The Enterprise AI Payoff: From Tokenmaxxing to Value per Token

Transcript
[0:10] Hello and welcome to the AI Supercycle Podcast, where we follow the capital

[0:14] flows and physical constraints shaping the largest industrial build-out of our generation. From semiconductors, AI factories and hyperscale data centers, to energy markets and power grids, critical materials, orbital compute, and embodied intelligence. This also extends through to the enterprise workflows and business models that ultimately have to pay for it. We track where the capital is being deployed, where the binding constraints are emerging, where revenue is being created, and whether the returns can justify the next wave of investment. Today is Thursday the 6th of August, 2026. I'm Tim Hardwick, founder of Quantumfields Market Intelligence. In the last episode, we looked at capital formation, the money that has to arrive before the chips, data centers, power plants, and networks can be built. We examined the shift from internal cash generation towards bonds, private credit and special purpose vehicles. And we ended with the feedback loop that governs the entire cycle. Strong returns on the capital already deployed make the next round easier to finance. Weak returns make capital more expensive, more selective and eventually more scarce. This week we follow the feedback loop to the other side of the equation because hyperscaler capital expenditure is not a return. A data centre is not a return and neither is a rack of GPUs.

AI Value Chain Begins
[1:39] Those are investments made in anticipation of future demand. The return only arrives when customers use that capacity and pay enough for it. And they have to keep paying for it at a level that earns more than the full cost of providing it.

[1:54] And consumers matter, of course. Hundreds of millions of people using AI assistants, search products and creative tools can support subscriptions and advertising. But the consumer paying £20 a month for an AI model on their own cannot underwrite an industrial build-out of this scale. Enterprises and governments have to become durable, expanding sources of revenue, and that creates a second dependency. If enterprises cannot show a return on their AI spending, their boards and finance teams will eventually slow the budgets that become hyperscalar revenue. The physical build-out may be supply-constrained today. Over time, though, enterprise value becomes the demand-side constraint. That is what this episode is about, the business case for AI and the shift away from what's being called token maxing. Token maxing simply means consuming more and more AI compute because usage itself is was treated as progress. This episode is about the shift away from that, toward token optimization and value per token. And stay with me to the end because I'm going to close with the specific things worth tracking this quarter.

From Compute to Revenue
[3:02] So let's start with the chain that we're trying to close. Capital markets provide money to a hyperscaler. The hyperscaler turns that money into land, buildings, servers, networking and power contracts.

[3:15] The infrastructure produces compute. Models turn that compute into tokens, predictions, images, code and agent actions. Enterprises buy those capabilities through cloud consumption. API calls, software licenses, and increasingly, some form of committed token capacity.

[3:34] Then they apply the technology to a business process. Only at the end of that chain do we discover whether economic value was created. For example, if an insurer processes claims faster without increasing customer complaints, value may have been created. If a manufacturer reduces unplanned downtime, value may have been created. If a bank improves fraud detection without generating an unmanageable volume of false positives, value may have been created. If a software company releases reliable products faster or a retailer converts more customer inquiries into profitable sales, value may have been created. But notice the language I'm using. May have been created. Technical output does not automatically create business value. Tokens generated are not value. Users provisioned are not value. Even hours saved are not necessarily value until the organization converts that released capacity into lower cost, greater throughput, better service, reduced risk, or additional revenue. Now that distinction sounds obvious. In practice, it is where a large share of enterprise AI business cases come apart.

[4:52] The AI industry has been really good at measuring supply. GPUs shipped, megawatts contracted, tokens processed, benchmark scores achieved and context windows expanded. Enterprises are now being forced to become equally good at measuring demand-side outcomes, not activity, outcomes. The first thing to say is that enterprise spending is not missing. The latest hyperscaler results provide strong evidence that customers are buying capacity and software at scale. Microsoft has just reported that Azure and other cloud services' revenue grew 43% year-on-year.

[5:32] Microsoft Cloud Revenue reached $59.3 billion for the quarter, and Microsoft 365 Copilot moved beyond 30 million paid seats. That is a significant commercial milestone. It tells us enterprises are willing to sign contracts and buy licenses, and that AI is moving beyond small experimental budgets. Alphabet reported Google Cloud Revenue of $24.8 billion, up 82% year-on-year, with operating income of $8.8 billion and an operating margin of 35.6%. Its cloud backlog reached $514 billion, with management attributing much of the increase to demand for enterprise AI offerings. Amazon Web Services reported quarterly sales of $42.2 billion, up 37%, and operating income of $16.6 billion. AWS is not merely growing, it remains an exceptionally profitable business while funding a much larger infrastructure program around it.

[6:36] And these numbers matter. They are evidence against the simplistic claim that hyperscalers are building into a demand vacuum. But they do not answer the enterprise ROI question. A paid co-pilot seat tells you that a license was purchased, but it does not tell you how often the product was used, whether the user changed the way they worked, whether the output was reliable, whether the time saved was captured, or whether the organization will expand or renew the contract on the same terms. A cloud backlog tells you that customers have made commitments. It does not tell you whether every workload funded by those commitments will create an acceptable return for the customer.

[7:21] Provider revenue and customer ROI can therefore diverge for a time. Enterprises can spend ahead of value while they experiment and train people and build data foundations. Long contracts can keep vendor revenue growing even when the buyer is still trying to prove the business case. The crucial test arrives later through production usage, renewal, expansion and the conversion of operational benefits into the P&L. The hyperscaler results say enterprise commitment is arriving. They do not yet prove that value realisation is broad, repeatable or mature.

ROI Surveys Diverge
[7:57] This is where the evidence becomes really interesting, because the major enterprise surveys appear at first sight to describe completely different realities. PwC's 2026 Global CEO survey, covering more than 4,400 chief execs, found that 56% had realised neither revenue nor cost benefits from AI. Only 12% reported both higher revenue and lower costs. That is a sobering result, particularly because it asks about financial outcomes rather than general satisfaction or productivity.

[8:32] McKinsey's latest global state of AI survey found that 88% of respondents were using AI regularly in at least one business function. Yet only 39% reported an enterprise-level impact on earnings before interest and tax. McKinsey classified around 6% as high performers, organizations attributing at least 5% of EBIT to AI, while also reporting significant value. Nearly two-thirds had not yet begun scaling AI across the enterprise. Deloitte's 2026 Enterprise Survey offers a more positive view of operational impact. Two-thirds of organizations reported productivity or efficiency gains. 40% reported cost reductions, but only 20% reported increased revenue. Three-quarters hoped AI would support revenue growth in the future. The aspiration is widespread. The realized top-line benefit is not. Then Google Cloud's July survey of more than 2,400 executives, found that 84% were seeing increasing financial returns from AI initiatives, with 26% saying that those returns were accelerating year-on-year. Google calls that 26% its group of AI ROI leaders.

[9:51] So which is it? Are most enterprises getting no financial return? Or are more than 8 out of 10 seeing returns increase? The honest answer is that these findings are not directly comparable. The surveys ask different questions of different people, and they use different thresholds and cover different populations. There is a big difference between a positive return on one initiative, increasing returns across a portfolio, and a material impact on the enterprise P&L. There is also a difference between a CEO answering for the whole company and a business unit leader answering for the initiatives closest to them.

[10:30] And vendor-sponsored research naturally has an interest in locating evidence of success, even when the methodology is sound. But this disagreement is in itself useful. It reveals how immature the measurement discipline remains. An organization can truthfully say that AI saves thousands of employee hours and also truthfully say that AI has not reduced cost. It can report a positive ROI on a carefully selected customer service application while the enterprise program as a whole remains loss-making. It can generate revenue through an AI-enabled product while spending even more on the models, data, integration, governance and human review. All of those statements can coexist. There is also a concentration effect. PwC's separate AI performance study Found that 20% of companies captured 74% Of the economic value attributed to AI, That looks less like a universal failure And more like a widening distribution A relatively small group is learning how to turn AI Into financial performance, While a much larger group remains in experimentation Local optimisation Or even in pilot mode.

[11:45] That distribution matters more than any single average. If value remains concentrated, AI spending is unlikely to collapse evenly across the economy. It is more likely to polarise. The leaders will scale, reinvest and consume more. The Lagarde will consolidate vendors, cancel weak pilots and demand harder evidence before approving the next budget. The jury is therefore not out on whether AI can create value because it clearly can, the jury is out on how broadly, how quickly and how reliably enterprises convert that capability into measured financial return.

Capturing Real AI Value
[12:25] The most common AI benefit is also the easiest to overstate, time saved. Suppose an AI assistant saves a professional two hours a week. Multiply that by 10,000 employees, then multiply again by an average hourly employment cost, and the spreadsheet produces an enormous annual benefit. It looks precise, but often it isn't. The calculation assumes every saved minute is productive additional capacity. It assumes the user adopts the tool consistently, that the output is accurate enough to avoid rework, and that the employee uses the release time on something more valuable, It may also assume the company can remove the corresponding labour cost, even when no headcount is actually reduced, and no future hiring is avoided, Time saved is a capacity benefit It becomes a financial benefit only through a capture mechanism Now there are several legitimate capture mechanisms.

[13:24] The same workforce may complete more cases, serve more customers or release products faster, The organisation may avoid future recruitment as volume grows Overtime and contractor spend may fall Service levels may improve, supporting retention or pricing, Employees may spend more time on sales, judgement, innovation or customer relationships, Any one of those can be valuable, but the mechanism needs to be named, measured and owned. Otherwise, the business case counts theoretical capacity as though it were cash. Now the same problem appears with qualitative benefits. Better decisions, faster learning, more resilient organisational knowledge and improved customer trust are all real sources of value.

[14:13] They shouldn't be dismissed merely because they are difficult to express in pounds or dollars, but they need observable proxies, such as decision cycle time, employee retention, customer satisfaction, or perhaps error rates, compliance findings, product development velocity, or the probability-weighted cost of a risk event. Of course, they were talking about risk reduction or risk mitigation.

[14:39] Qualitative does not mean immeasurable. Yes, it is intangible, and we do call them intangible benefits, but it means the relationship to financial performance is indirect and should be handled with that in mind. Quite often, these kinds of intangible benefits won't be included in the ROI. There is one more trap. AI can move cost rather than remove it. A call may leave the contact center but generate more complaints. A coding assistant may speed the first draft but increase testing or security remediation.

[15:16] An AI agent may automate a workflow while creating new expenditure on observability, evaluation and human approval. The correct unit of analysis is the end-to-end process, not the isolated task where the model happens to operate. This brings us to the business case. Token optimisation really does matter, but it comes later than many organisations think. The first optimisation decision is whether the use case should exist at all. A credible AI business case begins with the business outcome and a process owner, not with a model demonstration. What is the organisation trying to change? Is it reducing the cost to serve, increasing conversion, shortening the cycle time, improving quality or avoiding loss. Maybe it's increasing capacity or accelerating cash collection or creating a new source of revenue. If the answer is simply use AI, the initiative is not ready for investment. And I've worked in these environments and I've seen a lot of this before. And that is quite frankly, not a reason to invest.

[16:27] The second requirement is a baseline and also a counterfactual. What does the process cost today? How long does it take? What is the current error, escalation, or conversion rate? What would happen over the same period without the AI investment? Without a baseline improvement becomes anecdotal. Without a counterfactual every positive movement is attributed to AI, even when demand, staffing, pricing, or another transformation program may have caused it. The third requirement is an economic mechanism. How specifically does the operational change reach the P&L, the balance sheet, or the strategic objective?

[17:13] A shorter claim cycle may have lower handling cost. It also may improve customer retention and reduce outstanding reserves. Faster software development may produce more releases, but the value depends on whether those releases create revenue, or reduce risk, or avoid hiring. Better fraud detection creates value through expected loss avoided, net of false positives, investigation costs and customer friction. Fourth, the business case needs the full cost stack, Model inference is only one component There is data preparation, cloud infrastructure, retrieval, storage, integration, security, evaluation, licensing, engineering, change management, training, governance and ongoing human review, There is also the cost of failed calls, retries, hallucinations, escalations and downtime, A pilot might look cheap because the enterprise capabilities required to run it safely, at scale, have not been charged to it. Fifth, quality and risk thresholds must sit behind the financial target. A system that is cheaper but materially less accurate is not on optimisation. Nor is a system that produces a faster answer, but cannot be audited in a regulated process.

[18:34] The target is the lowest cost at the required level of quality, latency, security and control, not the lowest cost in isolation.

[18:43] Sixth, somebody must own benefit realisation after deployment. Technology teams can deliver a capability. They cannot, by themselves, redesign every operating process, change incentives and capture the released capacity. That ownership normally belongs with the business executive who controls the process and budget, working with technology, finance, risk and the workforce affected.

Measuring Across Three Levels
[19:10] Finally, the investment needs stage gates A discovery stage tests whether the problem is valuable or tractable, A pilot tests technical and operational performance Production tests adoption, reliability and unit economics Scaling requires evidence that marginal value remains above the marginal cost, At each gate, the organisation should be willing to expand, redesign or stop, This is not bureaucracy around AI It is the mechanism that prevents a large portfolio of exciting demonstrations from becoming a permanent cost centre, One reason ROI debates become confused is that organisations mix three different levels of measurement.

[19:57] The first level is technical and behavioural. Are people using the system? How many requests, tokens or agent actions are being generated? What is the latency, failure rate, accuracy and proportion of outputs that pass quality review? What percentage of licensed users are active? These measures will tell you whether the capability works and whether adoption is real They do not tell you whether the enterprise is better off The second level is operational Has cycle time fallen? Has throughput increased? Are fewer cases escalated? Has rework declined? Are employees able to handle more complex work? Has customer satisfaction improved? These are the bridge between the model and the business I'll see you next time.

[20:45] The third level is financial and strategic. Has avoidable cost fallen? Has gross profit increased? Has working capital improved? Has the organisation avoided hiring, reduced losses, retained more customers or created a new revenue stream? Has it increased strategic flexibility or reduced a material risk? The scorecard requires all three levels because each answers a different question. Technical metrics explain performance. Operational metrics explain process change. Financial and strategic metrics explain enterprise value. The connection between them is the important part. If active usage rises but operational performance does not improve, the tool may be popular but ineffective. If cycle time improves but financial performance does not, the organization may not be capturing the capacity. If profit improves, the baseline and counterfactual help determine how much of that change can reasonably be attributed to AI. This is still uncommon at the top of the enterprise. EY's 2026 CEO research found that only 11% of respondents linked AI impact directly to financial reporting and regular senior management review. Now this is a remarkably small number given the scale of the investment.

[22:11] The organizations most likely to sustain AI spending will be those that connect these levels routinely, from token, to task, to process, to P&L.

Token Economics Shift
[22:22] Now we can turn to tokens, because the enterprise conversation has shifted sharply in 2026. During the adoption phase, many organizations treated higher AI usage as an objective in itself.

[22:36] Employees were encouraged to experiment. More prompts, more seats, more tokens indicated engagement. Some people called this token maxing, deliberately increasing token consumption to build fluency, discover use cases and avoid falling behind. There was a rational argument behind it. If an organization applies tight cost controls before people understand the technology, it may suppress learning and miss the use cases that matter. Early experimentation needs room to breathe. But agentic AI changes the cost curve. A conventional assistant may respond once to a user. An agent can plan, retrieve data, call tools, consult another model, and validate the answer, and also retry a failed step and continue working without another human prompt. The user sees one completed task. The infrastructure may see dozens of model calls and a very large context moving through the workflow. The FinOps Foundation notes that a retrieval-based workflow using a reasoning model and several tool calls can consume one or two orders of magnitude more tokens than a direct prompt to a smaller model. System instructions, conversation, memory, retrieved documents, tool definitions, output length and retries all compound. That is why token costs have moved from the engineering team to the executive agenda.

[24:02] EY's latest US AI Pulse survey found that 82% of senior leaders at organizations investing in AI were concerned about token usage and related costs. 98% of those using token-based tools said the costs had caused them to reconsider their approach, yet only 64% actively monitored token usage with clear budget guardrails. The same survey contains an important counterpoint. more respondents were considering expanding the scope of their rollout than reducing it by 37% to 15%. Now this is not an AI retreat, it is a move from undisciplined adoption toward economic management. And this is where the language needs to be more specific. Token optimization does not mean using the fewest possible tokens. The FinOps Foundation puts this principle well. A model that uses 10 times as many tokens but creates 100 times the value is economically preferable to a cheaper model that produces an unusable answer.

[25:12] The objective is not minimum tokens, it is maximum value at an acceptable cost, quality and risk. That changes the unit of measurement. Cost per token is useful for explaining the bill, it is not enough for running the business. The mature measures are cost per successful customer interaction, cost per case resolved, cost per claim processed, cost per coding task accepted, cost per agent action, or cost per decision supported. The business side counterpart is revenue, margin, loss avoided, cycle time, or capacity created for a successful outcome. Put simply, a token is a unit of consumption. The economically meaningful unit is the outcome it helped to produce.

Optimizing for Outcomes
[26:04] So how does token optimization actually work? Well there are three layers to practical token optimization and they are selection, architecture and operation. Selection means using the right model for the task. Not every request needs the most capable frontier reasoning model. A smaller or specialized model may classify an email, extract a field or answer a routine question faster at a fraction of the cost. A routing layer can send simple tasks to a cheaper model and escalate only when confidence or complexity requires it.

[26:40] EY told Business Insider that an internal routing approach reduced token consumption by as much as 60% in the tools where it was deployed. Now the important point is not the exact percentage, It is that model choice can become dynamic rather than fixed. Architecture means controlling what enters and what surrounds the model. Long system prompts repeated on every call, entire documents retrieved when one paragraph is relevant, unnecessary conversation history, and oversized tool descriptions all add cost. Better retrieval, context compression, semantic caching, and structured tool interfaces can reduce consumption without reducing useful output. Repeated work can be cached. Non-urgent requests can be batched. Deterministic software can handle deterministic steps instead of asking a probabilistic model to rediscover the same answer every time. Operation means observing the system in production. Which team, application and business process generated the tokens? How many outputs were usable? How many calls were retried or abandoned? Did a model update change quality, latency or cost? Are budgets and alerts attached to a named owner? Can finance see the same consumption data as engineering?

[28:02] This is AI FinOps, but the best framing is broader than cost control. It is unit economics for machine intelligence.

[28:12] The optimization target should be cost per good outcome. Engineers sometimes call the usable part of a system output, good put, the results that arrive at the required quality and speed and can actually be used. A cheap system with a high failure rate may have poor economics because human review and retries overwhelm the saving. A more expensive system may be better value if it completes the process correctly the first time. There is also a portfolio decision. Some workloads fit AI embedded inside a software subscription. Others fit consumption through an API. Very large, stable workloads may eventually justify dedicated or self-hosted infrastructure.

[28:58] Each model has a different mix of fixed cost, variable cost, flexibility, transparency and utilisation risk. The cheapest option at small scale may become the most expensive at large scale, while owned infrastructure becomes ruinously expensive if it sits idle. Token optimisation therefore begins with the business outcome, continues through architecture, and ends with continuous economic measurement. Prompt engineering is only one small part of it.

Efficiency and Demand Rebound
[29:29] So what does token efficiency mean for the hyperscalers? Well, at first glance, token optimization sounds bearish for the infrastructure providers. If enterprises use fewer tokens for each task, perhaps they need less compute. Cloud revenue grows more slowly, and the case for all of that capex weakens. That is possible, but it's not the only outcome. The history of computing repeatedly shows the rebound effect. When the cost of a unit of computation falls, organizations find more uses for it. A cheaper inference can make a previously uneconomic workflow viable. Better routing and caching can free budget for more users, more agents and more complex tasks. A tenfold improvement in tokens per outcome can be overwhelmed by a hundredfold increase in the number of outcomes requested.

[30:24] The relevant equation for the hyperscalers is therefore not simply tokens per task, it is price per unit, multiplied by the number of units, multiplied by the breadth and persistence of the workloads. Efficiency is bearish if it reduces consumption faster than new demand appears. It is bullish if it lowers the cost of adoption and expands the market more quickly than it compresses the revenue from each task.

[30:53] This is why the enterprise ROI question sits above token optimisation. An enterprise that can prove value per outcome is more likely to expand usage, even as it becomes more efficient. An enterprise that cannot prove value will eventually cut spending, regardless of how low the model provider makes the price per token. It also changes where value may accrue. Smaller models, routing, orchestration and enterprise data could capture a larger share of the economics relative to a single frontier model. Per-seat software pricing may give way to hybrid structures combining licenses, committed consumption and outcome-based charges. Vendors that make cost and value transparent may gain share from those that hide token consumption inside a renewal. So the shift to token efficiency is not automatically the end of demand. It is the beginning of price discovery in the consumption layer. It forces providers to compete on cost to outcome, and it forces enterprises to decide which outcomes are worth buying.

Tracking the Key Signals
[32:04] So let's take a look at the signals that matter. And if we want to monitor this as investors and capital allocators, the dashboard needs to extend beyond hyperscaler capex.

[32:15] On the enterprise side, watch the share of AI initiatives reaching production, not merely entering a pilot. Watch active usage relative to paid seats, and watch renewal and expansion rates, the number of workflows redesigned, and the percentage of programs with a named business owner and finance-validated benefits. Watch cost per successful outcome, not just cost per token, and watch whether budgets are being cut in aggregate or reallocated from broad experimentation into a smaller number of proven use cases. On the provider side, watch for clearer disaggregation of AI revenue. Cloud growth and backlog are important, but they mix AI and non-AI workloads and may include very large commitments from a relatively small group of customers. We need the conversion of backlog into revenue, the conversion of revenue into gross profit, and eventually the return on capital supporting it.

[33:16] Watch cloud operating margins as depreciation, energy and inference usage rise. Watch paid seats alongside actual consumption, and watch the relationship between falling unit prices and rising token volume. Also, we can watch revenue or gross profit per megawatt of deployed capacity, utilization of the installed base, and the pace at which each new generation of hardware makes the previous one economically obsolete. And on the business model side, watch pricing. A move from flat per-seat licensing toward committed tokens or usage-based pricing transfers more consumption risk to the customer. Outcome-based pricing does the opposite. It asks the provider to share more of the performance risk. The direction of contracts will tell us which side has the stronger bargaining position. The most important leading indicator may be renewal behaviour through 2027.

[34:17] The first wave of enterprise AI contracts was often justified by strategic urgency and fear of missing out. The next wave will be judged against evidence. If active use, operational outcomes and financial benefits are visible, renewals can support another leg of infrastructure demand. Now if paid seats become shelfware and consumption is optimized without workload expansion, the revenue side of the supercycle will begin to look less secure.

Optimistic Case, Rising Demand
[34:45] Okay, so let me put the optimistic case properly, because, you know, the current data deserves a closer look. Azure, AWS and Google Cloud are accelerating. Backlogs are large. Cloud operating profits are substantial. More than 30 million paid Microsoft 365 co-pilot seats is not a lab experiment. Enterprises are still increasing budgets. and the EY survey shows that token cost scrutiny is causing more organisations to expand their rollouts than contract it.

[35:18] Productivity gains often arrive before they can be cleanly identified in corporate accounts, particularly when the benefit is faster decisions, better products or capacity for growth rather than immediate headcount reduction. On this view, the measurement gap is normal for a general-purpose technology at the beginning of organisational adoption. Now the infrastructure has to be built first, workflows, skills and management systems adapt later.

[35:47] Token optimization is simply the equivalent of the cloud FinOps discipline that emerged after enterprises learned how easily pay-as-you-go computing could produce an unexpected bill. That is a strong argument, and I agree with much of it. But it does not remove the timing risk. The capital is being committed now, depreciation will run for years, and enterprise budgets operate on annual cycles. A technology can be transformative over a decade and still produce an investment air pocket if revenue conversion arrives more slowly than the capacity built to serve it.

Selective Scaling Ahead
[36:24] Okay, so now let's take a look at the base case. My base case is not an enterprise AI spending collapse. It is a transition from adoption at almost any cost towards selective scaling under financial discipline. Over the next 12 to 18 months, I expect enterprises to reduce the number of disconnected pilots, consolidate some vendors, and move more budget toward applications with visible process ownership and measurable outcomes. That can look like retrenchment at the project level, while total AI spending continues to rise. The winners will receive larger deployments. The weak use cases will lose funding faster.

[37:05] I expect token optimization to reduce the cost of many individual workflows materially. Model routing, smaller specialized models, better retrieval, caching and tighter agent design will lower tokens per successful outcome. But I also expect those savings to make more workflows economically viable. Total token consumption can continue to rise even as token efficiency improves. The critical variable is workload elasticity. How much additional useful demand appears when the cost per outcome falls? If enterprise use expands faster than efficiency reduces consumption, hyperscaler revenue continues to compound and the CapEx case strengthens. If optimisation cuts consumption but workflows remain stuck in pilot, the revenue bridge becomes much harder to close I also expect the distribution of returns to remain wide A smaller cohort of AI mature enterprises will produce credible revenue, cost and margin gains.

[38:07] Many others will report productivity improvements without a clean P&L effect, That is not evidence that AI is useless it is evidence that technology access has moved faster than organisational redesign.

[38:20] So for hyperscalers, the near-term evidence remains constructive. Enterprise revenue is arriving. Cloud growth is strong and backlog provides visibility. But the standard of proof is rising. Investors will increasingly ask not only how much capacity is being built, but how much incremental gross profit and durable customer value each unit of capacity produces. The bull case is a virtuous cycle. Falling cost per outcome drives broader enterprise adoption. Roder adoption creates measurable value, that value supports renewals and larger consumption commitments, and those revenues validate the infrastructure build-out. The bear case is a broken bridge. Enterprises buy seats and capacity but fail to redesign work. Token optimization reduces the bill without expanding useful workloads. Renewals weaken, and hyperscaler depreciation rises into a slower revenue base. The QFMI base case sits between those extremes. The AI supercycle continues, but its next phase is more discriminating. The binding constraint begins to move from access to compute towards the ability to turn compute into enterprise value. That is the other side of the equation. Capital formation determines what can be built. Enterprise ROI determines what deserves to keep being built. That is the QFMI base case for enterprise ROI and token economics.

Closing Thoughts on the Cycle
[39:48] Okay, so let's bring this to a close and also take a look forward for the next week. The physical stack remains essential. Chips, memory, data centres, power, materials and networks. But none of it sits outside economics. The entire stack ultimately depends on useful work being produced at a cost customers are willing to pay and on those customers being able to show why they should keep paying it. Okay, so that's it for this week. I hope you found that of some value. Next week we'll be talking about physical ai robotics and embodied intelligence, then after this one we'll be taking a break for the summer and when we come back we'll be inviting guests on the show to talk about the ai super cycle, for more information head on over to qfmi.substack.com where you'll find all of our latest reports the market pulse and in the spotlight articles are free and always will be, the weekly outlook, the weekend debrief, and the monthly strategic research reports sit behind a paid subscription. Subscribe and you get the full picture. Thank you for listening to the AI Supercycle Podcast. See you next week.