The Circuit

Fifty disclosed payments, from an enterprise cloud contract to a household electricity bill and back again. Each figure sorted into what has actually been paid, what is contractually owed, and what has only been announced - because the difference between those three is the whole story.

Fifty payments, in order. The first is an enterprise signing a cloud contract. The fiftieth is a chipmaker investing in the customer that buys its chips. In between, the money passes through fifty companies on four continents, melts rock, buys uranium, pours concrete, raises a household electricity bill in Ohio, and is recognised as revenue perhaps a dozen times on the way.

The fiftieth payment lands on the first. That is why this is a circuit and not a chain.

That circuit is the subject of this piece. Not the companies individually (this paper has written about several of them one at a time), but the circuit: the order of the payments, the documented reason each payment exists, and what each payment was supposed to change about the payer's position.

Because that is the thing that is almost never set out. The AI build-out is reported as a series of announcements. It is not a series of announcements. It is a closed loop of obligations in which each participant's spending is another participant's revenue, and in which the total amount of revenue recognised around the loop is far larger than the amount of new money entering it. Understanding that is not a matter of opinion. It is a matter of reading the payments in order.

Figure 1 · The circuit

Fifty payments, ten stations, one closed loop

Stations, not companies: each box holds several of the fifty payments described below. The arrows are the direction of cash, not of goods, goods travel the other way.

The rules this piece follows, stated before it starts

One. Nobody can literally follow a single dollar. Money is fungible; a dollar paid into Microsoft is not the dollar Microsoft pays out. So this is a representative dollar, and the chain is what matters: every payment described here is a real, disclosed transaction between two named parties. The dollar is the device. The payments are the fact.

Two. The payments are numbered, not dated. The numbering is sequence - who pays whom, and in what order the money moves - and nothing else. Some of these transactions run on quarterly cycles, some on multi-year prepayments, some on twenty-year contracts. Where the real timing matters, and in the power and grid sections it matters enormously, the real timing is given explicitly. No number in this piece should be read as a date.

Three. Every figure carries its class, in the sentence. This is the rule that governs the rest, so it is stated with the tags used throughout:

booked: money that has been paid, received, or recognised as revenue in a filed account. It exists.
contracted, a binding obligation disclosed in an executed agreement. Owed, scheduled, enforceable; not yet paid.
announced, a stated intention. A framework, a letter of intent, an "up to" figure, a multi-year programme with no enforceable schedule. It may become contracted. It is not money.
estimate: a figure nobody inside the transaction published, compiled by an outside party. Useful, and not a fact.

These four are different kinds of object and this piece does not add them together. The most common error in coverage of this subject is to let an announced figure share a sentence with a booked one, and the error is never neutral: it always inflates. A reader who sees no tag on a number in this article should treat that as the article's failure, not as the number's promotion.

Four. Where nobody publishes the decisive number, this piece says nobody publishes it, instead of repeating a figure that originated in somebody's sales deck.

ACT ONE · The money enters

Payment 1 · An enterprise buys capacity it cannot build

The payment. A bank, an insurer, a retailer, a pharmaceutical company (the identity does not matter, there are tens of thousands of them), signs a multi-year commitment for AI capacity on Microsoft Azure. Individually these are invisible. Collectively they are the only reason the rest of this chain exists.

Why this expense exists. Because the alternative is building it. To run a frontier model in-house a company would need the accelerators, the power contract, the cooling, the network, the security accreditation and the people, and it would need them before knowing whether the application works. Renting converts an unbounded capital project into an operating line item that can be cancelled. That is the entire commercial proposition of cloud computing and it has not changed in twenty years; only the unit price has.

How it moves the payer's trajectory. It buys optionality. The enterprise is not buying compute, it is buying the right to find out whether AI changes its business before its competitors do, without betting the balance sheet on the answer. That is why these commitments are signed even by firms with no finished use case: the cost of being a year late is judged higher than the cost of a year of unused capacity.

This is the only hop in the entire chain where money enters the circuit from genuine end-demand: from somebody buying a thing in order to use it, rather than in order to sell it onward. Hold onto that. By Payment 50 it will be the whole point.

Payment 2 · Microsoft converts that promise into buildings and silicon

The payment. Microsoft is spending at a rate it has never approached in its history. The company called for roughly $190bn of capital spending in 2026 announced, a figure it later indicated would come down to around $175bn announced after it extended the assumed useful life of its data centres from fifteen to twenty-five years and reclassified some leases. The quarterly shape of fiscal 2026 was $34.9bn, $37.5bn, $31.9bn and then $41bn in the fourth quarter booked, up about 70% year on year. Microsoft brought 88 data centres online in fiscal 2026 booked, thirty-one of them in the final quarter alone, across five continents.

Why this expense exists. Because capacity sold is capacity that must exist. A cloud provider's forward commitments are contractual, and a commitment it cannot serve is a customer it hands to a competitor. Note the detail in the second-quarter disclosure: roughly two-thirds of the spending went to short-lived assets, principally GPUs and CPUs. That is not property. That is inventory with a depreciation schedule, and it is the majority of the bill.

How it moves the payer's trajectory. Microsoft's objective is not to sell GPUs by the hour. It is to be the default substrate on which enterprise AI is written, so that the models, the data, the tooling and the identity system are all in one place and moving away becomes a multi-year project. Every data centre is a brick in a switching cost. The useful-life extension is worth noticing too: lengthening depreciation from fifteen to twenty-five years spreads the cost of today's build across a longer stretch of future earnings, which is a judgement about how long this generation of buildings stays productive.

Payment 3 · Microsoft pays landlords for space it did not build

The payment. In the first quarter of calendar 2026 Microsoft spent $11.1bn on data centre leases alone booked. Not construction, leases.

Why this expense exists. Because building takes longer than demand will wait. Leasing from a specialist developer buys delivery dates that Microsoft's own construction pipeline cannot hit, in markets where Microsoft does not have land, power or permits. It is a speed purchase.

How it moves the payer's trajectory. It converts a construction constraint into a procurement one, which is the trade every hyperscaler is making right now. It also quietly moves risk off Microsoft's books and onto the developer's: and, as Act Seven shows, onto the developer's lenders. The entire data-centre leasing industry, and the private credit industry that stands behind it, exists because of the gap between how fast capacity is sold and how fast concrete cures.

Payment 4 · Microsoft pays Nvidia

The payment. The accelerators. Nvidia's fiscal 2026 revenue was $215.9bn, up 65% booked, with fourth-quarter revenue of $68.1bn and data centre revenue of $62.3bn in that quarter alone. The following two quarters went further: $81.6bn in the first quarter of fiscal 2027 and $96.2bn in the second, with data centre revenue of $89.0bn booked, up 117% year on year. A small number of very large customers account for most of it.

Why this expense exists. Because the performance per watt per dollar of a rack of current-generation Nvidia systems, taken together with the software that runs on it without modification, has no equivalent that can be bought in volume today. The competitive alternatives exist and are real, but they require porting work, and porting work costs model-training months.

How it moves the payer's trajectory. This is the purchase that converts Microsoft's capital into sellable capacity, which is the only thing that converts back into Payment 1. It also deepens a dependency Microsoft is openly trying to reduce, which is why the same company funds its own silicon programmes in parallel. Both expenses are rational at once: buy the thing that works now, fund the thing that might reduce the bill later.

Payment 5 · Nvidia pays the memory makers, at prices it does not control

The payment. High-bandwidth memory from SK hynix, Micron and Samsung. And in 2026, this is where the chain's pricing power visibly inverted. Samsung's average memory selling prices rose roughly 146% against the 2025 average in the first quarter of 2026; SK hynix's DRAM average selling price rose in the mid-60% range. Conventional DRAM contract prices rose 90-95% quarter on quarter, with DDR5 contract pricing passing $19.50 against roughly $7 a year earlier. SK hynix's HBM, DRAM and NAND capacity was described as essentially sold out for 2026; Micron exited consumer memory altogether to serve data centre customers; Samsung went into third-quarter negotiations seeking a further 20% and SK hynix reportedly removed price ceilings from long-term supply agreements.

Why this expense exists. Because an accelerator without memory bandwidth is a stalled engine. The computation is not the constraint; feeding it is. And because the memory makers reallocated capacity toward HBM, where margins are higher, the supply of ordinary memory tightened and the price of ordinary memory rose for everybody, which is the mechanism by which this chain reached into the price of consumer laptops and phones in 2026. It is also, directly, why Microsoft's 2026 capital guidance went up: the company named soaring memory prices as the reason.

How it moves the payer's trajectory. For Nvidia, securing memory allocation years forward is as strategically important as securing wafer allocation, and it is bought the same way: with volume commitments and prepayments that smaller buyers cannot match. Scale here is not a cost advantage. It is an access advantage. A competitor with a better design and no allocation ships nothing.

Payment 6 · Nvidia pays TSMC

The payment. Wafers, and increasingly the advanced packaging that sits around them. TSMC's revenue for January to August 2026 was NT$3,386.87bn, up 39.3% booked on the same period of 2025, with August alone at roughly NT$514.81bn, up 53.3% year on year.

Why this expense exists. Because Nvidia does not own a fab and has structured its entire existence around not owning one. A leading-edge fab is a $20bn-plus asset with a four-year lead time and a yield curve that punishes anyone learning on the job.

How it moves the payer's trajectory. Being fabless lets Nvidia spend on architecture and software instead of on buildings, and move to whichever node is best rather than defending the one it owns. The cost of that freedom is that its physical ceiling is set by somebody else's capacity decisions, taken in Hsinchu, two to three years before the product exists. Every wafer prepayment and capacity reservation is Nvidia buying a claim on a future it does not control.

Payment 7 · TSMC pays ASML, and Nvidia's dollar leaves Asia

The payment. Extreme ultraviolet lithography systems. Current-generation EUV scanners run to roughly $180m; the High-NA machines are priced at approximately $380-400m each estimate, print features down to about 8 nm and resolve features roughly 1.7 times smaller than the previous generation. ASML booked a record €13.2bn of orders contracted in the fourth quarter of 2025, more than double consensus, and planned to ship around ten High-NA systems in 2026.

Why this expense exists. Because there is exactly one company on Earth that sells a machine that can pattern leading-edge logic, and a foundry that does not buy it does not have a leading-edge node. This is the least negotiable purchase in the entire chain.

How it moves the payer's trajectory. A scanner bought today patterns wafers in three years. The order is therefore a bet on demand three years out, placed with money earned today, and a foundry that under-orders cannot catch up by spending more later: the queue is the constraint, not the cash. This is why foundry capital expenditure looks reckless in a downturn and insufficient in an upturn: it is always being set against a demand figure that does not exist yet.

ACT TWO · Down to the rock

The first seven payments move through companies most readers have heard of. The next ten do not, and they are where the chain stops being a market and starts being a geology lesson.

Payment 8 · ASML pays Carl Zeiss SMT, a supplier it had to buy a piece of

The payment. The optical column: the illuminator and the projection optics, a stack of multilayer mirrors figured to a tolerance that is the practical state of the art in the manufacture of physical objects. ASML does not merely buy these. In 2016 it took a 24.9% stake in Carl Zeiss SMT booked and committed to fund roughly a billion euros of research and development, specifically to get High-NA built.

Why this expense exists. Because at 13.5 nm no material refracts usefully, so the whole optical system has to be reflective, and each molybdenum-silicon multilayer mirror returns only around 70% of the light that hits it. Put ten of those in a path and most of the generated light is gone before it reaches the wafer. Every fraction of a percent of mirror reflectivity is therefore worth real money in throughput, and the figuring tolerances required are held by one supplier.

How it moves the payer's trajectory. This is the single most instructive payment in the chain, because it shows what a company does when its roadmap depends on a supplier it cannot replace: it stops being a customer and becomes a shareholder. Buying a quarter of Zeiss SMT did not give ASML control. It gave ASML a seat at the table where the optics roadmap is set, and it tied Zeiss's capital plan to ASML's product plan. When you cannot vertically integrate, you vertically commit.

Payment 9 · Zeiss pays the glassmakers

The payment. Ultra-low-expansion glass blanks (Corning's ULE and Schott's Zerodur being the reference materials), from which the mirror substrates are cut and figured.

Why this expense exists. Because a mirror in an EUV scanner cannot be allowed to change shape when it warms. The substrate has to have a coefficient of thermal expansion near zero over the operating range, which is a melt-chemistry problem solved by a very small number of glass houses, on timescales measured in months per blank.

How it moves the payer's trajectory. For Zeiss this is the bottom of its own supply pyramid, and the reason the optics schedule is not elastic. You cannot order next quarter's mirrors; you order the blanks years out, and the blank schedule sets the scanner schedule, which sets the fab schedule, which sets how many accelerators exist in 2029.

Payment 10 · ASML pays Trumpf for the laser that makes the light

The payment. The drive laser. Inside each scanner, roughly 50,000 microscopic tin droplets per second are fired into a vacuum chamber and struck by a two-pulse 30-kilowatt class CO₂ laser system built by Trumpf: the first pulse flattening the droplet, the second vaporising it into a plasma that radiates at 13.5 nm.

Why this expense exists. Because there is no lamp that emits usefully at 13.5 nm. The light has to be manufactured, droplet by droplet, hundreds of millions of times per wafer run. The industrial CO₂ laser capable of doing that reliably, for years, inside a production tool, is a Trumpf product.

How it moves the payer's trajectory. Source power is the throughput knob of the entire semiconductor industry. More watts at the plasma means more wafers per hour out of a $380m machine, which means a lower cost per wafer for the foundry, which means a lower cost per chip. ASML's spending with Trumpf is the industry buying its own productivity curve one kilowatt at a time.

Payment 11 · Trumpf pays Zeiss, and the first loop inside the loop closes

The payment. Beam transport and shaping optics for the laser system, bought from the same optics house ASML part-owns.

Why this expense exists. Because 30 kW of infrared has to be delivered to a moving target a few tens of microns across, without the optics degrading. The same metrology and figuring capability that makes the projection mirrors makes these.

How it moves the payer's trajectory. Note what has happened to the shape of the chain here. ASML pays Zeiss, ASML pays Trumpf, and Trumpf pays Zeiss. This is not a line. It is already a mesh, eleven payments in, and the same company's revenue is being fed by two different hops of the same dollar. By Act Seven this pattern will have a name and a controversy attached to it. At this level of the stack nobody calls it circular; they call it a supply chain.

Payment 12 · ASML pays its mechatronics suppliers, and the tin merchants

The payment. ASML outsources the great majority of the parts in its machines, assembling rather than manufacturing most of what it sells. Dutch and German engineering firms (VDL ETG, Neways, Prodrive and dozens of others), supply frames, stages, modules and subassemblies. And somewhere far below that, somebody is paid for refined tin for the droplet generator, from the producers in Indonesia, Peru and China who dominate the world's supply.

Why this expense exists. Because the company that designs the hardest machine in the world does not want to also run a machine shop; and because the fuel of EUV light is a soft metal mined from the ground.

How it moves the payer's trajectory. The outsourcing is what lets ASML scale output without scaling itself, but it also means ASML's delivery promises are the sum of several hundred smaller companies' delivery promises, most of them inside a few hundred kilometres of Eindhoven. That regional cluster is an industrial asset no amount of money reproduces quickly, which is the real reason the lithography monopoly has proved so durable. It is not a patent. It is a neighbourhood.

Payment 13 · TSMC pays the wafer makers

The payment. Polished 300 mm silicon wafers from SUMCO, Shin-Etsu Handotai, GlobalWafers and Siltronic: contracted years ahead, in long-term agreements, because wafer capacity is added in multi-year increments.

Why this expense exists. Because everything in this chain is printed on a disc of silicon grown as a single crystal, and the crystal has to be nearly perfect. Defects in the substrate become defects in the die, and at these geometries the tolerance for substrate imperfection is effectively zero.

How it moves the payer's trajectory. Wafer supply is one of the few inputs where a foundry's leverage is weak, because wafer makers add capacity cautiously after a decade of being punished for adding it eagerly. Long-term agreements are the foundry buying certainty, and paying for it.

Payment 14 · The wafer makers pay the crucible makers

The payment. Fused quartz crucibles from Shin-Etsu Quartz Products, Momentive Technologies and CoorsTek. Each crucible costs thousands of dollars, lasts roughly 400 hours of operation, and accounts for something like 30% of the cost of producing a silicon ingot.

Why this expense exists. Because to grow a single silicon crystal you must hold molten silicon at around 1,400 °C without the container contaminating the melt. A crucible is consumable: it is destroyed in the process of making the thing.

How it moves the payer's trajectory. This is the least glamorous and most under-appreciated line in the whole chain. A third of ingot cost is a pot that gets thrown away. Any improvement in crucible life goes straight to the cost of every chip made anywhere, which is why this unremarkable ceramic has its own research programmes inside companies whose names never appear in a technology headline.

Payment 15 · The crucible makers pay for rock from one American county

The payment. High-purity quartz. Roughly 70-90% of the world's supply estimate of semiconductor-grade high-purity quartz comes from the Spruce Pine mining district in Mitchell County, North Carolina (a town of about 2,200 people), worked by two private operators, Sibelco and The Quartz Corp.

Why this expense exists. Because the purity requirement is absolute. Semiconductor-grade quartz is silicon dioxide at 99.998% or better, with contamination measured in parts per billion, and at that level it is the only material known that will contain molten silicon without poisoning it. The Spruce Pine pegmatites formed roughly 380 million years ago in a continental collision, fifteen miles down, under conditions that produced a chemical purity for which no known geological equivalent has been found anywhere else on Earth.

How it moves the payer's trajectory. It does not move anybody's trajectory. It sets a ceiling on everybody's. This is the point in the chain where corporate strategy stops mattering and plate tectonics takes over: a handful of privately held operators in one Appalachian county sit underneath a semiconductor industry worth hundreds of billions a year and a solar industry worth hundreds of billions more. No amount of capital expenditure anywhere else in this chain changes that, and no participant in the chain has a plan for it beyond hoping the weather holds.

Payment 16 · TSMC pays for gases, by letting the gas company move in

The payment. Ultra-high-purity nitrogen, hydrogen, helium, argon and specialty gases, bought not by the tanker but by the plant. In July 2026 Linde announced a long-term agreement to supply one of the world's largest chipmakers in Phoenix, Arizona, investing $1bn announced to expand its on-site complex there with two new air separation units alongside three existing ones, one of Linde's largest electronics investments anywhere. Air Liquide committed over $160m to a new ultra-high-purity gas plant in Arizona, operational in 2028, including on-site hydrogen production with carbon capture.

Why this expense exists. Because a leading-edge fab consumes industrial gas on a scale that makes delivery absurd. The nitrogen alone, used to purge and blanket nearly every process, has to be generated at the fence line. So the customer does not buy gas; it buys a fifteen-to-twenty-year contract under which the supplier builds, owns and operates a chemical plant next door.

How it moves the payer's trajectory. It converts an impossible logistics problem into somebody else's capital project, and it welds the two companies together for two decades. It also quietly demonstrates something about the geography of this chain: when a fab moves to Arizona, the gas plant, the chemical plant and the hydrogen plant move to Arizona, and the industrial base of a region is reorganised by one customer's siting decision.

Payment 17 · TSMC pays the chemical houses

The payment. Photoresists and process chemicals from JSR, Tokyo Ohka Kogyo, Shin-Etsu, Sumitomo, Fujifilm and Merck's electronics division; precursors, slurries, filtration and ultra-clean handling from Entegris and its peers.

Why this expense exists. Because lithography is a chemical process as much as an optical one. The resist has to absorb EUV photons efficiently, release acid precisely, and develop with a line-edge roughness measured in a handful of atoms. At these doses the statistics of individual photons start to matter, and resist chemistry is one of the few remaining levers on resolution that is not priced at $400m a unit.

How it moves the payer's trajectory. Resist is a small line on a foundry's bill and a large determinant of its yield, which makes it one of the highest-return relationships in the industry and one of the most jealously guarded. These are decades-long joint development programmes, not purchase orders.

Figure 2 · The chokepoints

Where the chain narrows to one or two suppliers

Shares are as reported or as published by the firms themselves, except high-purity quartz, ABF substrate and wafer fab equipment, which are industry estimates and are given as ranges for that reason. The index-manager figure is the sum of three disclosed holdings as at 30 June 2026 and is shown for contrast: it is the only line here that confers no operational control at all.

ACT THREE · The tools that do the thousand steps

A modern logic wafer passes through something over a thousand process steps. Lithography is the famous one and it is a minority of them. The rest (deposition, etch, implant, planarisation, cleaning, and the measurement of all of it), is bought from four companies, and 2026 was the year they were paid more than they had ever been paid. Lam Research told investors it expected calendar 2026 wafer fab equipment spending in the low $150bn range estimate, raised from a prior $140bn; Bank of America took its 2026 forecast to $156bn and its 2027 forecast to $210bn estimate.

Payment 18 · TSMC pays Applied Materials

The payment. Deposition, implantation, and a great deal else. Applied reported record revenue of $9.1bn booked in a quarter, up 15% sequentially and 25% year on year, and its shares roughly doubled across 2026.

Why this expense exists. Because every layer in a chip has to be put there. Atomic layer deposition builds films a single atomic layer at a time, which is the only way to line a structure a few nanometres across conformally. At gate-all-around geometries this is not a refinement, it is the enabling step.

How it moves the payer's trajectory. Tool purchases are how a foundry converts money into the physical ability to run a node. The purchase order is placed years before the node earns anything, against a demand forecast, and it is the most consequential forecasting exercise in the industry.

Payment 19 · TSMC pays Lam Research

The payment. Etch, and the removal side of the process. Lam posted revenue of $6.72bn booked in a quarter, up 15% sequentially and 30% year on year, at a 52% gross margin, its highest in twenty years.

Why this expense exists. Because you cannot build a three-dimensional transistor by adding only. The high-aspect-ratio etches that make modern memory and modern logic possible (cutting straight-walled holes dozens of times deeper than they are wide), are among the hardest controlled processes in manufacturing.

How it moves the payer's trajectory. A twenty-year-high gross margin is the clearest single number in this entire chain about where pricing power sits. When a supplier's margin reaches a two-decade high in the same year its customers' capital budgets reach an all-time high, that is not a coincidence: it is the arithmetic of a sold-out order book.

Payment 20 · TSMC pays Tokyo Electron

The payment. Coaters and developers, etch, deposition, cleaning and wafer handling.

Why this expense exists. Because the resist has to be applied and developed with a uniformity matching the scanner's precision, and a scanner is worthless attached to an inferior track. Tokyo Electron's coater-developers are effectively paired with ASML's scanners in production lines worldwide.

How it moves the payer's trajectory. For a foundry, the trajectory effect is integration risk reduction: buying the combination that the rest of the industry has already debugged, rather than the theoretically optimal one. At these capital intensities, boring is a strategy.

Payment 21 · TSMC pays KLA

The payment. Inspection and metrology, the tools that find the defects and measure whether the previous nine hundred steps did what they were supposed to.

Why this expense exists. Because yield is the only economics that matters in a fab, and yield is unmanageable without measurement. The relationship is brutally simple: the probability a die survives falls exponentially with defect density and with area, and AI accelerators are enormous. Large dies are punished hardest by defects, which makes metrology worth more on exactly the products this chain is built to make.

How it moves the payer's trajectory. Metrology spending is how a fab shortens the learning curve on a new node, and the learning curve is the whole competitive difference between foundries. It is worth stating plainly, as this paper has before: no foundry discloses its yields. Every yield figure in circulation for a leading-edge node is a supply-chain estimate, and anyone treating one as fact is repeating a guess.

Payment 22 · The toolmakers pay the subsystem makers

The payment. Applied Materials, Lam and Tokyo Electron buy RF power generators, impedance matching, vacuum gauges, mass flow controllers, pressure control and ozone systems from MKS Instruments and Advanced Energy.

Why this expense exists. Because a plasma etcher is, functionally, a very well-instrumented radio transmitter aimed at a wafer, and the stability of the RF delivery determines whether the etch is repeatable. Toolmakers do not build these subsystems because specialists build them better and sell them to everybody.

How it moves the payer's trajectory. For the toolmakers this is the same bet ASML made: concentrate on the system and the process recipe, buy the physics modules. For the subsystem makers it is the quiet best position in the industry: selling to all competitors at once, insulated from which of them wins.

Payment 23 · The subsystem makers pay the integrators

The payment. Gas delivery panels, weldments, frames, modules and clean subassemblies from Ichor Systems, Ultra Clean Holdings and their peers.

Why this expense exists. Because a tool is mostly plumbing. Hundreds of metres of electropolished stainless steel, valves and fittings have to deliver corrosive and pyrophoric gases at controlled flow without contamination or leakage, and that is a separate manufacturing discipline from designing a process chamber.

How it moves the payer's trajectory. This layer is the chain's shock absorber, and its victim. Outsourced integrators carry the inventory and the hiring risk of the cycle on behalf of their customers: they ramp first and get cut first. The capital-intensity of the companies at the top of this chain is possible partly because the volatility has been pushed down to companies at the bottom of it.

ACT FOUR · The package, which is where the shortage actually is

For thirty years the hard part of making a chip was making the chip. That stopped being true. The binding constraint moved outward, into the business of attaching the die to a substrate, stacking memory beside it and getting the heat out: and the companies that do that work are, in 2026, the ones setting the ceiling on how many accelerators exist.

Payment 24 · TSMC pays itself, and then pays the assembly houses

The payment. Advanced packaging capacity (the interposer-based assembly that places logic and stacked memory on a single substrate), funded out of TSMC's own capital budget, with conventional assembly and test work placed with the outsourced houses: Amkor, ASE and SPIL.

Why this expense exists. Because an accelerator is not a chip. It is a module: several pieces of silicon from different processes, mounted together with a bandwidth between them that an ordinary package cannot deliver. The packaging step therefore became part of the foundry's product, not an afterthought handled downstream.

How it moves the payer's trajectory. Owning packaging turned TSMC from a supplier of wafers into a supplier of finished compute modules, which raises the value captured per wafer and makes the customer relationship far harder to unwind. A customer who buys wafers can take them elsewhere. A customer who buys assembled modules has bought a process flow.

Payment 25 · SK hynix pays Hanmi Semiconductor for the machines that stack memory

The payment. Thermal compression bonders. On 8 June 2026 SK hynix placed an order worth ₩44.2bn contracted (roughly $29-33m depending on the day's rate), with Hanmi Semiconductor for its newly developed TC Bonder 4.5 Griffin, built specifically for HBM4. At an estimated unit price of about ₩3bn, that is in the region of fifteen machines, with delivery, installation and acceptance contracted to complete by 2 September.

Why this expense exists. Because high-bandwidth memory is made by stacking memory dies on top of each other and bonding them with thousands of connections at once, under heat and pressure, without warping anything. The bonder is the machine that performs that step, and the number of bonders installed is a hard limit on how many HBM stacks can be produced.

How it moves the payer's trajectory. This is the single most precise illustration in the chain of how a small payment governs a large one. A ₩44.2bn equipment order (trivial against the sums elsewhere in this piece), gates the output of a product that is itself gating the output of a $96bn-a-quarter company. Concerns had been circulating that restrained capital spending would delay the HBM4 ramp; the order was read as answering them. Fifteen machines, and the market's view of next year's accelerator supply moved.

Payment 26 · SK hynix pays ASMPT as well, on purpose

The payment. A second order for thermal compression bonding equipment, placed with Singapore-headquartered ASMPT alongside the Hanmi order.

Why this expense exists. Because a sole-source dependency on the machine that gates your flagship product is an unacceptable risk, and because dual-sourcing creates competitive pressure on price and delivery where otherwise there would be none.

How it moves the payer's trajectory. This is the counter-move to everything described in Act Two. Where ASML responded to an irreplaceable supplier by buying a quarter of it, SK hynix responded by qualifying a second one. Both are rational; which one is available depends entirely on whether a second supplier exists. In lithography, none does. In bonding, one does, and the buyer's whole negotiating position flows from that single fact.

Payment 27 · Nvidia's package makers pay the substrate houses

The payment. ABF substrate (the rigid organic board the die sits on), from Unimicron, Ibiden and Shinko, who between them hold roughly three-quarters of the market estimate. From the second quarter of 2026 average quoted prices went up 5-10%, with spot prices on some products up over 30%. Ibiden and Unimicron are jointly investing over ¥800bn, about $5bn announced, to expand high-performance substrate capacity. Orders at the Taiwanese, Japanese and Korean makers are reported secured into 2030.

Why this expense exists. Because accelerator dies have grown so large that the substrate beneath them now fails before the silicon does: warping and thermal stress from expanding die sizes have made substrate performance a first-order engineering problem rather than a packaging detail. Analysts following the sector describe the supply-demand gap widening from around 8% in 2026 to 22% by 2030 estimate, with the shortage persisting through 2028.

How it moves the payer's trajectory. A $5bn capacity investment by two competitors at once tells you they have each independently concluded the demand is real and durable, and that the constraint is theirs to relieve or to profit from. For the buyers it is worse news than it looks: a shortage that is forecast to widen through 2030 means price is not the problem, allocation is, and allocation goes to whoever signed longest and earliest.

Payment 28 · The substrate houses pay a food company

The payment. The insulating film the substrate is built up from, bought from Ajinomoto: a Japanese firm whose principal business is seasonings and amino acids, and which holds the core patents and roughly 95% of the global market estimate for build-up film. In 2026 it raised prices by around 30%.

Why this expense exists. Because the film has the dielectric and thermal properties required, nobody has replicated it at volume, and the patents are held. There is no second film.

How it moves the payer's trajectory. Consider the position: a near-total monopoly on a material that every AI accelerator on Earth requires, held by a company most of the world knows for monosodium glutamate, able to raise prices 30% in a year without losing a customer. And consider the position of everybody upstream of it, which is that they paid. This is what a genuine chokepoint looks like in a financial statement, and it is one of four or five in this chain that no amount of capital can route around in under five years.

Payment 29 · Ajinomoto pays the chemical industry

The payment. Epoxy resins, hardeners, silica fillers and the rest of the formulation's inputs, bought from the global specialty chemical industry: which in turn buys from the petrochemical industry, which buys hydrocarbons.

Why this expense exists. Because build-up film is a filled polymer. Its properties come from the resin chemistry and the particle size distribution of the filler, and the inputs are commodity chemicals bought at commodity scale.

How it moves the payer's trajectory. Here the chain finally touches bottom on the materials side and hands off into the general industrial economy. It is worth marking the moment: twenty-nine payments down from an enterprise software contract, the money is buying barrels of oil and bags of silica. Everything above this line is value added. This is the line.

Payment 30 · Everybody pays the testers

The payment. Automated test equipment from Advantest and Teradyne, bought by the foundries, the memory makers and the assembly houses alike.

Why this expense exists. Because an accelerator module costs more than a car and cannot be sold untested, and because stacked memory has to be tested at several points in its assembly, a defect found after stacking has destroyed four good dies as well as the bad one. Test time per device has risen sharply with complexity, and test capacity has become a genuine constraint on output rather than a rounding error on cost.

How it moves the payer's trajectory. Test is the chain's last chance to avoid shipping a fault into a data centre where it will be found by a customer. Spending on it is spending on the right to charge premium prices, because the premium rests entirely on the claim that the thing works.

ACT FIVE · The box, the building, and the thing nobody can buy

Here the chain leaves the clean room, and the character of the constraints changes completely. Everything up to Payment 30 is limited by physics and patents. Everything from Payment 31 is limited by queues: by the fact that the world's capacity to build heavy electrical equipment was sized for a different century.

Payment 31 · Nvidia pays Foxconn to turn chips into racks

The payment. System integration and rack assembly. Foxconn's first-quarter 2026 revenue was NT$2.11 trillion, about $67bn, up 29.68% booked year on year. For the first time in the company's history, cloud and networking products reached 40% of total revenue, overtaking consumer electronics at 38%. The chairman confirmed AI server shipments were on track to double in 2026, with the company holding roughly 40% of global AI server rack assembly. Global production of the current rack generations ran at around 8,300 units in a single month estimate.

Why this expense exists. Because Nvidia no longer sells chips to be built into systems by somebody else; it sells the system. A 72-accelerator rack with a copper backplane, liquid cooling and its own power topology is a manufacturing problem, and Nvidia does not manufacture.

How it moves the payer's trajectory. Selling the rack instead of the chip raises Nvidia's revenue per deployment enormously and (more importantly), makes the interconnect, not the chip, the unit of competition. A rival selling a faster individual accelerator still has to answer the question of what it plugs into. That is the strategic purpose of the payment to Foxconn: it moves the contest to ground Nvidia chose.

Notice also what it did to Foxconn. A company built on assembling consumer electronics crossed over, in a single year, into being primarily an AI infrastructure company. That is the most dramatic trajectory change of any firm in this chain, and it was caused entirely by its customer's product decision.

Payment 32 · Foxconn pays Delta Electronics for the power path

The payment. Power supplies, power shelves, busbars and conversion. Delta reported record revenue on AI server power and liquid cooling demand.

Why this expense exists. Because a rack drawing well over a hundred kilowatts cannot be fed through conventional server power distribution. The conversion losses alone would be unaffordable, and the copper required at low voltage becomes physically unmanageable, which is why the industry is moving to 800-volt DC architectures.

How it moves the payer's trajectory. Power electronics companies spent decades competing on a few points of efficiency in commodity markets. AI racks turned efficiency into the binding constraint of a customer's entire business case, because every watt lost in conversion is a watt not sold as compute and a watt that must still be cooled. Delta's trajectory changed because the thing it was already good at suddenly became the thing that mattered most.

Payment 33 · Foxconn and Nvidia pay Vertiv to take the heat away

The payment. Coolant distribution units, heat rejection, power management and the thermal design of the room. Vertiv carried a backlog of around $15bn contracted and targeted revenue of $13.25-13.75bn announced for 2026, having worked with Nvidia and Foxconn on Taiwan's first current-generation AI data centre and on the first facilities using 800-volt DC architecture.

Why this expense exists. Because air cannot remove this much heat from this small a volume. Liquid cooling is not an efficiency option any more; it is the only way the hardware runs at all. The transition from air to liquid is the largest change in data centre mechanical design in thirty years, and it happened in about three.

How it moves the payer's trajectory. A backlog larger than a year's revenue is the cleanest possible statement of position: Vertiv is not selling into a market, it is rationing. For the buyers, the trajectory effect is that thermal capacity has become a scheduling input on a par with chip supply. A hyperscaler can have the accelerators and still not have a data centre.

Payment 34 · Vertiv and the rest pay for metal

The payment. Copper, aluminium and steel, bought through the industrial metals market from the global mining and smelting industry.

Why this expense exists. Because busbars, cold plates, windings and cabling are copper, and there is no substitute at these current densities. Electrical infrastructure is, in mass terms, mostly refined metal.

How it moves the payer's trajectory. For the miners this demand is a windfall arriving on top of the energy transition's demand, from a sector that did not exist in their forecasts a decade ago. For everybody downstream it is the chain's exposure to commodity prices and to mine permitting timelines, neither of which responds to how urgently a data centre is needed.

Payment 35 · Microsoft pays the builders

The payment. Construction management and general contracting for eighty-eight buildings a year, placed with the large American contractors who specialise in mission-critical work.

Why this expense exists. Because a data centre is a building, and buildings are built by builders. The specialised part is not the structure; it is the electrical and mechanical fit-out, the commissioning, and the ability to deliver on a date that was promised to a customer before ground was broken.

How it moves the payer's trajectory. Construction velocity has become a competitive weapon. The hyperscaler that commissions a gigawatt first gets to sell it first, and the contractors capable of running many simultaneous mission-critical projects are themselves a scarce resource being competed for.

Payment 36 · The builders pay for concrete and steel

The payment. Cement, aggregate, reinforcing bar and structural steel from the building materials industry.

Why this expense exists. Because of mass and because of time. Concrete cures at a rate nobody can accelerate with money, which makes it one of the few inputs in this chain that is genuinely immune to capital.

How it moves the payer's trajectory. For the materials industry this is ordinary demand in extraordinary volume, concentrated in a few counties. For the chain it is the first of the hard physical limits, and the smaller one. The larger one is next.

Payment 37 · Somebody pays for a transformer, and waits years

The payment. Transformers, switchgear and substation equipment from Hitachi Energy, Siemens Energy, GE Vernova, Eaton, Schneider Electric and ABB: bought by the hyperscalers, by the developers and by the utilities serving them.

Why this expense exists. Because a data centre campus is an industrial electrical load that has to be connected to a transmission network, and the connection is made of equipment that takes years to build. Standard large power transformers now average around 128 weeks from order to delivery estimate. Substation transformer lead times have stretched from roughly 140 weeks in 2023 to over 160 weeks. At the major manufacturers the picture is tighter still: Siemens Energy quoting 48-60 months with its backlog committed through 2030, Hitachi Energy likewise 48-60 months with capacity fully committed through 2029, and GE Vernova beyond 60 months, with some lines quoting delivery in 2031. Siemens Energy's grid order backlog reached a record €51bn, its total backlog roughly $168bn booked.

Why this is the most important payment in the chain. Because it is the one that cannot be hurried. A company can pay more for memory, outbid a rival for packaging allocation, charter aircraft for accelerators. It cannot buy a transformer that does not exist, and the factories that make them were sized for the replacement needs of a mature grid, not for the arrival of an entirely new category of load. The result is that the binding constraint on artificial intelligence in 2026 is not silicon. It is heavy electrical equipment, and it has a four-to-six year lead time.

How it moves the payer's trajectory. It converts corporate strategy into queue position. Whoever placed orders in 2023 and 2024 can build in 2027 and 2028; whoever is ordering now is planning for the end of the decade. Nothing in a quarterly earnings call changes that, and no amount of capital expenditure guidance moves a delivery date already assigned to somebody else.

Payment 38 · The utilities pay for turbines, and find them sold out

The payment. Gas turbines from GE Vernova, Siemens Energy and Mitsubishi Power. GE Vernova's gas turbine backlog reached 100 GW in the first quarter of 2026 and 116 GW of orders and slot reservations by the second contracted, up from 83 GW at the end of 2025, against a total company backlog of about $176bn booked. Siemens Energy sits near 70 GW. Lead times for a new combined-cycle plant have reached five years, up from three and a half in 2023, with costs up 49% over the same period, and some frames are sold into the next decade.

Why this expense exists. Because the electricity has to come from somewhere, and in the United States the dispatchable generation that can be built at scale on anything like a relevant timescale is gas.

How it moves the payer's trajectory. A turbine "slot reservation" is a new kind of financial instrument: a utility paying for the option to be allowed to build a power station at a given future date. When a manufacturer's order book becomes the scarce asset, access to it is what gets traded. And the cost rise (49% in three years), is paid by whoever ultimately buys the electricity, which is a point this piece returns to on Payment 45.

Figure 3 · The queue

What the chain actually waits for, in months

Transformer and turbine figures are the manufacturers' own quoted lead times and published backlog positions as reported through 2026; the greyed bar is the same combined-cycle plant quoted in 2023, shown for scale. The lithography line is the industry's working assumption for order-to-production-wafers and is an estimate, not a quoted figure. Everything above the bottom bar is heavy electrical equipment: which is to say, the constraint is no longer the chip.

ACT SIX · The power, and the people who did not order it

Payment 39 · The utilities pay for gas and for pipe

The payment. Natural gas from the American producers, and capacity on the pipelines that move it, bought under long-term supply arrangements by the utilities and independent power producers serving the new load.

Why this expense exists. Because the turbines bought on Payment 38 burn something, and because a data centre campus requires power at a constant rate, day and night, in a way that no other large industrial load quite does. A smelter can be curtailed. A training run can be, in principle, but the economics of idle accelerators make it the most expensive thing in the building.

How it moves the payer's trajectory. Long-term fuel contracts are the generator's hedge against the one risk it cannot pass on. And the demand itself has re-rated an entire industry: gas producers who spent a decade being told their asset base was stranded are now being asked to sign twenty-year supply agreements by counterparties with the strongest balance sheets in the world.

Payment 40 · Microsoft pays Constellation to restart a reactor

The payment. A twenty-year power purchase agreement contracted with Constellation supporting the restart of the 835 MW unit at the Crane Clean Energy Center in Pennsylvania, a $1.6bn restoration, backed by a $1bn US Department of Energy loan contracted with the first advance expected in the first quarter of 2026, and expected to generate again in 2027.

Why this expense exists. Because Microsoft needs firm, carbon-free power, at scale, on a timescale that new construction cannot meet, and a shut-down reactor with intact infrastructure is the only asset class that offers all three. The company itself described it as a once-in-a-lifetime opportunity, which is literally true: there are only so many recently retired reactors.

How it moves the payer's trajectory. This is the point where a technology company's procurement decision reaches into national energy policy. A twenty-year offtake agreement from a counterparty of Microsoft's credit quality is what made a reactor restart financeable; without the contract, the plant stays closed. The trajectory being altered here is not only Microsoft's. It is Pennsylvania's.

Payment 41 · Amazon and Google do the same thing differently

The payment. Amazon contracted with Talen Energy for 1,920 MW contracted of nuclear output from the Susquehanna station, running through 2042 with an extension option, and has since signed for power associated with an upgrade at Calvert Cliffs. Google took the other route: an agreement with Kairos Power to bring 500 MW announced of advanced reactors online by 2035, backing seven small reactors, with the Hermes 2 unit in Tennessee serving the first phase by around 2030, and with Samsung C&T committing roughly $100m in equity and construction services to build it.

Why these expenses exist. Because each company reached a different conclusion about the same problem. Amazon bought existing output, which is available now and expensive. Google bought a technology that does not exist at commercial scale yet, which is cheap now and risky.

How they move the payers' trajectories. Amazon purchased certainty and gave up any claim on the future cost curve. Google purchased a position in a new generation of reactor and accepted that it may not arrive. Both are buying the same thing in the end (the right to keep building after the grid says no), and the fact that the two largest cloud providers chose opposite instruments is the clearest available evidence that nobody knows which will work.

Payment 42 · The nuclear operators pay the fuel fabricators

The payment. Fabricated fuel assemblies from Westinghouse, Framatome and their peers.

Why this expense exists. Because a reactor restart needs a fuel contract before it needs anything else, and fuel fabrication capacity for a specific reactor design is not a spot market.

How it moves the payer's trajectory. For the operator, securing fuel is what converts a restart announcement into a schedule. For the fabricators, a decade of flat demand has become a decade of growth, underwritten by contracts signed by technology companies.

Payment 43 · The fabricators pay the miners and the enrichers

The payment. Uranium from Cameco, Kazatomprom and the other producers; conversion and enrichment services from Urenco, Orano and Centrus.

Why this expense exists. Because uranium has to be dug up, converted to a gas, spun in a cascade of centrifuges until the fissile fraction rises, and converted back. Each of those steps is a separate industry with its own multi-year capacity constraints and a heavily politicised geography.

How it moves the payer's trajectory. This is the second place in the chain (after Spruce Pine), where the constraint is not corporate but geological and diplomatic. And it is the point at which an enterprise software contract signed on Payment 1 has become a position in the global nuclear fuel cycle, which is a sentence worth reading twice.

Payment 44 · The utilities pay into the capacity market

The payment. Capacity payments in the regional electricity market. PJM (the grid operator covering thirteen states and the District of Columbia, and the densest concentration of data centre load in the world), cleared its auction for the June 2026 to May 2027 capacity year at a record $329.17 per megawatt-day booked, which was the price cap set by the operator and approved by federal regulators. The subsequent auction for the 2028/29 delivery year cleared at $325 per megawatt-day booked, again at the ceiling.

Why this expense exists. Because a capacity market pays generators to be available in advance, so that the lights stay on at peak. When demand forecasts rise faster than new generation can be built, the price of being available rises: and when it hits the administrative cap, it means the market has run out of ways to express the shortage.

How it moves the payer's trajectory. It does not move the payer's trajectory. The payer is a utility, and the utility passes it on.

Payment 45 · The bill arrives at a house in Ohio

The payment. This is the hop that makes the rest of the chain worth describing. In PJM's 2025/26 auction, analysis found data centres were responsible for 63% of the price increase, amounting to roughly $9.3bn in costs recovered from customers estimate across the region in higher electricity rates. In the 2028/29 auction, data centres accounted for $6.3bn, or 38% of total charges estimate. PJM expected the record prices to raise some customer bills by 1.5% to 5%. In practical terms, capacity prices were estimated to add about $18 a month estimate to the average residential bill in western Maryland and about $16 a month estimate in Ohio.

Why this expense exists. It does not exist for any reason the payer chose. A household in Ohio did not buy AI capacity, did not sign a cloud commitment, does not own the shares and will not see the model. It is paying because it is connected to the same wires.

How it moves the payer's trajectory. It reduces their disposable income by roughly two hundred dollars a year. That is the whole trajectory effect, and it is the only hop in the chain where the payer gets nothing in return except continuity of supply.

This is the first of the two places where real, external, non-circular money enters the circuit. It does not arrive from a customer. It arrives from a ratepayer.

Figure 4 · The bill

What the circuit charges people who bought nothing

Capacity auction clearing prices are PJM's own published results. The attribution of the increase to data centre load, and the household bill estimates, are analyses by outside bodies rather than figures published by the grid operator, and are presented as such. Capacity is one component of a residential bill among several.

ACT SEVEN · The money behind the money, and the loop closing

The first forty-five payments spend money. The last five explain where it came from, and they are the reason this piece is called a circuit rather than a chain.

Because the arithmetic does not work otherwise. Four companies plan $720-745bn of capital expenditure in 2026 announced, up roughly 75 to 80% from the previous year's record of about $410bn. No corporate cash flow in history has supported that. So the money is borrowed, and the structure through which it is borrowed is new.

Payment 46 · Meta sells 80% of its largest project to a lender

The payment. In October 2025 Meta formed a joint venture with funds managed by Blue Owl Capital worth $27bn contracted to fund and develop Hyperion, a gigawatt-scale campus in Richland Parish, Louisiana, built on a site about the size of 1,700 football fields and projected to deliver more than 2 GW, with construction finishing around 2030 as part of a project valued at over $50bn. Blue Owl's funds take 80% of the joint venture; Meta retains 20% and runs construction and property management. Blue Owl contributed roughly $7bn in cash and Meta received a one-time payout of about $3bn. Through a special purpose vehicle arranged by Morgan Stanley, the project issued $27bn of A+ rated debt and $2.5bn of equity contracted, anchored by PIMCO at around $18bn and BlackRock at around $3bn.

Why this expense exists. It is not an expense; it is the opposite, and that is the point. This is the mechanism by which the largest data centre project any company has attempted is built with debt that does not appear on that company's balance sheet. It has been described as an inflection point in the use of special purpose vehicles by large technology companies, and the largest such transaction completed to date.

How it moves the payer's trajectory. It lets Meta build at a scale its own credit rating and reported leverage would otherwise make expensive, and it transfers the asset risk to investors who want a long-dated, investment-grade, inflation-linked yield secured on a building with a single very creditworthy tenant. Both sides are getting precisely what they want. The question nobody can answer yet is what the building is worth in 2032 if the tenant's demand forecast was wrong: and the answer to that question sits with pensioners, not with Meta.

Payment 47 · Oracle borrows $16.3bn for one campus, and the banks are not the lender

The payment. Oracle closed a $16.3bn contracted financing for a single data centre campus in Saline Township, Michigan: described as the largest single-facility technology debt package ever assembled. PIMCO purchased approximately $10bn of the roughly $14bn bond tranche, stepping in as anchor after US banks pulled back from the deal.

Why this expense exists. Because Oracle contracted to supply an enormous amount of compute to a customer, and the buildings have to exist. The company's plans reportedly include spending in the region of $40bn to acquire around 400,000 top-tier accelerators estimate for those facilities.

How it moves the payer's trajectory. It is the most aggressive position any established company has taken in this cycle: Oracle is converting a very large customer contract into very large fixed assets funded with very large amounts of debt, on the thesis that it ends the decade as a top-tier compute provider rather than a database company. And the detail that the banks stepped back and a bond manager stepped in is the structurally important part. The marginal lender to the AI build-out is no longer a bank. It is a fixed-income fund managing other people's retirement savings.

Payment 48 · The lenders are paid by insurers and pension funds

The payment. The capital that Blue Owl, PIMCO, Apollo, Blackstone, Ares and their peers deploy is not theirs. It is raised from insurance companies, pension schemes, sovereign funds and wealthy individuals seeking yield. More than $200bn of private credit loans to AI-related companies estimate are already outstanding. Blue Owl, PIMCO and others have committed over $40bn to preleased data centres through unregistered bond sales in a little over a year. The Bank for International Settlements has estimated private credit to AI-related companies could reach $300bn to $600bn by 2030 estimate.

Why this expense exists. Because insurers and pension funds have long-dated liabilities and need long-dated, investment-grade income to match them, and a twenty-year lease to a hyperscaler looks, on paper, like one of the best matches available.

How it moves the payer's trajectory. It is the quiet transformation in this whole story. The AI build-out has become a fixed income asset class, and its ultimate funder is the retirement system. That is not a scandal and it is not hidden, it is disclosed in every one of these transactions. But it means the risk of this build-out is no longer concentrated in the share prices of a few technology companies, where only investors who chose that exposure hold it. It has been distributed into instruments held by people who never made a decision about artificial intelligence in their lives.

Payment 49 · The savers already own the companies at the start of the chain

The payment. The same savers' money, through index funds, buys the equity. As at 30 June 2026, BlackRock held about 8.06% of Nvidia, Vanguard about 6.39%, and State Street about 4.19% booked, roughly 18.6% between them, held overwhelmingly on behalf of passive index investors. The same pattern holds across the chain: at Microsoft, Vanguard held about 8.7% and State Street about 4.01%.

Why this expense exists. Because a worker contributing to a pension or a retirement account every month is buying, automatically and without choosing, a slice of every large company in the index, including every listed company named in this article.

How it moves the payer's trajectory. They have no trajectory. They have an automatic monthly instruction. And it produces the strangest fact in the entire circuit: the saver's money lends to the data centre, and the saver's money owns the company that fills it, and the saver's household pays the electricity surcharge that connects it. The same person appears at three separate points in this loop, in three different capacities, and is unaware of any of them.

Payment 50 · The loop closes, and a dollar arrives back where it started

The payment. In September 2025, Nvidia agreed to invest up to $100bn in OpenAI announced, which committed to purchasing and deploying millions of Nvidia accelerators in the data centres that money would fund. OpenAI separately contracted with Oracle for about $300bn announced, with AMD for tens of billions with a warrant arrangement making OpenAI potentially one of AMD's largest shareholders, with Broadcom for 10 gigawatts of custom accelerators announced to be deployed from the second half of 2026 through 2029, and with CoreWeave for around $22.4bn contracted in cumulative commitments. In total, OpenAI has committed roughly $1.15 trillion announced of infrastructure spending between 2025 and 2035 across seven vendors: Broadcom at about $350bn, Oracle $300bn, Microsoft $250bn, Nvidia $100bn, AMD $90bn, Amazon Web Services $38bn and CoreWeave $22bn. Nvidia also signed a $6.3bn order form under a take-or-pay capacity backstop running to April 2032 contracted, guaranteeing demand for capacity a partner had built.

OpenAI's revenue is running at approximately $24bn a year booked.

Why these expenses exist. Each one, taken alone, is defensible and even conservative. A chip company investing in its largest future customer is securing demand. A customer taking equity in a second chip supplier is creating competition for its primary one. A compute provider signing a backstop is making a buildout financeable for a partner who could not finance it alone. None of this is novel: vendor financing is as old as industrial capitalism, and the railways and the telephone networks were built the same way.

How it moves the payers' trajectories, and why critics object. Critics flagged the structure immediately as circular, money moving in a circle and arriving back as revenue. The loop they describe runs like this: Nvidia invests cash in OpenAI; OpenAI commits that money to Nvidia hardware and to cloud providers; those providers use the revenue to buy more Nvidia hardware; Nvidia books the sales. Sales, in other words, substantially funded by Nvidia's own investment, having gone round once. The combined value of OpenAI's agreements for chips and capacity has been put at over a trillion dollars.

And here Payment 50 meets Payment 1. The enterprise that signed a cloud commitment at Payment 1 is served in a data centre funded by a lender whose capital comes from the saver who owns the shares of the company that sold the accelerators, which were paid for in part by an investment from the company that makes them, in a customer whose own revenue is a fraction of what it has promised to spend.

That is the circuit. Every single link is disclosed. The question is not whether it is secret. The question is what it is made of.

Figure 5 · The three columns

The same scale, sorted by what the money actually is

Read the colours before the lengths. The five longest bars are all announced: intentions, frameworks and "up to" figures spanning up to a decade. The money that demonstrably exists is in the three dark bars, and the darkest fact on the chart is the bottom one: against $1.15tn of announced commitments, the company making most of them is collecting about $24bn a year. Two things this chart deliberately does not do: it does not add the bars together, and it does not claim the gap is unbridgeable. The hyperscalers fund most of their capital spending out of search, advertising, software and cloud businesses that are genuinely profitable today, which is real money and not circular at all. The gap is simply what the structure is betting on closing.

How fifty companies cohabit without destroying each other

Set the fifty payments side by side and an obvious question appears. Many of these firms hold positions over each other that, in any ordinary reading of competition, they ought to be exploiting far more aggressively than they do. Ajinomoto could price build-up film at whatever the traffic will bear. ASML could charge a great deal more than $400m. Sibelco and The Quartz Corp sit on top of the entire semiconductor industry. SK hynix could have taken more than the mid-sixties percent increase it took. None of them goes for the throat. Why not?

Because in this chain, killing your counterparty kills your own output. A supplier that prices its customer out of existence has destroyed its only market; there is no second AI industry to sell to. A customer that squeezes a sole-source supplier into losing money has destroyed the roadmap it depends on. This is not restraint, and it is certainly not sentiment. It is that the chain is so specialised that every participant's maximum long-run extraction is well below its maximum short-run extraction, and all of them know it.

Because dependency runs in both directions at every link. ASML cannot build scanners without Zeiss; Zeiss has no other customer of consequence for that capability. TSMC cannot operate without Linde's on-site plants; Linde built those plants for one customer under a twenty-year contract and cannot move them. Nvidia needs Foxconn's assembly; Foxconn reorganised its company around Nvidia's product. Each of these looks like leverage until you ask what happens the day after it is used.

Because where a second supplier exists, the buyer creates one: and where none exists, the buyer buys shares instead. Those are the only two available responses to dependency, and this chain contains textbook examples of both within days of each other. SK hynix dual-sourced its bonders between Hanmi and ASMPT because two competent suppliers existed. ASML took a quarter of Zeiss SMT because none did. Which strategy a company uses is not a matter of corporate culture. It is a matter of how many firms on Earth can do the thing.

Because the rivalries at the top of the chain share a single trunk at the bottom. This is the most important structural fact in the whole piece and the one most often missed. There is genuine, fierce competition in accelerators: Broadcom's custom silicon business, built for the hyperscalers' own designs, took fiscal 2026 AI revenue to around $58bn booked with supply secured toward roughly $115bn in fiscal 2027 and $230bn in fiscal 2028 announced, with Google's own accelerators alone accounting for something like 58% of unit shipments and 78% of revenue estimate in that programme, and with multi-gigawatt commitments from Meta and OpenAI behind it. That is a real competitive threat to the company at Payment 4, and it is why Payment 4's position is not safe.

But follow the money from Broadcom and it goes to the same foundry, which buys from the same lithography company, which buys from the same optics house. The custom accelerator sits on the same substrate, insulated with the same film from the same food company, stacked with memory from the same three vendors, assembled by the same integrators, cooled by the same thermal companies, powered through the same transformers that nobody can get. Competition is real at the leaves and absent at the root. Whoever wins the accelerator war, the trunk gets paid either way, and the trunk is where the chokepoints are.

And because, by Payment 49, the same institutions own both sides of nearly every rivalry in the chain. When three index managers hold around 18.6% of one company and comparable stakes across its suppliers, its customers and its competitors, the economic interest of the largest shareholder is not that any particular firm wins. It is that the sector's total profit grows. Nothing in that arrangement requires coordination, and nobody is suggesting any occurs: it is the arithmetic of passive indexation, which buys everything by construction. But it does mean the ownership structure sitting above this chain is indifferent to which participant captures the margin, while being highly exposed to whether the margin exists at all.

That is what cohabitation means here. Not harmony. Mutual hostage-taking, stabilised by shared ownership and enforced by lead times.

So what is the circuit actually made of?

Five doors let money into this loop. The critique of a piece like this one is always that it names the doors and walks on, so here is each one with its size, and with the specific event that shuts it.

Door one · End-demand

Size: not disclosed by anyone, and that absence is the finding. Not one company in this chain reports AI-attributable revenue as a separate line. Microsoft does not, Alphabet does not, Amazon does not. Every figure in circulation for "the size of the AI market" is therefore estimate, and most of them are built by assuming a conversion from capacity to revenue that is exactly the thing in question.

What shuts it: nothing. This door does not close; it widens. The risk here is not closure, it is rate: that end-demand grows more slowly than the depreciation schedules written against it. Which is a different problem, and a quieter one.

Door two · The hyperscalers' own profits, by far the largest

Size: capex guidance across the four is now in the $720-745bn range for 2026 announced, after Amazon raised its number to around $220bn, Alphabet to $195-205bn, Microsoft to about $175bn and Meta lifted its floor to $130-145bn. Against that, the cash the businesses actually threw off: Microsoft reported operating cash flow of $55.4bn in a quarter booked against $35.8bn of net capital expenditure, leaving $19.6bn of free cash flow. Alphabet generated $39.1bn of operating cash flow in the same quarter booked and reported negative free cash flow of $5.9bn, its first negative quarter since the company went public in 2004. Amazon's trailing twelve-month free cash flow fell to $1.2bn from $25.9bn a year earlier booked.

That is the most important paragraph in this article. The largest door is still open, the money behind it is real and earned, and it is visibly narrowing in the accounts of the companies holding it open. Free cash flow going negative at Alphabet for the first time in twenty-two years is not a forecast or an opinion. It is a filed number.

What shuts it: a capex guidance cut. This is the only door in the circuit that can be closed deliberately, by a board, in a single quarter, with no counterparty's permission. It is therefore the one that will close first if it closes at all. And the line to watch is not the capex guide, which is an intention; it is free cash flow, which is arithmetic.

Door three · Debt

Size: more than $200bn of private credit outstanding to AI-related companies estimate, over $40bn of preleased data centre paper placed through unregistered bond sales in about a year estimate, plus the specific executed packages: $27bn for Hyperion and $16.3bn for Oracle's Michigan campus contracted. The Bank for International Settlements has put the possible 2030 figure at $300bn to $600bn estimate.

What shuts it: a refinancing that does not clear, or spreads widening on preleased paper. This is the earliest observable signal in the entire circuit, and the reason to watch it rather than the share prices is that it is a credit judgement made by people whose return is capped and whose downside is not. Equity investors are paid to be optimistic. Lenders are not. When the lender who anchored a $14bn tranche declines the next one, that is information, and it arrives in public, in the pricing, before it arrives in any earnings call.

Door four · Index equity

Size: a continuous monthly inflow that nobody decides. Three managers held roughly 18.6% of Nvidia between them as at 30 June 2026 booked, and comparable positions across its suppliers, its customers and its competitors.

What shuts it: net outflows, or a rebalancing. And the structural point is that this door cannot be called. Nobody can demand their money back from an index position; it simply reprices. That makes index equity the softest claim in the circuit, which is precisely why it absorbs loss first and most quietly: a repricing is not an event, it is a Tuesday.

Door five · The ratepayer, and this one is already being shut

Size: roughly $9.3bn recovered from customers in a single auction cycle estimate, and $6.3bn, or 38% of total charges, in the 2028/29 auction estimate.

What shuts it: regulators, and they have started. As of June 2026, twenty-four states had approved at least one large-load tariff designed to make data centres carry the cost of the infrastructure they require estimate. The Oregon Public Utility Commission approved a framework for Portland General Electric that creates a dedicated rate class for large loads, requires those customers to cover 100% of the distribution upgrades their projects need, and imposes a 1 cent per kilowatt-hour surcharge on projects above 100 MW, with the proceeds funding programmes that offset residential costs and address low-income energy burden contracted. Oregon's POWER Act, signed in August 2025, directs the creation of a separate rate class for facilities of 20 MW and above. Pennsylvania's Public Utility Commission adopted a framework in April 2026 explicitly intended to shield existing ratepayers, assigning dedicated cost responsibility and minimum-demand commitments to the customers driving the build-out. Virginia's GS-5 tariff, Texas's SB 6 and Ohio's AEP model are the other developed examples. The principle being written into two dozen state rulebooks is a single sentence: if your load causes the cost, your rate class carries it.

This is the door the circuit was least entitled to and it is closing fastest. And it does not close harmlessly, because every dollar a regulator pushes back off a household lands on the project's own economics instead. The tariff wave is not a side-story about fairness. It is a direct, quantifiable, already-legislated increase in the cost base of every campus not yet built: and it arrives in the same years as the turbines.

What is recirculation and what is not

Everything in the chain that is not one of those five doors is recirculation, and here the accounting matters.

The same underlying dollar is recognised as revenue at Microsoft, at Nvidia, at TSMC, at ASML, at Zeiss, at Trumpf, and so on down. That is entirely normal; every supply chain on Earth reports this way, and nobody is doing anything improper. But it has a consequence that bites when a sector is growing this fast: the sum of revenue growth across a chain is not a measure of how much end-demand grew. It measures how much money moved. Those two track each other closely in a steady state and diverge sharply during a build-out, which is the condition here.

Then there is the specific structure at Payment 50, a supplier investing in a customer who commits to buying from the supplier. It is lawful, disclosed and ancient. Railways, telegraph networks, aircraft makers and telecoms equipment vendors were all built partly this way. It is also, historically, the structure that performs worst when demand disappoints, because it removes the one signal that would otherwise warn everybody early: an independent customer declining to buy.

And the asymmetry that governs all of it: the financial layer of this circuit can reprice in an afternoon. The physical layer cannot reprice at all. Share prices move the same day; private credit marks follow within quarters. But the turbine ordered for 2031 still arrives in 2031. The transformer still ships in four years. The reactor still restarts. The twenty-year gas contract still runs twenty years. The campus in Louisiana is still there in 2032, and the substrate capacity Ibiden and Unimicron are building still comes online into whatever world exists then.

What this establishes

One. The constraint on artificial intelligence in 2026 is not intelligence, and it is not silicon. It is a transformer with a 128-week lead time, a gas turbine sold into the next decade, a ceramic pot that lasts 400 hours, a film made by a seasoning company, and roughly eighty per cent of the world's semiconductor-grade quartz coming out of one county in North Carolina. The bottleneck is not computational. It is a queue, and a geology.

Two. The circuit is fully disclosed and almost entirely unexamined. Every figure in this piece came from a filing, a contract announcement, an auction result or a company's own statement. Nothing here required a leak. The structure is not hidden; it is merely never assembled in order, and assembling it in order changes what it means.

Three. The most consequential payments in the chain are the small ones. A ₩44.2bn order for fifteen machines governs the memory supply of a $96bn-a-quarter company. A 24.9% stake taken in an optics subsidiary in 2016 determines who can print a chip in 2029. A 30% price rise by a Japanese food company reaches every accelerator on Earth. Scale commands attention; chokepoints command outcomes.

Four. Three distinct groups of people are financing this who did not choose to. The ratepayer in Ohio, the pension saver whose fund bought A+ rated data centre debt, and the index investor who owns 18.6% of the whole thing by automatic monthly instruction. None of them appears in any announcement. All of them appear in the circuit, and in the only two places where risk ultimately rests.

Five. Every single participant in these fifty payments is behaving rationally and lawfully, and that is the finding, not an exoneration. Nobody in this chain is doing anything strange. Microsoft is right to buy capacity it has sold. Nvidia is right to secure its demand. ASML was right to buy a piece of Zeiss. SK hynix was right to dual-source. PIMCO is right to want long-dated investment-grade yield. Ajinomoto is right to charge what its position is worth. The circuit is not the product of anyone's bad intention. It is what you get when fifty rational actors each optimise locally inside a structure none of them designed, and that nobody is responsible for as a whole.

That is the pulse this paper exists to document: not a conspiracy, not a bubble, not a miracle: a structure, running right now, visible in public documents. And a structure can be tested, which is what the last section of this piece does.

The claim, stated so that it can be proved wrong

A piece that ends by saying nobody can know until the assets arrive has told the reader nothing he did not already suspect in the first paragraph. So here is the claim, with a date, a cohort and a named place where the consequence lands. It is the thesis of this article and it is falsifiable.

By the close of fiscal 2029, the cohort of data centre capacity financed during 2027 on terms set in 2026 must earn enough, per installed accelerator-year, to cover three things at once:

one, depreciation on assets whose assumed useful life has just been extended to as much as twenty-five years;

two, delivered power bought in capacity markets that have now cleared at the administrative price cap in consecutive auctions, in states that are simultaneously legislating to move that cost off households and onto the load that causes it;

three, debt service on paper placed at investment grade by lenders whose own funding is long-dated insurance and pension money.

If that cohort does not cover all three, the adjustment does not land on the chipmaker, which was paid in cash on delivery and books no exposure to the building. It does not land primarily on the hyperscalers, which funded most of the build from operating profits and can carry one bad cohort. It lands on three balance sheets: the private-credit and insurance holders of preleased data centre debt; the ratepayers inside capacity markets where the tariff reassignment has not yet happened; and the index investor, who holds the equity by automatic monthly instruction and will absorb the repricing without ever having been asked a question.

That claim is checkable, and the three places to check it are public today, without a leak and without a source:

The spread on preleased data centre paper. Not the share prices. Credit prices the downside; equity prices the hope. When the anchor lender on a $14bn tranche declines the next one at the same spread, the circuit has answered.

The free cash flow line of the four largest spenders. Not the capital expenditure guide, which is an intention, revised quarterly, and in this cycle revised upward every time. Free cash flow is arithmetic. One of the four has already printed a negative quarter for the first time since 2004.

The docket sheets of state utility commissions. Twenty-four states have already acted. Each large-load tariff approved is a transfer of cost from a household onto a project's own pro forma, and the cumulative effect of that transfer is the single most under-modelled input in every forecast written about this sector.

And if the revenue does cover all three, then this article is wrong in the only way that matters, the circuit closes on itself profitably, and the 2026 build-out will stand as the most productive infrastructure programme since the interstate highway system. That outcome is entirely possible. This paper is not predicting against it and has no position in the question.

What this paper does assert, narrowly and firmly, is this: whichever way the revenue goes, the identity of whoever absorbs the outcome is already fixed. It was fixed by the financing structure and not by the technology, it was fixed in contracts signed in 2025 and 2026, and the people it was fixed on are, in two of the three cases, unaware that it was fixed on them.

That is not a forecast. It is a reading of documents that already exist.

The circuit, in one table

No.	From → To	What it buys
1	Enterprise → Microsoft	AI capacity under multi-year commitment
2	Microsoft → construction & silicon	~$175bn of 2026 capex, 88 data centres
3	Microsoft → landlords	$11.1bn of data centre leases in one quarter
4	Microsoft → Nvidia	Accelerators
5	Nvidia → SK hynix, Micron, Samsung	High-bandwidth memory, at 2026 prices
6	Nvidia → TSMC	Wafers and advanced packaging
7	TSMC → ASML	EUV and High-NA scanners
8	ASML → Carl Zeiss SMT	Optics, plus a 24.9% stake
9	Zeiss → Corning, Schott	Ultra-low-expansion glass blanks
10	ASML → Trumpf	The 30 kW CO₂ drive laser
11	Trumpf → Zeiss	Beam transport optics
12	ASML → VDL, Neways, Prodrive; tin producers	Mechatronics, and the fuel of the light
13	TSMC → SUMCO, Shin-Etsu, GlobalWafers, Siltronic	300 mm wafers
14	Wafer makers → Shin-Etsu Quartz, Momentive, CoorsTek	Crucibles, 400 hours each
15	Crucible makers → Sibelco, The Quartz Corp	High-purity quartz from Spruce Pine
16	TSMC → Linde, Air Liquide	On-site gas plants, built and owned by the supplier
17	TSMC → JSR, TOK, Merck, Entegris	Photoresist, precursors, ultrapure handling
18	TSMC → Applied Materials	Deposition and implant
19	TSMC → Lam Research	Etch
20	TSMC → Tokyo Electron	Coat, develop, clean
21	TSMC → KLA	Inspection and metrology
22	Toolmakers → MKS, Advanced Energy	RF power, vacuum, flow control
23	Subsystem makers → Ichor, Ultra Clean	Gas panels and clean subassemblies
24	TSMC → itself, Amkor, ASE, SPIL	Advanced packaging and test
25	SK hynix → Hanmi Semiconductor	₩44.2bn of HBM4 bonders
26	SK hynix → ASMPT	The second source, on purpose
27	Packagers → Unimicron, Ibiden, Shinko	ABF substrate, booked into 2030
28	Substrate makers → Ajinomoto	The film, at 95% market share
29	Ajinomoto → specialty chemicals	Resin, hardener, silica filler
30	Everybody → Advantest, Teradyne	Test, because untested modules cannot ship
31	Nvidia → Foxconn, Quanta, Wistron	Rack assembly at ~8,300 racks a month
32	Foxconn → Delta Electronics	The power path
33	Foxconn, Nvidia → Vertiv	Liquid cooling, against a $15bn backlog
34	Vertiv et al → the miners	Copper, aluminium, steel
35	Microsoft → the contractors	Eighty-eight buildings a year
36	Contractors → cement and steel	Mass, and curing time nobody can buy
37	All of them → Hitachi Energy, Siemens Energy, GE Vernova, Eaton, Schneider	Transformers, with a four-to-six year queue
38	Utilities → GE Vernova, Siemens Energy, Mitsubishi Power	Gas turbines, 116 GW already booked
39	Utilities → gas producers and pipelines	Twenty-year fuel
40	Microsoft → Constellation	835 MW, a 20-year PPA, one reactor restarted
41	Amazon → Talen; Google → Kairos	1,920 MW now, or 500 MW later
42	Operators → Westinghouse, Framatome	Fuel assemblies
43	Fabricators → Cameco, Kazatomprom, Urenco, Centrus	Uranium, conversion, enrichment
44	Utilities → the PJM capacity market	Availability, at the price cap
45	A household in Ohio → the utility	Nothing. $16-18 a month.
46	Blue Owl, PIMCO, BlackRock → Meta's Hyperion JV	$27bn of A+ debt, 80% of the project
47	PIMCO → Oracle's Michigan campus	~$10bn of a $16.3bn package
48	Insurers and pension schemes → the private credit funds	The capital the lenders lend
49	Monthly pension contributions → index funds → the equity	~18.6% of Nvidia, held passively
50	Nvidia → OpenAI → Oracle, Microsoft, CoreWeave → Nvidia	The loop, closing on Payment 1

Sources

Documents that bind: company filings, auction results and contract announcements. Nvidia's reported results for fiscal 2026 and the first two quarters of fiscal 2027, including data centre revenue; Microsoft's fiscal 2026 quarterly capital expenditure disclosures, its data centre lease spending, its stated 2026 capital spending plan and its change to data centre useful life; TSMC's published monthly revenue reports for 2026; ASML's reported fourth-quarter 2025 bookings and its disclosed 2016 acquisition of a 24.9% interest in Carl Zeiss SMT with associated R&D funding; Applied Materials' and Lam Research's reported quarterly revenue and gross margin; Samsung's and SK hynix's disclosed average selling price movements for the first quarter of 2026; SK hynix's ₩44.2bn equipment order with Hanmi Semiconductor dated 8 June 2026; Ibiden's and Unimicron's announced joint capacity investment; Broadcom's reported fiscal 2026 AI semiconductor revenue and stated supply position for fiscal 2027 and 2028; Linde's July 2026 announcement of its $1bn Phoenix on-site expansion and Air Liquide's $160m Arizona investment; Constellation's power purchase agreement with Microsoft covering the Crane Clean Energy Center and the associated US Department of Energy loan; Talen Energy's agreement with Amazon for Susquehanna output; Google's agreement with Kairos Power and Samsung C&T's associated commitment; GE Vernova's and Siemens Energy's reported backlog and order positions; Meta's own announcement of the Blue Owl joint venture for Hyperion and the associated SPV financing; Oracle's Saline Township financing; PJM's published capacity auction clearing prices for the 2026/27 and 2028/29 delivery years; and the disclosed holdings of BlackRock, Vanguard and State Street as at 30 June 2026.

Press and analysis, labelled as such. Reporting and analysis on the circular structure of the OpenAI-Nvidia-Oracle-AMD-Broadcom-CoreWeave agreements and the aggregate value of those commitments; the compilation of OpenAI's $1.15tn of vendor commitments and its approximate annualised revenue; estimates of combined 2026 hyperscaler capital expenditure of roughly $725bn against about $410bn in 2025; reported transformer, switchgear and combined-cycle plant lead times and the manufacturers' quoted delivery windows; analyses attributing a share of PJM capacity price increases to data centre load and estimating the monthly effect on residential bills; estimates of market share in high-purity quartz, ABF substrate, ABF film and wafer fab equipment; estimates of the ABF substrate supply-demand gap through 2030; reported rack production volumes; Vertiv's stated backlog and revenue target; and the Bank for International Settlements' estimate of private credit exposure to AI-related companies by 2030.

What is not reported here, and why. No foundry publishes its yields for a leading-edge node, so this piece contains no yield figure; every such number in circulation is a supply-chain estimate. The price any individual customer pays for an accelerator, a wafer or a memory allocation is not disclosed by anyone, so no unit prices are given beyond the published list positions of lithography systems. The identity and value of individual enterprise cloud commitments are confidential, so Payment 1 is described structurally rather than named. Several of the market shares in Figure 2 are industry estimates rather than reported figures and are given as ranges for that reason. And the numbering of the fifty payments is sequence and not chronology, stated as such at the outset: the payments are real and documented, the order in which they are presented is this paper's, and no number in the piece is a date. Every figure above carries one of four tags - booked, contracted, announced, estimate - and where a figure appears without one, that is an omission by this paper and should be read as such rather than as a promotion of the number.

This piece builds on three earlier investigations by this paper: The Country Called NVIDIA, on the external and internal skeleton of the company at Payment 4; the examination of the foundry at Payment 6; and the piece on the lithography company at Payment 7. Where this one differs is that it does not stop at any single firm. The subject here is the circuit.