Welcome to Crack The Market

Crack The Market exists to help you own equities you truly understand. Not headlines, not narratives, but the actual structure of a business, where it sits in its value chain, and whether the forces around it are working for or against it.

I write it from the perspective of a former portfolio manager, with one objective: help you think like an owner of businesses and an institutional investor.

Who is writing this

I am a former portfolio manager who co-managed a multi-billion dollar fund, with years of experience investing across global equities, meeting management teams and analysing industries from a portfolio construction perspective.

I no longer run money for clients, but I still do the work the same way: I meet management teams, industry experts and other investors regularly, I read transcripts when they land, and I build the value chains myself. That is what you are reading here, translated into clear, actionable research.

A few examples of what that looks like in practice:

• I met Besi’s management before writing the deep dive on hybrid bonding, which is why the piece reads differently from coverage built on filings alone.

• Inside the Room is the write-up of several days spent with the people running the AI economy, the equipment makers, the foundries and the investors financing them, and it reads the way it does because I was in the room rather than reading the notes afterwards.

• When Nvidia, ASML, TSMC or the hyperscalers report, I read the transcripts the same morning and publish into fresh prints, not a week or a month later.

How I think about markets

Every piece on Crack The Market is built on three questions: does this business have a durable competitive advantage, where does it sit in its value chain, and how exposed is it to a megatrend that will still be running in ten years. Company analysis always connects back up to the structural picture, and thematic work always comes back down to specific names. I think like a portfolio manager and an asset owner, not a trader.

The research is long horizon and opinionated, and I hold myself to it: when I change my mind on something I have published, I say so in writing.

Where to start

If you only read three pieces, make it these. Two are free and one is paid, and between them they show what this publication does, from the structural map to the names I would actually own.

• Electricity is the New Oil. The load growth and grid investment supercycle. Free, structural, and the piece that convinced many of you to subscribe.

• How to Invest in the $6-7tn Race. The full AI value chain, from the power plant to the application layer, every layer priced per gigawatt and every company mapped. Free, and the map I continue to update going forward.

• My AI Selloff Shopping List. What I would actually buy, at what price, and why. Paid, and the clearest example of what a subscription delivers: the names and the verdict, not just the framework.

What I publish

More than 180 pieces so far, across four formats:

• Company deep dives, more than 80 published. Business model, value chain position, moat, capital allocation, financials, management, valuation, and a clear view. From Nvidia, ASML and TSMC to Besi, Coherent, Bloom Energy and Thermo Fisher, with the full list by theme below.

• Thematic and value chain primers. The definitive maps of a sector: Semiconductors, Semiconductor Equipment, the European and US Photonics Supercycles, The Memory Singularity, Token Economics, Who Powers AI, Healthcare at the Turning Point, Global Food Crisis and Fertilizers.

• Macro and event pieces, always with a positioning payoff. When something moves markets I write what it means for what you own: 3 Scenarios for the Middle East Crisis, The Worst Energy Crisis in History, The Great AI Reckoning, Is the AI Trade Financeable, Nvidia Just Repriced the Entire AI Value Chain.

• Actionable buy lists when the market gives a setup. The AI Selloff Shopping List, the Electrification Selloff Shopping List, Software Armageddon and what to buy, the Memory Basket with a verdict on every name. These are the pieces readers tell me they act on.

The six themes that organise everything

All content on Crack The Market fits into six enduring forces reshaping the global economy. These are not short-term themes, they are multi-year, often multi-decade shifts, and every company is analysed inside its ecosystem rather than in isolation.

Companies can be in multiple themes at the same time.

AI & Technology

The AI investment supercycle: semiconductors, equipment, memory, optical and networking, software, compute, data centers and digital infrastructure.

Start with: How to Invest in the $6-7tn Race, then Token Economics, The EU/US Photonics Supercycles, The Memory Singularity.

Companies covered (44), including: Nvidia, Broadcom, Alphabet, Microsoft, Oracle, OpenAI, SpaceX, Nebius, ASML, TSMC, Samsung, SK Hynix, Micron, SanDisk, ASM International, Besi, Technoprobe, Aixtron, Soitec, Siltronic, Coherent, Lumentum, Nokia, STMicroelectronics, Infineon, ARM, Cadence, Dassault Systemes, SAP, Salesforce, Shopify, Bentley Systems, Flex.

Electrification & Energy

The $30tn electrification and power buildout: grids, generation, nuclear, cables, turbines, solar, behind-the-meter power, and who actually powers AI.

Start with: Electricity is the New Oil, then Who Powers AI and the Electrification Selloff Shopping List.

Companies covered (29), including: GE Vernova, Siemens Energy, Constellation Energy, NextEra Energy, Iberdrola, Nexans, Prysmian, Quanta Services, Hubbell, Itron, First Solar, Schneider Electric, Siemens, Vertiv, Bloom Energy, Cameco, Arcadis and Hoffmann Green Cement, plus CTOS and ALTG on the small cap execution side, plus my nuclear, energy and copper work.

Reshoring & Sovereignty

Deglobalisation, industrial policy, defence, critical minerals, energy security and supply chains, and the companies that win in a world of tariffs.

Start with: 3 Themes and 40+ Stocks for a Deglobalizing World, then Global Food Crisis: Fertilizers and the Refining Crisis trilogy.

Companies covered (51), including: Nucor, Rheinmetall, Exail, Sopra Steria, MP Materials, Umicore, CF Industries, LKQ, Kurita, Valero, Equinor, Cheniere and Venture Global, plus the steel and aluminum, resources shock and seabed work.

China & Asia

The other half of the AI and industrial story, and the one Western investors most consistently misread: Chinese semis, memory, AI, and the export control fault line.

Start with: The $1 Trillion Memory Singularity, then CATL and Why India Is Such an Attractive Investment Destination.

Companies covered (9), including: TSMC, SK Hynix, Samsung, CATL, Kurita, alongside the India country piece.

Healthcare & Longevity

Pharma, life science tools, GLP-1, and a sector generalists abandoned that is now turning.

Start with: Healthcare: The Turning Point Is Now.

Companies covered (7), including: Eli Lilly, Novo Nordisk, Thermo Fisher, Danaher and Zoetis.

Resilience & Quality

Businesses with pricing power, net cash and durable moats, including the software names now recovering.

Start with: Software Armageddon, then RELX, the company to own forever, and the free Ingredients primer.

Companies covered (61), including: RELX, Atlas Copco, Roper, Halma, Copart, Waste Management, Air Liquide, FICO, CPKC, TJX, Trex, Knorr Bremse, Badger Meter, Veolia, Givaudan, DSM-Firmenich and Croda, plus the ten software names from Software Armageddon.

Join as a paid subscriber today to get access to institutional level research helping you own equities you understand and learn about the trends, technologies and companies that will underpin the value creation of the next decades.

Join Crack The Market

Track record, including what I got wrong

I think a research publication should be judged on its calls rather than its prose, so here are mine, dated, including the ones that did not work.

• Microsoft, 19th of April 2026. I published The Best Company In The World Is Too Cheap twelve days after the stock bottomed, into its worst quarter since 2008 and a consensus that AI capex had broken the story. It traded above $460 in late August on the print that proved AI monetisation was accelerating, which was the argument.

• The Google inference thesis, 23rd of November 2025. Google Is Breaking Nvidia’s AI Dominance argued the shift to inference favoured Google’s stack when consensus still had Alphabet as the AI loser. The stock is up roughly 65% over the twelve months since.

• Healthcare: The Turning Point Is Now, 1st of October 2025. A call on a sector generalists had abandoned. I published the follow-up nine weeks later, when the names were up 10% in November alone.

• The AI Selloff Shopping List, 19th of July 2026, published into the drawdown and the most read piece I have written, followed by the Electrification Selloff Shopping List on 31 July. Both are six weeks old. I will grade them at the six-month mark rather than claim a win on six weeks.

• Soitec. I flagged it as a photonics winner in The European Photonics Supercycle in May followed up in a piece pointing out that investors were pricing too much of the photonics growth and that I would own this photonics monopoly play only if it came down significantly. On the 27th of July, I laid out why the price finally made sense. Both updates are dated and public.

And the ones I got wrong, which matter more:

• Memory. In the September 2025 value chain piece, How to Invest in the $3-4tn Race, I under-weighted memory as a layer of the stack. It became the best performing part of the entire trade. I said so plainly in this year’s $6-7tn update and built Micron, The $1 Trillion Memory Singularity and The Memory Basket to correct it.

• Alphabet, 25th of July 2026. I had the thesis right in November and was late with the deep dive. By the time the full work was published, most of the re-rating had already happened. Early in a note and late with the deep dive is still late.

• Software Armageddon, 4th of February 2026. I argued the AI disruption thesis was right on direction and wrong on timing, and named ten stocks to play the rally. The direction held and software has led since May. The timing did not: the sector fell another leg into April and did not trough until mid-July. Right thesis, months early, which costs real money.

What do subscribers get?

Free subscribers:

• Selected research, including several of the flagship primers: Electricity is the New Oil, How to Invest in the $6-7tn Race, Semiconductors: The Most Important Sector in the World, and free deep dives like ASML, FICO, Constellation Energy, CTOS/ALTG, Umicore and Siltronic.

• Previews of every paid piece.

• Notes and market observations between pieces.

Paid subscribers:

• Full access to all company and megatrend deep dives, more than 180 pieces and growing 2 to 3 times a week.

• The actionable buy lists when the market gives a setup worth acting on.

• The subscriber chat, where I post real-time reads on prints, policy and moves between pieces, and answer questions directly.

• Coverage requests. Suggest a company or a theme and it moves up my queue. Several deep dives exist because a reader asked.

Founding members:

• Everything above.

• Priority on coverage requests. Your suggestion goes to the front of the queue and I tell you when it is scheduled.

• Direct answers. Reply to any piece with a question on a company or theme I cover and you will get a detailed answer.

• You are funding the long-term development of the publication, and I take that seriously.

Join Crack The Market

What I commit to

• Two to three deep dives every week.

• Actionable buy lists whenever the market gives a setup that justifies one.

• Real-time reads in the chat between pieces.

• Updates whenever I change my mind on something I have published.

Join as a paid subscriber to get access to institutional level research helping you own equities you understand and learn about the trends, technologies and companies that will underpin the value creation of the next decades.

If you want to become a better investor by understanding the businesses and trends shaping the next decades, you are in the right place.

Disclaimer: The information provided on this Substack is for general informational and educational purposes only, and should not be construed as investment advice. Nothing produced here should be considered a recommendation to buy or sell any particular security.

Twice a week, I will release deep dives into stocks and sectors that fit into the 6 themes that I see winning in the coming years and decades: AI & Technology, Electrification & Energy, Reshoring & Sovereignty, Healthcare & Longevity, China & Asia, Resilience & Quality.

Join Crack The Market

Join Black Line

Take advantage of this once in a generation opportunity to build long term wealth by investing in great stocks that will deliver returns for your portfolio for years to come.

The Engineering of Time

often discusses Physical Sovereignty. It is the principle that every digital ambition eventually collides with a physical requirement. Code needs chips. Chips need buildings. Buildings need cooling, cables, transformers and electricity. The cloud has never floated. It has always had a body.

By 2026, however, identifying that body is no longer enough.

Investors already understand that artificial intelligence requires more power. Utilities have rerated. Nuclear plants have become strategic infrastructure. Turbine manufacturers are selling production slots years in advance. Grid equipment has moved from an industrial afterthought to a national constraint. The market has discovered the electron.

The harder question is what happens after everyone discovers it.

Electricity is not valuable merely because it exists. It must exist in sufficient volume, with sufficient reliability, in a precise location, on a date that matches the deployment of the compute waiting for it. A megawatt arriving in five years cannot operate a cluster scheduled for next year. National abundance cannot energize a local site if transmission, generation or connection rights are missing. A cheap electron delivered too late may be economically worthless.

This changes the asset being sold.

The new scarcity is not only energy. It is certainty. More specifically, it is the ability to guarantee when firm power will become available.

Today we present a new collaboration. Black Line maps the physical system, the power mechanism and the transfer of strategic value. I dissect the company, its technology, its contracts, its economics and its valuation. The object is Bloom Energy, but the case is larger than a fuel cell manufacturer.

It is a case study in how the race for artificial intelligence is turning time into an energy product.

The Consensus Is Already Right

The first layer of the thesis is real. Global data center electricity consumption reached roughly 485 terawatt hours in 2025 and is projected to approach 950 terawatt hours by 2030 in the central outlook of the International Energy Agency. Electricity consumption from data centers focused on artificial intelligence could triple over the same period. The IEA also warns that the investment required has grown too large to be financed from technology company balance sheets alone.

The United States is the sharpest expression of the collision. Data centers consumed approximately 176 terawatt hours in 2023. The Lawrence Berkeley National Laboratory estimates that consumption could reach between 325 and 580 terawatt hours by 2028, equal to between 6.7% and 12% of total US electricity demand. The range is wide because adoption, efficiency, utilization and construction remain uncertain. The direction is clearer than the final number.

But “AI needs power” is no longer an investable discovery by itself.

It has become a category. The market speaks openly about the AI power trade. Nuclear generators, gas turbine manufacturers, electrical equipment suppliers and owners of powered land are routinely presented as derivatives of the same buildout. Bloom itself has become one of the most visible expressions of that narrative.

Repeating the consensus would confuse a correct observation with a new thesis.

The second layer begins with a distinction. Electricity demand is measured in energy and capacity. Economic damage is measured in time.

A data center may be physically completed while its electrical connection remains years away. The servers may be ordered. The financing may already be accruing interest. The customer may be waiting. Yet the entire structure remains unproductive because the site cannot receive firm power at scale.

The loss is not limited to an electricity bill. It includes delayed cloud revenue, idle accelerators, missed customer commitments and a weaker position in an industry where each generation of compute changes the competitive frontier.

Once the cost of waiting exceeds the premium for an alternative power system, generation behind the customer’s meter stops looking expensive. It begins to look like insurance against stranded compute.

The Scarcity of a Delivery Date

The power system and the technology system do not operate at the same speed.

The International Energy Agency estimates that a data center can move from development to operation in roughly two or three years. The wider energy system often requires longer periods to plan, permit and build new generation, transmission and connection infrastructure. Transformers, turbines and specialized electrical equipment have their own production constraints. Local utilities must model reliability. Regulators must decide who pays for upgrades. Communities must decide what they are willing to host.

The result is not one national queue. It is thousands of local mismatches between ambition and deliverability.

This distinction matters because a data center is an unusually concentrated load. A city can add homes, offices and electric vehicles gradually. A hyperscale campus can request hundreds of megawatts at one point on the map. The largest proposed campuses are measured in gigawatts. Their demand is also persistent. Servers do not wait for the afternoon peak or a favorable wholesale price. They are designed to consume power through the day and through the year.

The traditional response is to expand the grid. That remains necessary. It is also slow.

The alternative is to move generation to the load.

This is where the economics invert. A developer is no longer comparing only the levelized cost of two electrons. It is comparing two calendars. One option may offer cheaper grid power after a long and uncertain process. Another may cost more per unit but allow billions of dollars in compute to begin producing revenue earlier.

The relevant calculation becomes:

What is the value of each month removed from the path to operation?

That value creates what we can call the time to power premium.

The premium has three components. The first is speed. The second is reliability. The third is control. Speed brings the site online. Reliability protects the workload. Control reduces dependence on decisions made by utilities, grid operators and regulators whose incentives extend beyond one customer.

Bloom Energy is attempting to sell all three.

Its Energy Server uses solid oxide fuel cells to convert fuel into electricity through an electrochemical process. In most near term applications, that fuel is natural gas. The system is modular, located at or near the customer and designed to deliver continuous power without the conventional combustion process used by engines and turbines.

The technical details matter, and we will examine them. The macro significance is simpler. Bloom can place generation closer to the compute and deploy it on a calendar that can be radically shorter than a major grid expansion.

In 2025, the company delivered a fully operational system for Oracle in 55 days. In April 2026, Oracle expanded the relationship through a master agreement covering as much as 2.8GW. Importantly, 1.2GW are contracted. The remaining capacity is permitted by the agreement but should not be treated as a firm order.

That distinction is not a technical footnote. It separates the physical system being built from the narrative the market wants to believe.

The Private Power System

For most of the industrial age, large consumers purchased electricity from a public or regulated network. They might own backup generators, but the grid remained the primary system. Artificial intelligence is beginning to alter that hierarchy.

Data centers are becoming power projects with computers attached.

The land must be selected according to electrical capacity, gas access, water, fiber, permits and political tolerance. Generation must be designed alongside cooling and computing. Financing must account for the possibility that one part of the system arrives before another. The customer can no longer treat energy as an operating expense added after the building is designed.

Power moves upstream into the investment decision.

This does not mean data centers will abandon the grid. The more plausible outcome is a parallel architecture. A campus may begin with onsite generation, connect to the grid later, retain local power for resilience and add other sources as they become available. Solar, wind and batteries can supply part of the system. Existing nuclear and gas plants can support sites with favorable geography. Advanced nuclear may enter the mix after 2030. Demand response and flexible computing can reduce stress during certain hours.

The future is hybrid because the problem is multidimensional.

The IEA expects renewables to meet roughly half of the growth in global data center electricity demand through 2035. Natural gas remains central to the near term response, particularly in the United States. Nuclear becomes more important toward the end of the decade and beyond. No single source owns the future.

Bloom does not need to own the entire future. It needs to own a valuable part of the transition.

Its strongest position lies in the interval between a customer’s decision to build and the grid’s ability to serve that decision. If the local system remains useful after the grid arrives, the opportunity becomes larger. The Energy Server could move from bridge infrastructure to a permanent layer of power quality, resilience and operational control.

That is the first major question for the investment case.

Is Bloom selling temporary relief from a slow grid, or a durable redesign of how data centers power themselves?

The answer determines whether the company is monetizing a temporary bottleneck or participating in the creation of a second energy system.

The Finance of Speed

Technology alone does not create infrastructure. It must become financeable (which is the topic of one of my recent deep dives).

This is the deeper significance of Bloom’s relationship with Brookfield.

In October 2025, the two companies announced a framework of as much as $5bn for projects using Bloom’s fuel cells. On the 30th of June 2026, they expanded that framework to as much as $25bn. The structure is intended to finance power projects for artificial intelligence infrastructure.

The number requires discipline. It is not $25bn of Bloom revenue. It is not a guaranteed equipment order. It is not proof that every contemplated project will be built.

Its importance lies elsewhere.

Brookfield is attempting to combine compute, real estate, power equipment and capital from the beginning of the project. Bloom supplies a deployable generation platform. Brookfield supplies financing capacity, infrastructure expertise and a vehicle capable of owning projects over long periods. The customer can acquire power without carrying the entire initial cost in the same way it would purchase equipment directly.

This converts speed into something institutional capital can own.

The product is no longer a fuel cell stack. It is a financed path to an operating data center.

That path has value because capital is sensitive to time. Every delayed month changes interest expense, depreciation, customer delivery and the useful economic life of the compute. A power solution that advances operation can justify a premium even if its electricity is not the cheapest available over several decades.

The mechanism resembles the purchase of an option. The customer pays to reduce the probability that grid delay will strand the rest of the investment. Brookfield finances the infrastructure behind that option. Bloom manufactures the physical system that exercises it.

If this model scales, the competitive moat will not be defined by electrochemistry alone. It will be defined by the ability to offer technology, project development, installation, fuel access, maintenance and capital as one credible schedule.

That is harder to copy than a piece of equipment. It is also harder to execute.

The Geography of the Megawatt

The digital economy is global. Power remains local.

This is the geographic error hidden inside many forecasts. Analysts compare national generation with national demand, then conclude that enough capacity can be added. But the data center does not consume the national average. It consumes electricity at one substation, on one parcel of land, inside one regulatory jurisdiction.

A surplus in one region does not solve a shortage in another if transmission is unavailable. A gigawatt scheduled for 2031 does not energize a campus in 2027. A renewable project in the queue does not become firm power until generation, storage, transmission and market rules can deliver the complete service.

The Federal Energy Regulatory Commission has also warned that the demand pipeline itself contains noise. Developers can submit similar connection requests to several utilities while searching for the fastest site. This creates duplicate studies, distorts forecasts and can make speculative projects look like committed demand. In June 2026, FERC moved to impose stronger readiness requirements and reduce double counting.

This weakens the most aggressive demand headlines. It does not eliminate the local bottleneck.

Even after duplicate projects are removed, the surviving campuses require large blocks of firm power in fewer locations and on shorter schedules than the system was designed to provide. The scarce asset is not every announced megawatt. It is the deliverable megawatt attached to a real site, real financing and a real customer.

Bloom’s modularity matters in this geography. It allows power to be added in increments rather than through one enormous plant. It can reduce dependence on a distant transmission project. It can also move the constraint to another network.

Fuel must reach the site. Gas pipelines require capacity. Equipment needs permits. Communities may resist new industrial loads or local emissions. Maintenance must be available. A data center that escapes an electrical queue can still encounter a gas, land or regulatory queue.

Onsite generation does not abolish infrastructure dependence. It changes which infrastructure controls the calendar.

The Candidate Test

If time to power is the mechanism, Bloom is not the only candidate (as I’ve shown in my recent Who Powers AI series Part 2: Speed to Power). Behind the meter power is now the fastest growing power market in the world.

Caterpillar has announced an order supporting 2GW of fast response natural gas generation for a hyperscale campus, with delivery beginning in 2026. Cummins is supplying gas generators for a large data center project in West Texas. GE Vernova (which I covered in depth) reported more than $5bn of data center orders in the first half of 2026 across its electrification business, while its gas equipment backlog and slot reservations reached 116GW. Existing generators such as Constellation (which we covered at the beginning of the year) can place data centers near operating plants. FuelCell Energy is pursuing the same onsite fuel cell opportunity at a smaller scale. Brookfield can capture value through financing even if the winning equipment changes.

Each candidate owns part of the system.

GE Vernova is the broadest beneficiary because it can profit from both expansion of the grid and generation outside it. Caterpillar brings mature equipment and a global service network. Brookfield owns the financial architecture. Existing generators own immediate physical capacity. Advanced nuclear companies offer a cleaner long term destination, but not the fastest answer to a 2027 power requirement.

Bloom is the purest public expression of the immediate mechanism.

Its systems are modular. Its deployment speed has been demonstrated with a major customer. Its Oracle relationship provides hyperscaler validation. Brookfield provides a path to finance projects at a scale larger than Bloom’s own balance sheet could support. Its reported results show that demand has moved beyond presentation slides.

In the second quarter of 2026, Bloom recorded $1.065bn in revenue, an increase of 165.5% from the prior year. Product revenue grew 215.4%. Reported gross margin reached 34.3%, and management raised full year revenue guidance to between $3.9bn and $4.2bn. Those figures confirm acceleration. They do not establish durability.

Purity creates upside when the mechanism strengthens. It creates fragility when the mechanism weakens.

That is why the macro case cannot choose the stock by itself.

What Can Break the Premium

The first threat is efficiency.

Electricity use per unit of computation can fall even as total computation rises. Chips improve. Models become more efficient. Cooling systems waste less energy. Workloads can move across time and geography. The rebound effect may absorb these gains, but that outcome should not be assumed.

The IEA’s scenarios for 2035 range from roughly 700 to 1,700 terawatt hours of data center electricity demand. In its high efficiency case, demand is about 20% below the central path. The range is not a minor forecasting error. It represents different industrial futures.

The second threat is capital discipline.

The buildout depends on expectations of future AI revenue. This is one of the central questions of this AI buildout that I address in my Token Economics piece. If monetization disappoints, investors may force hyperscalers and developers to slow construction. Projects that exist in a utility forecast can disappear before equipment is ordered. Power demand remains structurally higher than it was, but the most aggressive premium collapses.

The third threat is institutional adaptation.

FERC and regional grid operators are not passive. They are changing rules for large loads, tightening project readiness requirements and searching for faster connection models. Utilities can build generation. Transmission portfolios can expand. Data centers can become more flexible. Every successful alternative reduces the scarcity Bloom is attempting to monetize.

Bloom benefits from the grid’s delay. Its success also gives regulators more reason to remove that delay.

The fourth threat is substitution.

Fuel cells compete against reciprocating engines, aeroderivative turbines, existing nuclear and gas plants, batteries, renewable portfolios and future advanced nuclear. Customers may accept higher local emissions for cheaper or more familiar equipment. Others may pay more for carbon free power. The winning architecture will vary by region, gas access, regulation and deadline.

The fifth threat is that Bloom solves the calendar but not the lifetime economics.

A customer may install fuel cells to begin operating, then reduce their role when the grid connection arrives. If the equipment becomes backup rather than primary generation, the addressable market and service economics change. The investment thesis requires evidence about what customers do after the bridge has been crossed.

The final threat is internal to the company.

Bloom’s 2026 growth has been highly concentrated. According to its second quarter filing, one customer accounted for approximately 73% of second quarter revenue. Two customers represented roughly 44% and 21% of first half revenue. The warrant issued in connection with Oracle also created substantial share based consideration that will be recognized partly as a reduction of revenue as systems are delivered. These facts do not invalidate demand. They change the quality of the evidence.

The same concentration that validates hyperscaler adoption also makes the growth story unusually dependent on a small number of deployment schedules.

The Asset Behind the Electron

Energy for artificial intelligence is no longer a hidden theme. The market has already begun to capitalize utilities, turbines, transformers, fuel cells and powered land as parts of the same industrial system.

The next repricing concerns certainty.

Who can deliver firm power in the right place? Who can finance it? Who can install it before the compute begins losing economic time? Who can preserve the customer relationship after the grid arrives? And who can do all of this without surrendering the economics through incentives, maintenance costs or competition?

Bloom Energy is the clearest test because it sits at the intersection of those questions.

It converts fuel into onsite electricity. It can deploy in modules. It has demonstrated an unusually short installation timeline. It has a major hyperscaler relationship. It has an infrastructure partner capable of financing projects far beyond the scale of a conventional equipment sale. Its revenue is accelerating at the same moment that the power bottleneck is becoming a board level problem for the technology industry.

But a system can be correct while a security is mispriced. A market can be enormous while its most visible supplier fails to capture attractive returns. A financing framework can remain potential. A large customer can become a concentration risk. A temporary bottleneck can disappear before an expanded factory earns back its cost.

The macro system establishes why Bloom’s opportunity exists. It cannot tell us how much of that opportunity belongs to the company, how durable its advantage will be or what price investors should pay for it.

That is where the investment case begins.

Now that you understand why the most valuable megawatt is the one that arrives on time, it is time to determine whether Bloom Energy can convert that advantage into durable shareholder value.

Welcome to the 83rd investment case, 44th AI & Technology, 29th Electrification & Energy and 51st Reshoring & Sovereignty idea on Crack The Market (and the most comprehensive Bloom Energy investment case you will find online)! Join us as we dissect the company selling time in the Age of AI, the solid oxide fuel cell pioneer that went from left for dead in early 2025 to the standard for onsite AI power, delivering its first billion dollar quarter and forcing the entire power complex to take behind the meter generation seriously.

Bloom Energy Investment Case

Table of content

• Business Description

• Investment Case

• Value Chain

• Growth Drivers

• Risks & Threats

• Financial Robustness & Capital Allocation

• Outlook/Guidance

• Governance & Management Quality

• Valuation

Business Description

Bloom Energy ($BE) is the global leader in onsite power generation, built on a proprietary high temperature solid oxide fuel cell (SOFC) platform that converts natural gas, biogas or hydrogen into electricity through an electrochemical process, without combustion and without moving parts.

Founded in 2001 by K.R. Sridhar out of technology originally developed for NASA’s Mars program, the company designs, manufactures and services the Bloom Energy Server, a modular power appliance deployable from hundreds of kilowatts to hundreds of megawatts, and the Bloom Electrolyzer, which uses the same solid oxide platform to produce hydrogen.

The company is headquartered in Silicon Valley, employs more than 2,000 people, manufactures entirely in the United States (cell printing in Fremont, California and assembly in Delaware) and has systems deployed across approximately 1,100 sites in 9 countries.

The revenue model has two layers:

• Product revenue (88% of Q2 26 sales) comes from selling Energy Server systems, either directly to end customers as a capex sale or, far more commonly, to a financier such as Brookfield or Industrial Development Funding who takes title and delivers power to the end customer under a PPA, capacity agreement or lease that Bloom originates.

• Service revenue is the long tail (6-7% of Q2 26 sales): multi year operations and maintenance agreements covering availability guarantees, remote monitoring and periodic stack replacement, an annuity that now runs above 20% gross margin after a decade in the red.

• This financier structure matters for how you read the filings: the “customer” in Bloom’s concentration disclosures can be the project finance affiliate buying the boxes, not the hyperscaler whose demand created the deal.

Bloom organizes its power generation markets into three categories:

• Data Centers, now the dominant growth engine. Bloom participates across hyperscalers, colocation operators, neoclouds, developers and infrastructure investors. All the major US hyperscalers and over a dozen US neoclouds, AI labs and colocation operators have validated and approved Bloom’s power solutions for their AI factories, a qualification process that took a decade in C&I and less than a year in AI.

• Commercial & Industrial, the historical core: advanced manufacturing, semiconductor fabs, healthcare, retail, telecom, education and other critical infrastructure. Around 80% of 2025 bookings were repeat orders, the clearest evidence of customer satisfaction in a business where the product sits on your site for 15+ years.

• Utilities, an emerging channel where capacity constrained utilities deploy Bloom systems at the point of distribution or directly at customer sites. The landmark November 2024 framework with American Electric Power covers up to 1GW, and management expects more utilities to follow.

Geographically the US dominates, with South Korea the second largest market at around 682MW deployed through the SK ecoplant and SK eternix partnerships (SK ecoplant also owns around 2.5% of the shares). Selective expansion is underway in Europe and Asia where gas infrastructure and market design favor distributed generation.

The numbers by segment show exactly where the AI inflection is landing:

• In 2025 Bloom generated $2,024m of revenue: product $1,531m (75% of total, up 41% yoy, at a 35% GAAP gross margin), installation $204m (up 67%, roughly breakeven by design as a pass-through), service $228m (at a 10% GAAP margin, still early in its journey) and electricity $60m (the small legacy PPA book, at a 46% margin).

• By Q2 26 the mix had shifted violently toward product: $935m of product revenue (88% of the total, at a 37.2% margin, up 291bps yoy), $51m of installation, $69m of service (at 22%, the first quarter above the long promised 20% threshold) and $10m of electricity, for $1,065m in total at a 34.3% non-GAAP gross margin.

The service line deserves a highlight: it is a $9bn+ backlog annuity attached to a growing fleet, and it just crossed into sustainable 20%+ territory after being a 21% money loser at IPO.

So who is the customer behind the 73% concentration?

• The Q2 filing discloses that a single customer, not a related party, was approximately 73% of quarterly revenue, while over the first half two customers were 44% and 21%, the second being a related party. Bloom does not name names but we can investigate ourselves.

• The related party at 21% is the Brookfield AI Fund joint ventures, the unconsolidated vehicles that buy Energy Servers under the $25bn framework, meaning roughly a fifth of first half revenue already flows through the Brookfield machine.

• The 73% counterparty is a financier, and investors think it is the Industrial Development Funding vehicle purchasing systems for the Nebius buildout, whose 328MW first project is being delivered this year (the CFO called the IDF structure a meaningful contributor to the quarter).

• It is very important to remember the definition of a customer for Bloom: in Bloom’s disclosures the customer is whoever takes title, so one financier can aggregate several end projects, and the end demand behind these line items spans Oracle, Nebius and others.

• For reference, 2025’s top three customers were 43% (SK ecoplant, the Korean distribution partner and a related party until July 2025), 13% and 12%.

• The concentration is real, but its composition rotates with delivery windows, it is not as problematic as some make it out to be.

Bloom Energy’s history - From Mars to the AI factory, a 25 year overnight success:

You cannot understand why Bloom leads this market without understanding where it came from, because the fuel cell sector was for two decades, a sector that investors tended to stay away from.

• K.R. Sridhar founded the company in 2001 as Ion America, reversing the oxygen generation technology his team had built for NASA’s Mars program (feed it electricity and it makes oxygen, feed it fuel and air and it makes electricity).

• It became Kleiner Perkins’ first cleantech investment (a legendary VC firm with John Doerr leading the investment, one of the most prolific venture capitalists in history who also backed Google, Intuit, Amazon, Sun Microsystems), and in 2010 the Bloom Box was unveiled on 60 Minutes at eBay’s campus with a wall of Silicon Valley celebrity behind it (Arnold Schwarzenegger, then-Governor of California, General Colin Powell, former Secretary of State and Chairman of the Joint Chiefs of Staff, who served on Bloom’s board, John Doerr from Kleiner Perkins and John Donahoe, the eBay CEO and leadership from other early anchor clients, including Google, Walmart, FedEx, Staples, and Coca-Cola), one of the most hyped product launches in energy history.

  

• Then came fifteen years of grind that killed nearly every peer. The fuel cell cohort became a graveyard of business models: PEM players chased vehicles and hydrogen hype, molten carbonate struggled with durability, and the 2020-21 hydrogen bubble inflated the whole complex (Bloom touched $44) before rates and reality collapsed it.

• Bloom itself flirted with irrelevance more than once: At the 2018 IPO, the street initiated coverage calling it an electricity grid in a box and projecting $1.5bn of revenue by 2021 at a 25% growth rate for a decade. The revenue arrived four years late, and then doubled in one.

• In between, the stock listed at $15, sank toward $3 in 2019, bled money on service contracts (minus 21% service margins at IPO) and carrying a hype cycle reputation (hydrogen, marine, international) that the market never let it forget. As late as February 2025 the stock traded at $23 with a $5.6bn market cap, analysts complaining about disappearing disclosures and a book and turn business with no visibility.

What separated Bloom from the fuel cell graveyard was three unfashionable choices:

• It picked solid oxide, the high temperature chemistry that runs on cheap natural gas at 50%+ electrical efficiency without platinum catalysts, rather than chasing the hydrogen dream directly.

• It stayed stationary and commercial, selling 24x7 power to companies that valued reliability, rather than chasing vehicles.

• And it vertically integrated US manufacturing and kept grinding costs down double digits a year for a decade, building the copy exact factory model and a fleet whose stack life improved more than 2.5x across generations.

• When the AI power crunch arrived in 2024-25, Bloom was the only fuel cell company on earth with proven multi hundred megawatt manufacturing, a 1,100 site operating fleet, a service organization, and a product that could ship in months.

• The AEP framework in November 2024 opened the utility door, Oracle in October 2025 opened the hyperscaler market, and a 25 year overnight success re-rated nearly 10x from the 2025 lows.

• The history matters for one more reason: this management team has survived hype cycles before, and so have the skeptics.

What makes the machine different is the absence of combustion:

• The Energy Server produces near zero NOx, SOx and particulates, consumes no water in steady state operation, runs quieter than air conditioning equipment, needs roughly 125 times less space than an equivalent solar installation, and delivers 99.9% fleet availability on non redundant configurations (99.999% with redundancy).

• It generates native DC power compatible with the emerging 800 VDC data center standard, and paired with supercapacitors it follows the violently oscillating load profile of AI training clusters with instantaneous step response, at least twice as fast as turbines and engines.

• It is, as management likes to say, not a faster horse. It is a car.

Investment Case

Read more

Five days ago, I published Is the AI Trade Financeable?, my deep dive into whether this monumental buildout can keep funding itself. Last night, in his prepared remarks, Jensen Huang described Nvidia as “the productive and financeable compute infrastructure”. The CEO of the largest company in the AI value chain reached for the exact word the credit market has been debating for weeks (clearly not by accident). Nvidia spent much of this print answering the funding question with a purpose built architecture, and that answer moves every layer of the value chain.

Two days ago, I published my definitive AI value chain map, How to Invest in the $6-7tn Race: eight layers from the power plant to the agent, priced per gigawatt, built on the through line that bottleneck rents migrate to the slowest expanding layer. These results are the first actual test of that map, delivered by the company that sits at the very center of this buildout.

As a reminder, here are all 8 layers of the AI value chain (from my deep dive two days ago), what they do, whether they are bottlenecks, where the durable economics sit and all the names (in bold the ones I have already covered).

So this piece is deliberately not another recap of the beat. Everyone will cover the beat.

Instead:

• I grade my own framework against the evidence and work through the read across for memory, power, optics, the buyer base, custom silicon and the financing layer, with names in each bucket. I find that this print strengthens the outlook for every single layer.

• I then update my Nvidia investment case, which I laid out in June when I called the stock on sale while the crowd chased the periphery. It remains, in my view, one of the most compelling ways to play this AI supercycle, with a valuation that is now completely decorrelated from its rapid and accelerating growth (the fastest amongst large cap tech names: 87% revenue growth in 2027, 58% in 2028).

Table of content

• The two structural signals: the first forward year guide in Nvidia’s history, and the $279bn supply commitment line that makes the balance sheet a second moat

• Memory: the margin bridge is a visible transfer to the memory makers, and why prices are heading higher again next year

• Value per gigawatt: Nvidia just marked my $47bn/GW framework up to $60bn, and what that pulls through in power, grid and cooling

• Networking and optics: another record quarter, plus the new third compute socket almost nobody noticed

• The buyer base: the concentration bear case shrank this quarter, neoclouds triple capacity, and physical AI arrives at Amazon

• The financing layer: Nvidia becomes the credit intermediary of the buildout, with a $500bn platform attached

• Custom silicon and the toll booths: why a supply constrained guide is bullish for Broadcom, TSMC and the equipment names

• My updated Nvidia investment case: near the lowest forward multiple since 2018, on estimates about to rise 20%

The July quarter

• Revenue of $96.2bn, up 106% year on year and 18% sequentially, the fourth consecutive quarter of accelerating growth, against a $92.3bn street. Data center $89.0bn, up 117%.

  

• On profitability: Gross margin 75.0%, flat sequentially. EPS of $2.22 versus $2.09 expected.

  

  

• Q3 guidance (October quarter) of $108bn against $104.6bn consensus, which makes the October quarter the first $100bn revenue quarter in corporate history, with data center alone around $100bn. China data center compute was less than 1% of the segment and the outlook assumes zero, so none of this leans on a China recovery.

  

• The beat and raise machine keeps working at a scale nobody has ever operated at. Despite this, the headline numbers are still not the most interesting part of these earnings.

The two structural signals

Signal one: the first forward year guide in Nvidia’s history.

Management expects FY28 revenue to grow approximately 70% year on year, near $680-700bn, adding close to $200bn to next year’s data center outlook against the roughly $1tn CY25-27 framework they had given before. This guidance is well ahead of anything investors were hoping for, consensus sat in the mid 40s and the buyside was likely positioned for 50-60%.

More importantly in my view, this is the first time ever that Nvidia has issued a formal full year forward framework. A company that has guided one quarter at a time through its entire history just gave us visibility into the whole of next year, underwritten by $279bn of secured supply and three year component visibility. That is a statement about the durability of this supercycle as much as about its size, and it changes how much confidence investors are entitled to have in the out years.

I want to focus for a moment on the fact that Jensen Huang talked about constrained supply. Customer forecasts point to demand that would support 100% growth if supply allowed, management expects supply to remain the bottleneck at least through the end of FY28, and the guide assumes zero China data center compute. So 70% is a floor built from secured components, with two free options on top: constraint relief and China (which I would not count on).

This is what we know about the scale of the AI demand:

• A cloud industry backlog now above $2tn, with top five hyperscaler capex heading from roughly $800bn this year to $1.3tn in 2027.

• Roughly 12GW of cumulative OpenAI commitments through 2030, plus an expanded AWS partnership adding 2 million GPUs from this quarter through early FY29 (which at $55k/GPU would be $110bn at today’s prices).

• Industry estimates put shipments around 20GW next year against roughly 13GW this year, before any constraint relief.

Off that base, a path to $1tn of revenue in FY29 is now credible, something the stock is clearly not pricing today.

The second order effect from this constrained supply is the growing opportunity for custom silicon (which I have been writing about since my Age of Inference piece last year). The bigger the constrained gap, the stronger the case for owning both sides of the compute layer, which I return to below.

Signal two: Nvidia’s balance sheet is now a further moat on top of its technology.

Supply and capacity commitments went from $119bn to $279bn in one quarter, and $366bn including cloud service agreements, data center leases and equity investments.

• The majority is memory, mostly HBM, with three year forward coverage and pricing already locked for the majority of FY28 needs (which is critical given we are in a Memory Singularity, this is very bullish for memory as I will address below).

• Add the guarantees and ecosystem investments disclosed alongside and the cumulative commitments approach half a trillion dollars.

• I think this is the single most important line in the release for the rest of the AI value chain: nobody else can pre buy the bottleneck at that scale, and rent flows to whoever controls the scarce input.

The prepayment behavior I flagged as the signature of a true bottleneck has also now appeared on Nvidia’s own balance sheet:

• Deferred revenue jumped from $1.4bn to $4.6bn in six months, including $2.8bn of customer advances versus $160m in January.

• Customers are pre paying the GPU maker while the GPU maker pre pays the memory makers.

• Nvidia’s balance sheet is now as much of a moat as its technology.

Below: the full read across for memory, power, optics, the buyer base, custom silicon and the financing layer, with names in each bucket, my updated Nvidia investment case at a decade low multiple, and what I am watching next

The read across layer by layer

Read more

We are living through one of the most profound technological revolutions in history, fueled by the explosive rise of artificial intelligence. What started as isolated breakthroughs has erupted into a full blown arms race, with the largest, most powerful and best capitalized companies the world has ever seen pouring over a trillion dollars a year into AI infrastructure. This battle is not just about who has the best models, it is about raw computing power, energy, data and the physical infrastructure behind it all.

Eleven months ago I published AI Investment Supercycle: How to Invest in the $3-4tn Race. It was the most read piece I had written to that point and it aged badly in the best possible way. The number was way too low and investors globally have kept being surprised by the speed and scale of this AI deployment, which is unlike anything we have ever witnessed. My map then was semis heavy, and the biggest bottleneck of the past year, memory, was only one list among many.

I have since covered more than 40 companies riding this AI megatrend, with dozens of thematic pieces on the biggest debates, selloffs, fast growing verticals and bottlenecks (memory, photonics, CPUs, compute, electrification), all accessible in my AI & Technology tab, each with its own podcast episode.

The AI Investment Supercycle is the most consequential megatrend of our lives, trillions are being poured into building the AI revolution, and this is the definitive map of how to invest in a once in a lifetime opportunity.

This piece has three jobs:

• It is the definitive structural map of the AI value chain as I see it today, one chain from power generation to the application layer rather than a compute theme and a power theme bolted together.

• It is the piece I want cited when someone asks what the buildout actually consists of.

• And it is the roadmap for my coverage over the next year, with every deep dive I have published linked in its place on the chain and every gap flagged as coming.

It deliberately comes out the day before Nvidia reports. Whatever the results say, the structure below is what the numbers will flow through.

Just some numbers to give you a sense of the scale:

• Hyperscalers spent roughly $443bn of AI capex in 2025. Guidance for 2026 aggregates to roughly $700-750bn, and 2027 consensus now sits around $1.1tn, a figure that was $480bn only a year ago.

• Total AI spend through 2030 is now expected at $6-7tn, double the number I published eleven months ago, with $2.6tn of it already contractually committed.

• A gigawatt of AI data center now costs roughly $47bn in the Vera Rubin era, up from $35bn, with Nvidia capturing roughly $32bn of it in rack content.

• Reasoning, agentic AI and the march toward superintelligence keep multiplying compute demand, up 100x to 1,000x per Nvidia.

• Meta alone has committed at least $600bn to US data centers and infrastructure by 2028.

Twice a week, I will release deep dives into stocks and sectors that fit into the 6 themes that I see winning in the coming years and decades: AI & Technology, Electrification & Energy, Reshoring & Sovereignty, Healthcare & Longevity, China & Asia, Resilience & Quality.

Join Crack The Market

Take advantage of this once in a generation opportunity to build long term wealth by investing in great stocks that will deliver returns for your portfolio for years to come.

What you will learn in this piece

• Why the cumulative buildout estimate has doubled to $6-7tn in eleven months and why cost per gigawatt keeps inflating even if the gigawatts eventually do not.

• The full chain in eight layers plus the financing overlay, what each layer does, whether it is a bottleneck, and every named public company in it.

• The through-line: bottleneck rents migrate to the slowest expanding layer, and the constraint is moving down the stack, from silicon toward physics.

• What I got wrong in last year’s version, starting with memory.

• Which layers are crowded, contested and overlooked on current positioning.

The Supercycle doubled in a year

The monumental speed and scale of the revision of this AI buildout needs to be acknowledged:

• In September 2025 I wrote that total AI spend through 2030 was expected at $3-4tn.

• The same value chain, measured the same way, now points to a cumulative $6-7tn by 2030.

• Roughly a fifth of that goes into power, energy and electrical equipment, which is why this version of the map starts at the power plant and not at the fab.

The trajectory underneath is steeper than the headline:

• Hyperscaler AI capex ran around $269bn in 2024 and roughly $443bn in 2025. Guidance for 2026 aggregates to roughly $700-750bn, and consensus for 2027 now sits around $1.1tn, a figure that was $480bn a year ago, a 127% upward revision in twelve months.

• On the broadest definitions, which now include SpaceX as a hyperscaler given the scale of its compute ambitions, 2026 growth approaches 89% and 2027 reaches $1.4tn.

• I use the narrower perimeter throughout this piece, but what matters is that every definition points the same direction.

Three more numbers frame the scale (that come from my Is the AI Trade Financeable piece):

• $2.6tn of future spend is already contractually committed across purchase obligations, construction commitments and leases not yet commenced. This is not sentiment, it is signed paper.

• Total investment in planned and in-construction data centers is around $1.2tn right now, with new announcements landing non stop: a $50bn flagship campus that was $27bn nine months ago, a 10GW neocloud target for 2030, an $85bn commitment to build 5GW in France.

• The group FCF swing is violent (FCF is now the new hot number, which is what has started to shake markets): from roughly +$208bn of free cash flow in 2024 to around -$440bn in 2027 on current consensus, a cumulative burn approaching $850bn across 2026-28. The most recent data point is Alphabet, which printed its first negative FCF quarter since becoming public in 2004. The trade has moved from self funding to funded, which is why credit is now a layer of this chain and gets its own section below.

Translate the dollars into physical units and the most striking fact of all appears: measured in gigawatts, this buildout is just getting started.

• AI data center capacity stood at only 22GW in 2024.

• It reaches 76GW by 2029 and 166GW by 2035 on current forecasts, a 7-8x expansion.

• Annual additions are accelerating from roughly 20GW in 2026 to 35GW by 2030, the trajectory I detailed in Who Powers AI - Part 1: The Anatomy of a Gigawatt.

• If you listen to Vertiv, who actually builds those facilities, those numbers are much much higher (which I detailed in Vertiv: The Tissue Between the Silicon and the Facility). The announced project pipeline sits at roughly 430GW, up from 140GW only eighteen months ago. The company sees 140GW of cumulative data center power capacity added over 2026-2030.

• Whatever the debate about any single quarter of capex, the physical trajectory accelerates from here.

And the subtlest driver of all: the cost of a gigawatt is inflating around 8-9% per hardware generation, driven by memory and power content. Capex grows even where the gigawatts do not. I will show you exactly where those dollars go.

The through-line: rents migrate to the slowest layer

The AI value chain that I map here converts electrons into tokens into revenue. Every layer adds a conversion step, and at any moment one layer is expanding more slowly than the demand flowing through it. That layer collects the rents: pricing power, prepayments, long term agreements, capacity booked years in advance.

Bottleneck rents migrate to the slowest expanding layer, and rent duration equals lead time. A GPU allocation can be given in quarters, a memory fab is a $20bn commitment with a two year build and qualification, a gas turbine slot is four to five years out, a grid connection takes a median of six years. The longer it takes to add capacity in your layer, the longer you keep the rent once demand finds you (and it is important to remember that demand is extremely difficult to model and has been consistently under appreciated throughout this cycle as we saw above).

Let’s recap how the constraint has moved throughout this AI supercycle and where it is likely to go next:

• In 2023 and 2024 the bottleneck was the GPU itself, and Nvidia collected the rent.

• Through 2025 and 2026 the constraint migrated to memory, where undersupply is now expected through late 2027, and to electrons, where firm capacity in the most contested grid region has repriced more than tenfold in two auctions.

• The next candidates are already visible: connectivity between chips and racks, advanced packaging and IC substrates, where customers are extending prepayments and capacity sharing agreements to secure supply, and the CPU, where agentic workloads are pulling attach ratios from one CPU per four GPUs toward one to one.

The direction of travel is consistent: from silicon toward physics.

• Design bottlenecks clear fast because they are solved with talent and money.

• Physical bottlenecks clear slowly because they are solved with permits, concrete, copper and fab tools.

That is why the economics of this chain keep migrating downward, and why the second half of this piece spends as much time on switchgear and substrates as on GPUs.

One fundamental point on which all of this value chain relies is the monetisation of AI:

• Lead times do determine who captures an outsized portion of the value in this cycle. But whether there are rents at all is decided at the top of the stack, by whether the tokens earn a return.

• The June correction proved it: a rack delay and a single Chinese model release were enough to knock the whole complex, including the layers with the longest lead times.

• I decomposed that episode in The Reckoning Arrives, published the actionable half of the pairing in My AI Selloff Shopping List, and stress tested the funding side in Is the AI Trade Financeable?. The framework there still stands: the positioning leg of the selloff has largely cleared, while the two fundamental questions, credit and commoditization, remain live.

• This map does not resolve those questions. But it does tell you where you are standing when they eventually get answered.

How to read the chain: the gigawatt is the unit of account

Last year I mapped a semiconductor stack and mentioned power. This year power is not a segment, it is the denominator of this entire supercycle. Every layer below the hyperscalers can be priced per gigawatt, and the gigawatt is what makes compute and power one single chain instead of two separate but tangential themes.

The cost per gigawatt of AI data center is rising, and fast:

• A gigawatt scale AI campus at current rack generations costs roughly $35bn all in, as I detailed in Who Powers AI - Part 1: The Anatomy of a Gigawatt: around $4bn of land and shell, a physical infrastructure layer dominated by power and cooling, and the racks, which are the overwhelming majority of the bill.

• In the Vera Rubin era that same gigawatt rises toward $47bn, roughly $32bn of racks at $9.1m each, 3,557 of them, plus $12-15bn of physical infrastructure.

• Cost per gigawatt is rising because the racks are getting denser and the memory inside them more expensive.

There is a crucial second curve underneath the cost curve, which keeps the technology adoption on track: compute per dollar is rising faster than cost per gigawatt.

• Delivered performance per billion dollars of capex roughly doubles from the GB200 generation to Vera Rubin.

• The machine is getting more expensive and much more productive at the same time, which is exactly what you would expect from a technology in its steep phase, and it is the reason the buildout keeps being rational at the unit level even as the aggregate numbers become uncomfortable.

Above the hyperscalers the metric changes, we talk in tokens:

• Models and applications do not price per gigawatt, they price per token.

• You can think of hyperscalers and neoclouds as the exchange rate between the two: they buy gigawatts and sell tokens.

• This is where the entire AI trade is playing out, and what determines if the trade will be successful or not is the Token Economics.

The AI value chain map consists of eight layers running from the power plant to the agent, plus a financing overlay.

The same AI value chain with every name in its place is below. Bold names have a published deep dive. This map represents both my AI coverage and the entire AI value chain trade in miniature.

The rest of this piece goes through the chain layer by layer, starting at the power plant and going all the way to the application and agentic layer.

For each layer I cover what it does, whether it is a bottleneck and why, where the durable economics sit, and the named companies, with my deep dives linked where they exist. Where I have not yet covered the sector or names, I flag it and tell you if it is on my roadmap for the year ahead.

Layer 1: Generation and fuel

What it does: Makes the electrons. Gas turbines, nuclear, solar plus storage, diesel and gas gensets, fuel cells. Nothing above this layer runs without it. These are the companies who power AI.

Is it a bottleneck? Yes, the binding one.

• A gigawatt campus is a $35bn asset that earns nothing until the power flows, and capital sitting idle waiting for electrons costs on the order of $300-400m per month.

• That impatience has created what I called the speed to power market in Who Powers AI - Part 2: Speed to Power: buyers rationally paying double the grid tariff for power that arrives two years earlier.

• The market signal is unambiguous: The capacity auction in PJM, the grid at the center of data center alley, cleared at $28.92/MW-day two auctions ago, $269.92 in the next one and roughly $329 in the latest, with the next auction capped at $696 and in my view likely to test it.

Where the economics sit: In firm, contracted power and in the machines that make it, on lead times that are literally impossible to shortcut:

• Gas turbines are sold out into the next decade at the big three. I mapped the oligopoly in The Turbine Trilogy: The pure players of the Golden Age of Power: GE Vernova (deep dive: GE Vernova: Let There Be Power), Siemens Energy (deep dive: Siemens Energy: All gas (& electricity), no brakes) and Mitsubishi Heavy Industries, with Doosan Enerbility riding the same backlog from Korea.

• Nuclear is the premium power product, firm and clean, and the executed hyperscaler deals, twenty year agreements clustering around $80-110/MWh, roughly double around the clock wholesale, define what that is worth. Constellation Energy (deep dive: Constellation Energy: The Federal Reserve of the Electron) is the reference asset, one hyperscaler bought a 960MW campus next to the Susquehanna plant outright and expanded to up to 1,920MW of supply through 2042. Upstream sits the fuel cycle I covered in Uranium: Entering a Nuclear Powered Supercycle with Cameco as the most actionable idea ahead of the coming IPO of Westinghouse (deep dive: Cameco: The Western Nuclear Champion), alongside Kazatomprom on the fuel side, with the strategic case in The Nuclear Renaissance and Why Nuclear Power Is No Longer Optional. SMRs are the option on the option and are becoming a reality much quicker than expected, coverage coming.

• Solar plus storage is the incremental electron on the impatience curve: LCOE down from around $150/MWh in 2011 to $30-60 today, and solar with four hours of storage landing around $81-135/MWh, competitive with gas engines on speed. First Solar (deep dive: First Solar: The rising Sun in a new electricity supercycle) and the complex I revisited in US Solar Is Investable Again. Nextpower, coverage coming. The storage half of the pair runs through CATL (deep dive: CATL: The battery titan with electrifying growth), the battery leader (in EVs and energy storage) behind the grid batteries now standard on every AI campus.

• Gensets and bridge power are the purest speed trade: data center genset revenue at the top three US suppliers has grown from around $2bn in 2023 toward $10bn expected this year, heading toward $20bn by 2030. Caterpillar, Cummins, Generac (first hyperscaler contract signed this June, data center backlog above $700m and more than doubling in a year), Rolls-Royce. Coverage coming on the group.

• Fuel cells are winning the behind the meter niche: Bloom Energy, coverage coming, Ceres Power in Europe.

• Utilities and IPPs own the existing fleet the queue jumpers are bidding for: NextEra (deep dive: Nextera Energy: How America’s Largest Utility Became An Electrification & AI Powerhouse), Iberdrola (deep dive: Iberdrola: Best integrated utility in the world), Vistra and Talen as the PJM torque, coverage coming.

Layer 2: Grid and electrical infrastructure

What it does: Moves and conditions the electrons.

• Transmission and distribution, high voltage cables, transformers, switchgear, busway, UPS, the entire gray space electrical spine.

• I made the structural case in Electricity is the new oil: Load growth and grid investment supercycle: the first sustained load growth in a generation colliding with decades of underinvestment, roughly $800bn a year of grid investment needed globally.

Is it a bottleneck? Yes.

• The median grid connection queue is six years, the longest lead time anywhere in the AI value chain, and transformer and switchgear lead times stretch to multiple years.

• US data centers alone need on the order of 100GW of new capacity by 2030 with credible estimates of a 45GW+ shortfall.

• Monthly US data center construction starts have grown roughly sixfold in two years, with 2026 starts spending approaching $116bn against $41bn in 2025.

Where the economics sit: This is, in my view, the best risk adjusted layer in the whole chain, for one underappreciated reason: the demand floor. At the electrical majors, orders grew 88% last quarter into a $176bn backlog, and only around 20% of it is tied to hyperscalers. Grid replacement, reshoring and electrification carry this layer even if AI disappoints, which is precisely why I built My Electrification Selloff Shopping List around demand floor versus AI exposure.

The key names, by sub-segment:

• The electrical majors: Schneider Electric (deep dive: Schneider Electric: The Electrification champion that keeps compounding), Siemens (deep dive: Siemens: The best conglomerate on Earth is sitting on EUR70bn of optionality), Eaton, Hitachi, ABB, Emerson, Legrand, nVent, Littelfuse. Eaton, Hitachi, Emerson, nVent, Littelfuse coverage coming.

• US electrification pure plays: Hubbell (deep dive: Hubbell: Pure play on US electrification demand driven by AI and grid investment cycle), Vertiv for the critical power and cooling interface (deep dive: Vertiv: The Tissue Between the Silicon and the Facility).

• Cables, the true capacity constrained oligopoly: Prysmian (deep dive: Prysmian: Global cable leader entering a golden age), Nexans (deep dive: Nexans: Electrification pure play with visibility into 2030), NKT, with the sector map in Cable manufacturers: The pure players of the Golden Age of Electricity & AI.

• Grid services and enablement: Quanta Services (deep dive: Quanta Services: Countercyclical compounder riding a multi decade infrastructure buildout), two thirds of revenue from utilities under multi year master service agreements with a record backlog above $53bn, and Itron (deep dive: Itron: Riding the grid investment super cycle and enabling smart grids). National Grid and Elia, coverage coming.

• The metal itself: copper runs through every layer of this chain, the case is in The Age of Copper.

Layer 3: The facility

What it does: Houses the machines. Land and shell, engineering and construction, thermal management, in building power delivery, and the leased data center market.

Is it a bottleneck? Partially, and increasingly.

• US data center construction starts are on track to roughly triple this year, from $41bn in 2025 toward $116bn, a sixfold move in two years, and the system building them was never sized for it.

• A gigawatt campus needs up to 4,000 workers on site at peak construction, and skilled electrical labor is the quiet constraint nobody models.

• Cooling is in transition: every watt the compute consumes must be removed as heat, and rack densities are forcing liquid cooling from exotic to default.

• And the power architecture inside the building is being redesigned completely for the next generation of AI data centers: the industry has locked in 400V/800VDC as the primary upgrade path, with power electronics migrating into separate sidecars from 2027, roughly 10% penetration on the Rubin generation and 30-40% on the coming custom silicon systems, enabling multi hundred kilowatt and eventually megawatt class racks.

• That transition reshuffles who supplies the power train. This will be the topic of part 3 of my Who Powers AI series, coming very soon.

Where the economics sit. Unevenly, and in my view it needs to be split into three groups:

• Thermal and power delivery content per rack is rising: power content roughly tripling per rack across generations and cooling now around 2% of a $9.1m rack. Vertiv sits exactly on this seam, with Munters, Modine, Delta Electronics, Trane, Johnson Controls and Carrier around it. A liquid cooling primer is coming.

• E&C and mechanical contractors get the revenue but historically thin economics, materials pass through and competitive bidding. Comfort Systems has been the exception and re-rated for it. EMCOR, Jacobs, WSP and MasTec carry exposure. Coverage coming on the group, alongside insulation and envelope names including Kingspan and Saint Gobain, while Halma (deep dive: Halma: High quality compounder in safety, health and environmental technologies) sits quietly across the fire, gas and safety systems every one of these buildings must install (on top of its hidden photonics exposure).

• Data center real estate is quietly having its best market in a decade: the addressable data center capex pool goes from roughly $145bn in 2025 to $250bn by 2029 on current forecasts, renewal spreads on large leases have turned firmly positive, and the leading platform’s pipeline could reach 8GW. Equinix and Digital Realty, coverage coming.

Layer 4: The rack

What it does: Converts electrons into compute. This is the white space: accelerators, CPUs, memory and storage, networking and optics, power semiconductors, and the companies that assemble it all.

It is also where most of the money lands, roughly $32bn of a $47bn Vera Rubin gigawatt, and the territory My AI Selloff Shopping List ranked along a quality to beta gradient into the correction.

Open the rack and the bill of materials tells you where the rents have moved:

Compute: GPUs, ASICs and the CPU renaissance

• Nvidia remains the center of gravity: 72 GPUs at roughly $55k each is nearly $4m of a $9.1m rack (44% of the total rack cost!), and the moat is the ecosystem, not the chip, the case I made in Nvidia: The backbone of the AI world is unstoppable and pressed again into the selloff in Nvidia/Broadcom: On Sale While the Crowd Chases the Periphery.

• The ASIC share shift is real and hardening: Custom silicon strips out generality for efficiency, consuming 50-80% less power per unit of work, at the cost of an 18-24 month design cycle and zero flexibility when architectures move. Broadcom has now locked in multi-generational custom silicon agreements extending through the decade with its anchor hyperscalers (deep dive: Broadcom: The $1.7tn AI underdog), and Marvell is ramping custom programs across all four major CSPs, with a recent transformative and wide ranging custom silicon and interconnect deal with Alphabet solidifying its positioning, coverage coming. I mapped the whole war in Age of Inference: ASICs vs GPUs and the Trillion $ Compute War to Make AI Profitable. Meanwhile, Cerebras, the wafer scale challenger, became the largest listing of the year (before SpaceX) in May with a $10bn frontier lab contract behind it, the first pure play token economics bet available to public investors.

• The CPU is the sleeper: Agentic workloads orchestrate, branch and call tools, work that runs on CPUs, and the attach ratio is shifting from one CPU per four GPUs toward one to one, with the server CPU TAM expanding toward $210bn. I made the call in CPUs Are the Next GPUs, with ARM as the expression (deep dive: ARM: The Best Business in the World, at the Hardest Price to Pay), Intel as the incumbent losing share and AMD the major name I have not yet covered, coverage coming. Qualcomm is the wildcard entrant.

Memory and storage: the bottleneck of this cycle

Memory is now around a third of the rack bill, roughly $3.2m of $9.1m, and it is the biggest single change in the AI value chain since last year’s version.

The numbers are extraordinary: NAND prices up 11.3x from the April 2023 trough, HBM4 pricing headed toward $53/GB by 2027 with a pass through mechanism that protects the GPU maker’s margin, HBM consuming 3-5x the wafer area of standard DRAM, 2026 HBM capacity fully booked at all three producers and most 2027 slots already committed, consensus undersupply through late 2027, and capex responding hard, Micron going from $25bn to $35bn+ in a single fiscal year and the big 3 (Micron, SK Hynix, Samsung) investing hundreds of billions into new fabs over the next years. Hyperscalers are booking memory years in advance, behavior you only see at a true bottleneck.

The memory stocks are up 4x (SK Hynix) to 32x (SanDisk) over the last year. By very far the biggest value creation over the period.

I have covered this layer in depth: the framework in The $1 Trillion Memory Singularity, every name adjudicated in The Memory Basket: Every Name, Every Verdict, and deep dives on SK Hynix (SK Hynix: The Flagship at Half Price), Micron (Micron: The Call Option on the Memory Singularity), Samsung (Samsung: More Profitable Than Nvidia, at a Fifth of the Price) and SanDisk (SanDisk: 17x 1 Year after Spin Off). Kioxia coverage coming. The interface silicon between memory and compute rides the same tightness, Rambus and Montage, a niche flagged for a primer. The unglamorous corner is hard drives, still roughly 80% of external storage capacity by gigabyte, Seagate and Western Digital, uncovered but noted for a future piece.

Networking and optics: the next bottleneck

Connectivity is becoming the new battleground, because a GPU is bottlenecked by how fast you can feed it data and return results.

The numbers are the steepest in the value chain: the interconnect content opportunity expands roughly 9x to around $154bn between the current NVL72 generation and the Rubin Ultra NVL576 systems arriving in 2028, spanning copper cables, pluggable optics, co-packaged optics and PCB midplanes. Scale-up bandwidth inside the rack, scale-out between racks, and the copper wall in between decide the architecture, with CPO penetration assumptions around 25% on the coming generations and co-packaged copper extending the incumbent’s life in the meantime.

• Optics, where I have gone deepest: Coherent and Lumentum (deep dive: Coherent & Lumentum: The Laser Duopoly That Rations the Supercycle), the system view in Optical & Networking Supercycle at the Heart of the AI Revolution, and the two regional maps, The European Photonics Supercycle and The US Photonics Supercycle, which cover the entire value chain including Soitec, Aixtron, STMicroelectronics, MACOM, Fabrinet.

• Switching and interconnect silicon, the gap in my coverage I am closing soon: Arista, Cisco, Credo, Astera Labs, Amphenol, Ciena. Coverage coming.

Power semiconductors and the 800VDC prize

Power content per rack is roughly tripling across generations, and the 800VDC transition redesigns the entire conversion chain from grid to chip. Infineon (deep dive: Infineon: Power semiconductor leader with strong structural drivers) and STMicroelectronics (deep dive: STMicroelectronics: Photonics, Satellites and Silicon Carbide) are the European expression, with Monolithic Power, Renesas, onsemi, NXP, Analog Devices, Texas Instruments, Mitsubishi Electric, Rohm, Navitas and Delta Electronics in the stack. A dedicated 800VDC winners piece is coming very soon as part 3 of my Who Powers AI series.

The assemblers: ODMs and OEMs

The server market crosses $1tn by 2028, with 8-GPU equivalent shipments growing around 48% this year. And yet this is structurally the thinnest margin link in the AI value chain: assembly economics are slim by design, the value sits in the silicon the assemblers integrate, and the AI mix actually dilutes their gross margins even as it inflates revenue.

That combination of huge revenue and thin economics is why I have covered the exception that is Flex (deep dive: Flex: The electronics manufacturing giant building Nvidia’s servers), which is climbing into power products where the margins live. Celestica and Hon Hai coverage coming, Dell, Supermicro, Quanta Computer, Wiwynn, Wistron and Jabil complete the map.

Layer 5: Making the silicon

What it does: Builds the machines that build the chips: foundry, semiconductor equipment, EDA and IP, wafers, substrates and materials. This layer is where the true monopolies of the chain live.

Is it a bottleneck? Structurally, yes. A leading edge fab is a two year, $20bn+ commitment. Advanced packaging is acute, HBM’s wafer appetite collides with packaging capacity. And the tool intensity of every roadmap keeps rising, which is why semiconductor equipment is the highest conviction second derivative of the whole AI buildout.

• Foundry: TSMC manufactures essentially every AI accelerator that matters (deep dive: TSMC: Manufacturing the World’s AI Compute, and the paired case in ASML & TSMC: Own the Toll Booths of the AI Supercycle). Intel is the political option on a second Western source, coverage is coming. GlobalFoundries, Tower Semiconductor and the silicon photonics foundry angle, coverage coming.

• Semiconductor Equipment, the layer I have covered most completely: sector framework in Semiconductor Equipment Primer: $1.3 Trillion of Market Cap, Monopoly and One Problem Nobody Wants to Talk About. Wafer fab equipment guides for 2026 were raised to $150-160bn, up 30%, and bottoms-up consensus now implies roughly $213bn in 2027 and $262bn in 2028, the strongest equipment cycle ever contemplated. ASML (deep dives: ASML: Five Machines, a Third Off and ASML: The most innovative company in the world will 2.5x its EPS by 2030), ASM International (deep dive: ASM International: A Pure-Play on Semiconductor Physics), Besi (deep dive: Besi: The Packaging Monopoly at the Heart of the 3D Chip Era), Technoprobe (deep dive: Technoprobe: The Hidden Picks-and-Shovels Monopoly of the AI Test Era), Aixtron (deep dive: Aixtron: The Epitaxy Monopoly Growing AI’s Lasers and Watts), with the European map in Six Monopolies, One Supercycle: EU Semicaps. Applied Materials, Lam Research, KLA, Tokyo Electron, Advantest, Teradyne, ASMPT, DISCO, Lasertec, SCREEN and VAT, coverage coming across the group.

• EDA and IP, the highest quality business model in semis, monetizing R&D rather than capex: Cadence (deep dive: Cadence Design: A defensive compounder in the semi space), Synopsys coverage coming as a sector piece.

• Wafers and engineered substrates, the late cycle squeeze: Siltronic (deep dive: Siltronic: Every AI Chip Starts Here), Soitec (deep dive: Soitec: The Photonics Monopoly, Now At A Price That Works), GlobalWafers, Shin-Etsu and SUMCO complete the map, with Entegris and Hoya on the materials and mask flank.

The blind spot I am correcting: IC substrates and AI PCB. This segment was absent from last year’s version and it is displaying every bottleneck signature you could possibly want to see: customers extending prepayments, capex sharing and long term agreements to secure capacity, the complex compounding revenue around 25% and EPS around 117% annually through 2027 on current forecasts, and a materials choke point where Ajinomoto, the Japanese food group, holds roughly 95% share of the ABF film every advanced substrate depends on. Ibiden, Unimicron, AT&S and Shinko carry the substrates, with TTM, Victory Giant and Shengyi on the boards feeding the custom silicon programs. A full substrate piece is coming, this is the most asymmetric uncovered segment on the map.

The parallel chain: China

In reality, there are now two distinct AI value chains, and the second one is being built at great speed.

• China’s IC volume self sufficiency reached roughly 70% by mid 2026, domestic equipment spend is compounding toward a $100bn cumulative opportunity.

• The leading domestic memory maker, CXMT, is on track to more than double capacity by 2030, and just completed its domestic listing (which was 570x oversubscribed) and is now the most valuable Chinese company. I will be breaking down the mechanics of China’s DRAM champion in the next installment of my Memory Singularity series.

• Domestic accelerators are absorbing a rapidly growing share of a national AI capex program in the hundreds of billions.

• SMIC, CXMT, YMTC (private but contemplating a listing), NAURA, AMEC, Cambricon and the Ascend ecosystem are the spine of this AI ecosystem.

• I flagged the strategic significance in The Great AI Reckoning and the China semicap primer is one of my next flagship commitments.

• But this is what you should never forget (as export restrictions keep coming): the West’s export control strategy has converted a customer into a competitor, and every layer of this AI map now has a Chinese equivalent.

Layer 6: Distributing the compute

What it does: Buys gigawatts, sells tokens. Hyperscalers, neoclouds and sovereign programs are the exchange rate between the bottom of this map and the top.

Is it a bottleneck? No, and that is the point. This layer absorbs every bottleneck below it and carries the FCF burn for the whole system. This is the layer that will have to prove the AI monetisation economics and which supports the entire AI trade with rising AI capex.

• Hyperscalers: Microsoft (deep dive: Microsoft: The Best Company In The World Is Too Cheap), Alphabet (deep dive: Alphabet: The Hyperscaler That Already Pays For Itself), Oracle, which changed the shape of the market by turning backlog into the industry’s scoreboard (deep dives: Oracle: The Oracle of the AI Age and Oracle has changed everything), Amazon and Meta, both coverage coming. The Chinese demand stack runs in parallel, Alibaba, Tencent, ByteDance and Baidu, spending against a national program in the hundreds of billions.

• Neoclouds, the pure play token foundries: Nebius (deep dive: Nebius: The Purest Way to Play AI Compute), CoreWeave coverage coming, the converted miners IREN and Applied Digital, and SoftBank entering with a 10GW target by 2030.

• Sovereign: national programs are real money, Gulf gigawatt campuses, European buildouts, and SpaceX blurring the category entirely (deep dives: SpaceX: Inside the Trillion-Dollar Rocket and SpaceX: Is It Worth $1.7 Trillion?).

The monetization answer: read the backlog, not the capex

The most important development at this layer over the past two quarters is that the monetization question finally has a quantitative answer, and it is the backlog. Committed cloud backlog across the five largest buyers now totals around $2.4tn, and the trajectory at the two names where the debate was the most fierce has gone vertical:

• Google Cloud backlog reached $514bn at 2Q26, up 114% versus year end 2025, with more than half of it converting to revenue within two years, and cloud revenue growth accelerating to 82% in the quarter, exiting the year in the 90s. This is the proof of rapid monetization that investors feared would never show up.

• AWS backlog reached $496bn, up 103% versus year end, with a weighted average remaining contract life above six years, anchored by two of the largest compute commitments ever signed, a $100bn ten year agreement with Anthropic and a $138bn eight year agreement with OpenAI.

• This revenue is also a lot more visible: these are take or pay style multi year commitments from the fastest growing companies in history, not spot demand. The circularity critique, that the labs’ commitments are funded by the hyperscalers’ own investments in them, is partly fair and I acknowledge it, I stress tested it in Is the AI Trade Financeable?. But a $2.4tn contracted revenue base is a fundamentally different risk profile than what we had last year, which was the build it and they will come story.

The tension of the layer is now clear: record contracted revenue against record cash burn. Backlog proves that demand exists and the financing overlay shows the buildout can be funded. Both are true at once in my view, which is exactly what makes this layer so contested rather than simply cheap or simply expensive.

The financing overlay

This is the layer that did not exist in last year’s version, and its absence reflects just how much the scale and financing of this historic buildout has changed. The AI trade has moved from self funding to funded:

• The four heavy hyperscaler issuers have sold $223bn of bonds this year through the 20th of August 2026, more than double everything they issued in all of 2025, which was itself more than four times their 2020-24 average.

• Across the whole complex, $344bn of AI related issuance year to date against $136bn last year, spanning hyperscalers, data centers, chipmakers and the adjacent names. Nine individual raises above $20bn this year, a first bond deal in five years from the GPU maker at $25bn, even a century bond.

• My funding math from Is the AI Trade Financeable?: a roughly $411bn two year need against roughly $787bn of identifiable sources, a buffer near $375bn that covers another full year of FCF downgrades.

• The trade is financeable today but as I pointed out, you should watch the cost of it staying financeable, which now shows up in long term yields.

Credit is not a footnote to this chain anymore. It sits underneath the AI value chain, and it is the transmission mechanism through which a demand disappointment would propagate. That is question one of the two live questions from The Reckoning Arrives.

Layer 7: Models

What it does: Turns compute into intelligence, priced per token. This is where the unit of account switches from gigawatts to tokens, and the exchange rate is brutal and improving:

• The previous generation of systems produced a million tokens for around $4.00. The current rack scale generation produces the same million for $0.10-0.30, a cost collapse of roughly 95% per hardware generation, the framework I built in Token Economics: The Master Variable.

• Everything in the six layers below exists to push that number down, and everything above depends on demand growing faster than the price falls.

• So far it has: usage keeps expanding into every price cut, and the strongest recent signal is that the most aggressive price cutter in the market (Moonshot AI’s Kimi K3) raised API prices this summer (by 223% vs its previous model), the first hint that the price war is maturing into monetization.

And the demand side has so far won the race against the price collapse, spectacularly:

• When I published last year’s version, the AI native labs were generating a cumulative $18.5bn of annualized revenue.

• My estimate today is a run rate around $100bn, roughly 5x in under a year, with Anthropic and OpenAI respectively at approximately $45-50bn and $35-40bn ARR and the remainder across the hyperscalers’ model APIs.

• No industry has ever gone from $18.5bn to $100bn of run rate revenue in a year. That is what keeps the six layers below funded.

The capital markets have priced it:

• The two leading labs raised this year at valuations of roughly $850bn and $950bn+, together approaching $2tn of private market value, larger than all but a handful of public companies.

• Which sets up the event this layer is building toward: these companies are widely expected to come to public markets, and when they do they will be the largest listings in history.

• An IPO at this layer would do three things at once: hand public investors direct exposure to the token P&L for the first time, force a mark on every private AI position in the system, and give the whole AI value chain below a daily traded barometer of AI demand.

• I would treat the first frontier lab listing as the single most important liquidity event of this cycle, and I will publish on it the moment terms exist.

The named exposure until then is obviously thin given most of the layer is private: OpenAI (deep dive: OpenAI: The center of the AI revolution), Anthropic coverage coming, the hyperscalers’ in house models, Mistral carrying the European flag, and the Chinese frontier labs, DeepSeek and Moonshot AI, whose efficiency releases have twice now moved the entire public complex. That asymmetry, private companies setting the marginal price for trillions of dollars of public market capitalization, is itself a structural feature of this value chain and one reason volatility is not going away.

Layer 8: Applications and agents

What it does: Turns intelligence into workflows and revenue. This is the layer every other layer is a bet on, and honestly, it is the thinnest part of my own coverage, deliberately so until the economics settle.

This is the part of the AI value chain that has suffered tremendously over the last year:

The sector has shed over $1tn of market capitalization from its highs with the main index down 5% over the past year and flat YTD after a recent rally, while the physical layers of this chain re-rated 100-200%.

The market has voted for atoms over bits (and semis/power over software), on a wall of worry: that the frontier labs will take the stack and collapse application moats, that AI budgets are crowding out core software spend, and that the cost of creating software going to zero takes software pricing with it.

Some of that fear is rational but some of it is the setup, because a sector priced for disruption only needs to survive to re-rate (which is my argument in Software Armageddon: AI Fears, Capitulation + 10 Names to Play a Potential Rally). My conclusion was that the sector had been oversold.

What the evidence actually shows:

• Enterprise adoption is real and monetizing: roughly two thirds of enterprises now run AI in production, agent platforms are shipping across HR, finance and service workflows, and the architectural consensus is forming around the harness, the software that manages prompts, memory, tools and governance, staying separate from the model so the model can be swapped. I view the architecture as the moat: value migrates to proprietary data, distribution and workflow position, not to the model relationship.

• The resilient end of the spectrum is software with a physical world moat, which is where my coverage sits deliberately. The engineering and simulation platforms (vertical software) whose data and workflows describe atoms, not text: Bentley Systems (deep dive: Bentley Systems: The software compounder bringing infrastructure into the AI age) and Cadence. And the quality end of the stack, which features Salesforce (deep dive: Salesforce: A Value Tech Christmas Gift?), RELX (deep dive: RELX: The company to own forever & one of the very best businesses on Earth), Roper (deep dive: Roper: A resilient cashflow compounder in the software space with inflecting growth), Dassault Systemes (deep dive: Dassault Systemes: Why are investors overlooking this Software giant?), Shopify (deep dive: Shopify: How One Platform Took Over Modern Retail) and Sopra Steria for IT Services (deep dive: Sopra Steria: This boring IT business is a hidden Defense and Sovereignty play), with SAP as one of the highest quality and least disruptable names in global software (deep dive: SAP: The unstoppable growth engine behind every business). Palantir and the broader discussion around which software survives the AI labs and why software is becoming more and more compelling, coverage coming, it is one of the most consequential open questions on this entire map, and Datadog on observability, one of the cleanest usage based winners of agent traffic, remains in my backlog.

• Cybersecurity is the cleanest derivative demand story in software: autonomous agents expand the attack surface by construction, recent incidents keep proving it, and security budgets are the last line an enterprise cuts. Palo Alto Networks, CrowdStrike, Zscaler, SailPoint. A cybersecurity primer is on the roadmap.

And the wave after AI agents is already visible: physical AI, intelligence moving into robots, vehicles and machines, which turns the $50tn industrial economy into the next demand layer of this chain. I mapped the theme in The Physical AI Primer: When Intelligence Gets a Body and named my preferred exposure in The Physical AI Basket: Where the Value Actually Accrues, with the underappreciated route running through the engineering software above.

Physical AI is the strongest single argument that the AI value chain in this piece is a decade long structure rather than a cycle, because it resets the demand curve just as the current wave matures.

What changed, and what I got wrong

I wanted to take the time to go over what has changed since my September 2025 version and what I missed:

• I badly underweighted memory: Last year memory was one company list among many. It became the single biggest bottleneck in the chain and the best performing pocket in the market, memory, PCB and testers went from up from 100% to 32x year to date. I corrected this gap in my coverage with my very comprehensive Memory Singularity series, but the timing was late, and the cost of missing it was the largest alpha pool of the cycle.

• I treated power as an individual segment of the value chain rather than the common denominator: My old map had an energy section whereas my new map starts at the power plant, prices every layer per gigawatt and treats electrons as the binding constraint, which is what my ongoing Who Powers AI establishes.

• Substrates and PCB were absent entirely: This is a segment now receiving prepayments and capacity sharing from customers, the classic bottleneck signature in my view, and it did not appear on my map at all last year. It does now, and I am planning a dedicated piece and have already started covering the space with Soitec and Siltronic.

• The CPU was missing: The agentic shift toward one to one CPU attach and a $210bn server CPU market was not on last year’s map and is quite a recent topic of discussion in the market. I caught it in CPUs Are the Next GPUs, but slightly after the fact.

• I understated the ASIC share shift: Long term custom silicon commitments at the major hyperscalers have hardened materially since, and custom programs now run at all four major CSPs. The ASIC space remains one of the most compelling segments in the market as hyperscalers try to diversify away from Nvidia and I will continue to expand my coverage here.

• The financing layer did not exist: $344bn of AI related debt in eight months makes credit a structural and very recent part of the value chain, the rise of this massive borrowing as hyperscaler FCF goes negative has created a lot of debate and unease in the market.

What held up since last year: the infrastructure over models framing, the insistence that the real winners build and power the machine, Nvidia’s centrality in this revolution, and the electrification linkage that most AI maps still treat as a separate trade.

The core call, that this is the largest investment cycle in history and still accelerating, was right and remains right. It was just too small and continues to be revised up massively, quarter after quarter.

Positioning: crowded, contested, overlooked

The map tells you where value is created. Positioning tells you what you are paid to own today. I split it into three buckets, updated for the June to August correction:

• Crowded: Memory, PCB and testers, plus optics. The positioning leg of the selloff has largely cleared through these names, the leveraged vehicles have unwound, but crowding is a valuation condition, not just a flow condition, and these layers now need the fundamentals to keep delivering. They probably will into 2027 tightness, the entry just matters more than it did, and My AI Selloff Shopping List remains my ordering of quality versus beta for anyone re-entering the complex.

• Contested:

  • The two live questions from The Reckoning Arrives: AI model commoditisation and credit.

  • Nvidia versus the ASIC complex, which is really a question about where the compute margin pool settles.

  • The hyperscaler FCF burn against the debt window, which is a question about whether $1.1tn of 2027 capex gets funded at acceptable spreads.

  • I own a view on all of these, which you can find in my framework pieces linked above.

• Overlooked: The CPU complex, substrates before the market names them as a layer, the gray space electricals after the selloff, where the demand floor argument from My Electrification Selloff Shopping List still applies, regulated utilities and the PJM generators, EDA, and arguably the entire facility layer, where nothing has re-rated the way the rack has despite content per rack rising every generation.

The roadmap

This map doubles as my publishing plan for the AI megatrend. You can already find 42 names and dozens of thematic pieces across the AI value chain, all linked above in their places and searchable in my AI & Technology tab.

The gaps I have flagged as coming are, in rough priority order (with individual names on top of this): the Who Powers AI series continues (including the 800VDC power train), continued coverage of grid names, the China memory piece (part of my Memory Singularity piece), the China semicap primer, ASIC names, IC substrates and AI PCB, networking and interconnect, silicon photonics foundries, the server assemblers, Amazon and Meta to complete the hyperscaler set, AMD, the neocloud and real estate initiations, and my view on software in this AI age. If you want the deepest and clearest AI value chain coverage, this is the place.

A little request: If there is a name on it you want covered next, say so in the comments. The queue is ordered by my conviction, but subscriber demand moves it.

If this map is useful, please share it. It is free precisely so it can be widely read.

Share

The deep dives on every named company, the shopping lists and the frameworks behind this piece are for paid subscribers, alongside everything that comes next on the roadmap above.

Join Crack The Market

Disclaimer: The information provided on this Substack is for general informational and educational purposes only, and should not be construed as investment advice. Nothing produced here should be considered a recommendation to buy or sell any particular security.

Twice a week, I will release deep dives into stocks and sectors that fit into the 6 themes that I see winning in the coming years and decades: AI & Technology, Electrification & Energy, Reshoring & Sovereignty, Healthcare & Longevity, China & Asia, Resilience & Quality.

Join Crack The Market

Take advantage of this once in a generation opportunity to build long term wealth by investing in great stocks that will deliver returns for your portfolio for years to come.

After reading this article you will understand what Nokia actually is today, why the Infinera acquisition was the most consequential capital allocation decision this company has made in a decade, why the AI order backlog went from zero in Q1 2025 to €4.7bn at the end of Q2 2026 and roughly €5bn today, why the hyperscalers themselves are pushing volume toward Nokia, why Nvidia put $1bn into a Finnish telco vendor, and why the setup after the summer unwind is the most interesting entry this story has offered since the re-rating began.

In this article I go through:

• The exceptional Q2 2026 results, the 50% drawdown that made Nokia the worst performing optical stock in the world for a stretch, and why the opportunity is so compelling after the positioning reset

• The business, the two new segments, and how a 160 year old value trap became the only vertically integrated Western optical company outside the US

• The value chain position, why Google, Microsoft, Amazon and Meta chose Nokia as their second source, and why vertical integration in indium phosphide and DSPs is the scarcest asset in optical systems

• The financial history and the consensus reset, from a decade of shrinking revenue and flat earnings to a 15% EPS compounding path that the more constructive investors see closer to 25%

• Where the thesis breaks, the 2027 product transition, the mobile drag, the hyperscaler concentration, and the scenario analysis

Nokia Investment Case

Table of content

• Business Description

• Investment Case

• Innovation

• Value Chain

• Growth Drivers

• Risks & Threats

• Financial Robustness & Capital Allocation

• Outlook/Guidance

• Governance & Management Quality

• Valuation

Positioning, the continuation of my Photonics series

This deep dive continues the Photonics/Optical arc:

• I started with the Optical & Networking Supercycle primer in January

• Continued through The European Photonics Supercycle & The US Photonics Supercycle

• And the company deep dives on Soitec, STMicroelectronics, Aixtron, and Coherent and Lumentum.

In the European piece I ranked Nokia a Tier 2 weight and wrote that if you want vertically integrated European optical networking exposure with hyperscaler customer relationships, Nokia is the only credible name. That sentence has aged well. Since then the company has doubled its AI and cloud revenue, built a €5bn hyperscaler backlog, launched the industry’s first commercial AI-RAN platform, and bought a third fab. What has changed is not the thesis but the evidence behind it, and, thanks to the summer selloff I mapped in The Reckoning Arrives, the price.

My conclusion in The Reckoning was that we are living through a pause inside an intact supercycle, and that the discipline is to buy the businesses whose order books are accelerating into the drawdown rather than the ones whose narratives are. Nokia is now one of the single clearest European examples of that setup.

What just happened

Three things happened between April and August that together explain why I am writing this piece now.

• First, the backlog went vertical:

  • Until the middle of 2025, effectively 100% of Nokia’s network sales came from telcos, as they had for a decade.

  • Then the hyperscalers arrived: In Q1 2026, Nokia booked €1bn of orders from AI and cloud customers, roughly 40% of what it received in the whole of 2025, with a book-to-bill in that segment of around 3x. In Q2 it booked €2.8bn more, over half from optical and mainly pluggables. The result is an AI backlog that has gone from €1.7bn at the end of 2025 to €2.35bn at the end of Q1 to €4.7bn at the end of Q2, with the hyperscaler component alone now around €5bn against zero five quarters ago, and plausibly €6bn by year end at the current run rate.

    

  • Already a backlog 10x the current AI/Cloud revenue run rate: Management was very honest about the composition, roughly half of the Q2 intake converts to revenue within twelve months and the rest sits further out, because customers are placing long-dated orders to secure supply in a market where optical lead times run 12 to 18 months and parts of the industry describe themselves as sold out for multiple years. Orders like these are lumpy and will not repeat every quarter. But a company whose AI and cloud franchise generated €446m of revenue in Q2, up 105% yoy, now carries a backlog worth more than ten times that quarterly run rate.

AI and cloud orders, revenue and end-of-period backlog. €1.7bn of backlog at the end of 2025 became €4.7bn six months later. The gap between the bars is the visibility the market has never given Nokia credit for.

• Second, the numbers are starting to inflect:

  • Q2 net sales of €4.8bn grew 9%, gross margin expanded 70 basis points to 46%, operating margin reached 9%, and operating profit of €434m came in 16% above consensus.

  • Network Infrastructure grew 12% with optical up 20% and IP up 16%, and its gross margin expanded 240 basis points to 42.7% as Infinera synergies and scale came through.

    

    

  • The company is tracking somewhat above the midpoint of its operating profit guidance, re-based to €2.1bn to €2.6bn with the discontinued operations removed, while simultaneously funding three fab projects and around €800m of restructuring charges. This is what the early days of operating leverage look like.

• Third, the price finally moved against the story, violently:

  • The stock re-rated relentlessly from late 2025, from under €4 before the Nvidia announcement through a roughly 170% YTD move by early June that made it the fourth best performer in Europe, to a June peak near €14.5 where it traded at 38x forward earnings and roughly 25x forward EV/EBITDA.

  • Then the AI unwind/reckoning did its work: The shares fell nearly 50% in two months, at one point the worst performing optical stock in the world on a one month view, hit by the momentum factor unwind, by index-linked selling as the September blue chip index inclusion trade wobbled, and by a 5% drop after Q2 results when management, having just printed a record order quarter, declined to raise its AI outlook on the spot.

  • At €8.7 the stock sits roughly 40% below the June high, at 25x forward earnings and 13.8x forward EV/EBITDA, while 2027 earnings estimates rose throughout the entire drawdown.

  • Exactly as with Soitec in July, the market spent the summer de-rating a company whose earnings power was being revised up. That divergence is where I like to do my work, and the positioning reset underneath it, momentum money flushed, index money reduced, long-onlys who missed the first rally circling back, is precisely the change that precedes the next leg.

The estimates ground higher through the entire drawdown. The stock did not.

Business Description

Nokia is a Finnish network equipment and technology company headquartered in Espoo, with roughly €19.9bn of 2025 revenue, around 70,000 employees after the current restructuring, €4.9bn of annual R&D spend anchored by Bell Labs, and one of the largest patent portfolios in communications with more than 26,000 patent families, of which over 7,000 are declared essential to 5G.

From 2026 the company reports two segments, and the reorganisation is itself very instructive about what management thinks the equity story is:

• Network Infrastructure (NI), roughly €7.6bn of revenue, spanning optical networks (transport systems, coherent pluggables and, since Infinera, the underlying indium phosphide photonic integrated circuits and DSPs), IP networks (edge routing where Nokia is the global number one, plus a growing data centre switching franchise), and fixed networks (fibre access, where Nokia leads in optical line terminals). This is the segment that carries the AI thesis, selling to hyperscalers, neoclouds, sovereign clouds and telcos building the transport layer of the AI build out.

  

• Mobile Infrastructure (MI), roughly €11.4bn of revenue, combining radio access networks, cloud-native core software, and Nokia Technologies, the patent licensing unit with a €1.4bn contracted revenue run rate and more than €800m per year locked in through 2030. Radio is the structurally challenged business while core software and licensing are the stable profit engines that fund the transition, together contributing the bulk of the segment’s €1.5bn of annual operating profit.

  

  

Customer mix is the single most important lens on this company:

• In 2024, telecom operators were roughly 80% of revenue, a customer base growing 1% to 2% a year at best. Management expects that share to fall below 60% by 2028 as AI and cloud, mission critical enterprise and defence-adjacent demand grow into the mix.

  

  

  

  

  

• The combined optical plus IP franchise alone is on track to go from roughly a quarter of group revenue in 2025 to around 40% by 2027 on the more constructive estimates.

• The equity story is the speed of that change, exactly the same mix arithmetic that transformed the Soitec story in a single quarter, playing out here on a €20bn revenue base.

The same mix arithmetic that transformed the Soitec story, on a €20bn revenue base.

A brief history of Nokia:

• 1865: Founded as a paper mill on the Nokianvirta river. Over the next century it diversifies into rubber, cables and electronics.

• 1990s to 2007: Nokia becomes the mobile phone. At the peak it holds around 40% global handset share, is Europe’s most valuable company and represents a meaningful share of Finland’s GDP and exports.

• 2007 to 2013: The iPhone and Android destroy the franchise. The handset business is sold to Microsoft in 2014, one of the great value destructions in corporate history.

• 2014 to 2016: Nokia reinvents itself as a network equipment vendor, buying out Siemens from their joint venture and then acquiring Alcatel-Lucent, which brings Bell Labs and the IP routing and optical businesses.

• 2016 to 2023: These are Nokia’s lost years. 5G product missteps around in-house silicon force a painful catch-up, market share leaks to Ericsson and Huawei, and the AT&T RAN loss in late 2023 marks the low point. The stock spends years between €3 and €5 and the shareholder base becomes a graveyard of European value hopes.

• June 2024: Nokia announces the acquisition of Infinera, a vertically integrated US optical vendor with its own indium phosphide fab in San Jose, for €2.5bn. The deal closes on the 28th February 2025.

• February 2025: Justin Hotard, then head of Intel’s Data Center & AI group and previously a senior HPE executive, is announced as CEO. He starts on the 1st April 2025. The signal is unambiguous, the board hired a data centre executive, not a telco guy.

• 28th of October 2025: Nvidia announces a $1bn equity investment, taking a 2.9% stake, alongside an AI-RAN and AI networking partnership.

• 19th of November 2025: The CMD unveils a new strategy under the banner of connecting intelligence, new segments, and a 2028 operating profit target of €2.7bn to €3.2bn (a target that will likely be overshot) against roughly €2bn delivered in the trailing twelve months.

• 2026: The evidence arrives with NI growth guidance raised from 6% to 8% up to 12% to 14% at Q1. €3.8bn of AI and cloud orders in the first half and a hyperscaler backlog near €5bn. The first commercial AI-RAN platform launched in July. Three optical capacity investments announced. The stock triples from its 2025 lows to nearly €14.5 in June, then gives back roughly a third of its value in the AI unwind.

Investment Case

Read more

My verdict has not changed and I still see this as a pause inside an intact supercycle. But the funding architecture underneath the trade now determines which layer of the stack would break first if anything breaks at all, and the short version is that the market is not running out of money, it is repricing it. That repricing is quietly sorting the complex into three groups (the self funded, the hybrid and the load bearing names) that equity investors still mistakenly lump together, which I think creates a compelling opportunity for investors over the next years, has consequences for the AI trade and which names you should own.

Why an equity investor should care about credit at all

Most of you reading this are equity investors, so I wanted to take come back to the basics of credit for a second to explain to you why you should care about this:

• The spread is the price of risk: When a company borrows in the bond market, the interest it pays is of course above Treasuries (risk free rate given the Treasury should never default on its debt), and the magnitude (in basis points) above is the price for the risk of not being paid back. If a company’s debt trades only at a 40bps premium, it means that default is close to unthinkable, if it trades 800bps above the risk free rate, it means that lenders want eight extra percentage of yield a year just to hold this debt.

• Spreads move for three reasons: how much paper/debt is being sold (supply), how the business is doing (fundamentals), and how much risk investors feel like holding (appetite).

• Credit tends to pay more attention to risk than equity: A lender’s best case is getting their money back with interest, so credit investors are extremely diligent about modeling the downside, and they reprice any changes to fundamentals or sentiment way before equity markets. This is why you see big moves in spreads and CDS markets (insurance against default) being discussed by investors despite the stock barely moving.

• Three dials matter for this piece:

  • Oversubscription measures dollars of orders per dollar of bonds sold, the purest reading of appetite.

  • Curve steepness compares 10 year against 30 year borrowing costs, and a steepening curve says investors trust the near future more than the far one.

  • The level itself of the debt, because the cost of debt now feeds into the price of cloud compute, into margins, and into the equity you own.

The marginal dollar being invested to build the AI infrastructure is increasingly a borrowed one.

The scale: AI credit is now a competitor of government debt

• A rapidly rising rate of borrowing: $223bn of bonds sold this year through 20th of August by the four heavy issuers among the hyperscalers, more than double everything they issued in all of 2025, which was itself more than four times their 2020 to 2024 average.

• Everyone in the ecosystem is raising debt: $344bn of AI related supply year to date across the whole complex against $136bn last year, spanning hyperscalers, data centers, chipmakers and the space adjacent names.

• Individual deals are become very large: This year we have already seen nine raises above $20bn, Nvidia sold $25bn in its first bond deal in five years, and there is even a century bond from Alphabet this year.

• Technology is becoming the biggest incremental debt issuer, something unthinkable a few years back: Tech has overtaken banks in several major investment grade indices for the first time.

Investors should also pay attention to the macro picture, which is more and more intertwined with this AI debt wave:

• The 30 year Treasury touched 5.33% this week, a 19 year high, and respected macro desks attribute part of that move to hyperscaler issuance arriving on top of a federal deficit running toward $2tn.

• High quality hyperscaler paper has become a direct competitor for government debt, hedged in the same sovereign markets, which is why the supply wave now shows up in Treasury yields themselves.

• Long rates are used as the discount rate for assets with long duration cashflows like hyperscalers (the stocks) but also their debt, which means that higher rates will be doubly a drag on their valuation.

• The September FOMC, which I flagged in the Reckoning as the first live test of a hike, is now a test for this trade specifically, with the debate being around whether rates will be held steady or hiked.

The funding math: a $375bn head start

The bear case rests on a simple equation:

• Consensus capex for the five hyperscalers in 2027 has been revised from $480bn a year ago to $1,088bn today, up 127%, while free cash flow estimates for the four ex Microsoft have swung to negative $145bn for 2027.

• What does that mean? Ever rising capex meets collapsing FCF, requires ever more debt, and eventually the market chokes on it.

• I agree with the first two parts of the equation (higher capex and lower FCF) but the third assumption does not agree with the reality.

What is absolutely critical to understand is that the FCF deterioration is the deliberate result of accelerated spending, not due to a profitability issue:

• Operating cash flow estimates rose 22% over the same twelve months and have been revised up at every reporting season, clearly a very bullish sign.

• The FCF deterioration is a spending decision rather than a profitability problem, and that distinction is everything in credit: Lenders fund spending decisions against stable cash generation all day long, while deteriorating fundamentals would close the debt market and would be a huge issue for hyperscalers.

Let’s look at the funding needs and sources of the four heavy spenders ex Microsoft (which I exclude because consensus has Microsoft FCF positive every single quarter through 2029) - The funding buffer covers 2026 & 2027:

• The need - around $400bn: Cumulative negative FCF of $231bn across 2026 and 2027, plus roughly $180bn of equity investments into the labs, gives a two year funding requirement of about $411bn.

• The sources - around $800bn available: $88bn of debt already raised in 2H 2025, roughly $294bn expected for full year 2026, the same run rate assumed for 2027, and $110bn of announced equity raises give total funding of about $787bn.

• The buffer - nearly $400bn, which covers another year of FCF downgrades: the group is funding roughly $375bn ahead of consensus cash burn. For scale, the entire negative revision to 2027 FCF over the past twelve months was $361bn, so the current pace absorbs another full year of downgrades at the same speed before anything is stressed.

There is no AI credit crisis:

• The hyperscalers are pre funding at investment grade spreads, years ahead of need, from a position where the marginal dollar of capex is still a choice.

• Alphabet is the cleanest case study, which I covered in depth in my deep dive: $242.5bn of cash and securities against $98.2bn of long term debt, still $144bn net cash after taking debt from $16bn to $100bn in a year and raising $49.4bn of equity and converts it did not strictly need. This to me reads less like distress and more like buying balance sheet strength and optionality while capital is cheap.

The forward program: $1.3tn of new debt needs to be raised by 2030

Read more

Part 1 of this series ended at the lead time wall: one gigawatt of AI capacity, roughly $35bn of capital, a shell that goes up in 18 months, and a grid connection that takes six years. The most creditworthy, most motivated, least price sensitive buyers in industrial history are being rationed by time rather than price.

This Part 2 follows the hyperscaler into the gap between demand and the grid, where the fastest growing power market in the world is being built behind the meter, and answers the question that was raised in Part 1: when a data center takes six years to be connected to the grid and a model needs to be trained next year, what does a hyperscaler actually do, and who gets paid for every month of speed.

Part 3, The Rewiring, will map the 800V DC and liquid cooling transition inside the data center.

Part 4, Electric China, will ask whether China is winning the power race and what breaks the whole trade.

Twice a week, I will release deep dives into stocks and sectors that fit into the 6 themes that I see winning in the coming years and decades: AI & Technology, Electrification & Energy, Reshoring & Sovereignty, Healthcare & Longevity, China & Asia, Resilience & Quality.

Join Crack The Market

Take advantage of this once in a generation opportunity to build long term wealth by investing in great stocks that will deliver returns for your portfolio for years to come.

Some numbers to frame what impatience is worth:

• Hyperscaler capex guidance for 2026 aggregates to roughly $700-750bn, up from around $443bn in 2025. Every dollar of it depends on power arriving on schedule.

• A gigawatt campus is a $35bn asset. At that scale, capital sitting idle waiting for electrons costs on the order of $300-400m per month before a single lost customer is counted.

• Roughly 121 GW of new US gas capacity is being built by 2030, and around 55 GW of it deliberately skips the grid, adding about 10 bcf/d of gas demand.

• Around 20% of US data center power delivery is expected to run behind the meter by 2030, from nearly nothing three years ago. And yet, on the grid side, renewables are the single largest source of new data center supply globally, the other side of this piece.

• The market is already repricing firm capacity, demonstrating just how valuable power has become: the PJM capacity auction went from $29 to $329 per MW-day in two auctions, a 10x move that lifted total system capacity costs from $2.2bn to $14.7bn per year.

The price of a month without power

Why have hyperscalers started investing so massively and quickly into on-prem energy sources for their data centers despite its higher cost? It boils down to some brutal economics:

• A hyperscaler committing $35bn to a gigawatt campus is really committing to a revenue stream that starts when the power flows. Delay the electrons by a year and you have lost a year of the entire asset: the racks, the models trained on them, the customers signed on the models, and the market position that compounds from all three.

• At current capex levels the crude carrying cost alone runs at hundreds of millions per gigawatt per month. The strategic cost, in a race where capability compounds, is larger and unquantifiable.

• The speed premium is therefore very rational. A buyer whose alternative costs $300m a month will rationally pay double the grid tariff for power that arrives two years earlier.

The question for investors isn’t whether the premium exists because we can clearly see it in the backlog of solution providers. It is which sellers capture it, for how long, and whether their scarcity is a moment or a decade, the test this series established in Part 1.

A thirty second detour: how an electron gets priced

One mechanism from power markets explains most of what follows. Wholesale power is priced at the margin: at every moment, the grid dispatches generators from cheapest to most expensive until demand is met, and the cost of the last unit needed sets the price every generator receives.

• Now add gigawatts of data center demand that is almost perfectly inelastic, a buyer that will not switch off whatever the price, and the mechanism runs one way: ever more expensive marginal machines get dispatched, and the price they set flows to every megawatt on the system.

• This is why AI demand moves entire regional power prices (which has become a massive political football), why the owners of cheap existing generation are the silent winners of the buildout, and why a data center that brings its own generation is really buying an exit from someone else’s marginal price.

• One more consequence of the doubling math: US data center demand is currently doubling roughly every two years. Run that forward naively and data centers alone would in theory outgrow total US generation by the mid 2030s. That will not happen, something breaks first, and the routes in the next section are the market’s answer to what.

For readers who want the full mechanics, from locational pricing to spark spreads, the recently published free primer Electricity Pricing in the Age of AI, which was kindly brought to my attention by a subscriber, is the best short introduction I have found, and I recommend it.

There are five ways to skip the grid queue

When the interconnection queue is six years long (and rising), the hyperscalers have essentially five alternatives, and each one is a different industry getting paid. Every step toward faster power (from right to left on the graph) costs more per megawatt hour.

The impatience curve. Indicative costs and timelines by procurement route.

• The grid: Hyperscalers can join the queue, wait the median six years and pay the industrial tariff of $82-93/MWh. This guarantees the cheapest electrons but also the slowest compute. This remains everyone’s stated long term preference, which will matter for Part 4.

• Co-location: Hyperscalers can contract the output of an existing plant, overwhelmingly nuclear, next to which they build. This guarantees firm, carbon free, and available power in one to three years because the generation already exists. The executed 20 year deals cluster around $80-110/MWh, a structural premium to wholesale that defines what firm clean power is now worth. But there is only so much existing capacity that can be contracted.

• New build gas: Hyperscalers can build their own combined cycle plant on site. This takes wwo and a half to four years, and the timeline is gated by the turbine slots covered in the Turbine Trilogy. This setup is efficient once running.

• Behind the meter: Hyperscalers can use reciprocating gas engines and aeroderivative turbines, which are modular units installed in under 18 months. These options give hyperscalers the most expensive permanent electrons and have created the fastest growing power market in the world.

• On-site alternatives: There are now a few alternatives that can be used to spin up a data center. Fuel cells at the premium end, temporary diesel at the desperate end, and in sun rich grids, solar plus storage riding alongside as a fuel saver and, increasingly, as capacity.

The rest of this piece explores the fast half of these options (behind the meter, and the other side of the meter), the half that did not exist as an investable market three years ago, and tells you how to play this race of speed to power.

Behind the meter: the fastest growing power market in the world

Read more

Welcome to the 81st investment case, 42nd AI & Technology and 49th Reshoring & Sovereignty idea on Crack the Market (and the most comprehensive Siltronic investment case you will find online)! Join me in uncovering one of the most important company nobody watches: one of only five producers on earth of the hyperpure silicon wafers beneath every AI chip, the only one headquartered in the West, and the cheapest way into a wafer cycle that has already turned. Four years of downturn are ending with volumes inflecting and prices rising, while an income statement buried under the depreciation of a freshly built fab hides the whole thing from every screen on the market.

Along with my friend , we are releasing this deep dive into a German mid-cap the market rediscovered, repriced and half-abandoned in a single year: a seventy-year-old survivor of an oligopoly that consolidated from twenty players to five, trading near the book value of a freshly built fab, with two thirds of its volumes locked in prepaid long-term contracts and a strategic bidder that is still a shareholder.

Twice a week, I will release deep dives into stocks and sectors that fit into the 6 themes that I see winning in the coming years and decades: AI & Technology, Electrification & Energy, Reshoring & Sovereignty, Healthcare & Longevity, China & Asia, Resilience & Quality.

Join Crack The Market

Take advantage of this once in a generation opportunity to build long term wealth by investing in great stocks that will deliver returns for your portfolio for years to come.

Siltronic Investment Case

Table of content

• Business Description

• Why Wafers

• Investment Case

• Value Chain

• Growth Drivers

• Risks & Threats

• Financial Robustness & Capital Allocation

• Outlook/Guidance

• Governance & Management Quality

• Peer Comparison

• Valuation

Business Description

Siltronic (ETR: WAF) makes one product: hyperpure silicon wafers, the polished discs on which every semiconductor is built. Before TSMC prints a single transistor of an AI accelerator, a CPU or a memory chip, it needs an almost perfectly engineered platform to build on. Siltronic melts polysilicon into metre-long single crystals, slices them into discs around one millimetre thick, and polishes them to a tolerance where a single defect the size of a bacterium on a tennis court is out of spec. The precision matters because in the semiconductor world, errors are not tolerated and they compound. Imagine building a skyscraper on foundations that are off by a few millimetres and by the top floor you have the Leaning Tower of Pisa. Print billions of transistors on a wafer with microscopic defects and yields collapse, costing millions of lost revenue for chipmakers.

The product family runs from polished wafers, the core product, through epitaxial wafers, where an additional ultra-pure silicon layer is grown on top for demanding logic and power applications, to specialty lines such as Ultimate Silicon for leading-edge memory. The industry standardised on the 300mm format because it offers 2.25 times the surface area of a 200mm wafer at a similar processing cost per pass, and virtually every leading-edge logic and memory chip is built on one. And wafers are not the commodity the word suggests: each one is qualified to a specific customer process, tuned for dopants, oxygen concentration and defect density, and a qualification can take months. Customers do not switch suppliers on a running production line unless something has gone badly wrong.

The market is small, concentrated and disproportionately important. Global semiconductor wafers were worth roughly a $11.4bn market in 2025. They feed the >$1tn semiconductor device industry which in turn feeds more than $5tn of electronics end markets. Five producers supply roughly 90% of global demand: Shin-Etsu and SUMCO of Japan with around half the market between them, GlobalWafers of Taiwan, SK Siltron of Korea, and Siltronic, whose share has held between 11% and 14% for over a decade. Siltronic is the only one of the five headquartered in the West. The customer list is confidential, but analyst reporting consistently places TSMC, Samsung, Micron, SK hynix, GlobalFoundries, Infineon and STMicroelectronics at the core of the book. The fabless designers everyone associates with AI never buy a wafer, their foundries do given they are the ones actually building the chips.

Production runs in Freiberg and Burghausen in Germany, Portland in the US, and Singapore, where the existing 300mm fab has been joined by FabNext, a new highly automated facility that represents the largest investment in the company’s history at over €2bn and Siltronic’s gamble to catch up to its peers. Legacy small-diameter production was closed in July 2025. Roughly two thirds of 300mm volumes benefit from long-term agreements, many carrying customer prepayments, with the remainder exposed to the spot market.

Siltronic’s sales are largely concentrated in Asia at >75% of total sales: 36% to Taiwan and China, 25% to South Korea, 14% to other Asia, 12% Europe and 10% US and largelyandlargely invoiced in dollars.

Siltronic’s history - Seventy years of polishing:

Siltronic’s history matters because in this industry, survival is the moat.

• Wacker Chemie (the German chemical company) began researching hyperpure silicon in 1953 and produced its first wafer in 1962, three years before Moore stated his now famous law. The business expanded to Portland in 1978, absorbed Freiberg after reunification in 1995, took its current form as Siltronic AG in 2004, entered Singapore through a joint venture in 2006, and listed in Frankfurt in 2015 as Wacker Chemie sold down.

• In 2020 GlobalWafers bid €145 per share for the whole company, an offer that lapsed in early 2022 due to German regulatory timing and not valuation, leaving its parent Sino-American Silicon holding 12.4% to this day.

• The FabNext decision followed, over €2bn committed at the top of the cycle and ramped through the worst of the downturn, and 2026 brought a €273m capital increase and a return to the MDAX (Germany’s midcap index).

• The wider industry consolidated from more than twenty wafer suppliers in the 1990s to five today. Siltronic is one of the survivors, and the survivors inherited an oligopoly.

The 2025 numbers show a company at the bottom of a cycle: sales of €1,346.7m, down 4.7%, EBITDA of €316.9m at a 23.5% margin, EBIT of minus €26.4m and a net loss of €77.9m as depreciation on the new Singapore capacity began to flow. H1 2026 sales were €628.1m with EBITDA of €134.5m at 21.4% and EBIT of minus €104.3m.

At €87.7 the company has a market cap of about €2.9bn, with €0.7bn of net debt following the June capital increase. Wacker Chemie holds 21.7% after selling 7% into strength in May 2026, HAL Trust holds 13.7%, Sino-American Silicon holds 12.4%, and the free float is about 52%.

There are three things that you need to understand this company: the oligopoly, the contract book, and the fab in Singapore that is depressing reported earnings today and will drive them tomorrow.

Why Wafers: The Toll Booth Nobody Watches

Before the investment case, a detour worth taking, because most readers will never have looked at this segment, and that fact is itself part of the thesis.

Why this is an attractive business:

The case for owning wafer capacity is the simplest in semiconductors: the silicon content of human life only goes up. Every megatrend I have covered runs on wafers. AI datacentres are wafers by the million. An electric car carries a multiple of the semiconductor content of a combustion one. Industrial automation, defense electronics, satellites, the grid itself: each wave of adoption adds chips, and every chip begins as a wafer.

Through every cycle of the past three decades, wafer area shipped has compounded at a mid single digit rate regardless of which end market led, and industry forecasts take the roughly $11.4bn market of 2025 beyond $20bn by 2030. Layer the structure on top of the growth:

• Five suppliers with 90% share, and market shares that have barely moved in a decade, because nobody can take business on price in this industry without destroying their own business.

• Qualification barriers measured in years, and switching costs measured in billions with customer yield at risk, which is an intolerable risk amidst a semiconductor supercycle and with AI chips costing much more than anything foundries had ever built before, making yield even more critical.

• Contract books with customer prepayments, a very comfortable financing structure.

You can think of the wafer industry as a small toll booth on a five trillion dollar road.

Why investors do not know it - Too small for global investors and too specialised for generalists:

Because the segment is structurally invisible. It sits two layers below anything with a brand: the consumer knows the phone, the investor knows the chip designer, almost nobody knows the wafer.

The five producers combined are worth a small fraction of a semicaps, let alone AI chipmakers like Nvidia, so the segment is too small for mega-cap mandates and too specialised for generalists.

Siltronic specifically is a German mid-cap with a confidential customer list and coverage from perhaps a dozen desks. And the word wafer sounds like a commodity, which repels quality investors who never look long enough to find the qualification moats and the contract structure underneath.

European Hidden Gem Stocks put it best: Investors look at semiconductors through the wrong end of the telescope, obsessing over the design layer where the margins are loudest and ignoring the material layer where the scarcity is. The corollary is opportunity. A €2.9bn company does not need to take share to matter in a market heading past $20bn.

Why wafers were late to the AI trade:

Because wafer makers sit last in the chain and feel everything with a lag, both on the way up and on the way down. During the 2021 to 2022 shortage, chipmakers double-ordered and wafer producers signed record long-term agreements at strong prices. When end demand cracked, customers spent 2023 through 2025 burning through wafer inventories rather than placing orders, so wafer volumes kept falling for three years while AI chip revenues exploded a few steps downstream.

AI itself was too small in wafer terms to move the needle at first: roughly 4% of 300mm volumes in 2023, because a few hundred thousand accelerator wafers did not move the needle inside a market shipping millions of wafers a month. The contract structure compounds the lag, since agreements that protect revenue on the way down also mute the price signal on the way up.

The result was the strangest divergence in the sector: Siltronic’s sales fell for three consecutive years while its end market entered the largest capacity build-out in semiconductor history. Being the lagging indicator is painful on the way down. It is the entire point on the way up, because the same lag now works in reverse: inventories are normalised, safety stocks are being rebuilt, AI reaches roughly 12% of 300mm volumes this year and approaches 18% next, and the orders that skipped the wafer layer for three years are arriving all at once.

AI reaches the wafer layer last. Share of 300mm wafer volumes attributable to AI applications. Source: Crack The Market, industry data.

Why the catch-up is just beginning:

The stock has already moved up nearly 100% in the first half of 2026 and, up nearly 100% in the first half of 2026and roughly 145% over a year, and it is fair to ask what is left.

Our answer: The re-rating so far has repaired the discount to survival, not priced the earnings. What drove the move was recognition of the volume inflection, the capital increase that removed the balance sheet question, and the MDAX return that brought index flows back. What has not yet happened is everything that moves earnings and cash:

• Pricing: Spot prices remain below contract levels and below reinvestment economics, and the contract book has not begun to reprice.

• Reported earnings: Consensus EPS stays negative through 2028, so every earnings screen and quant model remains structurally unable to own the name until 2029.

• Free cash flow: It turns positive only next year, as the capex cliff meets a rebuilding EBITDA line.

In each prior tightness, the wafer names delivered multi-year price appreciation, and the moves that mattered came after operating profit turned, not before. The first leg of this trade repaired the discount to the balance sheet. The second leg, the one this piece is written for, is the earnings arriving, which everyone so far has been ignoring.

Investment Case

The investment case rests on a single piece of mispricing: the market is valuing Siltronic off an income statement that is mechanically distorted, and the distortion is a timing problem, not a business problem.

The company will book €490m to 520m of depreciation in 2026 against guided EBITDA of roughly €260m to 320m. Reported operating losses this year and next are therefore locked in whatever demand does. But the losses are being manufactured by the accounting on a fab that is already built, already largely paid for, already running at high utilisation, and already serving leading-edge customers with the richest epitaxial mix in the company’s history. Depreciation is the echo of capex that has ended: capex itself is guided at €180m to 220m this year against half a billion of D&A and should remain at this level going forward.

Depreciation arrives as capex leaves. D&A overtakes EBITDA through the trough while capex falls to 40 cents per euro of depreciation. Source: Crack The Market, company guidance, consensus.

The cash economics and the earnings statement have decoupled, and the market is reading the wrong one in our view. On top of that, in the last twelve months, three things changed at the wafer layer:

• One: the volumes turned. After four years of downturn, wafer demand is inflecting. 300mm capacity is running at high utilisation across the group, with the legacy fabs in the low nineties, and 200mm demand is recovering in the second half after a long slide. Customers are requesting incremental volumes, opening discussions on new long-term agreements, and rebuilding safety stocks, particularly in memory. GlobalWafers just signed a ten-year supply agreement with Micron, and Shin-Etsu reported record 300mm shipments in June. The turn is not a forecast, it is in the shipment data.

• Two: pricing has a mechanism behind it. For the first time in three years, spot wafer prices are rising. They still sit below long-term agreement levels and, critically, below the level that justifies investment in new capacity. Management says this plainly and refuses to sign new agreements or commit further Singapore expansion at current terms. The peers behave identically, showing that the oligopoly is working exactly as it should: no one adds supply below reinvestment returns, wafer area demand grows around 7% this year driven by AI, and the clearing price has to rise. Some investors are starting to model an outright wafer shortage by late 2027 with repricing of more than 30% on new and renewed agreements. We do not need that to be right. But Siltronic, with roughly 67% of volumes under contract against closer to 80% at Shin-Etsu and SUMCO, has the most torque to a spot recovery of any of the majors.

• Three: the operating leverage is already visible. Gross margin improved from minus 8.5% to minus 4.5% in a single quarter on volume absorption alone. FabNext carries a targeted mid-term EBITDA margin above 50%, built on automation, a high epitaxial share and scale economics, and management expects the site to reach group margin levels around year-end. Every incremental wafer from here drops through at a rate that the market has forgotten after a long four year downcycle.

We are paying 7x consensus 2028 EBITDA for the cheapest pure play in a five-company oligopoly, with the balance sheet repaired in June and the contract book yet to reprice. The downside is an asset base at book value. The upside is the cycle plus a repricing that the market does not expect. That is the definition of asymmetric to us.

We want to be straight that this is not a hidden gem at the bottom of the cycle. The stock started the year at €51, touched €104.70 in May, absorbed a €273m equity placement at €91 in June, dipped to €68 after the Q2 print and trades at €87.7 as we write. The easy trade, buying at €32 when the market feared the balance sheet, is gone. What is left is a different and to us more interesting question: with the balance sheet fixed, volumes inflecting and pricing starting to move, is the 2028 earnings power in the price? The most cautious investors are saying yes to this question right now. The most constructive say the street is still underestimating 2028. That disagreement is the whole stock debate, and the rest of this piece is our position in it.

The market rediscovered the story, then handed part of it back. Share price milestones through 2026. Source: Crack The Market, company reports.

Value Chain

Nobody asks where chips are actually made, the same way nobody wants to know how sausages are made. The market is captivated by the companies designing accelerators and the foundries printing them, and it yawns at the picks-and-shovels supplier at the very top of the chain. Yet every advanced processor on the planet begins life as a 300mm polished disc from one of five companies. Roughly $1.4bn of electronic-grade silicon becomes $11.4bn of wafers, becomes >$1tn of semiconductor devices, becomes more than $5tn of electronics. Siltronic sells the blank canvas to the artists.

We ask one question of every holding: if a competitor showed up tomorrow with a blank cheque, could they take the business? European Hidden Gem Stocks answered it in nine words: only five companies in the world can do this. There are a few reasons for this:

• Entry is effectively closed: The industry consolidated from more than twenty suppliers in the 1990s to five producers with 90% of the market. The irony is that silicon is one of the most abundant elements on earth, and anyone could start a wafer business tomorrow. What they could not do is produce defect-free wafers at atomic-level precision, at scale, qualified into the most unforgiving manufacturing processes on the planet. That capability rests on decades of accumulated flowsheet knowledge, roughly 2,000 patents and patent applications, and crystal-growing know-how that lives in the operators rather than in any textbook. The practical answer to a blank cheque is a decade of losses with no guarantee of a single major customer.

  

• Switching costs are real and asymmetric: A chipmaker that switches wafer supplier on a running process must requalify, re-optimise yield, and accept risk on a production line where a single bad batch costs more than the wafers ever will. Customers do not take that risk to save single-digit percentages. This is why market shares are stable and why two thirds of Siltronic’s 300mm book sits in multi-year agreements with prepayments.

• But the structure did not protect prices in the downturn: Wafer prices fell for three consecutive years and remain below levels that justify new capacity. An oligopoly with switching costs that cannot hold price into oversupply has a barrier to entry but no immunity to the cycle. What the structure does deliver is discipline on the way down, no one chased share with price and shares did not move, and speed on the way up: in the 2016 to 2019 tightness, average selling prices compounded at double digits for three years, and the 2020 to 2022 upcycle added mid single digits a year on top. Concentration means that when the market tightens, repricing is fast and general, which is what we are buying with Siltronic.

• On substitution we are relaxed: Silicon has had sixty years of manufacturing optimisation behind it. Silicon carbide serves electric vehicle power electronics, gallium nitride serves RF and fast charging, gallium arsenide earns its keep in satellites where cost is no object, diamond is theoretically wonderful and commercially absurd, and graphene, for all its electrical properties, does not naturally switch off, which is a problem for a switch. None of them replaces silicon for logic or memory, and the foundries are not changing the substrate under the largest capacity race in the industry’s history.

• The European angle is particularly compelling: Siltronic is the only major wafer producer headquartered in the West, which makes it structurally relevant to the European Chips Act and to any serious attempt at supply resilience. It is also part of why GlobalWafers was willing to pay €145 per share for the company in 2022. But we have to remember that the incremental capacity is in Singapore, not Germany/Europe, so sovereignty is an option embedded in the shareholder and the asset base, not a near-term earnings driver.

Growth Drivers

If you screen Siltronic for growth, it looks broken, with sales down for two years and guidance pointing slightly lower again in 2026. That is a trap in our view. The growth is in wafer area, price and drop-through, and all three are turning at once.

Driver one - AI turns wafers back into a volume story:

For two decades wafer demand compounded with general-purpose computing and smartphones. The accelerated computing build-out changes the slope. AI’s share of 300mm wafer volumes was around 4% in 2023, is estimated at roughly 12% this year, and approaches 18% in 2027, more than offsetting soft smartphone and PC demand. Server-related silicon, already 18% of Siltronic’s volumes. is growing around 46% in 2026, and total wafer area consumption around 7% before inventory effects. High-bandwidth memory is a silent multiplier: it consumes disproportionate wafer area per bit, which is why memory makers are allocating constrained capacity to AI and starving consumer end markets. An important point is that the weakness in phones and PCs is a symptom of tightness in memory, not of weak demand, which I covered in my Memory Singularity series. Volumes lead prices in this industry, and volumes are already moving.

Driver two - The pricing mechanism, the core of the thesis:

Wafer prices have declined for three consecutive years and sit below two thresholds that matter. They are below existing long-term agreement levels, and below the level at which any producer earns an adequate return on new capacity. Management states publicly that current prices do not justify further Singapore expansion and that it remains selective on new agreements pending appropriate commercial terms. The peers behave identically, so supply additions have effectively stopped while demand grows and mix shifts toward leading edge. We want to emphasise that this is how every wafer upcycle has started.

Importantly, contract structure sets the timing. Roughly two thirds of Siltronic’s volumes are locked in agreements struck at yesterday’s prices, so the near-term benefit of rising spot is modest. But agreements roll, and renewals over the next two years will be negotiated into a tightening market by producers who need materially higher prices to sanction the next expansion. Some investors are looking for renewal repricing of more than 30% and an outright shortage by late 2027. History supports the shape if not the exact number, and Siltronic’s lower contract coverage means it captures spot improvement first. We are not underwriting a price spike but we are underwriting the end of three years of declines and a positive pricing contribution from 2027, which on its own transforms the earnings of a business with this fixed cost base.

Driver three - The Singapore drop-through:

FabNext is the largest investment in the company’s history and it is doing in 2026 exactly what a ramping fab does: depressing every reported margin while building the earnings power. The site is highly automated, weighted to epitaxial and leading-edge product, and targeted at a mid-term EBITDA margin above 50% against a group margin of 21% today. The absorption math is already visible: group gross margin improved four full points in a single quarter on higher volumes alone, and management expects the site to approach group margin level around year-end. The fixed costs are sunk, the depreciation is running at full rate, and material brownfield space remains for the next phase once pricing justifies it. When a business guides EBITDA margins of 20 to 24% at the bottom of its cycle with a half-ramped flagship fab, the question is what the same cost base prints at full utilisation and recovered pricing. Consensus answers 33% by 2028, and the constructive desks sit above that.

Risks & Threats

For Siltronic there are four ways to break the thesis, two cyclical and two structural to the equity story. We wrote down a signpost for each so we know when the thesis is broken versus when we are just early.

1. The recovery stalls (cyclical, the big one).

• The thesis needs wafer demand to keep compounding into 2027. If AI infrastructure spending digests, or the memory allocation squeeze resolves into genuine end-demand weakness, utilisation flattens and the pricing mechanism never engages.

• In that world Siltronic is a loss-making cyclical with €500m of annual depreciation for two more years.

• How we’ll know: 300mm utilisation commentary softening, the safety stock rebuild pausing, any walk-back of the volume language at Q3. The mitigant is the contract book, two thirds of volumes with prepayments, a floor under revenue that previous downcycles never had at this coverage.

2. Pricing never clears the reinvestment threshold (cyclical).

• Volumes can recover while prices stay pinned: spot improves on small volumes but stalls below agreement levels, and renewals reset flat or lower as contracts struck in the 2021-22 tightness roll off. This is the quiet version of the bear case according to cautious investors and it is a fair one.

• Chinese 300mm entrants building for domestic fabs could cap the commodity end of the recovery even if they cannot qualify at leading edge.

• How we’ll know: renewal commentary at the full-year results and the first 2027 pricing indications. If renewals reprice down into a tightening market, the mechanism at the centre of this piece is wrong and we will say so.

3. Dilution round two (structural).

• The June raise took the imminent balance sheet question off the table, but a full build-out of the remaining Singapore brownfield is estimated at a further €800m to 1bn. If demand accelerates, management will want to build.

• Funding expansion into an upcycle with more equity is exactly how cyclical shareholders get hurt at the top. The discipline test is simple: no major expansion without repriced agreements and customer prepayments to match.

• How we’ll know: a capex guide moving materially above €250m, or another placement, without contract economics disclosed alongside. Either one and we trim.

4. The dollar (structural).

• Wafers are priced in dollars and the cost base is heavily euro and Singapore dollar. Guidance assumes 1.18 for the second half, and the 2025 results showed what an adverse move does to the margin bridge.

• How we’ll know: sustained euro strength above 1.25 takes a visible bite out of the 2027 margin path, and our base case comes down with it.

Another important watch item is the shareholder base. Wacker placed 7% into strength in May and retains 21.7% as a financial holding, and Sino-American Silicon still owns 12.4% with no board seat. More than a third of the company can be put for sale at some point, and blocks will occasionally hit the price. But the same shareholder base that creates risk also creates strategic optionality, as GlobalWafers’ 2022 bid at €145 demonstrated. Volatility around these sales would be an opportunity.

The risks we are not worried about: substrate substitution, for the reasons in the value chain section. Customer concentration, diversified across logic, memory and analog. And the balance sheet itself post-raise, which we deal with next. If you can live with the four risks above, you get paid to wait. If you can’t, this isn’t your name.

Financial Robustness & Capital Allocation

The table below tells you the whole story: a €1.8bn peak in 2022, three years of decline into the 2025-26 trough, a capex programme that peaked at €1.3bn in 2023 and collapses to €200m now, depreciation that arrives exactly as capex leaves, and a net debt line that round-trips from €573m of net cash to €837m of debt and back down. Reported earnings stay negative through 2028 on consensus. This is one of the most important facts of this investment case: the equity works through EBITDA, cash flow and deleveraging years before the net income line catches up.

2026e to 2028e per consensus, net cash flow and net debt per CTM estimates. Historical figures from company reports.

Two peaks, two troughs, one turn. Sales and EBITDA margin through the cycle. Source: Crack The Market, company reports, consensus.

The balance sheet was repaired in June. Siltronic was never a solvency story, but by early 2026 it was becoming a flexibility story: net debt of €836.5m at year-end against €316.9m of EBITDA, with the trough still running.

• The capital increase: 3m new shares, 10% of the capital, placed at €91 for gross proceeds of €273m, with HAL Trust participating meaningfully. Management had said in March that no raise was planned, so the market rightly read the move as opportunistic. We read it as opportunistic and protective at once, and we approve: taking 10% dilution within sight of the all-time high is the cheapest insurance this company has bought in years.

  

• The result: Net debt of €691.9m at June, equity ratio 45.4%, €650m of cash and securities plus an undrawn €127m facility, a balanced maturity profile. Leverage of 2.4x trailing EBITDA is elevated for a cyclical but will fall thanks to a rising denominator.

The balance sheet round trip. Net cash into the build, deleveraging out of it. Source: Crack The Market, company reports, CTM estimates.

Cash flow and the capex cliff which is the underappreciated half of the story.

• Capex is guided at €180m to 220m for 2026, down from €369m in 2025 and a €1.3bn peak during the build. Depreciation runs at €490m to 520m. When a capital-intensive business spends 40 cents of capex per euro of depreciation, it is now in harvesting mode.

• H1 net cash flow was minus €116.5m, but the quarterly shape tells you that the direction is positive: minus €89.5m in Q1 narrowing to minus €27m in Q2, with operating cash flow already positive.

• Customer prepayments under the long-term agreements provided roughly €400m of funding through the build years, a structural advantage of the contract model.

If you hold capex near these levels and let EBITDA rebuild toward consensus, free cash flow inflects in 2027 the way volumes inflected in 2026. No dividend was paid for 2025 and none should be expected for 2026. With the income statement in engineered losses and the brownfield option live, cash belongs on the balance sheet.

Outlook & Guidance

The 2026 guidance, refined slightly upward at the Q2 results, is the picture of a trough year with the turn visible inside it:

• Sales in the low to mid single digit percent range below 2025 on a 1.18 dollar assumption, flat to slightly up excluding currency and the discontinued small-diameter business.

• EBITDA margin unchanged at 20 to 24%, the better volume picture absorbed by higher freight and energy costs.

• Depreciation raised to €490m to 520m, EBIT down significantly, capex cut to €180m to 220m, net cash flow around last year’s level of minus €85m.

The commentary is very interesting: Management flags 300mm capacity at high utilisation throughout the year at both the existing sites and Singapore, a 200mm recovery beginning in the second half, growing customer interest in securing medium and long-term volumes, and, in the sentence we would frame, spot prices improving selectively but not yet at levels that justify further capacity investment, which is why the company remains selective on new agreements. That is a supplier deliberately holding capacity back from a tightening market until the price is right and it is exactly what you want to read at this point in the cycle.

Beyond 2026, consensus now models EBITDA of €300m this year, €418m in 2027 and €559m in 2028, a near doubling in two years on volumes and mix alone. Cautious investors sit below that path and call the recovery priced. Constructive investors model repriced contracts and sit above €700m for 2028. Our base case, developed in the valuation section, takes consensus at its word, because the volume evidence is already here and the pricing mechanism has history behind it.

Governance & Management Quality

The executive team is new to the cycle but not to the company or the industry: CEO Michael Heckmeier took over in May 2023 and has spent his tenure ramping Singapore through the worst wafer downturn in a decade, alongside CFO Claudia Schmitt and COO Klaus Buchwald.

We focus on their track record so far: The 2025 guidance was met within every range in a brutal environment, the small-diameter closure was executed on schedule, the capex discipline is real and continuing, and the refusal to sign agreements at inadequate prices is the behaviour of a management team that understands its own cycle. The equity raise was timed into strength rather than necessity, which is when you want your management raising.

The Supervisory Board is exceptionally strong for a midcap company and further strengthens our thesis:

• Twelve members under German co-determination, chaired since 2013 by Dr. Tobias Ohler, the CFO of Wacker Chemie, who brings the former parent’s owner’s eye to capital allocation and recuses himself from anything touching the Wacker supply relationship.

• Beside him sits someone you would not expect to find at Siltronic but in our view gives the company a very powerful edge: Dr. Jos Benschop, Executive Vice President Technology at ASML, meaning the technology chief of the company whose lithography machines define what a leading-edge wafer must deliver, is reviewing Siltronic’s roadmap from inside the boardroom.

• The Audit Committee is chaired by Mariella Röhm-Kottmann, a former KPMG partner with a career auditing DAX and MDAX balance sheets and a sitting CFO herself, exactly the profile you want watching a company carrying covenants through a trough, and the board spent 2025 doing precisely that, with financial stability and covenant compliance a standing agenda item ahead of the June raise.

• Four of the six shareholder representatives are independent. This board is perfectly aligned with the questions that define our thesis: technology relevance at the leading edge, and discipline on the balance sheet.

  

The shareholder base tells its own story: Wacker holds 21.7% as an explicitly financial stake and demonstrated as much by selling 7% into the May rally, and its CFO chairs the supervisory board, which keeps an owner’s eye on capital allocation. HAL Trust added meaningfully in the June placement at €91. And Sino-American Silicon, GlobalWafers’ parent, still holds 12.4%, the residue of a €145 per share offer in 2022 that failed on regulatory timing rather than price. Sophisticated capital sits on this register at and above today’s levels, and a strategic buyer has already shown its hand once.

Incentives are sensibly built for a capital-intensive cyclical: long-term compensation keys to plan net cash flow and to plan EBIT, which charges management for depreciation and therefore for the returns on the capital they deploy. Given that the central risk to this thesis is a management team tempted to build into the upcycle, we take comfort that the incentive plan penalises exactly that.

Peer Comparison - The Cheapest Ticket to the Same Show

Five companies make the world’s wafers, but the investable set is smaller than it look:

• Shin-Etsu, the market leader, is a diversified chemicals group where wafers sit alongside PVC and silicones, so its group multiple tells you little about wafer value.

• SK Siltron is private, owned by the Doosan Group after its acquisition from SK, and cannot be bought at all.

• That leaves three clean listed pure plays: SUMCO, GlobalWafers and Siltronic.

The positioning differences are real:

• GlobalWafers is the industry consolidator, building greenfield capacity in the US with a fresh ten-year Micron agreement behind it, and the market pays a full growth premium for it.

• SUMCO is the sector reference: stable, the highest contract coverage in the group, and, tellingly, also loss-making at the net level in 2026 on its own depreciation cycle.

• Siltronic is the smallest of the three, with the newest asset base in the industry, the lowest contract coverage and therefore the most torque to a spot recovery, and the only Western headquarters.

• Three ways into the same inflection, at three very different prices.

Consensus estimates at current prices, 15 August 2026. Shin-Etsu figures are group-level including non-wafer businesses. P/E for Shin-Etsu is NTM, others 2028e. Siltronic P/E is not computable due to negative consensus EPS through 2028. Source: Crack The Market, consensus data.

How do we read the numbers:

• On next-twelve-month EBITDA, Siltronic at 11x screens more expensive than SUMCO at 9.5x, because a half-ramped flagship fab is sitting on its denominator.

• Move to 2028, the year the ramp and the pricing land, and it becomes the cheapest wafer asset money can buy: 7x against 8.5x at SUMCO, 10.3x at Shin-Etsu and 10.7x at GlobalWafers, while compounding revenue at 14% a year against 10% at SUMCO and 8% at Shin-Etsu, because its growth capacity is already built and paid for.

• Only GlobalWafers grows faster, at 27%, and the market charges 5x book and 20x near-term EBITDA for the privilege.

• On book value the gap is starkest: 1.6x against 2.5x, 2.6x and 5.0x, and Siltronic’s book is the newest in the industry, with FabNext carried at close to replacement cost.

Part of the discount is deserved of course given Siltronic carries more leverage than its peers, its returns on capital lag the group through the ramp, and it is the last of the five to return to positive earnings per share, in 2029 against 2027 for SUMCO.

But a valuation gap that exists because of depreciation timing and balance sheet history, rather than because of the asset or the demand, is precisely the kind we are paid to arbitrage.

If you want the wafer inflection: Shin-Etsu dilutes it, SK Siltron is unbuyable, GlobalWafers charges you three times book for it, and SUMCO gives you less growth for a higher multiple.

Siltronic is the cheapest play for this wafer upcycle, with the most operating leverage in the room.

Valuation

We can start by throwing away the P/E: Consensus EPS is minus €6.4 this year, minus €4.0 next and minus €0.8 in 2028, turning positive only in 2029, so any earnings multiple on this name is measuring the accounting rather than the business.

The screens agree, which is precisely the problem and precisely the opportunity: the name is invisible to earnings-based mandates through the whole recovery.

We value it on EV/EBITDA through the cycle, cross-checked against the asset base.

At €87.7 Siltronic’s equity is worth €2.9bn and the enterprise value roughly €3.6bn. That is 12x this year’s guided EBITDA midpoint and 11x the next twelve months, at or above the top of the historical range, which is the entire bear case: a cycle-peak multiple on trough earnings means the recovery is priced.

On consensus 2028 EBITDA of €559m the same enterprise value is 7x, and on the repriced numbers of constructive investors, it is below 6x. The stock is expensive this year and cheap on the recovery, which is what every cyclical looks like when they turn. The question is never the current-year multiple, it is whether the out-year earnings are real.

Two cross-checks anchor the downside:

• Enterprise value today sits within a few percent of the book value of property, plant and equipment at €3.4bn, meaning the market is paying approximately the depreciated cost of the asset base and receiving the oligopoly position, the contract book, the prepayments and the customer qualifications for nothing. Replacement cost sits far above book: FabNext alone absorbed over €2bn, and a greenfield entrant would need that plus a decade of qualification losses.

• GlobalWafers bid €145 per share in 2022 for a company that had not yet completed Singapore. Cycles differ and we do not lean on the bid as a target, but it tells you what this asset base is worth to the one class of buyer that thinks in decades.

We lay the scenarios out explicitly, anchored on 2028 over a 24 to 30 month horizon:

Importantly, what the bears require is not disappointment, but the cycle re-breaking, with pricing pinned below reinvestment levels for three further years while AI-driven wafer area compounds. Note also what the bounce has done to the entry: at €87.7 the spread is more balanced than it was three weeks ago (around €68), 31% to the base case and 65% to the bull against 43% to the bear. The asymmetry we want reappears on weakness.

At €70 to 75, where this stock traded at the end of July, the base case pays more than 50%. We would be buyers of dips in this name, not chasers of strength, and the volatility that the shareholder base and the quarterly earnings reliably supply is a tool for building the position.

We are not calling the recent bounce the start of the next leg. A soft quarter, a dollar move or a block sale from shareholders can put this stock back at €70, and at those levels, the setup would look even more asymmetric.

The target is €115 over 24 to 30 months, with the scenario table as the map and the signposts below as the trigger list. The June placement at €91, taken up meaningfully by HAL Trust, is a useful reference for where informed capital cleared size two months ago.

This is a cyclical value name with a structural AI kicker, not a compounder. It belongs in a barbell: own your secular growers elsewhere, and own this at book value with the cycle turning underneath it. If you can’t stomach a 30% drawdown when a quarter disappoints, this isn’t a name for you.

Conclusion - Buy the Inflection the Screens Cannot See

Siltronic spent four years in the wafer downturn and its income statement will spend two more there, because half a billion euros of annual depreciation on a newly built fab locks in reported losses into 2028 whatever demand does. That is exactly why the opportunity exists. Underneath the accounting, the recovery is already in the shipment data: volumes are inflecting, utilisation is high and rising, customers are queuing to secure long-term supply, and spot prices are climbing for the first time in three years while every producer refuses to add capacity below reinvestment economics. The revenue line turns in 2026, EBITDA compounds from 2027, and reported earnings, the only line the screens can read, turn last.

What tells us which path we are on, in order of importance:

• Singapore utilisation first, and Singapore margin progression toward group level second.

• Group margin recovery third, and 300mm demand fourth.

• The terms on renewed long-term agreements, because renewals repricing upward is the mechanism and renewals repricing downward is the refutation.

• Hyperscaler AI spending, the demand engine two layers downstream.

• Around those: the first 2027 guidance, capex discipline with no major brownfield commitment unless matched by repriced contracts and prepayments, the dollar, and the register, where blocks are volatility to be used.

The market rediscovered this story once already this year, handed part of the move back, and has begun taking it again. The rerating so far bought survival. The next leg belongs to the earnings, and the tape will hand us the entries along the way.

Every AI chip on the planet begins life as a polished silicon wafer. The market pays fortunes for the artists and yawns at the blank canvas. We will take the canvas, at book value, with the cycle turning underneath it.

Disclaimer: The information provided on this Substack is for general informational and educational purposes only, and should not be construed as investment advice. Nothing produced here should be considered a recommendation to buy or sell any particular security.

Every dollar of AI capex terminates in a requirement to move photons. GPUs are useless without the optical links that connect them, those links need lasers, and every laser at the speeds AI requires is built on indium phosphide (InP), a material only a handful of companies in the world can process at scale. After the industry spent a decade rationalizing capacity, InP fab capacity is now the binding constraint on the entire optical buildout.

Outside China, two American companies control the manufacturing of these lasers at the leading speeds: Coherent and Lumentum. They are the real winners at the heart of this bottleneck, converting what were slow-growth component businesses into two of the most strategically important franchises in AI infrastructure, with order books stretching to 2028, long-term agreements running through 2030, and Nvidia holding $2bn of equity in each to secure its own supply.

In this write up, which follows my US Photonics Supercycle piece, I dive into:

• Why the laser is the binding constraint of the optical supercycle and why the capacity cannot be added quickly

• Why the duopoly is more durable than the market appreciates

• What the recent earnings told us about demand, margins and visibility

• Why both companies are exceptional investments at different points on the risk/reward spectrum, with full profiles of each name and my verdict on which deserves the incremental dollar.

Just some numbers to give you a sense of what I am talking about:

• The datacom optical market is growing from roughly $20bn in 2025 to more than $70bn by 2030, a 28% CAGR, with 1.6T transceivers compounding at roughly 120% annually toward $40bn and each speed generation doubling prices at similar cost.

• Coherent’s book-to-bill ran above 4x last quarter. Purchase orders extend into 2028, long-term agreements run through 2030, and management sees the InP industry supply-constrained for a minimum of 12-24 months, possibly years once co-packaged optics demand arrives on top.

• Capacity is being added and capacity is revenue in this market. Coherent is doubling its InP capacity by the end of the September quarter, a quarter ahead of its own accelerated plan, and doubling it again by the end of 2027. Its $3bn quarterly revenue target for fiscal Q4 2027 implies a $12bn run-rate against the $7.1bn just reported for fiscal 2026.

• Lumentum grew revenue 109% year over year, printed a 50.4% gross margin, and guided operating margins to 40% with the long-term frame raised to 40-44%. Its pump lasers are sold out for the foreseeable future despite capacity expanding 40% in three quarters.

• The magnitude of the demand is truly mind boggling and explains why both names are so strategically important today. One hyperscaler’s network capacity connecting just two AI data center sites is double the total global backbone capacity that customer built over the entirety of the last decade.

• Nvidia invested $2bn of equity in each company in March 2026, attached multi-year purchase commitments in the multi-billion dollar range, and has locked up a considerable share of both companies’ InP capacity, with options to take more.

Twice a week, I will release deep dives into stocks and sectors that fit into the 6 themes that I see winning in the coming years and decades: AI & Technology, Electrification & Energy, Reshoring & Sovereignty, Healthcare & Longevity, China & Asia, Resilience & Quality.

Join Crack The Market

Take advantage of this once in a generation opportunity to build long term wealth by investing in great stocks that will deliver returns for your portfolio for years to come.

Table of content

• Why we are suddenly shipping so much optics

• Why the laser is the binding constraint

• The duopoly: two companies, two decades of process learning

• The supply-demand setup: capacity sold through 2028

• The margin story: two financial models on one bottleneck

• The new layers: OCS and CPO

• What the August earnings told us

• Coherent (market cap $67bn) - The complete stack and the capacity engine

• Lumentum (market cap $83bn) - The pure play on the scarcest layer

• Two interconnected ways to play the laser bottleneck

• Risks worth flagging

• Conclusion: the pair that rations the decade

Why we are suddenly shipping so much optics

For most of the last decade, datacom optics was a good business, not a great one:

• Transceivers, the small modules that convert electrical signals into light and back at each end of a fiber, were a competitive, deflationary market where a speed generation lasted three to five years and Chinese assemblers competed away the margin.

• That world ended when AI clusters changed the shape of network demand. I laid out the full market framework in the Optical & Networking Supercycle in January and mapped the 19 US-listed names in The US Photonics Supercycle a few days ago, which I highly recommend that you read, so here I will keep to the essentials and what is new.