Subscribe
Share
Share
Embed
The AI Supercycle podcast from QF-MI provides independent Capital Intelligence for the AI Industrial Economy. From semiconductors, AI factories and data centres to energy markets and power grids, critical materials, orbital compute and embodied intelligence, we track where capital is being deployed, where the binding constraints are emerging, and what it means for traders, investors and the broader economy.
Each episode examines the physical infrastructure underpinning artificial intelligence and the investment opportunities emerging from its industrialisation. Published weekly by Quantum Fields Market Intelligence (QF-MI).
[0:11]Hello and welcome to the AI Supercycle Podcast, where we follow the capital 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. We track where the capital is being deployed, where the binding constraints are emerging, and what it means for traders, investors, and the broader economy. And yes, if you listened to the podcast last week, you may have noticed we've changed the name. Cure for My Market Intelligence was always a working title while we settled on the right one. The AI Supercycle is really what this show is about, so that's what we're calling it. You'll also find the AI Supercycle newsletter on LinkedIn and our full research on Substack. Today is Friday the 19th of June, 2026. I'm Tim Hardwick, founder of QFMI. Last week we covered the first binding constraint in the AI supercycle, high bandwidth memory. That framework carries forward into everything we discussed today, and if you want the full picture, episode 2 covers this in depth.
[1:21]Delivered Power Constraint
[1:22]This week we move to the second binding constraint, delivered power, and I want to emphasize that word, delivered. It's not enough to generate electricity. Power that exists on the grid is not the same as power that can be delivered to the AI factory.
[1:37]That can be delivered to the data center. The binding constraint is not generation capacity, it is deliverable load capacity. Electricity has to reach the data center continuously and reliably in sufficient quantity to support hundreds of megawatts of compute load. That journey from generation through transmission lines, substations and local distribution networks to the data center itself is where the constraint actually bites. The last mile of the grid has become one of the defining bottlenecks of the AI industrial economy.
[2:09]Gulf Deal and Oil Flows
[2:09]But before we get into the topic of power constraints, let's take a look at some of the market developments over the last week. And the first on the radar is the US-Iran deal and energy flows. The biggest macro development over the last week, of course, has been around this US-Iran memorandum of understanding. And we've been tracking this situation in detail within QFMI and I want to give you the summary read of what it means for the power constraint we are discussing today. The 14-point MOU calls for a ceasefire across all fronts, the reopening of the Strait of Hormuz, a moratorium on uranium enrichment in Iran and a partial sanctions lift with a 60-day window for comprehensive nuclear negotiations. Both presidents have signed US Central Command confirmed yesterday that the blockade on all maritime traffic entering and exiting Iranian ports has been lifted On paper, the deal is in effect and the strait is open.
[3:08]But here is what is actually happening on the water. Maersk, the world's largest container shipping line, is still restricting cargo bookings across the Gulf. Cargo bookings for next week, through the strait, are not materialising.
[3:23]There are roughly 118 tankers stranded in the Gulf, according to Kepler. And analysts estimate 10-15 days to clear the backlog, and that will not amount to a full recovery. Mines are still in the water Insurance markets have not yet recertified The shipping industry's position is straightforward It's still too risky to transit, Meanwhile, the Switzerland talks that were scheduled for today have been postponed, Iran delayed sending its delegation Citing both the implementation sequencing of the MOU And Israel's ongoing military campaign in Lebanon Amen.
[4:04]The deal is in effect, but it is fragile. Fresh Israeli strikes in Lebanon today underscore just how easily this could unravel. The most likely path remains a slow, messy normalisation that relieves the acute energy crisis but leaves the nuclear question unresolved.
[4:21]But the range of outcomes from here is wide, and we are tracking this as a live situation. However, the financial markets have already priced the normalisation. Brent has fallen from around $120 at the peak of the crisis to roughly $79 this morning. But the physical reality has not caught up. The blockade is lifted on paper, but shipping lines are not sending their tankers through. That gap between the financial price and the physical situation on the water is exactly the kind of disconnect our framework is designed to identify. And it's a perfect illustration of the two-speed problem we're about to discuss in the context of delivered power. Now here is where this connects to today's topic. Cheaper oil is good news for data center operators in the near term. Energy is a major input cost and lower prices improve the return on investment calculus for every hyperscaler build-out. But cheaper oil does not solve the physical grid constraint. It does not add transmission capacity. It does not shorten interconnection queues It does not build transformers faster, If anything, cheaper energy makes data centre build-out more economically attractive which could actually accelerate demand and tighten the power constraint further.
[5:40]That is a subtle but important distinction. The most likely outcome is a slow normalization that eases energy costs but leaves the physical infrastructure gap exactly where it was. And that brings us to today's main subject.
[5:56] SpaceX, Intel, and Capital Flows
[5:56]But first, let's cover a few more of the developments in the markets this week. SpaceX went public a week ago at $135. It rose 19% on day one to close at $161. By Tuesday, it had surged to $225, surpassing the market cap of Amazon, but since then it has pulled back to around $185, which is still 37% above the IPO price, but 18% off its highs. Two things worth noting. First, the market absorbed $75 billion in new equity, and then bid it up another 30% even after the pullback. The capital absorption concern we flagged in episode 2 has been answered, at least for now. Investor appetite for AI infrastructure equity remains extraordinary. Second, SpaceX acquired Cursor this week, the AI coding tool, giving it a foothold in enterprise AI software development alongside its hardware and launch business.
[7:00]The next supply events to watch are the lock-up releases, and the structure here is unusual. Rather than a single event, SpaceX is using a staggered system. Up to 20% of the restricted shares can be sold after the Q2 earnings release, with an additional 10% contingent on the stock trading at least 30% above its offering price, a trigger that has already been met at current levels. Further tranches release in stages through to the main 180-day expiry on 8 December. Musk and other significant investors are locked for 366 days, one year and one day, with no early release provisions. Musk's 6.4 billion shares first become eligible on 12 June 2027. The staggered structure is designed to avoid a single supply cliff, but it means potential selling pressure is spread across the next six months, rather than concentrated on the one date, that is worth monitoring.
[8:00]And then yesterday, a development that connects directly to both the memory constraint we covered last week and the power constraint we're covering today, President Trump announced that Apple has agreed to work with Intel to design and manufacture chips in the United States. Think about what this means in the context of the diversification story that we've been tracking. In episode two, we covered Google ordering TPUs from Intel, NVIDIA evaluating Intel's Foundry, and Tesla on Intel's next-generation 14A process.
[8:31]Apple joining that list makes this an industry-wide shift. Every major hyperscaler and consumer technology company is now actively diversifying fabrication away from TSMC. Intel closed up over 10% yesterday, to all-time highs. The market is repricing Intel as a strategic US manufacturing asset, with government backing. This is industrial policy driving, supply chain reshoring. And here is the direct connection to today's episode. More US-based fabrication means more US power demand. Semiconductor fabs are extraordinarily power-intensive facilities. Intel's plants in Arizona, Ohio, and Oregon will need substantial grid capacity to serve Apple, Google, Nvidia, and Tesla simultaneously. Every new customer Intel lands increases the power draw on the US grid. The reshoring of chip manufacturing is not just a semiconductor story, it's a power story. So here's a quick update on the capital formation picture because the numbers just keep getting bigger.
[9:40]Anthropic closed a $65 billion Series H at $965 billion post-money valuation, surpassing OpenAI for the first time. They filed a confidential S1 with the SEC for the 1st of June, so that IPO is coming. SpaceX raised $75 billion. OpenAI's $122 billion funding round closed in March. Add it all up and the running total of fresh capital deployed into the AI complex is around $500 billion.
[10:13]The IPO pipeline ahead is significant. Anthropic and OpenAI have both filed with the SEC and could list this autumn. Databricks may also be eyeing a debut. The market absorbed SpaceX well, but the capital absorption test is far from over.
[10:29] Why Power Is Different
[10:30]Okay, so in the last episode we talked about memory, high bandwidth memory, and the memory constraint that the markets are pricing in. That supply chain is tight, but the industry is responding, fabs are expanding, capacity is being added, the timeline is measured in quarters. Power is different. Power is the constraint where the two-speed problem becomes most acute. Now I want to be precise about how this constraint actually works, because I think it is widely misunderstood. A hyperscaler building a data center does not need to wait for a new power plant. It connects to the existing grid. It signs a supply agreement with a utility and draws from the existing generation capacity. Data centers are being built right now, large ones, and they are doing exactly that. The constraint is not about any individual project, the constraint is cumulative.
[11:28]Every data center that connects to the grid reduces the available capacity for the next one. Energy prices rise for local residents and businesses because demand on the same infrastructure has increased. And we're seeing that communities are pushing back. And utility companies do not build new generation capacity because someone wants a data center.
[11:50]They plan generation investment years in advance based on projected demand growth. And AI data centres have arrived as a demand shock that was not in anyone's projections three years ago So let me put some numbers around this to give you a sense of scale, A single modern AI data centre campus can consume between 100 megawatts and a gigawatt of electricity, A gigawatt That is roughly the output of a large nuclear power station Or enough to power a city of 700,000 to a million people One data center and one city's worth of electricity. And the major hyperscalers are not building one of these. They are building dozens of these. Microsoft, Google, Amazon and Meta have collectively announced data center capacity that will require tens of gigawatts of new power over the next five years. The United States has added approximately 20 gigawatts of new generation capacity last year across all sources. The AI industry alone may need that much or more over the next few years, on top of everything else the grid already serves. Homes, businesses, industry, electric vehicles, and so on.
[13:04]So the first wave of data centers gets built because grid capacity exists today. They plug in and they work. But the second wave competes for tighter capacity. The third wave hits a wall. And that wall is where the two-speed problem becomes acute, the point where cumulative demand exceeds what the existing grid and existing generation can serve. And new capacity takes five to ten years to add. Now a reasonable person in listening to this might say, hang on, there is an enormous amount of new power generation being built around the world right now. Nearly 750 gigawatts of gas-fired capacity is in development globally. The renewable energy pipeline has reached nearly 5 terawatts. There is a nuclear renaissance underway across Europe, Asia, the Middle East and North America. Surely all of that is enough. And the answer requires you to look more carefully at four things. First, the new capacity is not being built for AI. It is replacing aging plants, meeting general demand growth, electrifying transport, powering reshored manufacturing, and fulfilling decarbonization commitments.
[14:17]AI data centers are incremental demand on top of all of that. The question is not whether total new capacity exceeds total AI demand. It is whether the surplus, after serving everything else, is sufficient. Second, geography matters. Much of the generation build-out is in China, the Middle East, India and Southeast Asia. The hyperscalar data center build-out is concentrated in the United States, with secondary hubs in Europe and select locations in Asia-Pacific. Global capacity does not help if it is not where the data centers are being built. Power does not ship across the oceans. Third, and I think this is the most revealing signal of all, some of the largest AI companies have given up waiting for the grid entirely. OpenAI and Oracle are constructing a natural gas facility in Texas specifically to power the Stargate AI data center complex. Now let's just let that sink in for a moment. A software company is building a power plant. Now, you don't do that because you want to. You do that because the grid cannot serve you. It's a business proposition. And fourth, timing. The nuclear renaissance is real.
[15:30]But most of that is in the second half of this decade. So Hinkley Point C in the UK is targeting 2029. Sizewell C is mid to late 2030s. TerraPower's first reactor in Wyoming is expected in 2030, Small modular reactors across Europe and North America are still in design and permitting phases The generation capacity is coming, but it is coming on a 5-15 year clock The AI data centre demand is arriving now.
[16:00]So yes, there is a massive global build-out of power generation underway, but when you account for what that capacity is actually serving, where it is located, the timeline on which it arrives, and the fact that the biggest AI companies are already building their own power plants because the grid cannot serve them. The supply and demand equation does not balance. The constraint is real. It's structural, and it's going to get worse before it gets better.
[16:27] Four Paths to More Power
[16:28]Okay, so where is this power going to come from? There are essentially four pathways, and each has its own constraints and timelines. The first is natural gas. This is the fastest to deploy at scale and will do most of the heavy lifting in the near term. Gas turbine plants can be built relatively quickly, and they are well-understood technology, and the United States has abundant domestic gas supply. But gas plants require pipeline infrastructure, they face permitting challenges, and they produce carbon emissions. For hyperscalers with net zero commitments, gas is a pragmatic bridge, but not a long-term answer. The second is nuclear. Nuclear is where the conversation has shifted dramatically in the last 12 to 18 months. The existing fleet of nuclear power stations in the United States and Europe is being repriced as a strategic asset. We are seeing hyperscalers signing long-term power purchase agreements directly with nuclear operators. Microsoft's deal with Constellation Energy to restart the three-mile island, Unit 1, was one of the first high-profile examples of this. Amazon has signed agreements with Talon Energy for nuclear-adjacent data center capacity at the Susquehanna plant, and Google has signed a purchase agreement for power from Kairos Power's small modular reactor technology. The nuclear story has two timelines. The existing fleet can be life-extended and in some cases restarted relatively quickly, within 2-4 years.
[17:56]But nuclear capacity that is new, including small modular reactors, remains 5-15 years away from meaningful scale. The technology is promising. The regulatory pathway is becoming clearer. But the deployment timeline is long. Nuclear is part of the answer, but it is not a near-term fix.
[18:16]The third pathway is renewables. Solar and wind are being deployed at record pace, and the cost curves continue to fall. But renewables have an intermittency problem that is particularly acute for data centers. AI workloads run 24 hours a day, 7 days a week. They require baseload power with very high reliability. Solar generates during daylight hours. Wind is variable. Battery storage is improving, but remains expensive at the scale required to back a gigawatt data center campus through the night. Renewables will contribute, but they cannot solve the baseload problem on their own without massive storage build-out. And then there is the pathway that sounds like science fiction, until you think about the engineering. Every constraint we have just discussed, the grid queues, the transformer shortages, the permitting timelines, the intermittency of terrestrial renewables, all of them are terrestrial problems. In orbit, solar power is continuous. There is no night cycle, no cloud cover, no atmosphere attenuating the signal. There is no grid to connect to. No permitting process. No five-year interconnection queue. The barrier has always been launch cost, and that is exactly what SpaceX has spent 25 years dismantling. Launch costs are falling by roughly two orders of magnitude.
[19:39]The company that just went public at a $1.75 trillion valuation is not just a space company. It's potentially an infrastructure company that offers a pathway around the most intractable bottleneck in the AI supercycle. And we're tracking this within QFMI as part of our space and orbital infrastructure research. And it's going to get its own episode in due course, because the intersection of space infrastructure and AI power demand is one of the most underappreciated themes in the market today.
[20:10]Each of these pathways is being pursued simultaneously and none of them are sufficient on their own. The realistic picture over the next five years is a combination of gas, nuclear, renewables and potentially orbital infrastructure heavily weighted towards gas in the near term with nuclear and renewables growing their share over time. The constraint is not that solutions do not exist, the constraint is that none of them can be deployed fast enough to match the pace of which AI compute demand is growing.
[20:40] Grid Bottlenecks Bite Hard
[20:41]There is a second bottleneck that most people are not paying enough attention to. The grid itself. And honestly, it's not hard to see why it gets overlooked. Transformers and transmission lines are not exactly the glamorous end of the AI trade, right? but they are binding. Generation is one thing, transmission is another. The electricity has to travel from where it is produced to where the data centre is located and the transmission infrastructure in the United States is ageing, congested and under-invested. The interconnection queue, the backlog of power projects waiting to be connected to the grid, now exceeds 2,000 gigawatts.
[21:20]Most of these projects will never be built, but the queue itself tells you something about the scale of demand and the difficulty of getting connected. The average wait from application to connection has stretched to roughly five years. For a hyperscaler that wants a data centre operational in two years, a five-year interconnection timeline is not a bottleneck, it's a showstopper. The regulators know it. Yesterday, the Federal Energy Regulatory Commission ordered six regional grids to justify or overhaul their processes for connecting large energy users like data centres. Grid operators have got 60 days to respond. The FERC chairman called this, and I'm quoting this directly, the biggest priority our country is facing at the moment. She also said, this is a race against time and we are going to win When the top energy regulator in the United States is using language like that You know the constraint is not theoretical It is operational and it is urgent.
[22:24]FERC is also encouraging frameworks for large energy users to bring their own power supplies, which tells you the regulator has accepted that the grid alone cannot solve this. And then there are the transformers themselves, large power transformers, the kind used in high-voltage transmission. They've got lead times of three to four years with only a handful of manufacturers globally. Without transformers, new generation cannot reach the grid and new data centres cannot receive power. This is reshaping where data centres get built. Siting has always been a multi-factor decision. Proximity to internet exchange points, water supply for cooling, real estate costs, permitting environments, tax incentives, labour, climate, natural disaster risk. But increasingly, power is the factor that gates everything else. You can have the perfect fibre, cheap land, fast permitting and abundant water. If the grid cannot deliver the megawatts, none of it matters. The first movers secure the best sites. The later entrants face a very different proposition.
[23:34] Investing in Power Assets
[23:35]So now let's turn our attention to the investment landscape. The power constraint has created a distinct investment landscape that looks very different from the semiconductor complex we discussed last week. In semiconductors, the investment opportunity is concentrated in a handful of names. NVIDIA, SK Hynix, TSMC, ASML, Micron. The supply chain is narrow and the equity expressions are well known.
[24:00]The power side is much broader, it's more fragmented and in many cases less well understood by the market. It spans several subsectors. Utilities and independent power producers are the most direct beneficiaries. Companies with large existing generation fleets, particularly those with nuclear assets, have been repriced significantly. Constellation Energy, Vistra, Talent Energy, NRG. These names have moved sharply higher as the market recognizes that their existing power generation capacity has become a strategic asset in the AI build-out. The nuclear and uranium supply chain is another layer. The spot price of uranium has moved substantially over the past two years, driven by a combination of supply deficits, utility restocking, and the AI-driven demand narrative. The uranium miners, such as Cameco, and the smaller producers, have repriced accordingly. Now, we're discussing uranium in this episode in the context of nuclear power generation. In the next episode, on critical materials, we'll return to uranium from the mining and supply side perspective, because the materials constraint is a distinct analytical question from the power constraint.
[25:12] Base Case: Constraint Deepens
[25:13]Okay, so let's look at the base case on delivered power. And this part of the episode is really where the framework that we use at QFMI commits to a reading. My base case on the power constraint over the next three to five years is that it deepens before it eases. Unlike memory, where supply is being added on a quarterly cadence, power supply operates on a multi-year cycle that cannot be compressed. The constraint is structural, not cyclical. and the market is still in the early stages of pricing its duration and severity. On the supply side, natural gas generation will do most of the near-term heavy lifting, but it faces permitting headwinds, pipeline constraints and carbon scrutiny. Nuclear restarts and life extensions will add meaningful baseload capacity over the next two to four years, but new nuclear builds, including small modular reactors, remains a second half of the decade's story, at the earliest. Renewables will grow, but cannot solve the baseline problem without a step change in battery storage economics. Grid transmission remains in the binding constraint behind all of this, with interconnection queues stretching to 5 years and transformer lead times at 3-4 years.
[26:25]On the demand side, the trajectory is unambiguous. Hyperscaler data center build-out is accelerating. Each new generation of AI accelerator consumes more power than the last. Inference workloads are scaling rapidly and they run continuously, unlike training runs which are episodic. The electrification of transport, the reshoring of manufacturing and the broader decarbonization agenda are all competing for the same grid capacity simultaneously. AI is not arriving into a grid with spare capacity. It is arriving into a grid that is already under strain. The PRISM framework that we use at QFMI assigns a probability of approximately 70% that the power constraint becomes the primary binding constraint on the AI supercycle by late 2027, and that is overtaking memory as the acute bottleneck. The reasoning is that memory supply, while tight, is being addressed through capacity expansion at SK Hynix, Samsung and Micron on a timeline of quarters.
[27:29]Power supply cannot be addressed on that timeline. The mismatch between the two-year demand clock and the five to ten-year supply clock is widening, not narrowing. The principal risk to this view would be a significant breakthrough in deployment speed for new generation or grid infrastructure, or a material slowdown in hyperscalar build-out that reduces power demand growth. On the breakthrough side, the pathway to watch is orbital compute. SpaceX disclosed in its IPO filing that it believes it is the only company with a viable path to orbital AI compute at scale. Reuters reported this week that executives outlined a roadmap for demonstration missions in 2027, ahead of the 2028 commercial timeline in the filing. The company has requested permission to launch up to 1 million space-based data center satellites. That is the long-dated option, not a near-term solution, but it is the only pathway that bypasses the grid entirely, and that makes it worth tracking. On the upside, geopolitical disruption to energy supply, whether through the Middle East, the Strait of Hormuz, or sanctions on energy exporters, could accelerate the constraint and drive power-related equities significantly higher.
[28:46]The indicators that I'm watching most closely are hyperscaler power procurement announcements, grid interconnection timelines, transformer order books, and uranium pricing. On the regulatory side, anything that accelerates permitting for new generation or transmission could shift the timeline meaningfully. And as with memory, we need to watch the capital formation side. The power build-out required for the AI supercycle will cost hundreds of billions of dollars. Whether returns justify that investment will determine how quickly the constraint can be addressed. This is the base case as it reads today. As new evidence arrives, we'll revise this.
[29:26] Power Wins the Bottleneck Race
[29:26]So that is the power constraint. Memory was the constraint the market noticed first. Power is the constraint that runs deeper. The two-speed problem means that this one cannot be solved with the urgency that the semiconductor industry is bringing to HBM.
[29:41]You cannot build a grid in quarters. The generation capacity does not yet exist. The grid was not designed for this, and the companies that own existing baseload power, particularly nuclear, are being repriced as strategic assets, because in the AI supercycle, the most valuable thing you could own might not be a chip, it might be a power station. The base case is that the constraint deepens before it eases, and the market is still in the early stages of pricing its severity. Okay, so that's about it for this week. 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 all sit behind a paid subscription. Next week in episode 4 we'll cover the latest market developments and take a look at the third binding constraint critical materials, this includes copper and uranium from a mining and supply chain perspective and the physical inputs that underpin everything else in the ai build out thank you for listening to the ai super cycle podcast see you next week.