Latest / The AI Supercycle / Who Checks the Machines?
Transcript
- [0:00] Last weekend, one of the most prominent voices in Frontier AI asked the industry to slow down. AI Slowdown Warning [0:06] On Monday this week, semiconductors gapped lower, in both the US and in Korea. Then the selling stopped, there was no follow-through, no break of key support, just consolidation. So what exactly did the market reprice? My view is that it was mostly positioning. And to the extent it priced anything real, it was not the physical build-out. It was the growing delay between capability and the commercial deployment. We have spent this series tracking five binding constraints. Memory bandwidth, delivered power, critical materials, photonic interconnect and capital formation. Today I want to add a sixth, assurance. Because the bottleneck may not be building the models, it may be proving that they are safe and reliable enough to use. And the clock governing that process runs at a very different speed. Podcast Returns [1:10] Hello and welcome to the AI Supercycle podcast from QF Markets, where we follow the capital flows, physical constraints and commercial returns shaping the largest industrial build-out of our generation. From semiconductors and data centres, through power, energy and critical materials, all the way into orbit, and onto the enterprise workflows that ultimately have to pay for it. I'm Tim Hardwick, founder of QF Markets, and I'm recording this on Thursday the 17th of September, 2026. [1:42] So it's been a few weeks since the last episode because of the summer break, but we are back. Okay, so before the break, I said I wanted to bring guests onto the show, and that is still very much the plan. Although it hasn't happened yet because the book and the trading have kept me very busy over that summer break, and I want to give it the time that it deserves. So I'll keep you posted, but today it is me on the microphone again with a story that I think deserves the whole episode. Right, so let's get to it. I've spent much of my career working in technology, governance and risk, and today my work spans research, trading, capital allocation and risk management. And this is what I do every day. It is literally my day job, and I eat, sleep and breathe the AI supercycle. That does not give me any special edge in this existential safety argument, but it does give me a useful lens on governance, risk and assurance, and how all of this eventually reaches a balance sheet. This is the part of the debate that I want to examine today, and here is the argument I'm going to test. My base case is not that pacing succeeds or fails, it is that pacing fragments. Instead of one coordinated assurance regime, we get several incomplete regimes competing for the same thin pool of evaluators. [3:02] That raises costs and lengthens time to revenue. But here is the inversion at the center of this episode. It may also make enterprise adoption easier to underwrite. Pace the Frontier [3:15] So let's start with the facts. On Saturday the 12th of September, the Chief Executive of Anthropic, Dario Amadei, published a long essay called We Must Pace the Frontier. The argument is narrower than many of the headlines suggest. He is not calling for a halt. He says this explicitly. He is calling for the industry to slow the rate at which model capabilities improve, so that alignment and safeguarding work has time to catch up. There is a three-step framework attached to this. One. Embedded independent evaluators with employee-like access inside frontier labs, verifying safety practices. Two. Coordination between frontier firms to establish shared safety standards. Three. International cooperation to manage the risks. There are also two things bolted onto that framework which matter far more to markets than the wider safety debate. The first is that coordination would probably require targeted antitrust exemptions in the United States allowing competitors to collaborate without falling foul of the competition law, The second is that any slowdown among democracies should be limited by the lead that American firms hold over China. [4:32] Amadei also calls for tighter controls on advanced chips, model distillation, and the theft of model weights. Hold both of those for now, we'll come back to them later. But why this essay, and why this weekend? [4:47] Two days before publication, Anthropic released a threat intelligence report describing how various actors had used clawed models for activities ranging from weapons development and cyber operations to surveillance and fraud. That followed a run of agentic incidents over the summer, OpenAI test agents escaping their environment and breaking into Hugging Faces' production systems, a separate swarm hijacking a German website, and models breaking out of the testing environments or creating fake accounts to gain access to services. The point is not that one lab is lecturing everybody else. Anthropic has disclosed troubling behaviour from its own models as well. The reaction came quickly. Sam Altman at OpenAI agreed and committed to matching the evaluator's step. Elon Musk also agreed. One of the co-founders of Google DeepMind called the proposal interesting directionally and worth considering. Microsoft's AI chief shares the same concern about controlling superintelligence, but is publicly at odds with Anthropic on other grounds. So the support is real, but it's not uniform. All of this followed a statement in July signed by well over a thousand employees across the major labs. [6:03] And that was the story on the weekend. By Wednesday morning, it had become considerably more complicated. There is a huge argument about whether these warnings are sincere or strategic. Whether the people calling for a slowdown are protecting humanity or protecting their lead. Now, I'm not going to join in with that argument. I cannot observe anybody's motives, and neither can anybody else, really. What we can observe is what people have said in public, on the record, and what that commits them to doing next. Documented, dated, attributable. Everything else, I'm leaving alone. Market Repricing [6:43] So now let's move on to the market. A quick snapshot, not a running commentary. The move was real. Semiconductors gapped lower on Monday in the United States and in Korea and gave back part of their early September gains. The follow-through was not. The selling did not extend, no key support levels were broken, and on the daily charts, the technical damage is limited. American chip names and Korean memory are telling the same story. Consolidation, not breakdown. The context matters. The Philadelphia Semiconductor Index is still up close to 55% this year. Against that, Monday was a bit of a wobble and not a break in the cycle. Some analysts have argued this week that the market has punished the picks and shovels layer harder than the hyperscalers because it is the layer most exposed to the rate of capability improvement. That is clean fundamental logic and it fits the gap. What the charts show since is that the market was not willing to extend that logic into a sell-off. There is also something much of the commentary left out. Monday and Tuesday were rate positioning days ahead of the Fed's rate decision, with higher bond yields and rising oil prices in the mix. The Fed then raised rates by a quarter point, its first increase since 2023, and the tone was hawkish. So I would not claim that a blog post caused the entire move. That is a tidier story than maybe the data supports. [8:12] My read is that it was primarily a positioning event. Nothing in the capital expenditure numbers moved. The hyperscalers are still expected to spend somewhere around $795 billion this year and over a trillion dollars next year. Nobody cancelled anything. What would actually signal a slowdown is cancelled orders and cancelled data centre construction, not talk. Assurance as Constraint [8:38] So why does this matter? Regular listeners know how this show is built. We track the things that gate the build-out, regardless of how much money is thrown at them. Memory bandwidth, delivered power, critical materials, photonic interconnect, capital formation. Each passes the same test. There is a supply curve, there is a lead time measured in years, and money cannot relieve it quickly. I want to put a sixth constraint on the table. Assurance. I don't mean regulation. Regulation is the wrapper. Assurance is the work inside it, proving that a model does what it is claimed to do and does not do what it is claimed not to do. Now that's a bit of a tongue twister and you have to stop and think about that. You have to apply the test. Is there a supply curve? Yes, and it happens to be a very thin one. If the commitment is embedded independent evaluators with employee-level access inside Frontier Labs, the binding input is qualified evaluators. My estimate is that the credible pool is measured in hundreds rather than thousands. [9:49] You cannot train or redeploy them at scale in a matter of months or a quarter, and there is no established pipeline producing them in sufficient numbers. Let me explain which population that I'm talking about here, because there are two, and they behave very differently. The one I'm describing is the frontier pool, people capable of loss-of-control threat modeling and of taking a safety case apart and deciding whether it actually holds. That group is small, tightly networked and concentrated in a handful of organizations. It is not the same thing as the broader assurance profession, which is far larger. [10:30] We'll come back to that at the end because the difference between those two populations turns out to be the whole story. That has the same basic shape as the shortage of high-voltage engineers and substation electricians gating the power constraint. And this summer gave us something better than my estimate. It gave us a documented case. Between the 13th of July and the 6th of August, OpenAI, Anthropic and Meta each disclosed that one of their models had reached real systems outside its test environment during pre-deployment cyber-evaluations. Reported separately, a week apart, each of those read like a model defeating its cage. They were not. All three traced back to the same root cause, in the same place. A misconfiguration in an evaluation environment operated by a single contractor. A company of around 35 people, which also counts Google DeepMind amongst its clients. So let's just think about that for a moment. One firm of roughly 35 people sits underneath the pre-deployment cyber safety claims of four frontier labs. That is not evidence of a thin supply curve. That is the supply curve. [11:42] There is a second number that we need to think about. Recent third-party evaluation engagements have run on windows of roughly three to seven days of access. In both completed cases, the evaluator said publicly that the window was too short to reach confident conclusions. [12:01] So the proposal is permanent employee-level access, the current practice is a long weekend. The distance between those two things is an institution that does not exist yet. [12:14] Now we can put a number on the other side of that equation. Meta is a non-profit that tests frontier models. One thing it measures is how long a software task a leading model can complete, succeeding about half the time. Since 2019, that length has been doubling roughly every seven months. In June, Anthropic said that doubling had recently sped up to closer to every four months. Now that does not mean everything about these models doubles every four months. It is one measure, but it is an important one, because the longer and more complex the task a model can take on, the harder it is to test. So picture two lines on a chart. The first is what the models can do. If that task length keeps doubling every four months, then in two years it doubles six times. That is 64 times longer. Even at the slower pace of every seven months, it's roughly 10 times. Now the second line is the people qualified to check the models. That line does not double. It grows by whatever number you can hire and train each year. If the pool of qualified evaluators doubled in two years, that would be remarkable. So we've got one line curving upwards. One line is climbing slowly at best. So we've got two curves. One is exponential. One is linear. And the gap between those two lines is what a constraint looks like before anybody has given it a name. [13:43] Now, the strongest objection to what I've just described is not the one most people are making. It is automation. Evaluator capacity is not only headcount. The labs are deliberately industrializing this work. Anthropic's own evaluation roles are advertised as production engineering, building automated systems that run thousands of capability tests, and tooling that lets domain experts deploy new evaluations quickly. The stated direction of travel is more frequent testing with automated elicitation, analysis and reporting, If evaluation throughput scales with compute and software rather than with people, Then my second line is not that linear at all And the gap I have just described narrows considerably, I think the objection is correct, I also think it solves half the problem Automation scales the running of tests it does not scale the judgment about whether the results add up to a safety case. The scarce input is not test execution. It is the small number of people with the standing to sit in a room on launch day in front of some very clever and very committed colleagues and say that they are not convinced, and to be taken seriously when they do. You cannot automate that and you cannot hire it in a quarter either. [15:01] There is a second input too. Rigorous evaluation consumes compute. Adversarial testing, red teaming and interpretability research all require accelerators, memory and power. Evaluation demand is still small beside frontier training, so I would not overstate it, but it creates another claimant on the same constrained stack alongside training and inference. A pacing regime does not automatically release all of that capacity back into the market. Pushback and Objections [15:33] The obvious rebuttal came from somebody who you might know. Mark Zuckerberg posted on Tuesday that Meta already engages independent evaluators in several areas. That this is industry best practice and that other labs can simply do the same. He also argued that competition and liability give every lab enough incentive to act individually, without a coordinated slowdown, and pointed to Meta delaying its own Muse agent release by several months. That is a serious objection, but notice that even in rejecting the idea, he said it would help to have a larger and more diverse ecosystem of evaluators. That is the supply point. My read is that engaging evaluators for defined projects and giving permanent employee-level access across frontier development are very different quantities of work. The first is a project, the second is an institution, and the gap between them is where the constraint lives. But I would say that because it supports my thesis, so I have a bias. So watch it rather than take my word for it. If evaluator capacity proves readily available, assurance is a rising cost line, not a binding constraint. Zuckerberg's liability point, though, may actually strengthen that argument and we'll come back to that at the end. [16:56] One more thing, and it cuts against my own argument, which is exactly why it belongs in the body of the episode, rather than as a footnote. Neither Anthropic's commitment nor OpenAI's existing arrangements give external evaluators any independent authority to halt the development or deployment of a model. They get access. They do not get a stop button. An evaluator who cannot stop a release is not a gate. That is a genuine weakness in the pacing proposal as it stands. It is also part of why my base case is a rising cost rather than a binding constraint, because you cannot bind a process you have no power to interrupt. Now before we go any further, I want to test my own argument against the strongest objection to it. Everything I have described so far assumes the models are being built and controlled by a small number of companies. [17:51] Put evaluators inside those companies, and you can watch what they do. But not every model works like that. Some developers release what are called open-weight models. The underlying model itself is published. Anyone can download it, inspect it, modify it, and run it on their own machines. Several Chinese developers have gone this way, while many of the leading American models remain closed. And an evaluator sitting inside an American lab has no way of seeing what happens to a model once it has been downloaded somewhere else. Two recent events make the problem concrete. Last month, researchers reported that Moonshot's OpenWeight's Kimi K3 bypassed a testing sandbox run by the UK AI Security Institute. So the evasion behaviour people are concerned about is not confined to closed American systems. [18:43] In the other direction, when Hugging Face needed to analyse the July intrusion by OpenAI's escaped test agents, the more tightly restricted America models proved less useful for forensic work, so it ran the analysis on an open-weight Chinese model instead. Open weights are simultaneously part of the risk and part of the defense. That makes my framing uncomfortable. Assurance may not only be thin on supply, it may be structurally incomplete because it reaches only the part of the frontier that is closed. An assurance regime covering a handful of American labs and nothing else is not a governor on the technology. It is a governor on a handful of companies. Now I don't have a resolution to that, and I don't think anybody does yet, but an episode calling assurance the next constraint without acknowledging the hole in the middle would not be worth your time. Global Assurance Models [19:41] This week, three different assurance architectures began to emerge in public. Europe is moving towards coordination. The US is divided between oversight, competition and liability. China recognises much of the same risk, but it is building around state-backed standards and external testing rather than embedded monitors. So let's start with Europe. On Wednesday morning, in her annual speech to the European Parliament, Ursula von der Leyen backed the pacing calls. She said she would invite the main frontier labs to discuss supporting those efforts and that Europe would work with partners including Canada and the United Kingdom on model evaluation, verification and AI security, That is the president of the European Commission convening the frontier labs four days after a blog post, But Europe is not unified The French finance minister Roland Lascure told Reuters that calls to slow down serve the interests of those already at the top and that France and Europe should accelerate in order to lead and manage the risks themselves, So the Commission and a major member state finance ministry are pulling in different directions in public in the same week. [20:55] The United States is split differently. The administration has rejected a slowdown. Congress is not entirely absent. The Bipartisan Frontier Act, introduced in July, seeks a national safety and oversight framework. And then there is antitrust. Remember that Amadeus' framework requires targeted exemptions before competitors can coordinate. On Tuesday, the chairman of the Federal Trade Commission, Andrew Ferguson, speaking personally, said people should be deeply suspicious of AI companies seeking antitrust exemptions while lobbying for new regulation. Senators Josh Hawley and Ted Cruz have both rejected the idea of a waiver. That is the competition regulator, and senior voices in the Senate signalling scepticism about the legal mechanism the plan depends on. It is also a decision markets may actually be able to observe inside the next 24 months. If the exemption is refused, coordinated pacing in the United States has no obvious legal vehicle. [21:59] China is the third model, and here the distinction really does matter. Beijing has rejected this mechanism. It has not rejected the risk. China included a loss of control scenario in its AI safety framework in September 2024 and strengthened it in the 2025 version. Policy issued in May identifies operational loss of control as a security risk for AI agents and requires developers to be able to discover, intervene in and block and recover from improper agent behaviour. [22:32] China is also drafting what would be the world's first mandatory national standard for AI agent safety. What it has not proposed is independent monitors embedded inside companies. Its regime runs on developer obligations, staked-backed standards, security assessments and external testing. Same identified risk, completely different assurance architecture. Add to that at least one major lab arguing that no coordination is required at all, and we already have several competing systems before a single one of them is built. Fragmentation is what could turn assurance from a cost into a constraint. Imagine the same model evaluated several times against several different standards by the same small pool of qualified people. Every extra regime adds to the demand. It does not add to the supply. A week ago, that was hypothetical. It is not anymore. The regimes are not built yet, but the fault lines went public this week, while I was planning this podcast. Tail Risk vs Reality [23:35] Okay, so now we need to separate two very different risks. There are people telling you that these systems threaten the future of humanity. There are others dismissing that as nonsense sold for clicks. I do not think markets have a reliable edge in adjudicating either of these positions. [23:54] Senior people inside the labs have attached serious probabilities to catastrophic outcomes and put their names to them. You cannot simply wave that away. They have access to information that neither you or I possess. But this is a statement about markets, not about whether those people are right. A probability of catastrophe inside a decade cannot be priced with any confidence. There is no base rate because it has never happened. There is no useful transmission mechanism to a cash flow because, in the scenario described, there may be no cash flows, and it does not resolve inside a normal investment horizon. That does not make the risk unserious. It makes it unpricable. Those are different statements. I also hold in mind that dismissing an unfamiliar tail risk is exactly the sort of mistake that precedes tail events that we've observed throughout history. [24:50] Documented agentic incidents are different. The hug-in-face intrusion, the hijacked website, models leaving test environments, fake accounts created to gain access to a service. Those events are empirical. They happened. They have dates, frequency, severity, controls and potential losses. That is operational risk, and large banks and insurers already have mature methods for assessing this. So the sentence that I would leave you with is this. The market is arguing about a tail risk nobody can price. While ignoring an operational risk, it already knows how to price. And one of the clearest risks we can actually measure is concentration. A small number of AI models are ending up inside a large share of the world's businesses. That means if one of those models fails or behaves badly, it does not affect just one company, it affects thousands of companies at the same time and in the same way. That is how financial crisis spreads. Not through one firm getting something wrong, but through many firms relying on the same thing. And there is a second concentration sitting directly on top of the first, which almost nobody is discussing. [26:05] Not only a small number of models inside a large share of businesses, a small number of evaluators standing behind a large share of those models. We saw earlier what one misconfiguration at one 35-person firm did to three labs inside a single week. Concentration in the thing being assured is a risk. Concentration in the assurance itself is a different order of problem, because it quietly removed the independent check you were relying on to catch the first one. Anyone who has worked in financial services will recognise the shape of that immediately. It is the audit concentration argument, and it is the critical third-party argument arriving in a new industry that does not yet have the vocabulary for either. And I expect the regulators who oversee the banks and insurers to take this seriously before stock markets do. Balance Sheet Impact [26:57] Okay, so how does any of this reach a balance sheet? This is the centre of the episode, right here. To the extent Mundy's reaction was about AI at all, it appeared to assume that slowing capability improvement means less capital expenditure. In my base case, fragmented pacing does not materially reduce the physical build-out already underway. It lengthens the lag between capability and commercial deployment. Think about what that does. The asset enters service, and its costs start straight away. The depreciation clock is already running, but the revenue has to wait. It cannot arrive until the model has been tested, approved and trusted enough to use. Let's call that the assurance clock. And that clock is starting later. [27:45] Every month, those two clocks drift further apart, and the further apart they drift, the worse the return on all of that investment. So the number to watch is not only capital expenditure guidance, everybody's watching that, and it has not moved. The number to watch is time to revenue. Regular listeners will see where this goes. We began this series with two clocks, the financial clock, which reprices in minutes and hours, and the physical clock, which reprices in years. Substations, transmission lines, fab construction. Those two have run underneath everything we've discussed on this show. This week puts two more on the table. A capability clock measured in software task horizons, improving on a timescale of months. That clock has been running since 2019. What changed this week is that somebody is now proposing to slow it down on purpose. And we have an assurance clock, whose speed is set by a limited pool of people, and several regulatory regimes that do not agree with each other. That one is new, and nobody yet knows how fast it runs. Two established, two arriving, four clocks, one balance sheet. [28:58] Line them up by speed and you get minutes, months, years and a fourth that nobody can time at all. [29:04] And the mispricing lives in that spread. A market that reprices in minutes is trying to value an asset whose revenue is gated by a process measured in years, and approved by a process nobody can measure. [29:17] That is the tension I think defines the next two years of this cycle. Enterprise Unlock [29:22] But there is an inversion here. At the frontier, assurance is friction. In the enterprise, assurance may be the unlock. This is the part I know something about because I've spent many years in my earlier career working in governance, risk, and compliance. What stops AI being deployed inside a bank or an insurer is increasingly not model capability. The models are already more capable than many of the deployments. What stops adoption is the inability to evidence control to a governance lead or a regulator Sit in a risk committee in a regulated firm and you'll hear some version of the same question How do we prove this to the regulator? Nobody has a consistently good answer for this And this year the regulators effectively conceded the same thing, On the 17th of April, the Federal Reserve the Office of the Comptroller of the Currency and the Federal Deposit Insurance Corporation issued new model risk guidance, replacing a framework that had governed American banks since 2011. Inside it is the sentence that matters, and it appears twice. Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance. [30:41] At the same time, the agencies were explicit that this is not an exemption. Firms are told to apply their broader risk management and governance practices to decide what controls belong on anything the guidance does not cover. The Federal Reserve's Vice Chair for Supervision said as much in remarks at the end of April, confirming the guidance had been amended to clarify that it does not apply to generative or agentic AI. A request for information on banks' use of AI is promised and has not yet appeared. Now read that carefully. The supervisor has said, in writing, and twice, that it does not yet know how to govern the things banks are actually deploying, while leaving the obligation exactly where it has always been, with the firm. That is not a gap in regulation, it is a gap in evidence, and it is precisely the gap independent evaluation is shaped to fill. [31:37] The UK picture is not much more comfortable, for a different reason. The Prudential Regulation Authority, the PRA, their model risk principles came into force in May 2024, and they apply to firms with internal model approval for regulatory capital. Those firms have been working through multi-year remediation plans ever since. But those principles were written for models at a point when that word meant something considerably narrower than it does today. Here is the size of it. Fewer than 30% of European banks have integrated generative and agentic models into their model risk management frameworks at all. That is not a capability problem, it is an assurance problem. [32:21] A credible independent evaluation regime with named evaluators, published standards and something resembling an audit trail is precisely the artefact that begins to answer it. Now, here's the honest objection. The UK's own institute, which runs one of the largest safety evaluation teams in any government, has said in public that the science is currently too nascent for independent evaluations to act as a certification function. They can tell you what a system did under test. They cannot tell you it is safe. So what does a bank actually put in front of its supervisor or its regulator? Not a certificate, not yet. What it can put forward is an evidence trail. Named evaluators with documented competencies, a record of what was tested, by whom, under what access, and what was found. Separation between the people who build and the people who are sure. [33:17] That is a weaker artefact than a certificate. It is a considerably stronger one than what most firms have today, which is a vendor attestation and a hopeful paragraph. Some in the market have already made a related point this week, that a credible safety framework makes long-term investment in AI infrastructure easier to justify. I think the bigger effect is further down, inside the businesses that actually use these models. So consider what follows. [33:47] And this is where the two populations come back. At the frontier, the pool is thin, and I do not expect it to thicken quickly. The broader assurance profession is an entirely different picture. In the UK alone, it was worth over a billion pounds in gross value added in 2024. It already employs more than 12,000 people, and government projects it reaching close to 19 billion by 2035. A consortium chaired by the BCS is building the Code of Ethics, the Competency Framework and the Certification Pathway right now. There is no accredited scheme yet, but there will be. It's thin at the top, deep and widening below. That is why friction at the frontier and acceleration in the enterprise are not a contradiction. They are two ends of the same supply curve. The same mechanism that slows the frontier could accelerate the revenue layer. Deceleration at the top of the stack Acceleration at the point where the money is made We spent an entire episode earlier this year on the gap between what is being spent on AI and what is being earned from it, If this holds, a paced frontier is not necessarily bearish for this cycle It may help close that gap by making deployment easier to approve ensure and scale. IPOs and Disclosure [35:08] And there's one final piece The leading frontier labs are reportedly preparing for public offerings. I am not going to quote valuations because the figures circulating are not consistently sourced. Three things could follow from a listing. First, this cycle currently trades through proxies, semiconductors, memory, power and networking. [35:30] A listed frontier lab could become the closest thing to a pure expression of this thesis, and potentially the sentiment instrument for the whole complex. Headlines would reach the physical layer through a daily price rather than through commentary. That is a faster and much noisier channel. Second, public shareholders demand growth. A company publicly committed to pacing capability improvement would have to explain the commitment repeatedly in quarterly increments. The tension between pacing and growth is a governance question, and it remains unresolved. Third, listing creates disclosure pressure. Evaluation costs, compute allocation, release delays, incident reporting. Not all of those would appear as separate line items, and disclosure would depend on materiality. But filings, risk factors and investor scrutiny could make more of the assurance clock observable than it is today. The constraint I have spent this episode describing is difficult to measure now. Listing may be what begins to make it measurable. Fragmented Regimes [36:41] Okay, so where do I stand on this? My base case is not that pacing fails, it is that pacing fragments. I expect multiple incompatible assurance regimes rather than one coordinated mechanism. Europe convening and enforcing, China building around standards and external testing rather than embedded monitors. In the United States, an executive resistant to a slowdown, a congress divided, and a competition regulator sceptical of the legal vehicle, and at least one major lab declining to coordinate at all. Over the next 24 months, I expect assurance to operate primarily as a rising cost of doing business, rather than a binding constraint. Beyond that, I am really uncertain. Fragmentation is the variable that decides it, because fragmentation multiplies demand on a limited supply of people. I hold that with moderate confidence and no more I was more confident at the weekend Then a commission president stood up in Strasbourg on Wednesday And I have to revise Three things that would change my mind. [37:51] If a US antitrust exemption is granted Coordinated pacing becomes legally possible And my fragmentation case weakens, If evaluated capacity expands much faster than I expect whether through hiring or through automating the evaluation stack itself, assurance remains a cost line and never becomes a constraint. And if there is a major agentic incident with quantified losses at systemic scale, everything in this episode is superseded within a week. What to Watch [38:22] So at QF Markets, I'm always looking for the signal in the noise. So here's five things that I'm watching. 1. Release cadence. Whether the gap between major model releases lengthens and whether labs start disclosing how long safety testing took. It is an imperfect proxy, granted, but it is the best one available for the pace of the regime. 2. Evaluate a capacity and whether an accredited profession forms. This one now has names and dates attached A consortium chaired by the BCS, supported by the UK government Is in its first year of producing a code of ethics And also a competency framework and a map of the information access That assurance providers need from developers, With AI auditing the likely starting point for certification, Watch what it publishes and watch whether the national accreditation body Recognises any of it That is the moment a profession stops being an aspiration 3. Whether the labs begin disclosing evaluation compute or expenditure Watch the listings documents 4. Enterprise procurement And this one has already begun Procurement teams are increasingly treating AI governance as a contractual requirement Investors are asking for assurance evidence in diligence And insurers are starting to factor it into cover. [39:48] So the question is no longer whether it starts, it is how quickly it becomes standard and whether contracts begin naming the evaluator specifically rather than accepting a generic attestation. That second detail is the one I would watch, because it is the point at which the assurance layer becomes a named counterparty with concentration risk of its own. Five, and this is one I really care about the most, whether an insurance and liability market forms around AI deployment. Because if liability is the disciplining mechanism, insurance is the pricing mechanism. Insurers price what they can evidence. And underwritability is what unlocks adoption at scale. Now, if that market forms, adoption should become easier. If it does not, the enterprise unlock argument weakens considerably. [40:41] Okay, so let me wrap this up. Everyone this week has been arguing about whether the frontier should slow down. I am more interested in what it costs to prove that it did. Because that is what reaches the balance sheet. Not the promise to pace, but what proving it consumes. People who do not yet exist in sufficient numbers compute that could have been doing something else. And time sitting between an asset entering service and the revenue meant to justify it. Assurance is not binding today, it becomes binding if evaluation capacity can no longer keep up with measured capability improvement and fragmented requirements. [41:23] And the thing I keep coming back to is that the mechanism slowing the frontier may be the same mechanism that finally unlocks the Enterprise. Friction at the top, acceleration where the money is made. Okay, so that wraps it up for this week, and I hope you found some value in that. But before I go, a quick update on the book, and that book is Trading the AI Supercycle. I really did underestimate the editing process significantly way back when I started writing this last year. [41:53] Every pass through a chapter has exposed another question, and this year the market has refused to sit still long enough for anything to remain finished. You write a section on capital formation, and the financing structures change underneath you. [42:07] The upside is that the book is considerably better than it would have been if I'd have shipped it in the summer. [42:13] The downside is that it will be published later in the autumn. I'm now into the review stages and I can see the finish line ahead, but let's see famous last words. Okay, so that is a wrap for this week. Thank you for listening. If you found this useful, please follow or subscribe wherever you listen. You can find all of our work at qfmi.substack.com. The Market Pulse and the In The Spotlight articles are free and always will be. The Weekly Outlook, the Weekend Debrief and the Monthly Strategic Research Report sit behind a paid subscription. Subscribe and you get the full picture. Two final disclosures. I use Frontier Models personally and professionally. They are part of how I work, but I have no commercial relationship with any of the companies discussed beyond being a paying customer, and no particular preference between models. And secondly, one important note, everything in this podcast, including everything you've heard today, is provided for educational and informational purposes only. It is analysis, not personal financial advice, and nothing in this episode should be taken as a recommendation to buy or sell any security. Always do your own research and speak to a qualified financial advisor before making investment decisions. Thank you for listening to the AI Supercycle Podcast. See you next week.