This post is about the structural layer underneath it. Of all the “open” AI work happening anywhere right now, how much of it is actually structurally durable - and how much is one strategic-priority shift away from being recaptured?

The answer jumped out at me in Bill Gurley’s piece, in the difference between two of his examples. I should have made the distinction sharper in my own post on Open Source as Structural Defence. Let me fix that now.

Subscribe now

Android was not Kubernetes

Android and Kubernetes were both Google open source plays. Both were strategically brilliant. Both worked. They are not the same, and the difference is everything.

Android launched in 2007 through the Open Handset Alliance - a coalition of handset makers and carriers determined to prevent Apple from dominating mobile the way Microsoft dominated PCs in the 1990s. The Alliance never moved Android into a neutral foundation. No referee. No independent governance. And so, over time, Google “recaptured” Android by tying access to Google Play, Maps, Gmail, and YouTube to a Google-approved build. Ship a competing version? Lose the apps that matter. Android today is open in name and Google’s in practice everywhere except China.

Kubernetes - the software that runs most of the world’s cloud applications today - was different. Google open-sourced it in 2014 and almost immediately donated it to the newly-formed Cloud Native Computing Foundation, a neutral body sitting under the Linux Foundation. No single company controls it. AWS - the very company Kubernetes was designed to commoditise - now contributes to it actively, because its customers demanded Kubernetes support. The architecture made unilateral capture impractical.

Same company. Same era. Both open source. Wildly different structural outcomes.

The lesson is sharp: the presence or absence of a neutral foundation is what determines whether “open” stays open. Without a referee, open is conditional. Revocable.

The recapture risk we keep pretending isn’t there

Now apply that test to every open-weight AI model released in the last three years.

Llama - once marketed as America’s leading open-weight model - was Android. Released by a single company, on a license that company controls, with the strategic posture maintained as long as it served that company’s interests. Then Llama 4 underperformed. Llama 4 Behemoth got shelved. Zuckerberg confirmed in July 2025 that Meta would no longer release superintelligence-capable models openly. Meta Superintelligence Labs released Muse Spark - their first frontier-class model - closed. Their 2024 manifesto titled Open Source AI is the Path Forward reads very differently today.

This was always going to happen. Not because Zuckerberg is bad, but because the architecture made any other outcome economically irrational. There was no foundation. There was no referee. There was no structural defence against the single company shifting its strategic priorities. The “open” was conditional from day one.

Gemma is in the same position. Mistral too. So are DeepSeek, Qwen, GLM, Kimi - same structural pattern, just headquartered somewhere else and with a national government as the strategic sponsor.

There is currently no Kubernetes-equivalent for frontier AI weights. Not in the US. Not in China. Not in Europe. Every open-weight frontier model in existence today is held by a single sponsor whose continued openness is a matter of ongoing choice, not architectural commitment.

That’s the gap. That’s what’s missing. The entire open-weight ecosystem sits on Android-style infrastructure - open until somebody decides it isn’t.

The toll booth, revisited

In my last post, I argued that open doesn’t have to win the race - just circumvent the toll booth. The market is already settling that question in open’s favour, with Chinese models doing the heavy lifting.

But there’s a second question hiding underneath: who owns the road that the toll booth sits on?

Linux works as infrastructure because no single company can revoke it. Kubernetes works because no single company can revoke it. The reason these projects make their respective layers non-rentable is not that they exist as open source - it’s that they exist as open source under a structure that prevents recapture. The foundation isn’t a nice-to-have. The foundation is the entire point.

Open weights released by single companies don’t pass this test. They’re toll booths in waiting. Useful today because their sponsors find openness strategically convenient. But the road belongs to the sponsor, and the sponsor can put up a toll booth whenever the economics change.

“Llama is open” was always a present-tense statement, not a structural one. The same is true of every other open-weight frontier model in 2026.

Autonomous vehicles: the test case

Gurley devotes a long section to autonomous vehicles, and it’s worth reading in full. His argument is that AVs are the cleanest possible candidate for an Open Source Strategy play. Waymo, owned by Alphabet (Google’s parent), is the runaway leader. Tesla has a parallel proprietary stack. Cruise burned $10 billion and folded. Argo went to zero.

But there are fifty-plus companies - every global automaker, every ride-hailing platform, every delivery and logistics company, every retailer with a fleet - whose business depends on AV technology and who cannot individually win the proprietary path. The capital math is brutal. An open AV consortium under a neutral foundation is the only economically rational move for the rest of the industry.

This argument is almost tautologically convincing - because it makes concrete one specific instance of the broader displacement story from my post on what comes after the jobs disappear. Among the many sectors AI will reshape, with AVs, the abstract debate about open infrastructure becomes a physical question about who controls the safety logic of vehicles operating in our cities.

If the AV stack stays proprietary, we get a world where two or three companies set the implicit value-of-a-life calculus for every road in the West - regulator-by-trade-secret, with no other vendor able to inspect or improve on the safety logic. We’ve seen this before. Social media platforms have spent the last fifteen years setting their own implicit calculus on the cost of harm to children against the value of engagement - calculations that only became public when Frances Haugen leaked Meta’s internal research in 2021. Proprietary stacks, proprietary tradeoffs, no external audit until the whistleblower arrives. The AV version of that pattern operates with metal and momentum.

Gurley’s argument is that the non-Waymo, non-Tesla players have every strategic reason to run Google’s own playbook against Alphabet - Waymo’s parent. And, critically, that an open AV stack done right would be Kubernetes-shaped, not Android-shaped - because the consortium of also-rans is large enough, well-capitalised enough, and aligned enough to put it under a neutral foundation from day one.

Whether anyone has the chutzpah to actually do it is the open question. Gurley’s word, not mine. But it’s the right word.

The agentic layer is next

Here’s what comes after AVs.

The same structural argument applies to the agentic coordination layer - the protocols and infrastructure that let AI agents discover, trust, and transact with each other. If you take Gurley’s framework seriously, this layer is exactly where the next leader’s moat will compound: a small number of frontier vendors will define how agents authenticate, how they exchange value, how they delegate tasks, how they resolve conflicts. Whoever owns those protocols owns the toll booth on every agent-to-agent interaction in the global economy.

This is the Overture-equivalent for AI - the layer where neutral foundation work needs to happen now, not in five years when recapture is already complete. It’s exactly why we started AAIF.

The pattern recurs at every layer: weights, vehicles, agents. The question is never whether open is possible at that layer. The question is whether open is structurally durable at that layer - and that depends entirely on whether a neutral foundation gets there before the leader’s moat finishes forming.

The convergence

Gurley wrote the corporate-strategy version. I’ve been writing the political-economy version. The two arguments converge on the same single prescription: build the foundation now - before the moat hardens.

This is rare. Most policy fights pit public interest against incumbent interest and get settled by lobbying budgets. Open AI infrastructure is the unusual case where the public interest aligns with the strategic interest of roughly fifty of the world’s largest non-frontier companies. That alignment is the most underexploited political resource in this entire debate.

Gurley’s piece should be on every Fortune 500 CEO’s desk this quarter. The political-economy version should be on every legislator’s. They’re the same argument.

It’s a brilliant piece, and I want to spend two posts responding, expanding on my political-economy version of Gurley’s corporate-strategy analysis. The arguments converge in ways that are unexpected and important. That convergence is unusual and it matters.

Most policy fights are zero-sum between public interest and incumbent interest. Workers want one thing, capital wants another. Citizens want one thing, platforms want another. The two sides line up against each other and the fight gets settled in the favour of whoever has more lobbying budget. Open AI infrastructure is one rare case where the public-interest argument and the corporate-strategy argument point to the same prescription - because they’re driven by the same underlying economics. That alignment is an underexploited political resource in this debate - and it shows up in how badly Western policy is currently misreading China.

Subscribe now

The follower’s weapon

Gurley’s sharpest observation lands an almost throwaway line: “when you are the follower in a technology race, open source is your most powerful weapon.” He calls it the follower’s weapon, and it’s the whole story of RISC-V. China couldn’t buy ARM licenses, couldn’t access advanced US silicon, and couldn’t credibly out-innovate Intel. So they embraced an open instruction set that no one can sanction - because the project itself is governed in Switzerland, and the spec is, well, open. Gurley quotes an analyst calling RISC-V “sanction-proof.”

DeepSeek, Qwen, Kimi, GLM, MiniMax. The Chinese open-weight frontier is real, improving fast, and enshrined in two consecutive national Five-Year Plans. As Demis Hassabis, CEO of Google DeepMind, points out, Chinese models are leading in open source. Anysphere built Cursor’s Composer 2 on Kimi. Airbnb’s customer service agent runs on Qwen. Brian Chesky calls Qwen “very good”, “fast and cheap”. He picked Qwen for AirBnB, not to make a geopolitical statement, but because the economics made sense.

This is the follower’s weapon, working exactly as Gurley’s framework predicts.

Commoditise your complement

The cleanest way to read China’s open-weights strategy is microeconomic, not ideological. When the price of a complement to your product falls, demand for your product rises. So if you want to grow a market, drive the price of its complements toward zero.

Every major open source play in Gurley’s piece is running this playbook. Google made Android open source to commoditise the complement to search and ads - phones got cheap, search queries multiplied, ad revenue followed. IBM put a billion dollars into Linux to commoditise the complement to its services and hardware businesses - operating systems got free, IBM’s consulting business flourished.

China is now running the same play at a national-policy scale. The complement to an AI model is the value-added application built on top of it - every product, every service, every workflow, every agent. By making frontier-grade models freely available, China is driving the price of that complement toward zero for its entire economy. Every Chinese startup, every Chinese enterprise, every Chinese researcher gets to build on a stack without a rent floor.

And once the foundations of your application layer have no rent floor, you don’t just dominate at home - you export. Chinese AI-powered services, built on free Chinese models, competing in global markets against Western services built on metered API calls to a handful of proprietary vendors. That’s not just domestic industrial policy, that’s a global trade strategy. And China has been playing that game for a long time.

It’s also why the public-interest argument and the corporate-strategy argument converge. Cheap inputs grow the market for everyone who uses the input. Expensive inputs concentrate the market for everyone who sells the input. That’s not politics. That’s economics. And it works.

The Crypto Wars, again

We’ve been here before. In the early 1990s, the US government tried to lock down strong cryptography on national-security grounds. Phil Zimmermann’s release of PGP in 1991 made strong encryption freely available worldwide. In response, the US government classified cryptographic software as a munition under ITAR, threatened Zimmermann with a federal investigation, and pushed the Clipper Chip - a government-backdoored encryption standard intended to become the default for commercial communications. The framing was familiar: terrorists, child predators, foreign adversaries. The actual effect: American companies were locked out of foreign markets where strong crypto was already freely available, while the US public was offered weakened encryption with a built-in government key that never really took off to begin with.

The Crypto Wars ended around 2000 with the open ecosystem winning decisively. Unenforceable export controls collapsed - and PGP and foreign-made crypto were already everywhere. The Clipper Chip died because no one trusted it. And every secure thing you do online today - banking, encrypted messaging, e-commerce, password managers - exists because that fight was won.

Now revisit the AI policy debate of 2026 with that history in mind. The closed-stack incumbents - what Eric S. Raymond called “Cathedrals” in his 1997 essay on why open development always beats proprietary development - have powerful incentives to use China-hawk framing to push regulators toward restricting open competition. Because if you can’t beat the open ecosystem on quality and price, you have to beat it in Washington. On April 29, 2026, two House committees sent letters to Anysphere and Airbnb demanding information about their use of Chinese open-weight models. The companies chose those models on pure price-performance grounds. That didn’t matter. The fight isn’t actually about national security. It’s about whether American companies are allowed to use cheap, capable, open alternatives to the domestic proprietary vendors.

The Crypto War playbook, dusted off thirty years later.

The uncomfortable reframe

A simple, if inconvenient, conclusion follows: Chinese open-weight models are currently the only thing preventing a US proprietary oligopoly from becoming the global default.

Not Mistral. Capable, but a tier below the absolute frontier, and a single European company carrying the entire Western open-weight load.

Not Llama. Zuckerberg confirmed Meta would no longer release superintelligence-capable variants openly. Meta’s 2024 manifesto on open AI reads very differently today.

And not Gemma. Gurley’s piece is clear-eyed on this: Gemma is open only at small sizes, with Gemini firmly closed at the frontier. There’s also a Google-shaped precedent for what happens when this company runs an “open” project without a neutral foundation governing it - read Gurley on how Android got quietly recaptured. Treating it as structurally different from Llama because it has Google’s brand name on it is naive.

If you remove the Chinese open frontier from the picture, what’s left of the open-weight ecosystem capable of running an actual economy? Almost nothing. The follower’s weapon is working - but it’s working for every follower around the world, not just China. Every European nation. Every African and Southeast Asian government. Every Latin American startup. Every academic lab. Every American developer who can’t afford OpenAI’s enterprise pricing. They are all riding on infrastructure that Chinese labs are paying to develop and openly releasing - and Western policy is increasingly geared towards shutting that infrastructure down.

If Washington bans Chinese open weights on national security grounds, the models will still exist and still get used everywhere outside the US. Six billion people - three-quarters of the planet - will still pick the stack that’s free, capable, self-hostable. The US will end up technologically isolated from the majority of the world’s AI users, with a handful of proprietary models serving the domestic market. Those proprietary models may well sit at the top of the leaderboards. But that won’t matter - because the global infrastructure layer will have been built on someone else’s stack. An historically epic own-goal.

The framing

Press and analyst coverage frames the open weights debate as “can open beat closed at the frontier?“ - but that’s the wrong question. The market doesn’t need open to beat closed at the frontier. It needs open to be good enough that closed can’t charge rent. The market is already settling this question in open’s favour. Chesky didn’t pick Qwen because it beats GPT-5.5. He picked it because it’s good enough and it’s cheap. Anysphere didn’t pick Kimi because it’s the world’s best model. They picked it because it’s the world’s best model for their price point.

Open doesn’t have to win the race. It just has to circumvent the toll booth.

That’s the entire game. Linux didn’t beat Windows on consumer desktop, and it didn’t need to. It made operating-system infrastructure non-rentable, and the world’s economy now runs on top of it. Kubernetes didn’t beat AWS on cloud market share, and it didn’t need to. It made container orchestration non-rentable, and even AWS now contributes to it. The bar isn’t outperformance. The bar is preventing one company owning the layer everyone depends on. Then commoditise the complement, grow the surrounding economy, let the value capture happen at a different layer.

Apply that to AI. Open weights don’t need to outperform GPT-5.5 or Mythos. They need to exist in a form capable enough, cheap enough, and self-hostable enough that no proprietary oligopoly can charge platform-tax rent on the entire global economy. That bar is much lower than “beat the frontier.” That bar is already being cleared - by Chinese labs, mostly - and Western policy is actively trying to shut them out of the American market on grounds the Crypto Wars showed us, three decades earlier, were unenforceable and self-defeating.

The choice

There is a version of the next decade in which America protects four companies and loses the global architecture. There is another in which it recognises that the follower’s weapon is also America’s weapon - that the open ecosystem is the only thing keeping its own startups, researchers, and developers from being permanent renters of someone else’s AI infrastructure.

That choice is still open. But the window is closing.

In the next post: why who owns the bottleneck is the question that ultimately governs the economic benefits of AI - and why most of what’s being called ‘open’ today is only open… until it isn’t.

Read Part 2

The short version

Over the past few months, I’ve been writing a series on the structural risks of AI - economic concentration, geopolitical weaponisation, the cautionary tale of social media, and the case for governed open infrastructure as the defence. Each post followed the argument one step further.

Subscribe now

I walked away from a role I loved, at a company I deeply respect, working with people I’d built genuine trust with over many years (a quarter of my life, in fact). That wasn’t easy. But the argument kept arriving at the same place: the most important layer in AI - the protocols governing how AI agents communicate, coordinate, and act across systems, including the underlying infrastructure and tools - is being built right now. Whether that layer ends up open or captured will shape the next generation of technology, and the economy that runs on it. The window to get this right is finite. And it’s a layer I might know a little bit about.

Block - and my entire career - wouldn’t have been possible without the open infrastructure that others chose to build. Every system my coworkers and I built, every product we shipped, ran on foundations that existed because someone, years earlier, chose sharing over rent-seeking and control.

Open source has been part of my work from the start - I’ve contributed to and led projects in that space all my professional life. What’s new is the urgency. The agentic layer of AI is where those same principles need to land now.

In the last post, I asked engineers and companies building with AI to choose openness, and to demand it. This is me taking my own advice.

What the work looks like

The Agentic AI Foundation, housed under the Linux Foundation, is the place for open, vendor-neutral standards infrastructure, and tooling for AI agents. As CTO, I’ll be helping shape the technical direction - working with member companies, growing the contributor community, fostering ideas, innovation, collaboration - and setting up the governance scaffolding that protects what matters: freedom of choice, transparency, broad access, and healthy, open competition.

It’s hands-on work. The kind that doesn’t show up in headlines and probably never will. When infrastructure does its job, you stop noticing it. Nobody thinks about the protocols that make email work, or the standards that let you switch cloud providers. How often do you think about Linux kernel maintainers? That’s the goal.

The space is moving fast, and what we’re building today will evolve. Priorities will shift, projects will change, the technical landscape will look different in a year. That comes with working at the frontier of a new layer. What stays the same is the mission: keeping this layer open, governed, and hard to capture.

How to get involved

If you’re an engineer: contribute to open agentic infrastructure. Choose open protocols, open standards, and open implementations in what you build. In fact, demand them off your vendors. And give back to the commons. You’re following in the footsteps of the people who gave us Linux, open internet protocols and web standards, and much, much more.

If you’re a company building with AI: become a member. Participate in governance. Open standards mean the whole industry moves faster. Ideas circulate, trust is earned through transparency, and you compete on what you actually build, not on who captured the plumbing first.

If you want to follow along: I’ll be writing more. About what we’re building, what we’re learning, and where the hard parts are. Follow me here.

This isn’t the end of the conversation. It’s the beginning of the work. Come build with us.

Subscribe now

There’s a particular kind of decision that isn’t really a decision - where you’ve followed an argument far enough that there’s only one conclusion available to you. This is that post.

Where we are

Most AI today is call-and-response. You type a prompt, you get an answer. But AI is rapidly moving toward systems that act - that plan, delegate, and execute across multiple services and tools on your behalf. This is agentic AI - AI that has agency. In this world, the critical question isn’t which AI is best. It’s whether the systems those AIs operate in are transparent, open, and built on shared foundations - or whether they become an incomprehensible monolith that no one else can fathom, inspect, or leave.

Think about email. The protocols that make email possible are open. No one owns it. Gmail, Outlook, and Proton Mail all implement it differently and compete on speed, privacy, features, and user experience. You can switch between them without losing your contacts or your history. The standard is shared. The implementation is where the value lives.

Now think about social media. No shared protocol. No portability. No choice. The social media corporations own the layer, and everyone else plays by their rules.

The agentic layer of AI - the protocols governing how AI agents talk to each other, delegate tasks, work across systems, and take real-world actions - is being built right now. Whether it looks like email or social media is still an open question. But not for long.

Why this has to happen now

Imagine two futures.

In the first, the agentic layer becomes a commons - like Linux, like the internet protocols. Companies compete on the quality of their agents, their applications, their services. The best products win because customers choose them, not because switching is impossible. Anyone can run AI on their own terms, on their own infrastructure. Transparency is preserved because the protocols are open and inspectable. Governments can regulate what they can see. New entrants can build without asking permission. The value AI creates flows to whoever delivers the most benefit - not whoever captured the infrastructure first.

In the second, one or two companies set the standards before anyone else can. The coordination layer becomes proprietary. Agent interoperability exists only on the gatekeeper’s terms. The benefits of AI productivity compound upward into the hands of whoever captured the infrastructure first. The governance arguments from this series become moot - not because they were wrong, but because the architecture was already locked before anyone acted on them. The social media outcome, at larger scale, with higher stakes. In Post 2, I wrote about Leo Szilard - a man with the right instincts about governing a transformative technology, outmanoeuvred by the logic of national competition. That’s what failure looks like.

The classic tragedy of the commons is that shared resources get overexploited because everybody acts in their own self interest. The tragedy here is even worse: the commons never gets built in the first place, and the resource is captured before anyone realises it should have been shared. Unless we actively choose to build this, the default isn’t neutrality. It’s concentration. Momentum alone will carry us toward a company-dominated outcome - not through conspiracy, but through the same financial incentives that turned photo filter apps into machines for harvesting attention.

The question isn’t whether AI transforms the world. It will. The question is whether the transformation concentrates power - or distributes it so that all of humanity benefits. That question has an infrastructure answer. But infrastructure doesn’t build itself.

What can be done

It can be tempting to think this is a problem for someone else - for governments, for tech executives, for people closer to the machinery. But if you’re reading this, you likely have more influence than you think.

If you follow this space closely, you have a responsibility that goes beyond awareness. A few concrete things you can do today:

Choose with intention. Use platforms built on open standards. When you move away from closed ones, make the reason visible.

Talk to people in your orbit. Colleagues, founders, investors - most of them haven’t thought about infrastructure lock-in. You can change that.

Amplify the right signals. Share writing that names the problem clearly. Engage with it publicly. The window on AI infrastructure openness is still being set - and public discourse shapes what decision-makers think is politically and commercially viable.

If you’re an engineer, your choices carry structural weight. When you architect AI into your products, choose interoperability. Evaluate vendors on whether their systems use open standards or proprietary ones. Contribute to open protocol design. The engineers who built the internet’s open protocols chose resilience over proprietary control, and that choice shaped the next fifty years. Choose openness. Demand it.

If you’re a company building with AI, open standards are in your interest. Proprietary lock-in feels like a moat until the company on the other side of it is bigger than you - and then it’s a cage. Open, vendor-neutral standards mean you compete on the quality of what you build, not on who captured the plumbing first. The way web browsers compete today. That’s a market that rewards the best product, not the most entrenched.

The Agentic AI Infrastructure Foundation exists to build exactly this - open, vendor-neutral standards for agent communication and coordination, housed under the Linux Foundation. The same institutional home that governs Linux and the open cloud infrastructure that the technology industry already depends on. Open to contributors from any jurisdiction. Structured to resist capture by any single company or government. Get involved.

Following the argument

This series began with David Ricardo - an economist who built the framework for technological optimism, looked at it more carefully, and arrived somewhere uncomfortable. He followed the argument even when it led to an inconvenient conclusion.

That’s the spirit in which I’ve approached this problem - not just in writing about it, but in building the infrastructure to address it. I co-founded AAIF because the gap I kept writing about - the ungoverned layer above the models, the tier that social media proved was the most dangerous to leave undefended - is a gap I can help close. Not because I have all the answers. But because the structural foundation matters, the window to build it is finite, and the cost of getting it wrong is a world where the most powerful technology in human history is owned by the few and rented to the rest.

I followed the argument. It led here. And going forward, I’ll be getting more deeply involved - in the foundation, in advocacy, and in the work of making sure this layer stays open.

More on that - and my new role - in the next post.

Part 7

Subscribe now

In 1991, a Finnish student posted a message to a mailing list: “I’m doing a (free) operating system (just a hobby, won’t be big and professional like gnu).” That hobby now runs the majority of the world’s servers, all Android phones, and almost all of the cloud infrastructure on earth.

Post 4 ended with a claim: the open foundation that social media never got can still be built for AI. It’s not theoretical - this pattern has been working for decades, in plain sight.

What open infrastructure actually achieved

Two precedents. Different problems.

The first is the open internet. The protocols beneath every website, every email, every online service are open standards that no one owns. That’s why you can switch email providers and keep your history. That’s why a startup can build a website without asking permission. The power of open infrastructure isn’t that it’s free - free as in speech, not free as in beer. It’s that no one can take it away from you.

The second is the Linux Foundation. What it built isn’t just an operating system - it’s a governance model that removed core computing infrastructure from the competitive battlefield. Every company, government, and individual builds on the same foundation. Competition happens above it. The foundation itself is a global commons.

Together, these precedents address the risks explored earlier in this series. Open protocols preserve choice and prevent economic capture. Governed foundations prevent geopolitical weaponisation - and the anarchy of openness without accountability. Applying both principles higher up the stack - to applications and services - reduces the risk of the outcomes we saw with social media.

There’s an irony worth sitting with. The internet protocols were born as a US military project. ARPANET was funded by the Department of Defense in the late 1960s - the same institution that later tried to force Anthropic to remove ethical guardrails from its AI. The DoD inadvertently built the most open, globally shared infrastructure in human history. It wasn’t designed to become a commons. It became one because the engineers who built it chose resilience over proprietary control - and by the time anyone thought to capture it, the architecture was already open.

Architecture, established early enough, can outlast the intentions of its creators. That’s a promise when the architecture is open - and a warning when it isn’t.

That choice is available to every engineer building AI infrastructure today - and to every engineer consuming it. Choose openness. Demand it.

The CNCF extended this logic up the stack, building shared open infrastructure that means AWS, Azure, Google Cloud, Alibaba Cloud, and Tencent Cloud all compete on price and performance rather than lock-in. The infrastructure most of the technology industry depends on exists because of governance choices made when the technology was still young and malleable enough.

Open source and AI: what’s already working

The software and model layers of AI are already trending toward openness.

DeepSeek’s R1, released in January 2025, demonstrated near-frontier capability at a fraction of the assumed cost. What followed was a wave of open models competing directly with proprietary systems: Z.ai’s GLM rivals Claude on coding benchmarks, and Alibaba’s Qwen pushes boundaries in reasoning and code. The gap between open and proprietary is roughly half a generation, closing further with every release cycle.

The most striking signal: NVIDIA has committed $26 billion over five years to developing open models, and launched the Nemotron Coalition to build them collaboratively. When the dominant hardware company invests that heavily in openness, the momentum is hard to reverse.

The entire toolchain used to create these models is already open source - the infrastructure every lab, including the proprietary ones, uses to train and deploy AI. And AI labs across the world - including Chinese ones - are converging on shared standards for how software talks to AI models, meaning no single provider is irreplaceable. Switching costs are collapsing.

Any company can run a competitive open model on its own infrastructure today, without asking permission from the company that built it. That was unthinkable two years ago. The open source defence at the software layer isn’t a hope. It’s a fact on the ground.

Where it gets harder

None of this means the problem is solved. As we saw with social media, open foundations don’t prevent capture at the layers above - but they do make alternatives possible. Open foundations are the floor, not the ceiling. Necessary, but not sufficient.

There are challenges that open source alone doesn’t resolve. Manufacturing concentration: you can open-source a chip design, but someone still has to fabricate it - and right now that means TSMC in Taiwan, a single point of failure in one of the most geopolitically contested regions on earth. The physical layer: robots don’t copy at zero marginal cost. And a frontier capability gap, driven by funding asymmetry and financial incentives that currently favour proprietary development. These are real constraints. But they don’t invalidate the model. They demand answers that build on open foundations and go beyond them.

What a more robust defence looks like

Tristan Harris, in his 2025 TED talk, frames AI’s future as a choice between chaos and dystopia - decentralisation without responsibility versus centralisation without accountability. The narrow path runs through governed open infrastructure: open enough to prevent capture, governed enough to prevent chaos.

Governance is one half. But what about the financial incentives? The same forces that turned social media into an attention-harvesting machine are present in AI. Open infrastructure doesn’t change the profit motive. It changes where profit can be extracted. If the protocol layer is a commons, you can’t make money by owning the gate - you have to provide value. This isn’t anti-business. It’s pro-competition. The companies that build the best products, that deliver the most value, should win - because customers choose them, not because switching is impossible. Open infrastructure preserves transparency and choice.

That doesn’t eliminate bad incentives. It makes them contestable. And contestable is the precondition for everything else - regulation, competition, transparency, accountability - to work. You can’t regulate what you can’t see. You can’t enforce portability on a proprietary protocol. You can’t hold a system accountable if one company controls the entire stack.

Public compute - the equivalent of public roads for AI - is one piece. Antitrust applied to the AI stack is another. But the most urgent gap is in the protocols governing how AI agents talk to each other, delegate tasks, and work across systems - what technologists call the agentic coordination layer. This is where enormous value will flow. The standards aren’t set. The layer is still young enough to shape. This is the CNCF moment for AI - except the stakes are higher and the window is shorter.

If this layer gets captured before open standards take hold, we’re back to lock-in seen with social media. If it becomes a governed commons, we preserve the structural defence that open source built for computing - and extend it into the tier that social media proved was the most dangerous to leave undefended.

This is why I co-founded the Agentic AI Infrastructure Foundation.

The hobby Linus Torvalds posted about in 1991 became the foundation the world builds on. Just like Linux, when this work succeeds, it becomes invisible - and nobody goes to war over it. That invisibility is the achievement. And the aspiration.

In the next post: what the AAIF is building, and why.

🚧 Part 6 Coming Soon 🚧

Subscribe now

For most people today, AI is just a chatbot. That’s accurate… for now. But Facebook was once just a campus directory, and that evolved into something much more consequential. The choices made in those early years of social media locked in outcomes that are proving nearly impossible to reverse. What can we learn from social media, and apply to AI?

It was just a photo filter app

Facebook launched as a way for college students to connect. YouTube was a place to share home videos. Twitter was a way to tell people what you were having for lunch. That’s all these platforms were, at their inception.

No one intended what they became, their goals evolving in line with simple financial incentives like get people to sign up, keep them engaged, sell ads. But each of those simple goals compounded. Engagement optimisation became attention capture. Attention capture became behavioural manipulation. The dangerous emergent behaviour - anxiety, polarisation, the erosion of shared reality - wasn’t part of a dastardly plan. It was the logical outcome of simple incentives operating at scale with no structural accountability.

Each platform was designed as a proprietary walled garden, optimised to keep you hemmed in. Your followers, your content, your social graph - none of it portable. You couldn’t take your Instagram audience to a competitor any more than you could take your WhatsApp contacts to a different messaging app. The walls were the product, and there was no practical way to opt out. Your friends were there. Your photos were there. Your professional network was there. Leaving meant losing all of it. As Cory Doctorow put it: network effects attract users, switching costs take them hostage.

The business model that grew inside those platforms - free to use, funded by ads - produced the attention economy, one of the most consequential emergent forces of the last twenty years. When revenue depends on time spent rather than value provided, the incentive is to trigger compulsion, not satisfy need. The algorithm doesn’t care whether the hour spent made you feel better or worse, or helped you achieve your goals. It just cares that you spent the hour.

Most people reading this have felt the dismal effects. The app you open without deciding to. The argument that left you angry for hours. The moment you looked up and realised forty minutes had simply disappeared. What we imagine as personal failings are, in fact, a system doing its job.

They knew. It didn’t matter.

Even when the builders of these systems recognised the problem, the architecture made it nearly impossible to fix. Tristan Harris, a former Google design ethicist who co-founded the Center for Humane Technology, has documented how platforms exploit psychological vulnerabilities to capture attention - and awareness alone has failed to produce structural change. His 2025 TED talk makes an argument strikingly similar to this series: that we need to learn from social media’s failures before repeating them with AI. He’s right - and this series is an attempt to take that argument one step further, from diagnosis and a highly compelling call for awareness to infrastructure and governance.

Harris’s work, alongside the revelations of Meta whistleblower Frances Haugen, made clear that people inside these companies understood exactly what they were doing to teenagers, to democracy, to public health - and kept building anyway, because the architecture made any other choice economically irrational. The problem was never awareness. It was structure.

In 2022, Instagram’s Adam Mosseri stood on a TED stage and argued that blockchain-enabled technologies could remove platforms as middlemen, letting creators own their relationships with audiences directly. It was the most concrete vision anyone at Meta had offered for decentralising the platform. Less than a year later, Meta shut down the entire initiative. The explanation: they needed to “focus on other ways to support creators” - which boiled down to ads on Reels.

Mosseri’s talk was sincere. But decentralisation undermines the walled garden, where the revenue lives. The walls don’t come down because someone gives an inspiring talk. They come down when the incentives change. For social media, they never have.

Now imagine it replaces you

There’s a distinction between a technology that mediates human experience and one that substitutes for it. Social media was the former - it sat between you and the world. AI is the latter. It doesn’t sit between you and the world. It replaces pieces of the world itself.

The doctor-influencer you followed online is now the AI giving you a tailored diagnosis. The creative work you watched artists make is now generated on demand. The expertise you paid a lawyer for is replaced by a conversation with AI. Each genuinely valuable. Cheaper access to expertise, healthcare - real goods, especially for those who previously couldn’t afford them.

The problem isn’t the substitution. It’s who owns the system doing the substituting. Social media captured your attention, your time, and it turned out, much more. AI captures your function. When a proprietary system replaces what you do, what you create, what you know - the leverage isn’t behavioural. It’s existential. Same script, bigger explosion.

The walls are going up again

Decisions being made today about ownership of AI’s foundational infrastructure are this generation’s equivalent of the social media story. Except that the technology is more powerful, the transition faster, the concentration risk higher.

Without deliberate structural intervention, AI will follow the same path. Not because the people building it are malicious. Because there is too much money to be made building walls. The same incentives that turned a photo filter app into a machine for harvesting attention are present in AI, at a larger scale, with higher stakes.

Awareness hasn’t fixed social media. Good intentions haven’t fixed it. Congressional hearings haven’t fixed it. What was never built, when there was still a chance, was structural intervention - open infrastructure, portable data, shared governance, business models where you pay for the service rather than being the product. Think email: you can switch providers, keep your history, and no one owns the protocol. The same can’t be said for any of your social media feeds.

It can still be built for AI. I wouldn’t be writing this series if I didn’t believe the possibility of intervention exists. There’s a decades-long track record of it in computing, it’s already being applied to AI, and I’ve put my own career behind it. The answer starts with open infrastructure: systems no single company or government can capture, that anyone can build on, and that provide the transparency and choice that social media’s architecture never had. But open infrastructure alone isn’t enough - without shared governance, open becomes a different kind of unaccountable. The imperative is to build both.

In the next post, I want to get concrete: what does intervention look like, who is leading it, and why the work happening in open source AI foundations right now may be the most important effort most people have never heard of.

Read Part 5

Subscribe now

My first two posts laid out the economic risk and the geopolitical risk of AI. Both are structural, accelerating, and have historical parallels that should make us uncomfortable. But both made the same case: that there’s a window in which structural choices can still be made, and it’s closing.

This post is about that window - why I think it exists, why I see it as narrower than most people realise, and what I believe closes it.

The weavers who never saw the upside

The comforting story about technological disruption goes like this: transitions are painful, but institutions catch up. The gains redistribute. It works out.

I don’t buy it. Here’s why. Acemoglu and Johnson - from Post 1 - documented the impact of the Industrial Revolution on English handloom weavers. Real wages halved in barely a decade. The economic benefits of industrialisation didn’t reach workers in the span of an average working lifetime. The people who bore the cost never lived to see the upside.

But there’s a second act. Over decades, institutions caught up. Factory acts. Unions. Public education. Progressive taxation. The gains redistributed - not quickly, and not necessarily fairly, but they did. The pie grew, and some of those benefits were realised by all.

Why? Not goodwill. Those institutions formed because labour was still necessary. Factory owners needed workers. That gave workers leverage - to strike, to bargain, to make production stop. Every mechanism that forced redistribution traced back to the same truth: capital could not produce without labour.

What happens when that’s no longer true?

The standard rebuttal is that technology creates new jobs to replace the ones it destroys. The loom killed the weaver but created the factory mechanic. The internet killed the travel agent but created the web developer. True - but every previous revolution automated specific tasks, shifting demand to new human skills. AI is different. It doesn’t automate a task. It automates the capacity to do tasks. When the next new job category emerges - whatever the AI-age equivalent of “web designer” or “SEO consultant” turns out to be - AI will be able to do that too. The escape hatch for generations, retrain into the new thing, closes when the new thing is also automatable on arrival.

I build AI products and watch this play out almost weekly. Recently, I built and shipped a complex application - back-end services, mobile and web front-ends, geolocation, payments, audio and video - in days. I’m not a mobile developer. I’m a back-end and infrastructure engineer. I’d never built a multi-platform mobile app before. But with AI, I was able to ship something more polished than many professional apps on the market. Work that would have taken a team of five several months. Research that used to take me a week now takes under an hour. Understanding an unfamiliar codebase - weeks of reading - now takes hours. I know good engineers who’ve lost their jobs because a team of twenty can now be done by five - and the five ship faster than the twenty ever did. That’s not a comfortable thing to witness.

Each of these is a data point on a trajectory where the value of human labour trends toward zero. Not everywhere at once, but directionally - and faster than I would have imagined, even a year ago.

As a techno-optimist (yes, I am!) it’s tempting to see the good in this trend. Faster drug discovery. Cures for diseases, cheaper than ever. Maybe even a solution to climate change. AI can and will bring about enormous scientific advancements. But our existing economic systems will siphon the financial gains from those advancements to the few who own the AI systems.

Some - including Acemoglu himself - argue that AI’s near-term economic impact will be more modest than the hype suggests. He may be right about the pace. But the question I’m asking isn’t how fast the productivity gains arrive - it’s who captures them when they do. I could be wrong about the timeline. But I don’t think I’m wrong about the direction. If you work in tech, you’re probably already seeing it. If you don’t, ask someone who does how their job has changed in the last year. The answer will be more dramatic than you expect.

Three phases - and what we lose in each

The AI transition has three phases, defined not by what gets automated but by how much leverage labour retains.

Phase 1: Partial automation. This is now. AI handles cognitive tasks at the margins. Jobs don’t disappear, they compress. Fewer people do more work. The result is layoffs - we’re seeing a lot of them right now - and predicted productivity gains that accrue to the firms doing the cutting. Ownership of AI capability concentrates in a handful of companies. But labour is still necessary, and the traditional mechanisms of democratic pressure - unions, politics, the threat of withdrawal - still function, even if at diminished capacity. The trap is complacency. The numbers look fine while the groundwork for what comes next is being laid by the companies doing the disrupting - not by governments, not by unions, not by the public.

Phase 2: Broad displacement. Robotics catches up with cognitive AI. White-collar displacement accelerates. The tax base erodes as labour income falls. History is consistent: mass displacement without adequate response produces political radicalisation. Leverage to create change erodes. Strikes lose their teeth when you aren’t needed anyway - go ahead and strike, who cares? The trap is policy lag. By the time legislators grasp the scale, the leverage to respond may be gone.

Phase 3: Near-zero marginal cost production. Robots build robots. Solar powers everything. The cost of goods collapses. If productive infrastructure is privately owned, you get neo-feudal lock-in - the scenario from Post 1. If it’s publicly governed or broadly distributed, you get something closer to post-scarcity: basic material needs met at near-zero cost, human effort freed for work that’s chosen rather than coerced. The trap is irreversibility. Because labour is optional, strikes have no effect. No demand for labour means no labour market to tighten. The only remaining leverage is political - and by this point, the neo-feudal lords have already captured the political system entirely.

We’re in Phase 1, and soon entering Phase 2 and I expect AI-driven layoffs to continue. This isn’t just the period where structural choices can still be made. It may be the last time workers have the power to demand them.

The speed asymmetry

Even within the window, there’s a problem: everything that needs to happen operates on different timescales.

The AI capabilities I build with today are dramatically more powerful than six months ago. In three months, they’ll be more powerful again. Acceleration is exponential. Meanwhile, retraining takes five to ten years. Safety net reform takes ten to twenty. International governance takes twenty to thirty. The gap between how fast the technology moves and how fast institutions can respond is the core structural problem.

It’s the same mismatch that made the Industrial Revolution brutal for weavers, just faster.

Consider the early internet. Section 230, passed in 1996 - twenty-six words tucked into a telecoms bill, almost no public debate, forty million people online - gave platforms immunity from liability for what their users posted. Those twenty-six words determined the structure of the digital economy for decades, enabled the unchecked rise of social media giants, and we’re still arguing about the consequences thirty years later. The decisions being made right now about AI infrastructure are this generation’s Section 230 moment. Some are happening in the right places: standards bodies and open source foundations that are transparent and democratic by design. But many are being made in cloud computing contracts and chip export policies - behind closed doors, no public input, no oversight.

Taxing the feudal lord

Standard policy responses - corporate tax reform, UBI, robot taxes - are still necessary. But they redistribute gains after concentration has occurred. And they depend on ongoing political will, which is the very thing that erodes as labour leverage disappears. A tax can be repealed. A subsidy cut. A UBI pilot defunded. Each requires someone in power to keep choosing redistribution, every year, against the interests of those with the most resources to lobby against it.

Infrastructure-layer interventions are different. Who owns the machines. Who governs the protocols. What becomes open standard. These operate before concentration, and they’re structural rather than discretionary. An open protocol doesn’t need a politician to defend it every budget cycle. A commons, once established - like the Linux kernel, or TCP/IP - is harder to capture than a tax is to repeal.

Open infrastructure alone doesn’t feed anyone. If AI eventually produces everything at near-zero cost but no one has a job, an open protocol doesn’t put food on the table. But because the protocols are open, no one needs permission to build, to compete, or to participate. Compare that to mobile app stores, where two companies control the only distribution channels and take a cut of everything - that’s what a proprietary infrastructure looks like. Open infrastructure keeps the system auditable, governable, and critically, it provides choice - a government can run sovereign AI infrastructure for public services without depending on a foreign corporation. Without that foundation, taxation can be evaded, regulation lobbied away, and democratic oversight has no surface to grip. With it, those mechanisms have a fighting chance.

Policy asks how we redistribute what the system produces. Infrastructure asks who controls the system. Taxing the feudal lord is better than not taxing the feudal lord. But it’s not the same as preventing feudalism.

The transition window

The weavers had no agency over the transition that destroyed their livelihoods. But the institutions that eventually protected them formed because labour still had leverage: the ability to withdraw work and make the system stop.

We have one thing the weavers didn’t: hindsight. We can see where this trajectory leads. And like the weavers, we still have labour leverage - but unlike them, ours has an expiration date. The window in which ordinary people retain enough economic relevance to demand a different architecture is finite and closing. Just a little more every day.

In the next post, I want to look at what an infrastructure-layer response actually looks like. It turns out someone already built one - and you definitely use it every day without knowing it.

In Part 1, Ricardo followed the argument to an uncomfortable conclusion. In Part 2, Szilard tried to govern the knowledge before the race began. The weavers waited for institutions to catch up and paid the price of a generation. We don’t have a generation. We might have a decade. How are you preparing?

Read Part 4

In my last post, I laid out the economic risk: the structural problem of wages and productivity diverging, accelerating as AI automates cognitive labour. But at Policy Week Sydney, I shared two fears that keep me up at night. The second is geopolitical: that weaponised AI becomes a tool of national competition and state power rather than a shared resource. I find it hard to look away from.

Subscribe now

The physicist who saw it first

Leo Szilard was living in Bloomsbury, Central London, when, in September 1933, pausing to cross Southampton Row, he first conceived of the nuclear chain reaction. He had read H.G. Wells’ The World Set Free, which imagined atomic bombs and the destruction they would bring. Wells wrote it as a warning. Szilard read it as a blueprint that physics was about to make real.

Szilard’s first instinct was not to build the weapon. It was to govern the knowledge. Filing a patent on the chain reaction, and assigning it to the British Admiralty to keep it out of the public eye. He was trying to prevent an arms race.

It didn’t work. By 1939, with war approaching, Szilard co-authored Einstein’s letter to President Roosevelt that launched the Manhattan Project. He spent the rest of his life in the wreckage of that decision - not because it was wrong, but because the governance architecture he’d hoped for was never built. After Hiroshima, he argued that secrecy and national monopoly over existential capability doesn’t produce security. It produces escalation.

Szilard was a pacifist by instinct - yet he spent his career arguing against those instincts, because at each turn, the alternative was worse. Build the bomb before Germany does. Warn Roosevelt before it’s too late. Accept the standoff because unilateral disarmament was unserious. The logic of competition overriding the logic of peace, and each concession making the next one harder to avoid.

He was right. The Soviet Union didn’t refrain from building a bomb because the US kept the secret. They built one faster, because the secret confirmed the threat. The logic of mutually assured destruction created a world of blocs and borders, of us versus them. Monopoly over existential capability created a race to parity - and the race itself is where the danger lived.

But it wasn’t enough to be right. He was outmanoeuvred by the logic of national competition before governance structures could be built. Not because shared governance was impractical, but because it was politically inconvenient - and by the time the political will existed, the dynamics were already locked in.

Leopold Aschenbrenner also made the Szilard comparison, in his influential 2024 essay series on AI’s trajectory, Situational Awareness. He compellingly argued that people building frontier AI today may be in Szilard’s position: seeing clearly what’s coming while the rest of the world hasn’t caught up. Where I part ways with Aschenbrenner are the lessons I take from Szilard’s story. Aschenbrenner draws the nuclear parallel to argue for national securitisation - that the US must race to maintain its AI advantage. I follow the same parallel to the opposite conclusion: that the race itself is the primary danger. Not that the good guys need to win the race, but that the race, once started, produces bad outcomes for everyone.

With AI we’re already deep into the cascading logic of an arms race.

Anthropic and the Department of War

In July 2025, Anthropic (the company behind Claude) signed a $200 million contract with the Pentagon. Claude became the first frontier AI model approved for use on classified military networks. The contract contained two explicit red lines: no autonomous lethal weapons, no mass surveillance of US citizens.

A month later, the Pentagon rebranded as the Department of War.

In January 2026, Defense Secretary Pete Hegseth directed all DoW AI contracts to adopt “any lawful use” language - overriding Anthropic’s restrictions. When Anthropic held firm, the Trump administration ordered all federal agencies to cease using Anthropic products. Hegseth then designated Anthropic a “supply chain risk” - a classification historically reserved for foreign adversaries - meaning any company doing business with the US military must certify it doesn’t use Anthropic products. Given that eight of the ten largest US companies use Claude, the blast radius is enormous.

OpenAI stepped in to fill the gap, accepting the terms Anthropic had refused. 1.5 million users allegedly cancelled their ChatGPT subscriptions in response. Anthropic sued, calling the designation “unprecedented and unlawful.”

I’m telling this story not to take sides in a commercial dispute, but because it illustrates the structural problem. A company drew ethical red lines on how its technology could be used. The state moved to crush it. A competitor stepped in. As the EFF has pointed out, if ethical protections depend on the commercial decisions of individual companies, they’re revocable the moment a government applies enough pressure. The market signal is unmistakable: ethical limits on AI’s use by governments are a competitive disadvantage to be punished, not a standard to be upheld.

This is the Szilard question, live and in court. Can private actors set ethical limits on how the state uses transformative technology? Or will the logic of national security override every guardrail, just as it did with nuclear weapons? What precedent will that set for other state actors?

The arms race we’re building

Zoom out and the pattern is broader still. The US has pursued AI superiority through export controls - restricting advanced chip sales to China since 2022. The intended effect was to slow Chinese AI development. The actual effect: constrained by hardware limitations, Chinese labs were forced into more efficient architectures. In January 2025, DeepSeek released R1 - near-frontier capability at a fraction of the assumed compute cost. A control mechanism produced the opposite of its intended result.

This is the nuclear dynamic repeating. Each side’s attempt to maintain superiority validates the other’s fear and accelerates the competition. We’ve seen this story before, we know how it plays out and how the stakes escalate.

At each turn, the space for cooperation narrows. The most dangerous endpoint is geopolitical fragmentation of AI infrastructure: incompatible systems, no shared standards, no verification mechanisms, no shared foundation on which to build trust.

The counter-model

Szilard’s instinct - shared governance of transformative technology - failed for nuclear weapons. But it didn’t fail everywhere.

Open source software has quietly become the neutral ground on which geopolitical competitors build shared infrastructure. TCP/IP, HTTP, SMTP - the protocols the Internet runs on - are shared standards, used by every nation regardless of political alignment. The Linux Foundation and CNCF extended this logic into computing and cloud infrastructure, creating vendor-neutral governance around the technologies the modern economy runs on. Competition happens above the foundation - in applications, services, features - not at the foundation itself.

What makes this remarkable is who participates. Huawei - a company the US sanctioned and restricted from buying American chips - has been a top-five contributor to the Linux kernel for years, and at times the single largest corporate contributor. China ranks second in the world for contributions to CNCF projects. Alibaba, ByteDance, and Huawei are all active participants in Kubernetes development. These aren’t peripheral contributions. It’s the core infrastructure that runs the world’s cloud computing, and engineers from rival nations are building it together.

And critically: no government has weaponised this layer. Despite escalating tensions over trade, chips, and AI, neither the US nor China has attempted to fork the Linux kernel along geopolitical lines, or to deny adversaries access to Kubernetes. The shared infrastructure remains shared - because both sides depend on it, and because the governance structures make unilateral capture impractical. This is exactly the dynamic Szilard wanted for nuclear technology and never got.

Shared AI infrastructure standards - at the protocol layer, for example - could play a similar stabilising role. Not preventing competition at the application layer. And, powerfully, creating something the nuclear era never had: a technical foundation for verification, interoperability, and shared governance. The preconditions for trust, even between adversaries.

This isn’t a hypothetical for me. It’s why I co-founded the Agentic AI Infrastructure Foundation alongside several major AI companies - to try to build exactly this kind of shared, open infrastructure for the AI layer where agents communicate and coordinate.

The window

Szilard’s instincts were right. His timing was wrong. By the time the international community was ready to talk seriously about nuclear governance - the Non-Proliferation Treaty wasn’t signed until 1968, twenty-three years after Hiroshima - the dynamics were already entrenched.

Will the same logic outmanoeuvre us again? AI infrastructure is being built right now - protocols, standards, ownership structures. They pass through a brief window where the architecture is still malleable, and then lock in - and what looked like a choice becomes a permanent condition. As students of nuclear arms control can tell, negotiation becomes increasingly difficult beyond that point.

That window is narrower than most people realise. In the next post, I want to lay out why the next few years may be the only years that matter, and what might happen on the other side.

Szilard spent his last decades trying to put a genie back in a bottle. We might not have to. But only if we act before the genie escapes.

Read Part 3

“What’s your biggest fear with AI?”

Subscribe now

My response was twofold. My first fear is economic: the short-term disruption that’s already underway, and the risk that long-term policy responses fall short of what the situation demands. The second is geopolitical: that AI becomes weaponised; a tool of national competition and state power rather than a shared resource that benefits everyone. I didn’t have to reach for hypotheticals - but more on this in a future post.

What follows here is the first in a series of posts elaborating on both fears - not as abstract anxieties, but specific, structural problems with historical precedents and, I believe, potential solutions.

Let’s start with the economic risk.

The economist who changed his mind

In 1817, David Ricardo published Principles of Political Economy, establishing much of the intellectual foundation of modern economics. He argued - as most economists have argued since - that technological progress is ultimately good for workers. Machinery displaces some jobs, but it creates new ones, raising productivity and living standards overall. This went on to become something close to economic orthodoxy.

What’s less well known is that subsequently Ricardo quietly changed his mind.

In a new chapter added to the third edition in 1821, he revised his position, concluding that the introduction of machinery could harm the working class. That productivity gains didn’t automatically translate into shared prosperity. That the displacement of labour by capital was a real and persistent problem, not just a temporary adjustment. He’d worked through the numbers and arrived somewhere… uncomfortable.

He never trumpeted his reversal, but economists Daron Acemoglu and Simon Johnson recently highlighted it as one of the more significant - and underappreciated - moments in the history of economic thought. A man who had built the framework that justified technological optimism, looking at the same framework more carefully, concluded the exact opposite.

I find this story important because it models a willingness to follow an argument even when it leads somewhere inconvenient. That’s the spirit in which I want to approach this topic.

The divergence that’s already happening

On AI’s economic disruption: it isn’t coming. It’s already here. And it has been accelerating for decades.

Since the 1970s, productivity and wages in most developed economies have sharply diverged. GDP has grown. The economy has grown. Corporate profits have grown. The value produced per worker has grown. And yet median real wages - what the typical worker actually takes home after inflation - have been largely flat. The gains have accrued overwhelmingly to capital owners, not labour.

The evidence is there. It is measurable in the data of every major Western economy. The extent of divergence is sometimes debated, but the direction is not. The standard response is that this divergence has non-technological causes - globalisation, declining union membership, policy choices. All true. But when people claim “AI will create new jobs to replace the ones it displaces”, the last fifty years of evidence urge caution. Even if the jobs are created, the gains won’t automatically distribute themselves to the people who need them.

Previous waves of automation have largely displaced routine physical labour - repetitive and rule-based - freeing people to engage in higher value work. What’s different about this wave is that it displaces cognitive labour: the analysis, the coding, the diagnosis, the legal research. The jobs that the middle class retrained into after the last round of automation are now the jobs being automated. The escalator that was supposed to carry displaced workers upward is being removed.

A note on where the blame doesn’t lie

None of this is an argument against companies adopting AI. A company that can deliver better work, faster, at lower cost, is doing what companies are supposed to do. The problem isn’t that companies are making rational decisions. The problem is that the system surrounding those decisions has no mechanism to distribute the resulting gains. That’s not a failure of any individual companies and their decisions. It’s a failure of policy and institutional design. And it’s the level at which the response needs to operate.

The neo-feudal scenario

Follow this trajectory to its logical endpoint and you arrive somewhere economists don’t have great language for, but which we might call neo-feudalism.

Automation drives productivity gains that accrue to whoever owns the automated capital: the models, the robots, the infrastructure. That ownership concentrates rapidly in a small number of corporations and individuals. Concentrated capital compounds: the returns from owning productive infrastructure are reinvested into more productive infrastructure. Meanwhile, the labour income that used to distribute purchasing power broadly across the population shrinks. The middle hollows out. Not the very bottom, where some physical and care work remains hard to automate, and not the very top, where ownership confers returns regardless of personal productivity. But the broad middle, credited with its stabilising effect on democratic societies.

The natural political response is to raise taxes on corporate profits and redistribute accordingly. That may certainly be a necessary part of any response, but it faces serious structural headwinds. Corporate profits are the most mobile form of taxable income: companies can relocate IP, engage in transfer pricing, domicile earnings in low-tax jurisdictions (remember the Double Irish Dutch sandwich?). The OECD spent a decade negotiating a global minimum corporate tax rate of 15% and still faces significant holdouts. Getting from 15% to something that meaningfully funds a post-labour social contract requires a level of international coordination that has historically proven extremely difficult to achieve.

But the problem is deeper than the collection problem. Corporate tax revenue distributed through government programs isn’t the same thing as wages. Wages have historically acted as an automatic distribution mechanism, continuously putting purchasing power into the hands of a broad population. Replace wages with government transfers and you’ve introduced a political dependency that concentrated wealth tends, over time, to reshape in its favour. Political capture by self-interested elites happens all the time, eroding protections and safeguards offered to the average person.

This is the scenario I’m worried about. Not the old B-movie trope of robots enslaving humanity, but a quieter, more insidious shift with a small ownership class capturing the productive gains of automation, while the rest becomes economically peripheral. Not oppressed exactly, but just irrelevant. History is rich with examples of what societies look like when the economic foundation of the middle class is removed.

Here we have neo-feudalism.

Why the window matters

None of this is inevitable. But the window for structural solutions is narrowing. Decisions being made now - about who owns AI’s foundational infrastructure, about how it’s governed, about what becomes a commons and what becomes proprietary - will be much harder to reverse in a few years than they are today.

Ricardo publicly changed his mind in 1821, even though it was inconvenient. The question worth sitting with is whether we’re willing to do the same. To look at where the current trajectory actually leads, rather than where we’d like it to go, and to act on what we find.

There’s obviously a lot more to be said, so I’m splitting my thoughts into a few posts. Next, the geopolitical risk - and the story of a physicist imagining nuclear fission on a pre-war London street corner and spending the rest of his life trying to grapple with its implications. In some regards, AI may find its only parallel in the atomic bomb, so we might have a lot to learn from his story.

Read Part Two