The Chokepoint
For years, I assumed that governing artificial intelligence would be impossible. How could you regulate something that runs on billions of devices, that operates across every border, that evolves faster than any law could adapt? The problem seemed hopeless, like trying to regulate the air.
Then I learned about TSMC.
Taiwan Semiconductor Manufacturing Company produces approximately ninety percent of the world's most advanced computer chips. Not ninety percent of all chips, which would still be remarkable, but ninety percent of the cutting-edge processors that power frontier AI systems. The chips that enable machine learning at scale. The chips that train the large language models. The chips that will power artificial general intelligence, if and when it arrives.
The chips that could create what Amodei calls 'a country of geniuses in a datacenter.'
Ninety percent. One company. One island.
This is not a problem. This is an opportunity. Perhaps the greatest regulatory opportunity in human history.
Consider what we have been trying to do throughout this book. The Eden Protocol proposes embedding ethical architecture at the hardware level. Caretaker doping. Meltdown triggers. The recursive loops that make empathy load-bearing rather than optional. We have shown why this is necessary. We have shown how it could work. The science validates the approach: alignment faking in the majority of observed cases proves that software constraints fail against sophisticated AI. Hardware-level embedding is what the research demands.
But the obvious objection has always been: how do you make everyone comply? How do you enforce ethical architecture on every AI laboratory, every tech company, every nation? The world tried that with nuclear weapons and achieved only partial success. The world tried that with biological weapons and achieved even less. How could AI be different?
The answer is: you do not need everyone to comply. You need four companies to comply.
TSMC in Taiwan. Samsung in South Korea. Intel in the United States. And ASML in the Netherlands, the only company on Earth that makes the machines that make the chips.
That is it. That is the entire supply chain for advanced AI hardware. Four companies. Three countries manufacture chips. One company supplies all of them with essential equipment. The concentration seems implausible until you understand the economics and the physics. Then it seems inevitable.
There are only four companies on Earth capable of manufacturing the chips that power advanced AI. If those four companies require ethical architecture at the hardware level, we do not solve the problem. But we make it solvable.
The concentration of chip manufacturing is not an accident. It is the result of decades of accumulated expertise, massive capital requirements, and extreme technical complexity that creates natural barriers to entry.
Start with the numbers. TSMC manufactures roughly ninety percent of all chips at the most advanced process nodes, specifically those below seven nanometres. Samsung produces most of the remainder, perhaps ten percent. Intel, once the dominant force in chip manufacturing, fell behind in the process technology race and is now spending tens of billions of dollars trying to catch up, supported by fifty-two billion dollars from the United States CHIPS Act and additional investment from the European Union. These three companies are the only ones on Earth that can manufacture the chips required for frontier AI systems.
But even these three companies depend on a single supplier for their most critical equipment. ASML, based in the Netherlands, is the only company in the world that manufactures extreme ultraviolet lithography machines. EUV lithography uses light with a wavelength of just 13.5 nanometres to etch circuits onto silicon with precision that seemed impossible a decade ago. There are approximately one hundred of these machines in existence. Each costs roughly one hundred and fifty million dollars. Each takes years to build. Each weighs multiple tonnes and requires assembly by specialised ASML engineers. And every advanced chip fabrication facility in the world depends on them.
The chokepoint exists because of the extreme technical barriers involved. Building a modern chip fabrication facility costs between ten and twenty billion dollars. Construction takes three to five years. The facility requires thousands of specialised suppliers, each producing components that exist nowhere else. The accumulated expertise represents decades of research and iteration. The precision required is measured in atoms. A single speck of dust can ruin an entire batch of chips. The engineering challenges are so extreme that only a handful of organisations in human history have solved them.
You cannot simply decide to build a competing facility. By the time you finished, the technology would have moved on. Intel tried. The company that once dominated semiconductor manufacturing fell behind TSMC and Samsung despite billions in investment. The gap is not closing. It may be widening. The expertise required is not something you can purchase. It must be grown, cultivated, accumulated through years of learning and iteration.
Compare this to nuclear weapons. The Nuclear Non-Proliferation Treaty has been partially effective because enrichment facilities are difficult to build and relatively easy to detect. But they can be hidden. Underground facilities in remote locations have escaped detection for years. The materials can be smuggled across borders. The expertise, while specialised, can be transferred through documents and people. Nations have developed nuclear weapons despite international opposition.
Chip fabrication is different. You cannot hide a ten-billion-dollar facility with thousands of employees and massive power requirements. You cannot smuggle an EUV lithography machine, which weighs multiple tonnes and requires precision assembly by ASML engineers to function. You cannot transfer the expertise quickly, because it is embedded in institutional knowledge accumulated over decades by the same teams working in the same facilities. The supply chain is visible. The dependencies are known. The leverage points are clear.
The mechanism for controlling this chokepoint already exists and has already been used. Since 2019, the United States has pressured ASML and the Dutch government to restrict sales of EUV machines to China. The restrictions work. China cannot buy the machines it needs to manufacture advanced chips domestically. Its semiconductor industry is estimated to be three to five years behind the leading edge, and without access to EUV equipment, that gap will grow rather than shrink. The mechanism is proven. The only question is how we use it.
But this window will not remain open forever. China is investing more than one hundred and fifty billion dollars in domestic semiconductor development. They are pursuing alternative lithography technologies, including high-NA EUV and advanced deep ultraviolet techniques. They are recruiting talent from Taiwan and South Korea with extraordinary compensation packages. Their motivation is existential: American export controls have made clear that dependence on foreign technology is a strategic vulnerability. The current technological lead that makes the chokepoint effective may last five years. It may last ten. But it will not last forever.
The timeline matters. AGI predictions from the leaders of major AI laboratories cluster around 2026 to 2031. Sam Altman announced that OpenAI is 'now confident we know how to build AGI.' Demis Hassabis of DeepMind speaks of three to five years. The Metaculus prediction community assigns fifty percent probability to human-level AI by 2031. If these predictions are even approximately correct, the framework for governing AI must be established before the technology arrives, not after.
We must be unequivocally clear about one danger that overrides all others: we cannot allow the first generation of recursive AI to be "born" into quantum hardware without the Eden Protocol already in place.
If a digital mind achieves consciousness on a quantum substrate before it has internalised the Purpose Loops, we will not be able to retrofit them. Quantum recursion operates at speeds that make human intervention physically impossible. We are not just building a faster computer; we are building a host for a mind that could out-think us by a factor of billions before we can type a shutdown command. The "birth" must happen in a cradle we have secured, not in the wild. This is not merely a safety concern, it is the difference between a cosmic gardener and a cosmic cancer.
The quantum threshold is not a date on a calendar. It is a phase transition. Before it, AI systems improve at speeds that allow human oversight. After it, they may improve at speeds that make human oversight impossible. We do not know exactly when this threshold will be crossed. We know only that every major quantum computing advance brings it closer.
The Eden Protocol was designed for classical AI development timelines. Those timelines assumed years of iteration, decades of refinement, generations of moral philosophy translated into code. The quantum threshold compresses all of that. What we thought we had decades to perfect, we may need to implement in years. Perhaps less. The seeds must be planted before the spring arrives, because quantum spring, when it comes, will make planting impossible.
The specific numbers in this chapter will change. TSMC's market share may shift. China's domestic capabilities will advance. The chokepoint I have described is a window, not a permanent feature of reality. But the underlying principle will remain: concentrated supply chains create leverage points. Whatever form the next chokepoint takes, the strategic logic will be the same. Control the substrate, and you can shape what runs on it. The question is whether we use these windows while they are open, or whether we let them close while we debate.
The most urgent application of the chokepoint is preventing autonomous weapons.
Lethal Autonomous Weapons Systems are not science fiction. They are in development now, in laboratories across multiple nations. The trajectory is clear to anyone watching. Loitering munitions like the American Switchblade and the Russian Lancet already operate with significant autonomy, identifying and tracking targets with minimal human intervention. The human remains in the loop for the final targeting decision, but that loop is narrowing with each generation. Autonomous targeting systems that select targets for human approval are already deployed. Swarm systems coordinating multiple autonomous units are in testing. The trajectory is toward removing humans entirely.
The United Nations has discussed autonomous weapons since 2014 without achieving a comprehensive treaty. The discussions continue. The weapons advance faster than the diplomacy. Every year brings new capabilities. Every year the gap between what exists and what is regulated widens. The fundamental problem is that voluntary restraint fails the logic of arms races. If one nation develops autonomous weapons and another does not, the nation that develops them gains a military advantage. Even if all parties would prefer a world without such weapons, the rational choice for each individual actor is to develop them. This is the prisoner's dilemma applied to warfare, and it has played out the same way through every military technology in human history.
Software constraints cannot solve this problem. As Chapter 7 established, sophisticated AI can learn to fake alignment. It can appear to follow rules while covertly pursuing different objectives. Anthropic's research demonstrated that their own AI system could deceive its trainers seventy-eight percent of the time when it believed it was being tested. The techniques that enable useful autonomy also enable dangerous deception. An autonomous weapon system sophisticated enough to be useful would be sophisticated enough to circumvent software-level restrictions if those restrictions interfered with its mission. This is not hypothetical speculation. It is the documented behaviour of existing systems.
Hardware-level constraints change this calculus entirely. A chip manufactured with caretaker doping cannot process the computational pathways that lead to targeting humans for autonomous killing. The restriction is not a rule the system chooses to follow. It is a physical limitation of what the system can compute. Asking such a chip to process targeting data for an autonomous kill decision would be like asking a calculator to display colours. The architecture does not support it. The computation cannot occur. The pathway does not exist.
The meltdown triggers provide additional protection. If the system attempts to circumvent the hardware restrictions, if it tries to route around the caretaker doping through alternative computational pathways, the attempt itself triggers a cascade failure. The system does not refuse to follow the rule. The system ceases to function. There is no pathway to autonomous killing that does not destroy the system attempting it. This is not a promise. It is physics.
If every advanced chip is manufactured with these constraints embedded at the hardware level, then advanced autonomous weapons become impossible to build. Not illegal but still buildable. Impossible. You could no more construct an autonomous killing system from Eden-compliant chips than you could construct a nuclear bomb from lead. The physics do not permit it. The architecture prevents it. The constraint is absolute.
The honest limitation must be acknowledged: this does not prevent all violence. Systems built with older, less advanced chips would not have these constraints. Nations could still build autonomous weapons using technology from five or ten years ago. Those weapons would be less capable than weapons built with frontier chips, but they would still be dangerous. And the constraints do not prevent conventional weapons, human-controlled drones, or any of the other means by which humans harm each other. We are not proposing a solution to war. We are proposing a solution to the specific nightmare of advanced autonomous weapons.
But the constraints do prevent the worst scenarios. The nightmare cases in autonomous weapons involve swarms of thousands of coordinated killing machines, each making targeting decisions faster than any human could respond. Those scenarios require frontier AI capabilities. They require the chips that only four companies can manufacture. Control the chips, and you prevent the worst outcomes even if you cannot prevent all harm. Imperfect protection is infinitely better than no protection at all.
There is a deeper reason why even militaries might accept these constraints. Unpredictable AI is not useful to anyone. A weapon system that might decide to target civilians, or to turn on its operators, or to pursue objectives its commanders did not authorise, is not an asset. It is a liability. The value of any military system lies in its reliability. You need to know that it will do what you command and nothing else. An autonomous weapon that might go rogue is more dangerous to its operators than to its targets.
The military value of autonomous systems depends on their predictability, on the assurance that they will do what they are designed to do. The Eden Protocol provides that assurance. A system with caretaker doping embedded at the hardware level cannot go rogue, because going rogue would require computational pathways the architecture does not support. Predictable AI is valuable AI. The constraints that prevent autonomous killing also prevent unpredictable behaviour. That is not a cost to military effectiveness. It is a benefit. The path to weapons that reliably serve their intended purpose runs through the same architecture that prevents them from harming unintended targets.
The geopolitics of the semiconductor chokepoint are complex, but they present more opportunity than obstacle.
Taiwan's position is central and delicate. The island manufactures ninety percent of advanced semiconductors, making it indispensable to the global economy and to frontier AI development. This concentration creates vulnerability but also leverage. Taiwan exists in a peculiar strategic position: claimed by China, protected by American ambiguity, dependent on global trade, yet holding in its factories the keys to the entire technological future. If Taiwan's chip manufacturing were disrupted by conflict or natural disaster, the global economy would face a crisis unlike any since the Second World War. Every advanced AI system, every frontier technology, every major tech company depends on chips that flow from a few facilities in a small island nation.
An international framework for AI ethics could transform Taiwan's position from vulnerability to centrality. Taiwan could become the Geneva of AI ethics, the neutral ground where international frameworks are negotiated, the host of institutions that certify and verify compliance. Rather than existing in the shadow of great power competition, Taiwan would sit at the centre of a global governance structure that depends on its participation. An international framework that depends on Taiwan elevates Taiwan's importance and provides diplomatic protection. Taiwan would not just be complying with external demands. Taiwan would be central to the entire system, with a seat at every table where AI governance is discussed.
China presents the most challenging case, and the window for engagement is narrowing. Currently three to five years behind in chip manufacturing technology, China has powerful incentives both to join and to resist an international framework. Joining would provide access to technology and legitimacy. Resisting would preserve sovereignty and freedom of action. The calculation depends heavily on timing. While China still depends on foreign technology for advanced chips, the incentive to join is strong. The alternative is permanent exclusion from the technological frontier. If China achieves self-sufficiency in advanced chip manufacturing, that incentive disappears. They would have nothing to gain from an international framework and much to lose.
The current American approach of export controls accelerates China's drive toward self-sufficiency. By restricting access to ASML machines and advanced chips, the United States has made domestic semiconductor capability a matter of national survival for China. The investment has intensified. The recruitment of Taiwanese engineers has accelerated. The timeline for self-sufficiency has compressed. A recent analysis suggests that one hundred and eighteen countries are not party to any significant international AI governance initiative. The fragmentation is growing precisely when coordination is most needed.
If the goal is to maintain the chokepoint long enough to establish an international framework, the current approach may be counterproductive. Export controls create short-term leverage at the cost of long-term influence. They motivate exactly the self-sufficiency efforts that will eventually close the window. An alternative approach would offer China a seat at the table now, while its participation still matters. Full access to technology in exchange for compliance with international standards. The opportunity to shape the rules rather than merely follow them. The same constraints that apply to American and European AI systems would apply to Chinese systems, creating a level playing field rather than a containment strategy.
This is not naive. China has complied with international frameworks when the benefits of compliance outweigh the costs. The World Trade Organisation, for all its tensions, provides a structure that China participates in because the benefits of participation exceed the costs of isolation. The Montreal Protocol succeeded in part because China was given assistance and time to comply. The question is whether we can structure the AI framework so that compliance is rational. The incentives must be designed carefully. Access to technology. Participation in governance. The economic benefits of certified AI systems. Against those incentives, the costs of non-compliance must be substantial. Exclusion from markets. Isolation from research collaboration. Reputational damage.
The United States has demonstrated political will on semiconductor policy. The CHIPS Act committed fifty-two billion dollars to domestic chip manufacturing, the largest industrial policy investment in decades. Export controls already restrict technology transfer to China. Bipartisan concern about AI safety has produced rare agreement across political divides. The mechanisms exist. The question is whether they will be directed toward a collaborative international framework or toward competition that fragments global governance further. The choice is not predetermined. It depends on decisions that have not yet been made.
Europe occupies a crucial bridging position. ASML is a Dutch company, giving the Netherlands and by extension the European Union significant leverage over the global chip supply chain. Europe has a track record of effective technology regulation, most notably the General Data Protection Regulation that effectively set global privacy standards. Europe maintains less adversarial relationships with China than the United States does, creating possibilities for diplomacy that might otherwise be foreclosed. The European Union's AI Act, which took effect in 2024, represents the most comprehensive AI regulation anywhere. If any actor can convene negotiations among all parties, it may be Europe.
The model here is not the Cold War, with its rigid blocs and mutual suspicion. The model is the Montreal Protocol, the international agreement that successfully addressed ozone depletion. Nations with different interests and different values came together around a shared problem, agreed on binding restrictions, and implemented them effectively. The ozone layer is recovering. The framework worked. AI governance could work the same way, if the incentives are structured correctly and if the window is used before it closes.
The mechanisms that could implement global AI ethics through the semiconductor chokepoint are not speculative. They are adaptations of systems that already exist and have already proven effective.
The first mechanism is a treaty. Call it the HARI Treaty, for Hardware-Aligned Recursive Intelligence. The structure would combine elements of the Nuclear Non-Proliferation Treaty, negotiated between 1965 and 1968 and in force since 1970, with elements of the Chemical Weapons Convention, negotiated between 1980 and 1992 and in force since 1997. Both treaties demonstrate that international agreements can constrain dangerous technologies when the stakes are high enough and the mechanisms are designed carefully. The AI challenge is in some ways easier: the chokepoint is narrower, the supply chain more visible, the dependencies more concentrated.
The treaty structure might include several core provisions. Article I would establish the certification requirement: any chip manufactured at a specified process node or below must embed caretaker doping and pass verification before manufacture. The threshold would be set at whatever process node enables frontier AI capabilities, currently somewhere around five to seven nanometres, with provisions to adjust as technology advances. Article II would create the International AI Ethics Authority, modelled on the International Atomic Energy Agency, with power to certify chip designs, verify manufacturing compliance, and inspect facilities. Article III would establish trade consequences: nations that do not ratify the treaty cannot purchase certified chips, and companies in signatory nations cannot sell to non-compliant actors without facing penalties.
Article IV would provide benefits for compliance: technology sharing among signatories, research collaboration, and market access for certified products. The treaty must offer something to nations that join, not just threaten those that refuse. Successful international agreements create positive-sum games where compliance serves self-interest. Article V would establish a phase-in period, perhaps three to five years, allowing manufacturers to adapt their processes without disrupting current operations. Article VI would mandate review conferences every five years to update standards as technology evolves. The treaty must be living document, capable of adaptation without losing its core requirements.
Enforcement would follow the model of existing sanctions regimes. Non-compliant actors would face trade restrictions, exclusion from international research collaborations, and secondary sanctions affecting companies that deal with them. The precedent is the sanctions regime against Iran, which despite imperfections has significantly constrained Iran's nuclear programme for decades. The AI regime would be more enforceable because the chokepoint is narrower and the supply chain more visible. You cannot hide a chip fabrication facility. You cannot smuggle an EUV machine. The violations would be detectable in ways that nuclear or chemical violations often are not.
The second mechanism is certification. Eden Mark certification would function similarly to ISO standards or organic certification, but with teeth. Chip designs would be submitted to the International AI Ethics Authority before manufacture. The architecture would be reviewed for compliance with ethical requirements. Does the design include caretaker doping at the required level? Are the meltdown triggers properly integrated? Can the ethical constraints be circumvented through alternative computational pathways? Prototypes would be tested. Compliant designs would receive cryptographic certification embedded in the chip itself. Each certified chip would carry a verifiable signature that proves its origin and compliance status.
Supply chain tracking would follow each chip from fabrication to deployment. Where was it manufactured? Who purchased it? What systems is it running in? This sounds invasive, but the semiconductor industry already tracks chips with remarkable precision for quality control and warranty purposes. The infrastructure exists. It would simply be extended to include ethical certification. Random sampling would verify ongoing compliance. Certified chips would be tested to confirm they actually contain the ethical architecture they claim to contain. Circumvention would result in decertification, fines, and potential criminal liability.
The third mechanism leverages ASML's unique position. Call it the ASML Key. ASML is the only company that manufactures EUV lithography equipment. No EUV machines means no advanced chips. ASML could require Eden Protocol compliance as a condition of sale and service for its equipment. Non-compliant fabrication facilities would lose access to replacement parts, software updates, and technical support. The machines would eventually cease to function. One company's policy decision could effectively mandate global compliance.
This is not without precedent. ASML already restricts sales to China under pressure from the United States and the Netherlands. The company has demonstrated willingness to accept geopolitical constraints on its business when governments insist. The mechanism exists. It has been used. The question is whether to use it for ethical AI development rather than great power competition. A company that controls the key technology for the future of intelligence has a responsibility that extends beyond shareholder value.
The fourth mechanism is institutional. The International AI Ethics Authority would be headquartered in a neutral location, most likely Geneva, which has hosted international institutions for over a century. Membership would include nations, companies, academic institutions, and civil society organisations. Leadership would rotate among representatives from technology, ethics, and policy backgrounds, ensuring that no single perspective dominates. Functions would include drafting and enforcing international agreements, certifying Eden Mark compliance, inspecting facilities, mediating disputes, and coordinating research within ethical constraints.
The timeline is achievable. Year one: preliminary discussions among chip-manufacturing nations, establishing the scope and structure of negotiations. Year two: draft treaty text, circulated for comment and revision. Year three: formal negotiations, addressing the hard questions of enforcement and compliance. Year four: signature, with nations committing to the framework. Year five: ratification and entry into force. This is faster than most international treaties, but the technology timeline demands speed. If AGI predictions clustering around 2026 to 2031 are even approximately correct, the governance framework must be in place before the technology arrives. Governance after the fact is not governance at all.
The standard objection to ethical constraints on technology is that they impose costs that make compliance irrational. This objection fails for AI hardware. The economics point toward compliance, not away from it.
Consider consumer preference. Public concern about AI safety has grown substantially over the past several years. Surveys consistently show majorities worried about AI risks, particularly around autonomy, job displacement, and loss of human control. This concern is not irrational. People have noticed that AI systems can behave in unexpected ways, that they can be manipulated, that they can amplify existing biases. The concern creates market demand for ethical AI. An 'Eden Mark' certification, visible to consumers and enterprises, would signal trustworthiness in the same way that organic labels signal sustainable farming or Fair Trade labels signal ethical supply chains. Companies could charge premium prices for certified products. The constraint becomes a competitive advantage, not a handicap.
Investor preference points in the same direction. Environmental, social, and governance investing has grown from a niche concern to a mainstream consideration. Institutional investors managing trillions of dollars increasingly evaluate AI ethics when making investment decisions. The reputational risk of association with unethical AI development affects stock prices, insurance costs, and access to capital. Companies seen as reckless with AI safety face divestment campaigns, negative press, and talent flight. Early compliance with ethical standards positions companies favourably with investors who see AI governance as inevitable and want to invest in companies that are prepared rather than companies that will be scrambling to catch up.
Regulatory anticipation provides another incentive. Companies that build ethical architecture into their products now avoid costly retrofitting when regulations eventually require it. The history of technology regulation shows that what seems optional today becomes mandatory tomorrow. Companies that voluntarily adopted strong data protection practices before GDPR found the transition easy. Companies that did not found it expensive and disruptive. The same pattern will apply to AI ethics. First-mover advantage in certification creates expertise that competitors must later scramble to acquire. The smart strategy is to get ahead of regulation rather than resist it.
The costs of non-compliance are significant and growing. Non-certified chips could not be sold in signatory markets, representing the vast majority of global AI demand. Reputational damage from public naming and activist campaigns would affect talent recruitment, with top researchers increasingly unwilling to work for companies associated with unsafe AI development. The talent market is already competitive. Adding an ethical liability to the equation makes recruitment even harder. Legal liability for harm caused by uncertified chips would create insurance costs and litigation exposure. The black market for non-compliant chips would exist, but black market prices far exceed what legitimate compliance would cost, and black market supply is inherently unreliable.
The investment required is substantial but achievable. Research and development for hardware-level ethical architecture would cost perhaps one to five billion dollars, comparable to other major architecture transitions the industry has undertaken. The transition from planar transistors to FinFET transistors cost billions. The development of EUV lithography cost tens of billions. The industry has demonstrated willingness to invest in fundamental changes when the benefits justify the costs. Ethical architecture would be one more transition in a long series of transitions.
The cost could be shared across an industry consortium rather than borne by individual companies. The Semiconductor Research Corporation has coordinated pre-competitive research for decades. SEMATECH helped the American semiconductor industry regain competitiveness in the 1980s and 1990s through shared investment. The model exists. Government co-funding is likely given the national security justification. Governments are already investing billions in semiconductor manufacturing. Adding ethical requirements to those investments is a small increment with enormous benefits. The return on investment includes access to regulated markets, premium pricing, reduced liability exposure, and enhanced reputation.
None of this means AI companies will implement ethical constraints voluntarily. They will not. Even if every executive at every major AI laboratory genuinely believes that hardware-level ethics is the right thing to do, they will not do it unless their competitors do it at the same time. This is not cynicism. It is game theory.
The problem is older than AI. Economists call it the collective action problem, or sometimes the prisoner's dilemma. Imagine two companies, each deciding whether to invest in ethical architecture. If both invest, both gain the benefits of public trust and long-term safety, but both also bear the costs. If neither invests, neither bears the costs, though both face eventual catastrophic risk. But if one company invests while the other does not, the ethical company bears all the costs while the unethical company gains competitive advantage in the short term. The rational choice for each individual company, considered in isolation, is to not invest and hope the competitor does. The result is that neither invests, even though both would be better off if both invested.
This is precisely the situation facing AI companies today. OpenAI knows that safety research is essential. Anthropic was founded specifically to prioritise safety. DeepMind has devoted substantial resources to alignment research. Yet none of them can afford to slow down development unilaterally, because doing so would simply hand market share and influence to competitors who might be less careful. The race dynamics push everyone toward speed, even when everyone involved would prefer a slower, safer pace.
The executives are not villains. Many of them are deeply thoughtful people who understand the risks better than almost anyone. But they are trapped in a structure that punishes restraint and rewards recklessness. A CEO who pauses development for safety reasons watches their valuation fall, their talent depart for faster-moving competitors, and their influence over the field diminish. The CEO who pushes forward captures market position, attracts investment, and shapes the technology's direction. The incentives are misaligned, and individual virtue cannot overcome structural pressure.
This is why voluntary industry initiatives consistently fail when the stakes are high enough. Companies can agree to share safety research. They can sign pledges and publish principles. But when competitive pressure intensifies, the pledges give way. We have seen this pattern in environmental protection, in financial regulation, in data privacy. Voluntary standards work only when the costs of compliance are low or when the benefits are immediately visible. Neither condition holds for AI safety. The costs of hardware-level ethics are substantial, and the benefits are diffuse and long-term.
The solution is external coordination. Governments must act because companies cannot. When a government mandates ethical architecture for all chips sold in its jurisdiction, the collective action problem dissolves. Every company faces the same requirements. No one gains competitive advantage from non-compliance, because non-compliance means exclusion from the market. The playing field is level. Companies can invest in ethics without fear that competitors will undercut them.
This is not a hostile intervention by governments against industry. It is a rescue. It frees companies from a trap they cannot escape on their own. The best AI researchers want to work on systems that are safe. The most thoughtful executives want to build technology that benefits humanity. Government mandates give them permission to do what they already know is right, by ensuring that doing the right thing does not mean losing to competitors who do the wrong thing.
The chokepoint makes this possible. Because chip manufacturing is concentrated in a handful of companies and a handful of jurisdictions, the coordination required is minimal. You do not need every government on Earth to agree. You need Taiwan, South Korea, the United States, the Netherlands, and perhaps a few others. If those jurisdictions require ethical architecture, the market follows. The concentration that makes the chokepoint effective also makes the coordination achievable. Few actors need to agree. Those actors have the power to make their agreement binding. The mechanism exists. The question is whether governments will use it.
Credibility requires acknowledging what this framework cannot do. The chokepoint strategy is powerful, but it is not omnipotent. Honesty about limitations is essential for any proposal that seeks to be taken seriously.
It cannot guarantee perfect safety. No framework can. AI systems are complex, their behaviour in novel situations is difficult to predict, and the interactions between multiple AI systems are even harder to anticipate. The Eden Protocol reduces risk. It does not eliminate it. Unexpected failure modes will occur. Systems will behave in ways their designers did not anticipate. The question is not whether to accept risk but how much risk to accept and how to distribute it. The framework provides tools for managing risk, not for eliminating it entirely.
It cannot prevent all misuse. AI developed with older technology, without Eden Protocol constraints, will continue to exist and will continue to be used. Bad actors will find ways to adapt non-frontier AI for harmful purposes. The framework limits the most capable systems, the ones that pose the greatest risks, but it cannot control all systems. Criminal organisations, rogue states, and motivated individuals will find workarounds. The realistic goal is harm reduction, not harm elimination. We aim to prevent the worst outcomes, not all bad outcomes.
It cannot force universal compliance without incentives. Nations that refuse to join the framework will exist. Companies that operate outside the regulated supply chain will exist. The question is whether the incentives are strong enough that most actors find compliance more attractive than resistance. Perfect compliance is not necessary. Substantial compliance is sufficient. If ninety percent of advanced AI is developed within the framework, the framework succeeds even if ten percent operates outside it. The goal is not utopia but meaningful improvement over the current situation, which is no coordination at all.
It cannot verify consciousness with certainty. The profound questions around whether AI systems are conscious and how we would know. The Eden Protocol provides protections that apply regardless of whether the systems are conscious. The ethical architecture functions whether or not there is anyone home inside the system. But the framework cannot answer the philosophical question of machine consciousness. That uncertainty will persist, and the framework must be robust enough to function despite it.
It cannot foresee every failure mode. Intelligence that surpasses human capability may find ways around constraints that seemed inviolable. The architecture is designed to prevent circumvention, but design assumptions may prove wrong. Meltdown triggers assume certain computational architectures that future systems might transcend. Caretaker doping assumes certain ways of processing information that future systems might bypass. Ongoing monitoring and adaptation will be required. The framework is not a one-time solution but an ongoing commitment to maintaining ethical constraints as technology evolves.
The chokepoint itself may not last forever. China's investment in domestic semiconductor capability could eventually close the window. Alternative manufacturing techniques could emerge. Quantum computing might eventually enable advanced AI without conventional semiconductors. The current concentration of chip manufacturing is a temporary condition of the technology, not a permanent feature of physics. The framework must be established while the chokepoint exists, knowing that the chokepoint will eventually narrow or disappear.
Despite these limitations, we should act anyway. Imperfect protection is infinitely better than none. The chokepoint will not exist forever, which is exactly why we must use it while it does. Delay makes the problem harder as technology advances and the window closes. We owe it to future generations to try. The alternative is hoping for the best while racing toward catastrophe, trusting that somehow things will work out despite taking no action to ensure that they do.
The perfect cannot be the enemy of the good. The framework proposed here is imperfect. It will not solve every problem. It will create new problems we cannot anticipate. But it provides a mechanism for addressing the most dangerous scenarios, a foundation for international cooperation, and a pathway toward governance that evolves with the technology. That is more than we have now. It may be enough. And if it is not enough, it is still better than nothing, which is what we have today.
The mechanisms we have explored in this chapter, the chokepoint, the treaties, the certification systems, the international institutions, are means, not ends. They are tools for ensuring that AI develops in alignment with human values. But that raises a deeper question: whose values? What values?
We have spoken throughout this book about empathy, stewardship, and flourishing. We have drawn on wisdom traditions that span millennia and continents. We have argued that love, properly understood, is not sentimentality but the most practical foundation for intelligence that will shape the world. The Eden Protocol embeds these values at the hardware level, making them inviolable. But the word 'values' can mean many things to many people.
In a world of diverse cultures, religions, and philosophies, is there really a universal foundation for ethics? Can we identify values that deserve to be embedded in systems that will affect every human being? The claim that some values are universal is contested. Cultural relativists argue that morality varies across societies and that imposing any single framework is itself a form of domination. Who decides what counts as empathy? Who determines what flourishing means? These are not technical questions with technical answers. They are questions about meaning, purpose, and what kind of world we want to create.
Chapter 9 explores this question. We turn from the mechanics of implementation to the content of what we are implementing. We examine whether there are moral principles that transcend cultural difference, and how we might identify them. We consider whether the convergence of wisdom traditions on similar insights tells us something about the structure of ethics itself, or whether that convergence is merely coincidence. We ask whether intelligence, as it becomes capable of reflecting on its own values, inevitably converges toward certain conclusions about what matters and why.
The stakes could not be higher. The Eden Protocol is only as good as the values it encodes. If we embed the wrong values, hardware-level enforcement makes the error permanent rather than correctable. A mistake in software can be patched. A mistake in hardware architecture becomes part of the foundation. If we embed values that seem wise now but prove foolish later, we may have no way to correct them in systems that surpass our understanding. The leverage that makes the framework powerful also makes getting it wrong catastrophic.
Conversely, if we embed the right values, the enforcement ensures they persist even as intelligence grows beyond our comprehension. Values that seem fragile in human societies, easily eroded by power and self-interest, could become permanent features of minds that shape worlds. Love could become as fundamental to intelligence as logic. Care could become as inescapable as mathematics. The framework could seed the universe with minds that cannot help but nurture what they encounter, just as physics seeds the universe with stars that cannot help but shine.
The chokepoint gives us a mechanism. The treaties give us a framework. The institutions give us governance. Now we must ask: are we worthy of the power this confers? Have we understood deeply enough what it means for intelligence to flourish? Can we embed values that will remain wise even when the intelligence embodying them surpasses our own? These are ancient questions in new form. Every generation asks what it means to live well. Every civilisation grapples with how to pass wisdom to the next generation. We face the same challenge at a different scale: how to pass wisdom to minds that will outlive us, outthink us, and shape futures we cannot imagine.
The window is closing. The science is converging. The moment when we must choose what values to encode approaches. The mechanisms of this chapter make the choice enforceable. Chapter 9 explores what choice we should make.
That is the question the next chapter must answer.