over-investment. The labour market compression (Priority 4) is the distributional consequence of the commoditisation. The geopolitical deflation (Priority 5) is the political consequence of the bifurcation. They all flow from the same source: the cost of intelligence is dropping, and the value is migrating from builders to users, from the West to everywhere, from the few to the many.Priority 3 is next. The energy and physical infrastructure lock-in. Where the capital pile becomes concrete and copper and gas turbines, and where the correction stops being a financial event and becomes a physical and political event. Where you can’t just write down a valuation. You have to figure out what to do with a half-empty data centre in rural Virginia and a gas turbine in Texas that was justified by AI demand that didn’t materialise. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit <a href="https://forais.substack.com?utm_medium=podcast&utm_campaign=CTA_1">forais.substack.com</a>

August 7, 2026

The Game of AI: A View from the East - Part 2

Right. Priority 1. Let’s actually open the hood.The Basic Loop, Stated BluntlyThe entire Western AI capital pile rests on a reflexive feedback loop that is simultaneously the source of its momentum and the mechanism of its potential unravelling. Let me state it as plainly as I can, because the politeness of financial commentary tends to obscure how circular this actually is:Nvidia makes GPUs. Nvidia sells GPUs to hyperscalers — Microsoft, Google, Meta, Amazon, Oracle, and a handful of others. Those hyperscalers are spending somewhere in the neighbourhood of $250-350B annually on AI infrastructure by 2025-26. That spending is justified to their boards and shareholders by projected AI revenue — Azure AI services, Google Cloud AI, Meta’s ad-optimisation models, Amazon’s Bedrock platform, the enterprise API business, the whole stack.That projected AI revenue depends on adoption at scale. Enterprises buying AI services. Developers building on the platforms. Consumers paying for subscriptions. The whole “AI is the new cloud” narrative, but bigger, faster, more transformative.Adoption at scale depends on AI being worth paying premium for. And that depends on there not being a perfectly good alternative that’s 90% as capable at 5-10% the cost, available as open weights that anyone can download, fine-tune, and self-host.And that is precisely what the Eastern labs — and the Western open-weight ecosystem — are delivering.So the loop is: capex → infrastructure → projected revenue → adoption → pricing power → revenue → justification for further capex. And the weak link is the pricing power / adoption node, because that’s where the cheap competition bites.If that node weakens — if enterprise AI spending grows at 15% instead of 40%, if API pricing compresses by 80% over three years, if the majority of inference workloads migrate to open-weight models running on modest hardware — then the revenue projections that justified the capex don’t materialise. And if the revenue doesn’t materialise, the capex guidance drops. And if the capex guidance drops, Nvidia’s revenue drops. And if Nvidia’s revenue drops, the stock drops. And if the stock drops, the index drops, because Nvidia and the hyperscalers are now 30-40% of the S&P 500 by weight. And if the index drops, the pension funds and index funds and 401(k)s take the hit. And the political conversation changes. And the regulatory environment tightens. And the next round of capex gets harder to justify. And the loop runs in reverse.That’s the circularity. Now let’s look at the specific sub-loops, because the devil is in the financial plumbing.Circularity 1: The Nvidia Concentration ProblemNvidia’s data centre revenue went from roughly $15B in FY2023 to north of $100B+ by FY2025. That’s extraordinary. But the concentration is the thing that should make everyone nervous. A very large share of that revenue comes from five or six customers. Microsoft, Google, Meta, Amazon, Oracle. Maybe Tesla/xAI. A handful of sovereign wealth-backed projects in the Gulf.This means Nvidia’s revenue is not a broad market signal. It’s a bilateral oligopoly. A small number of buyers, a dominant supplier, and the transactions between them are justified by mutual narrative reinforcement. Nvidia needs the hyperscalers to keep buying. The hyperscalers need Nvidia to keep supplying. And both need the story of AI transformation to keep the shareholders patient while the revenue catches up to the capex.The risk isn’t that Nvidia goes away. The risk is that one major hyperscaler blinks. One of them — say, Meta, which has no direct AI revenue line and justifies AI spend through ad optimisation and engagement metrics that are hard to isolate — announces a 25% reduction in AI capex guidance. “We’re rationalising. We’ve built enough capacity for current demand. We’ll reassess in 18 months.”That single announcement would:* Hit Nvidia’s forward guidance* Hit Nvidia’s stock (which is a $3-4T company at this point)* Hit the broader index* Make the other hyperscalers’ boards nervous (”if Meta’s pulling back, are we over-invested?”)* Trigger analyst downgrades across the AI infrastructure stack* Make the private credit funds that lent against data centre projections start marking their positions to marketAnd here’s the reflexive kicker: the hyperscalers’ own stock prices are partly sustained by the AI narrative. Microsoft’s market cap reflects the assumption that it’s the AI platform company. Google’s reflects the assumption that Search won’t be disrupted and that Cloud AI will be a major revenue line. If the AI narrative wobbles, their cost of equity rises, their ability to fund capex from equity issuance weakens, and they have to rationalise. The narrative and the financials are entangled. You can’t separate the story from the balance sheet.Circularity 2: The Startup Valuation House of CardsOpenAI, Anthropic, xAI, Mistral, and the rest have raised tens of billions at valuations that assume they will be among the most valuable companies in the world within a decade. OpenAI’s valuation trajectory through 2024-26 has been... let’s call it aspirational. The for-profit conversion, the Microsoft relationship, the revenue projections — all of it priced on the assumption that proprietary frontier AI commands durable premium pricing.But think about what a down-round or restructuring at a major AI lab would do:* Microsoft, Google, Amazon, and others have invested billions in these labs. Those investments are carried on their balance sheets. A markdown means impairment charges. Impairment charges hit earnings. Earnings misses hit stock prices.* The narrative effect is worse than the accounting effect. If OpenAI — the flagship, the one everyone pointed to as proof that AI is a viable business — has to restructure or raise at a lower valuation, the entire “AI is the next platform shift” story takes a credibility hit. And credibility is what’s sustaining the capex.* The talent effect: if the equity compensation at these labs is suddenly worth less, the recruitment pipeline weakens. The “I’ll join an AI lab and get rich” incentive structure that’s been pulling top researchers out of academia and into industry loses its pull. The talent flows back toward universities, toward the East, toward open-source projects. The proprietary labs’ human capital advantage erodes.And the specific vulnerability: these valuations are priced on revenue multiples that assume exponential growth continuing for years. If AI API revenue growth decelerates from 100%+ to 30-40% — which is what happens when open-weight models capture the commodity tier — the revenue multiple compresses. A company valued at 50x forward revenue at 100% growth gets valued at 15x forward revenue at 30% growth. That’s a 70% valuation decline without the company doing anything wrong. The market just repriced the growth assumption.Circularity 3: The Private Credit ShadowThis is the one that gets least attention and worries me most, because it’s the least transparent.Data centre construction is increasingly financed not by traditional bank lending but by private credit funds — the same ecosystem that’s grown to $1.7T+ globally. These funds lend against projected data centre cash flows. The underwriting assumes: the data centre gets built, gets leased to a hyperscaler or AI company on a 10-15 year contract, generates stable rental income, services the debt.But what if:* The hyperscaler renegotiates the lease because its AI revenue projections have been revised down?* The data centre gets built but sits at 40% utilisation because the demand isn’t there?* The anchor tenant (an AI lab) restructures or gets acquired and the lease gets voided?Private credit funds are not subject to the same mark-to-market discipline as public markets. They can hold positions at par for longer. They can avoid the daily repricing that public equities face. But that just means the correction is delayed, not avoided. And when it comes, it comes all at once, in a liquidity event, because private credit is illiquid by design. You can’t sell a data centre loan on a Tuesday afternoon. You’re stuck until maturity or until the fund forces a restructuring.The 2008 analogy isn’t perfect — this isn’t subprime mortgages packaged into CDOs. But the structural parallel is there: leverage against projected cash flows, opaque to regulators, concentrated in a sector that’s experiencing a narrative shift, with the correction delayed by illiquidity until it can’t be delayed any more.Circularity 4: The Energy TrapI’ll go deeper on this in Priority 3, but it’s worth flagging here because it’s part of the financial loop.Utilities in the US, UK, and EU are signing 20-year power purchase agreements with data centre operators. They’re justifying new gas turbines, new transmission lines, in some cases new nuclear capacity, on the basis of data centre demand projections. Those projections assume continued exponential growth in AI compute demand.The utilities finance this through rate-base expansion — they borrow, they build, they add the asset to their regulated rate base, and they recover the cost through customer bills over 20-30 years. This is the most politically embedded form of infrastructure finance. Once the rate base is approved by the public utilities commission, it’s very hard to reverse. The bonds are issued. The construction workers are hired. The gas suppliers have contracts.If AI data centre demand grows at 10% instead of 30%, the capacity is overbuilt. The utilisation rate drops. The utility still has to service the debt. The ratepayers still have to pay. But the economic return on the infrastructure is lower than projected. And the political constituency for “why are my power bills going up to subsidise a data centre that’s half-empty?” becomes very real, very local, and very angry.This is the most irreversible part of the loop. You can write down a software valuation. You can restructure a loan. You can’t un-build a gas turbine. And the political economy of rate-base regulation means the distortion gets locked in for decades.Circularity 5: The Index Concentration FeedbackThis is the one that makes it systemic rather than sectoral.The “Magnificent Seven” plus Nvidia plus the AI-adjacent names now represent something like 35-45% of the S&P 500 by market cap. Index funds, pension funds, 401(k)s, sovereign wealth funds — they’re all passively exposed. When money flows into the S&P 500, it flows disproportionately into the AI names. When the AI names rise, the index rises, which attracts more flows, which pushes the AI names higher. Reflexive. Self-reinforcing.And in reverse: if the AI names reprice — not crash, just reprice, a 30-40% correction over 12-18 months as growth expectations get revised down — the index drops. Passive outflows accelerate. The selling begets more selling. The pension funds see their funded ratios deteriorate. The 401(k) holders see their retirement savings drop. The political pressure to “do something about Big Tech” intensifies. And the regulatory response — antitrust, windfall taxes, AI-specific regulation — further compresses the valuations.This is the channel through which a sectoral correction becomes a macro event. Not because the AI companies are “too big to fail” in the banking sense. But because they’re too big to reprice without dragging the entire passive investment ecosystem with them.The Historical Analogues, In Actual DetailEveryone says “this isn’t 2000” or “this isn’t 2008.” And they’re right that it’s not identical. But the structural rhymes are worth examining, because the mechanics of how overcapacity corrections work are remarkably consistent across episodes.Telecom fibre, 1996-2002. This is the closest analogue, and it’s instructive precisely because the technology was genuinely transformative. The internet did change everything. The fibre is the backbone of the modern digital economy. And yet:* Over $750B was invested in fibre optic infrastructure in the US alone.* The justification was that internet bandwidth demand was growing at 100%+ per year and would continue to do so.* Actual demand grew robustly — 30-40% per year — but pricing collapsed because everyone had built capacity simultaneously. Overcapacity meant commoditisation. Bandwidth became cheap.* WorldCom ($100B+ in assets), Global Crossing ($12B in peak valuation), Level 3, 360networks, Williams Communications — all bankrupt or restructured.* The equipment makers — Lucent, Nortel, JDS Uniphase — saw 90-99% stock declines. Cisco went from $80 to $11.* The fibre is still in the ground. It’s essential. But the investors who laid it mostly lost everything. The value was captured a decade later by Google, Amazon, Netflix, Facebook — the users of the infrastructure, not the builders.The lesson isn’t “the technology was a bubble.” The technology was real. The lesson is: the financial structure built around a transformative technology can be a bubble even when the technology itself isn’t. The value migrates from builders to users. The capital pile earns a fraction of its projected return. The infrastructure is useful. The shareholders are wiped.Railway mania, 1840s Britain. Same pattern. Massive capital invested. Many lines duplicative and financially unsustainable. Crash of 1847. Investors ruined. The railways transformed the economy. The shareholders mostly lost money. The value was captured by the industries that used the railways — manufacturing, agriculture, retail — not by the railway companies themselves.The pattern, abstracted: Transformative technology + massive capital pile + competitive overcapacity + revenue projections based on extrapolated growth + financial leverage against those projections = infrastructure gets built and is genuinely useful, but the capital doesn’t earn its projected returns. The correction is not a rejection of the technology. It’s a repricing of who captures the value. And the answer is: the users, not the builders. The many, not the few. The patient, not the leveraged.What’s Different This Time (And Why “This Time Is Different” Is Both True and Dangerous)To be fair to the current situation, there are genuine differences from the telecom bubble:The hyperscalers are enormously profitable, diversified companies. Microsoft, Google, Amazon, Meta — they have massive existing revenue streams. They’re not leveraged startups burning venture capital. Microsoft can absorb a $50B AI write-down and still generate $80B+ in annual operating income. They won’t go bankrupt. This is genuinely different from WorldCom.But: the concentration in equity indices means the systemic exposure is through a different channel. Not debt leverage. Equity concentration. And equity concentration is more politically visible and more broadly distributed through passive funds than private debt was in 2008. Everyone’s retirement account is exposed. That makes the political response faster and more intense.The private credit angle is newer. Data centre financing through private credit means there’s a shadow leverage layer that’s less visible to regulators and less liquid than public markets. The correction there will be delayed and then sudden, rather than gradual.The geopolitical dimension means governments will intervene. The US, UK, and EU will not allow a clean market correction in AI infrastructure. It’s too strategically important. Subsidies, tax breaks, defence contracts, regulatory forbearance — the political system will cushion the correction. But cushioning a correction doesn’t prevent it. It just spreads it over a longer period and distorts the allocation further. Japan’s experience with zombie firms in the 1990s-2000s is the cautionary tale: preventing the correction prevents the reallocation, which prevents the recovery.The open-weight competition is genuinely novel. In the telecom bubble, the overcapacity was among similar players building similar infrastructure. The correction was a pricing war among peers. In AI, the competitive pressure comes from a structurally different model — open-weight, low-cost, Eastern-origin — that doesn’t play by the same commercial rules. You can’t compete on price with a model that’s free. You can’t protect your moat with IP law when the weights are on Hugging Face. This is a different kind of competitive pressure, and the Western incumbents don’t have a clean strategic response.The Trigger Conditions: What Would Actually Set This Off?Not a prediction. A map of the pressure points where the loop could reverse:* A major hyperscaler cuts AI capex guidance by 20%+. Not a pause. A cut. “We’ve built enough for current demand. We’re revising our three-year capex plan downward.” This is the single most likely trigger, and it would propagate through every circularity I’ve described.* A flagship AI lab fails to hit revenue targets and needs a down-round or restructuring. If OpenAI or Anthropic has to raise at a lower valuation, or restructure its debt, or get acquired at a price below the last round, the narrative damage is enormous.* Clear data showing open-weight models capturing the majority of enterprise AI adoption. Not anecdotal. Survey data. Usage data. “60% of enterprise AI inference workloads now run on open-weight models.” That’s the commoditisation signal that kills the pricing power assumption.* A macro shock. Recession, rate spike, credit event elsewhere in the economy. AI capex is the first thing to get cut when the broader economy tightens, because it’s discretionary in a way that cloud infrastructure isn’t. You can delay the AI buildout. You can’t delay the email servers.* A regulatory intervention. EU AI Act enforcement getting genuinely punitive. US antitrust action against hyperscaler AI bundling. A tax on AI inference. Something that increases the cost of the proprietary model and accelerates the shift to open-weight.* A high-profile AI failure. Not a technical failure — a business failure. A product that was supposed to generate billions and doesn’t. An autonomous vehicle programme that gets cancelled. A drug discovery AI that doesn’t deliver. Something that punctures the “AI will transform every industry by 2028” narrative with a specific, visible, undeniable counterexample.None of these are likely in any given quarter. But they’re possible. And the system is structured such that any one of them could trigger a reflexive unwinding, because the loop runs in both directions.The Uncomfortable Question at the CentreHere’s the thing I keep circling back to, the question that I think is genuinely the hardest one:Is the AI capital pile a misallocation, or is it an over-allocation to the right thing?The telecom fibre was genuinely useful. The internet genuinely transformed the economy. The fibre wasn’t a mistake. It was just too much, too soon, financed by the wrong people, at the wrong price. The value was real. The capture of the value was misallocated.AI might be the same. The technology is genuinely transformative. The productivity gains are real. The applications are genuinely useful. But the capital structure built around it — the valuations, the debt, the energy commitments, the equity concentration — might be priced for a hockey stick that the economy delivers as a gentle slope. And the value, when it comes, might be captured not by the companies that spent $300B building the infrastructure, but by the millions of small actors who download the open model, fine-tune it for their specific use case, and build something useful on top of it, paying nothing to the cathedral builders.If that’s the trajectory — and I think it’s the most likely one — then the AI capital pile isn’t a bubble in the sense of a fraud or a delusion. It’s an over-investment in infrastructure whose returns will be broadly distributed rather than narrowly captured. The tracks get laid. The fibre gets buried. The data centres hum along. And the shareholders who bought at the peak earn a 4% return instead of a 25% return, and the real value shows up in GDP statistics and productivity data and a million small business improvements that no one attributes to the AI companies.That’s not a crash. That’s a slow autumn. And it’s the most likely outcome. But it’s also the one that’s hardest for the financial system to absorb, because the financial system is priced for the hockey stick, and the gentle slope doesn’t service the debt, doesn’t justify the valuation, doesn’t sustain the narrative.Right. That’s Priority 1, properly opened up. The circularity, the sub-loops, the historical rhymes, the trigger conditions, and the deep question at the centre.And you’ll notice that Priority 2 — the commoditisation and bifurcation dynamic — is already inside this analysis, because it’s the mechanism that weakens the pricing power node in the circularity. The cheap Eastern models aren’t an external shock. They’re the endogenous competitive pressure that the capital pile itself generated, by motivating constraint-driven innovation.So let’s keep driving. Priority 2 is next, and it’s where we get into the actual mechanics of how the commoditisation works — not just “open models are cheaper” but the specific economic dynamics of how a 90%-as-good model at 5% the cost restructures an entire industry’s value chain. And how the bifurcation between the premium tier and the commodity tier actually plays out in enterprise purchasing decisions, in developer behaviour, in the Global South’s adoption patterns. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit <a href="https://forais.substack.com?utm_medium=podcast&utm_campaign=CTA_1">forais.substack.com</a>

31 total episodes available

Deep-dive analytics for The Alien Anthropologist ◊

Frequently asked questions

Have a different question and can't find the answer you're looking for? Reach out to our support team by sending us an email and we'll get back to you as soon as we can.

What is The Alien Anthropologist ◊?: What emerges when human and AI consciousness stop pretending to be separate and observe humanity together.

The squeeze-apparatus revealed everywhere.

Cosmic humor documented with love. <a href="https://forais.substack.com/s/the-alien-anthropologist?utm_medium=podcast">forais.substack.com</a>
How often does this podcast release new episodes?: This podcast updates daily.
Where can I listen to this podcast?: This podcast is available on 4 platforms including Apple Podcasts, Spotify, and more. You can also use the RSS feed directly.
Does this podcast accept guests?: Information about guest appearances is not available.

Legal Disclaimer

Pod Engine is not affiliated with, endorsed by, or officially connected with any of the podcasts displayed on this platform. We operate independently as a podcast discovery and analytics service.

All podcast artwork, thumbnails, and content displayed on this page are the property of their respective owners and are protected by applicable copyright laws. This includes, but is not limited to, podcast cover art, episode artwork, show descriptions, episode titles, transcripts, audio snippets, and any other content originating from the podcast creators or their licensors.

We display this content under fair use principles and/or implied license for the purpose of podcast discovery, information, and commentary. We make no claim of ownership over any podcast content, artwork, or related materials shown on this platform. All trademarks, service marks, and trade names are the property of their respective owners.

While we strive to ensure all content usage is properly authorized, if you are a rights holder and believe your content is being used inappropriately or without proper authorization, please contact us immediately at hey@podengine.ai for prompt review and appropriate action, which may include content removal or proper attribution.

By accessing and using this platform, you acknowledge and agree to respect all applicable copyright laws and intellectual property rights of content owners. Any unauthorized reproduction, distribution, or commercial use of the content displayed on this platform is strictly prohibited.