The Coming Wave: Technology, Power, and the Twenty-first Century's Greatest Dilemma

What the book is actually arguing

Mustafa Suleyman co-founded DeepMind, sold it to Google, launched Inflection AI, and now runs Microsoft AI. He has watched the frontier labs from the inside for more than a decade. That vantage point is the reason to read The Coming Wave, and it’s also the reason to read it with one eyebrow up. The author warning us about the wave is the same author helping build it, and will be very rich when it breaks.

The thesis, stripped to the bone: a cluster of technologies — transformer-based AI, synthetic biology, robotics, quantum computing — is about to spread faster and cheaper than anything in history. We cannot halt that spread, because the incentives (strategic, commercial, scientific, humanitarian) all push forward. Suleyman calls the task of keeping powerful tools under human control the containment problem, and treats it as the defining challenge of the century. Fail, and you get either catastrophe (engineered pandemics, autonomous cyberweapons, cascade-failure infrastructure attacks) or dystopia, which is the surveillance apparatus needed to prevent catastrophe. The book is an argument that a narrow path between those poles exists, that we need to build it right now, and — this is the quieter line — that the author is not actually sure it’s possible.

We think this is the most useful framing of AI risk to come from someone currently shipping the technology. It is also less original than his reviewers claim. Nick Bostrom, Stuart Russell, and a decade of biosecurity literature have made most of these points before. What Suleyman adds is the insider’s texture: he was in the room, he knows these people, and he is plainly trying to atone on the page for what he has helped build.

Four features that make this wave different, and why three of them hold up

Suleyman argues that this particular wave is uncontainable because of four traits. Asymmetry: small actors can wield tools that once required a nation-state. Hyper-evolution: the cycle from research paper to deployed product has collapsed from decades to months. Omni-use: one general-purpose model can draft a legal brief, generate a virus candidate, or write malware — same weights, different prompt. Autonomy: the tools increasingly act without a human in the loop.

Three of these are genuinely persuasive. Asymmetry in particular is the point most of the AI-safety conversation refuses to sit with. When the marginal cost of a benchmark-level model drops toward the price of a laptop, the comforting assumption that only well-resourced labs can do this goes away, and with it most of the quiet hope that AI can be governed through licensing a dozen big companies. Suleyman is honest about this, to his credit, and his own ten-step containment plan includes choke points on chips and compute precisely because everything downstream is leaking.

The autonomy argument we find weaker. Suleyman slides between “agentic systems that book your calendar” and “systems that self-replicate and resist shutdown,” and the slide is too fast. The former is real and boring. The latter is speculative and has been speculative for fifteen years. Treating them as points on the same curve makes the near-term case feel more apocalyptic than the evidence warrants, and makes the long-term case feel more tractable than it is. We’d have preferred he picked a lane.

The omni-use point, though, is where The Coming Wave lands its cleanest punch. Every previous technology that caused containment problems (nuclear fission, aerosolized pathogens, centrifuges) had a narrow weaponization path you could gate on physical inputs. A large language model has no physical input. It is weights in a file. Suleyman’s observation that you cannot write an export control on a file that can be emailed, while governments are writing export controls on H100 GPUs, is the kind of specific, uncomfortable point that earns the book its reputation.

The nuclear analogy, and where he half-breaks it

Most books on AI governance reach for the nuclear arms playbook: treaties, inspections, the logic of mutual deterrence. The Coming Wave reaches for it too, and then spends a surprisingly honest chapter explaining why it won’t transfer.

Nuclear weapons, Suleyman argues, were partially contained by three accidents of physics. Enriched uranium is expensive. Weapons-grade facilities are large, hot, and detectable from orbit. The cost-benefit of using one is so catastrophic that mutual deterrence can hold a line. None of those conditions apply to the next wave. A gain-of-function experiment fits in a university lab. A frontier model fits on a hard drive. And the cost-benefit of use is not symmetric between superpowers — a small group can cause disproportionate damage without having anything for a superpower to retaliate against.

This is the book’s best sustained section, and it is also where it stops being purely an AI book. Synthetic biology is arguably a larger part of the argument than AI is, and the pages on CRISPR, benchtop DNA printers, and the collapsing cost curve of gene synthesis are sharper than the pages on AlphaGo and GPT. Suleyman seems more worried about a motivated graduate student with a DNA synthesizer than he is about a misaligned language model, and the book is better when he admits it.

That said, the analogy breaks down in a direction he doesn’t push hard enough. Nuclear weapons only matter when deployed. AI is deployed every time someone sends an API request. The containment problem for nuclear fission is a problem of discrete events. The containment problem for AI is a problem of continuous operation. Those require completely different governance models, and Suleyman reaches for treaty-and-inspection language that fits the first but not really the second.

Where the book hits hardest: the state is cracking at the worst possible time

The middle of The Coming Wave pivots away from technology and onto politics, and it is the middle that stayed with us longest after closing the book. Suleyman’s argument is that the liberal democratic state is in structural decline — fiscal strain, collapsing institutional trust, fragmentation of the shared information environment — at exactly the moment it needs to absorb an unprecedented governance load.

He is not subtle about this. He recounts trying to brief senior politicians on automation risk and watching them nod politely while visibly not absorbing a word. He describes the Copenhagen climate talks, where he was in the room, as a case study in why the modern state cannot coordinate on problems that don’t fit an election cycle. The implication lands hard: the political class that failed at carbon emissions, where we had forty years of warning and a clear metric, is the one we’re asking to govern a technology we don’t have forty years or a clear metric for.

This is where Suleyman’s inside-the-tent perspective is most valuable. He has sat with both rooms — the lab that builds the thing and the ministry that would theoretically govern it — and he can describe, with specificity, how far apart they are. The scene where a DeepMind ethics board he himself set up quietly dissolved under corporate pressure is one of the most useful pages in the book about why voluntary industry governance is theater. He built the thing, then he admits the safeguards he built don’t work. That kind of self-incrimination is rare.

We’d push further than Suleyman does, though. The book treats state fragility as a backdrop; we think it’s the main event. If AI and biotech are uncontainable by weak states, the first-order question is how to strengthen states, not how to design technology safety features. Suleyman gestures at this and then retreats. It’s the book’s biggest unforced error.

The ten steps, and why they read thinner than the diagnosis

The last third of The Coming Wave is a ten-point program for containment: an Apollo-scale safety R&D push, mandatory independent audits, choke points on strategic inputs, responsible builders inside the lab, new corporate forms, government capacity-building, international treaties, cultural norms around failure reporting, mass movements, and a synthesis he calls the narrow path.

Reading this list is like watching someone who has just spent three hundred pages arguing that containment is structurally impossible propose that we contain it via audits and public-benefit corporations. The diagnosis is systemic; the prescription is a policy wishlist. Many of the items are individually sensible — we agree that safety R&D is underfunded, that audits need teeth, that chip-supply choke points are load-bearing. But the book spends most of its length arguing that the incentives make coordinated action nearly impossible, and then fifty pages proposing coordinated action, and the gap between those sections is where the argument loses its nerve.

Suleyman half-admits this. He frames the ten steps as necessary but not sufficient, which is the kind of honesty we respect and also the kind of honesty that should have prompted a rewrite. If your own plan is necessary-but-not-sufficient, the book’s closing move ought to be an argument about what the sufficient part looks like, not a literary flourish about narrow paths.

The strongest item, for what it’s worth, is his insistence that safety-worried technologists should work from inside the labs, not outside. He argues that outside critics cannot shape a production system, only ornament one. This will annoy half his readership and is probably correct. He is describing his own career, which gives the argument both weight and obvious self-interest.

What the book leaves on the table

Three gaps we noticed.

First, almost nothing on labor. A book about technology that will reshape every industry, containing maybe ten pages on what happens to the people currently doing those industries, is a book missing a chapter. The only serious proposal we could find is a glancing reference to a robot tax. For an argument that fragile states cannot absorb shocks, the shock of mass labor displacement gets a shrug.

Second, thin on China. Suleyman acknowledges the US-China race and notes that any Western containment regime dies the moment Beijing decides not to participate. He then moves on. Given that his whole thesis depends on coordinated international action, and given that he has spent a decade close to both ecosystems, we expected a harder look at whether the Chinese Communist Party actually has different incentives on AI safety than a Western AI lab does, or the same ones dressed in different language. The book flinches.

Third, no serious engagement with open-source release. By 2023, when the book went to press, it was already clear that frontier-class model weights would leak, be re-released, and be impossible to recall. Suleyman barely names this. It is the single biggest hole in the containment thesis, and the book’s willingness to skip over it suggests the author was writing from the perspective of someone whose own labs ship closed-weight models.

Who should read it

If you are technical, impatient, and want the pessimism case made by someone who won’t embarrass you at dinner: read the first three parts and skim the fourth. It will take three hours, and the diagnosis is worth those hours even if you already think you’ve heard the argument. Suleyman is not as original as his reviewers claim, but he is more specific, and specificity is what the AI-risk conversation has been starving for.

If you are looking for a policy blueprint, read Dan Hendrycks or Helen Toner instead. The Coming Wave names the problem; it does not solve it. But naming problems precisely is the scarce good in this field right now, and on that measure Suleyman has earned the bestseller list. We think he is wrong about a few things and honest about more of them than most of his peers, and a book where the author visibly argues with himself is a better book than a confident one.