Capture / layer 1 · The value chain

Compute & physical

The chip makers, the big cloud companies and the newer rented-computing firms, data centres and the power they draw, memory and networking. Dependency tier 1.

How the economics move

The site's judgement of each part of this layer, read against the readings below and the theory on the value chain page: what is getting cheap and interchangeable, what stays scarce, and what would prove it wrong.

What the powers mean

Hamilton Helmer, a strategy consultant and investor who studied why some firms keep earning more than their investors could get elsewhere, found that an advantage lasts only when a benefit sits behind a barrier that stops rivals copying it, and named the kinds that recur:

a cornered resource: preferential access to something others need: a patent, a licence, a scarce input, a rare team
branding: buyers pay more for the same thing because of what they believe about who made it
counter-positioning: a newcomer's model the incumbent could copy only by damaging its own business
network economies: each user makes the product more valuable to the others, so the leader's lead feeds itself
process power: a way of working that improves cost or quality and takes a rival years to learn
scale economies: unit costs fall as volume grows, and a rival cannot match them without matching the volume
switching costs: a customer would lose money, time or work by moving to a rival

Holding

The chip designer's lead is a shortage profit; the foundry's and the toolmaker's are harder to copy

Three kinds of firm make AI chips. Designers such as NVIDIA, AMD and Broadcom draw them; a foundry (a factory that makes chips to other firms' designs), above all TSMC, prints them onto wafers (thin discs of silicon, each later cut into many chips); and toolmakers such as ASML build the machines that do the printing. Eliyahu Goldratt, a physicist who studied factories, found that a system runs at the pace of its slowest part. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. So while buyers cannot get enough chips, the designer they most want takes the rent. This site judges that as more chips arrive and the biggest buyers design their own, that rent is exposed, while the foundry and the toolmaker keep theirs for as long as no rival can match what they do.

Getting cheap and interchangeable: Chips made on older processes are cheap and interchangeable; TSMC's own filing says rivals in those processes are numerous. This site judges that the cost of a unit of computing keeps falling with each chip generation, and that the general-purpose AI chip is being copied from below: every large buyer that designs its own chip, often with Broadcom's help, turns part of NVIDIA's market into a price to negotiate. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers as lower prices and better products.
Staying scarce: This site judges that capacity at the newest processes, and the machines that print it, stay scarce. A new factory for the newest process takes years to build and the skill to run it takes longer. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier, and named the kinds that recur. Two apply here: process power (a way of working that takes rivals years to learn) and scale economies (unit costs that fall as volume grows, so a rival must match the volume to match the cost). The EUV machines have one maker, fed by one optics supplier: a cornered resource, Helmer's term for preferential access to something others need, which does not ease when chips do.
What would make it last: For the designer: its software (for NVIDIA, CUDA, the platform developers write AI code against) and its whole-rack systems with their own links between chips keep buyers even when rival chips are plentiful and cheaper, turning a shortage rent into switching costs, what a customer would lose by moving to a rival. For the foundry and the toolmaker: they keep a lead at each new process, so their scale and process power outlive the AI shortage.
What would prove this wrong: NVIDIA keeping both its share of AI chip supply and its gross margin while AMD's chips and the custom chips ship in volume and wafer supply at the newest processes keeps growing fast; or a second foundry making AI chips at the newest process in volume for a major designer.
Powers here today: a cornered resource, scale economies, process power, switching costs
Binds on (how tight): AI chips (moderate), Chip foundries (moderate), Chip-making equipment (no reading; a model judges it probably moderate), Advanced chip packaging (no reading)
The rival view: Some link in the chip chain proves easier to expand

Holding

High-bandwidth memory is a cyclical commodity earning a shortage's price

An AI chip is only as fast as the memory that feeds it. High-bandwidth memory, or HBM, is DRAM (the working memory a processor reads and writes constantly) cut into dies (single chips), stacked and placed beside the processor so data reaches it quickly. This site counts three makers: SK hynix, Samsung and Micron. Micron's filing says HBM needs more wafers (the silicon discs chips are cut from) and more factory space than ordinary DRAM for the same amount of memory, so when AI demand runs ahead the makers profit, and if it weakens they can turn the same factories back to ordinary memory. Networking, the switches, cables and light-based links that join many chips into one machine, splits in two: NVIDIA's own links, which tie buyers to its systems, and gear built to open standards, which any maker can supply.

Getting cheap and interchangeable: Ordinary memory, the kind in phones and laptops, is a textbook commodity: identical parts, several makers, and prices that swing with each wave of new factories. Networking built to open standards such as Ethernet, which Broadcom sells to the largest cloud and model companies, is interchangeable by design. Eliyahu Goldratt, a physicist who studied factories, found that a system runs at the pace of its slowest part, and David Ricardo, an English economist who studied why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. On their account the rent in memory lasts only until the makers build their way out, and SK hynix says it is building new factories in Korea.
Staying scarce: This site judges that what stays scarce is the skill to stack dies with good yield (the share of units that work) and to be qualified (tested and approved) by a chip designer for its next product; a newcomer would need both, and this site judges they take longer to learn than one product generation. NVIDIA's own links between chips stay scarce because only NVIDIA makes them, though its filing says it now lets other firms' chips connect to them.
What would make it last: Custom HBM, memory built to one buyer's design, which SK hynix calls a key differentiator, becomes the norm, so each chip generation is locked to one maker and switching costs replace the shortage. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than rivals, named switching costs, what a customer would lose by moving to a rival, as one of the lasting advantages he found. In networking, the rent lasts only if NVIDIA's own links stay faster than open standards by enough that buyers accept being tied to one vendor.
What would prove this wrong: Memory makers keeping record margins through a year in which all three add HBM capacity faster than AI chip shipments grow and ordinary memory prices fall.
Powers here today: scale economies, process power, a cornered resource
Binds on (how tight): Memory (no reading), Networking and optics (no reading; a model judges it probably moderate), Advanced chip packaging (no reading)
The rival view: Computing cannot flood the market overnight

Tightening

A site already wired to the grid is scarcer than the building or the money

A data centre needs three things: a building, money, and electricity delivered through a grid connection. This site judges that buildings and money can be had quickly, while a new connection means waiting for the utility to add lines and transformers, and new power stations take years. Eliyahu Goldratt, a physicist who studied factories, found that a system runs at the pace of its slowest part; David Ricardo, an English economist who studied why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. So, in this site's judgement, the rent goes to whoever already holds land with power and a connection. Much of that input sits with utilities whose prices a regulator sets, so they cannot charge what the shortage is worth; this site judges that the rent leaks instead to owners of connected sites and of power stations already running.

Getting cheap and interchangeable: This site judges that the building itself, the shell, the racks and much of the cooling, follows standard designs that many contractors can put up, and that money for the build-out is plentiful. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs an innovation needs to reach its customers complementary assets; generic ones, which anyone can buy, earn an ordinary return, not a rent.
Staying scarce: This site judges that grid connections, heavy electrical gear such as large transformers and turbines, and the electricians and other trades who install them stay scarce. A connection is close to a licence: a utility grants it under a regulator's rules and it cannot be copied, which makes it what Hamilton Helmer, a strategy consultant who studied lasting advantages, calls a cornered resource: preferential access to something others need. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, who studied contracts and found that a buyer who depends on an asset built for it can be squeezed by whoever owns it, would expect the largest buyers to own such assets or lock them up in long contracts.
What would make it last: Holders of connected sites and running power stations lock their buyers into long contracts at prices set during the shortage, or regulators let utilities earn a guaranteed return on new lines built for data centres, turning a shortage rent into a regulated one. It fails to convert if power built on site, beside the data centre and off the grid, proves quick and cheap.
What would prove this wrong: New large sites being connected to the grid about as fast as they are built, with the rent for powered data-centre space falling while use keeps rising.
Powers here today: a cornered resource, scale economies
Binds on (how tight): Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), Data-centre buildings (tight)
The rival view: The spending runs ahead of the revenue, and investors take the loss

Holding

The big clouds keep a rent from customers who stay; the neoclouds earn the shortage's price on borrowed money

A cloud rents out computing by the hour or by the contract. The big clouds, often called hyperscalers (Microsoft's Azure, Amazon's AWS, Google Cloud and Oracle), sell AI chips alongside storage, databases and the software businesses already run on them. The neoclouds (newer firms such as CoreWeave, Nebius, Crusoe, Lambda and Fluidstack that rent out mostly NVIDIA's graphics chips, GPUs, and little else) sell the chips almost alone. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. This site judges that a neocloud's margin is that kind of rent, the gap between what a scarce chip rents for and what it costs to buy, power and borrow against, and that it lasts while chips and powered buildings are short; a big cloud's margin rests more on customers whose data and systems already sit inside it.

Getting cheap and interchangeable: This site judges that an hour on a given GPU is becoming a commodity: the same NVIDIA chips are rented by many firms, Nebius names CoreWeave, Crusoe and Lambda as well as the big clouds among its rivals, and CoreWeave says its customers specify NVIDIA's chips, so the chip, not the cloud, is what they are buying. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers as lower prices. On that finding, the rent on each generation of chips falls towards cost as those chips become plentiful. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its customers complementary assets. This site judges that a neocloud's rack of rented chips is a generic one, which anyone with money, power and a supply of chips can assemble; the scarce inputs behind it belong to others, NVIDIA's chips on one side and the demand of a few big buyers on the other.
Staying scarce: This site judges that what stays scarce is powered, connected data-centre space and, for the big clouds, the customer relationships and the data already stored with them. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. Scale economies and switching costs apply to the big clouds' existing business today: scale economies (unit costs that fall as volume grows, so a rival must match the volume to match the cost; Google offers its own AI chips, TPUs, beside GPUs) and switching costs (what a customer would lose by moving to a rival, here the work of moving data and systems built around one cloud). This site judges that neither is yet proven for AI work in particular, which is younger and easier to move than the systems firms have run in the cloud for years.
What would make it last: The big clouds turn chip rental into services customers build on, such as their own chips, their own model catalogues and the tools that run AI inside a firm's data, so switching costs hold AI work as they hold older work. For a neocloud: long take-or-pay contracts with buyers who can pay, signed during the shortage, outlast it, and it wins powered sites others cannot get, a cornered resource in Helmer's term (preferential access to something others need). Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, who found that a buyer that depends on an asset built for it can be squeezed by its owner, would expect the biggest buyers to build their own capacity once they can, which is the risk a neocloud's contracts must outlast.
What would prove this wrong: Big-cloud margins falling to the neoclouds' level as AI becomes most of what they sell, with customers moving AI work between clouds on price alone; or a neocloud keeping its margin through a year in which GPU rents fall steeply and its largest customer leaves at the end of its contract.
Powers here today: scale economies, switching costs
Binds on (how tight): Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), AI chips (moderate), Capital (slack)
The rival view: The builders overbuild and rent out the rest

The companies that matter here

Candidate to be worth a trillion dollars

ASML

Show the profile

Market: ASML sells lithography machines, the tools that print a chip's circuit pattern with light onto a wafer (a thin silicon disc later cut into chips), plus the service, upgrades and software that keep them running. Its most advanced line uses extreme ultraviolet light (EUV, light of a very short wavelength that can print the finest features). Its newest EUV design, High-NA (optics that gather more light, to print still finer features), has gone mostly to customers' research sites, and the filing says ASML has now released its first full-specification High-NA machine to its first customer. The filing says ASML is currently the world's only maker of EUV machines, and names Canon and Nikon as rivals only in the older deep-ultraviolet machines. Its buyers are chip factories making logic (processors) and memory. In 2025 its logic sales grew, driven by foundries (factories that make chips to other firms' designs) building for AI demand, while its memory sales fell; sales of older machines in China were stronger than it expected. ASML files a 20-F (the annual report foreign companies file with the US regulator) with US GAAP accounts; this site's filing reader does not yet read it, so it records no figures for it.
Customers (concentrated): A small group of chip makers: the leading logic foundries and the memory makers. The filing does not name them, but says its largest customer's share of its sales rose in 2025 and that its sales lean heavily on its two largest customers. The rankings shift from year to year.
Powers: a cornered resource: An exclusive arrangement with Carl Zeiss SMT, its sole supplier of optics, in which ASML holds a stake, and no EUV rival named in its own filing; this site judges the benefit is pricing and the barrier is the long joint optics work a rival would have to repeat.
scale economies: This site judges that each EUV generation takes years of research spread over every machine sold, so a rival would need the same volume of sales to carry the same cost.
switching costs: Customers design their factory processes around its machines and buy its service and upgrades for years after; its service and upgrade sales grew faster than machine sales in 2025, though the filing says part of that came from upgrades to installed EUV machines being counted there instead of as machine sales.
Under the site's rent rule: a monopoly-like profit, kept by the company itself. Depends on: Chip-making equipment (no reading; a model judges it probably moderate), Chip foundries (moderate), Memory (no reading), Capital (slack), Critical minerals (no reading; a model judges it probably moderate).

What would have to hold by 2035

The world keeps building factories for the newest processes (the latest generations of chip-making, which print the smallest features) in logic and memory faster than old ones retire, so demand for EUV machines grows rather than just replacing worn tools.
Chip makers keep shrinking features with lithography rather than with other processes that would need fewer of its machines; the filing names competition from alternative technologies or manufacturing processes only in general terms.
High-NA EUV reaches mass production and customers adopt it on the finest layers, so the next generation is also ASML's alone.
Zeiss and ASML's other single-source suppliers can grow output fast enough that their capacity does not cap its sales for years at a time.
Export rules between the United States, the Netherlands and China do not shut it out of a large block of buyers while a state-backed rival learns to build competing machines.

What would show it failing

A rival, including a state-backed Chinese maker, shipping working EUV-class machines to a paying customer's factory.
Leading chip makers delaying or skipping High-NA for the finest layers and staying with older EUV plus extra patterning steps (printing the same layer more than once).
Its largest customers cutting orders for new EUV machines while AI chip demand still rises, showing the factories were built ahead of need.
New export rules removing China from its buyers without other regions making up the older-machine sales.

Drafted by claude-opus-5.5 from ASML annual report (20-F), 25 Feb 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Candidate to be worth a trillion dollars

Micron

Show the profile

Market: Micron makes memory chips: DRAM (the working memory a processor reads and writes constantly), NAND (the storage that keeps data when power is off) and high-bandwidth memory, or HBM (DRAM dies, single chips cut from a silicon wafer, stacked and placed beside an AI chip so they can feed it data very fast). It sells to cloud firms, server makers, phone and PC makers and car makers, and reorganised in 2025 around business units, one of which serves the largest cloud buyers and holds all its HBM sales to data centres. Memory has long been a cyclical, commodity market: its 10-K (the annual report US companies file with their regulator) warns that an oversupplied market pushes prices down, and that heavy investment, including by Chinese state-backed makers, threatens oversupply. This site's filed figures show the swing: Micron made a gross loss (sales below the direct cost of making them) in 2023, and has since swung from losses to a gross margin, the share of sales kept after those costs, among the highest this site has recorded for it. How big the market becomes depends on whether AI keeps memory short, and for how long.
Customers (concentrated): Its ten largest customers account for a large part of revenue, and the 10-K singles out one unnamed customer, mostly served by the cloud memory unit, as a large buyer. Customers are generally reluctant to sign long fixed-price contracts, so prices and volumes are renegotiated as the market moves. A 2023 Chinese government ruling bars operators of critical information infrastructure in China from buying its products.
Powers: scale economies: Emerging, not proven. This site judges that a memory factory at the newest process costs so much that only a few firms can build one; the benefit holds against newcomers, but not against Samsung or SK hynix, which build at the same scale or larger.
Under the site's rent rule: a thin profit, kept by incumbent firms. Depends on: Memory (no reading), AI chips (moderate), Advanced chip packaging (no reading), Chip-making equipment (no reading; a model judges it probably moderate), Capital (slack).

What would have to hold by 2035

AI keeps raising the amount of memory each unit of computing needs, so memory demand grows faster than the makers can add capacity, cycle after cycle.
HBM stays a different product from ordinary DRAM, bought on stacking skill and a long qualification (testing and approval) with each chip designer rather than on price per bit, the smallest unit of stored data.
Micron keeps pace with SK hynix on each new HBM generation, including HBM4, instead of arriving after the best prices are gone.
The memory makers do not all add so much capacity at once that the old glut returns.
Chinese state-backed makers such as CXMT do not reach the newest processes and flood ordinary DRAM with cheap supply.

What would show it failing

Memory prices and Micron's gross margin falling back toward their 2023 lows while AI spending is still rising.
Chip designers buying HBM on price alone and moving volume freely between suppliers from one generation to the next.
Micron's HBM4 reaching volume later than its rivals' or failing a major customer's qualification.
The memory makers' combined spending on new factories rising far faster than the growth in memory bits shipped.

Drafted by claude-opus-5.5 from Micron annual report (10-K), 3 Oct 2025 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Candidate to be worth a trillion dollars

SK Hynix

Show the profile

Market: SK hynix makes DRAM (the working memory a processor reads and writes constantly), NAND (storage that keeps data when power is off) and high-bandwidth memory, or HBM (DRAM dies, single chips cut from a silicon wafer, stacked and placed beside an AI chip so they can feed it fast). Its own results post for 2025 says record results came from AI memory, that its HBM revenue grew sharply, that demand for ordinary server memory also rose sharply, and that it was the first in the industry to complete the preparations to mass-produce HBM4, the next generation, while still supplying the previous one. It is building new factories in Korea and advanced packaging plants (where chips and memory are mounted together) in Korea and Indiana. It files no reports with the US regulator, so this site records no filed figures for it; its only source here is that short post, which is the company describing itself, not an audited filing. This site judges memory to be historically cyclical and commoditised, and a record year says nothing on its own about the next downturn.
Customers (concentrated): The results post names no customers and gives no concentration figures. It describes demand from AI memory and from server memory and storage. That a few AI chip designers buy most HBM is this site's judgement from the market's shape, not from this source.
Powers: process power: Emerging, not proven. Its post says it was the first in the industry to be ready to mass-produce HBM4 and supplies both current generations at once; this site judges the benefit is earlier orders at the best prices, and the barrier is how long it takes to learn to stack dies with good yield (the share of units that work).
scale economies: Emerging, not proven. This site judges it one of three firms with DRAM factories at the newest processes; the cost of a new factory keeps newcomers out, though not its two peers.
Under the site's rent rule: a thin profit, kept by incumbent firms. Depends on: Memory (no reading), AI chips (moderate), Advanced chip packaging (no reading), Chip-making equipment (no reading; a model judges it probably moderate), Capital (slack), Electricity generation (tight).

What would have to hold by 2035

Its lead in HBM survives each new generation, so it keeps being among the first suppliers qualified (tested and approved) by the biggest AI chip designers.
Custom HBM, memory built to one buyer's design, which its post calls a key differentiator, becomes common, turning a lead in yield into switching costs that outlast any single shortage.
AI keeps memory demand ahead of the makers' capacity, so memory prices do not fall back to their old cycle.
Its new factories and packaging plants come online without it and its rivals overbuilding at the same time.
Its largest AI chip customers keep growing and do not spread their orders evenly across the memory makers.

What would show it failing

Samsung or Micron qualifying HBM4 with the largest AI chip buyer at the same time and in similar volume.
HBM prices per unit falling from one generation to the next while demand still grows.
Its operating margin, the share of sales left after running costs, falling back to the levels of an ordinary memory downturn while AI spending is still rising.
Custom HBM staying rare, with buyers treating HBM as interchangeable between makers.

Drafted by claude-opus-5.5 from SK Hynix announcement, 28 Jan 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Candidate to be worth a trillion dollars

Samsung Electronics

Show the profile

Market: Samsung Electronics is a conglomerate. Its chip division makes DRAM (working memory), NAND (storage) and high-bandwidth memory, or HBM (DRAM dies, single chips, stacked beside an AI chip so they can feed it fast), designs phone processors and camera sensors, and runs a foundry (a factory that makes chips designed by others). It also makes phones, televisions, home appliances, displays and car electronics. Its business report for 2025 says that, with low inventory and supply constraints in memory, it expanded HBM sales, and that it aims to supply HBM4 for new AI chips, both graphics processors and custom ones. The same report, citing the research firm DRAMeXchange, shows its share of world DRAM revenue falling in both 2024 and 2025. It signed a foundry contract with Tesla running to 2033. Samsung files no reports with the US regulator, so this site records no filed figures for it; its source here is its English business report.
Customers (diffuse): The report lists its major customers across the whole company, alphabetically: Alphabet, Apple, Deutsche Telekom, Hong Kong Techtronics and Supreme Electronics. It does not break out the buyers of its memory or HBM, which this site judges are far more concentrated than the company as a whole. The one long contract it lists is Tesla's, for foundry work.
Powers: scale economies: Emerging, not proven. This site judges it one of three firms with DRAM factories at the newest processes, with cash from phones and appliances to keep building through a downturn; the barrier holds against newcomers, less so against SK hynix.
Under the site's rent rule: a thin profit, kept by incumbent firms. Depends on: Memory (no reading), Chip foundries (moderate), AI chips (moderate), Advanced chip packaging (no reading), Chip-making equipment (no reading; a model judges it probably moderate), Capital (slack).

What would have to hold by 2035

It closes the HBM gap with SK hynix and is qualified (tested and approved) in volume by the largest AI chip designers for HBM4 and after.
AI keeps memory short for long enough that its large capacity earns scarcity prices rather than dragging prices down.
Its foundry wins more outside designs at the newest processes (the latest generations of chip-making), like the Tesla contract, so that it becomes a real second source, an alternative supplier buyers can switch orders to, to TSMC for AI chips.
Buyers value getting memory, foundry work and packaging (mounting chips and memory together) from one firm enough to pay for it.
Ordinary consumer electronics, a large part of its sales, do not absorb the gains from chips.

What would show it failing

Its share of world DRAM revenue falling further while memory is still short.
The largest AI chip designers qualifying HBM4 from SK hynix and Micron but taking little or none from Samsung.
Its foundry failing to win another large outside customer for its newest processes.
Memory prices falling back to their old cycle while AI spending is still rising, as its own large capacity comes online.

Drafted by claude-opus-5.5 from Samsung Electronics business report, year to 31 Dec 2025 (opens in a new tab), SK Hynix announcement, 28 Jan 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Already worth over a trillion dollars

NVIDIA Corp

Show the profile

Market: NVIDIA designs accelerators, chips built to do the huge number of simple sums that training and running AI models needs, mostly graphics processors (GPUs). It sells them as chips and as whole server racks, with the networking gear that ties many of them together and the software developers write their code against (CUDA, its programming platform). It does not make the chips itself: foundries, factories that make chips to other firms' designs, make its wafers (the silicon discs chips are cut from). Its filing says it brings out a new data-centre architecture every year. How big it can become depends, in this site's judgement, on how long the big buyers keep raising their spending on AI data centres.
Customers (concentrated): The filing lists cloud providers, newer specialist clouds it calls neoclouds, model makers, companies and governments. Most chips reach them through server makers, assemblers and distributors, though some cloud providers and model makers also buy directly. The filing also estimates that one AI research and deployment company contributed a meaningful amount of revenue by buying cloud services from its customers. In the year to January 2026 its two largest direct customers each took a large part of revenue, and in the first half of the following year three did; and the filing says some indirect buyers, reached through those channels, were also each large enough to disclose. This site judges that those end buyers are largely the same few companies designing chips of their own.
Powers: switching costs: Years of AI code, tools and staff skills are written for CUDA, so moving to another chip means rewriting and retesting; this site's filed figures show the benefit in a gross margin (the share of sales left after the direct cost of making them) well above AMD's even as rivals shipped. In 2026 it agreed to buy Hugging Face, a platform where developers share and run open models, and committed to keep it open to other chip makers.
scale economies: This site judges that as one of the largest buyers of wafers, advanced packaging and memory it spreads a very large research budget over the most units, and that a new architecture every year is a pace a smaller rival cannot fund from its own sales.
Under the site's rent rule: a moderate profit, kept by the company itself. Depends on: Chip foundries (moderate), Advanced chip packaging (no reading), Memory (no reading), AI chips (moderate), Networking and optics (no reading; a model judges it probably moderate), Electricity generation (tight), Data-centre buildings (tight), Capital (slack).

What would have to hold by 2035

Spending on AI data centres by the largest cloud and platform companies keeps rising into the 2030s instead of being cut back after a stretch where AI revenue fails to cover it.
Chips the big buyers design themselves (Google's, Amazon's, Meta's, and those Broadcom builds for them) stay a minority of AI computing shipped, so NVIDIA's buyers cannot replace it with their own designs.
NVIDIA's gross margin stays well above AMD's and Broadcom's after supply of chips, packaging and memory catches up with demand, which would show the rent (profit above what it takes to keep a business going, earned because something holds rivals off) rests on software lock-in and not only on shortage.
Software that lets the same model code run on any chip does not make the choice of chip a matter of price alone.
It keeps first call on the leading foundry's newest process and on advanced packaging (mounting the chip and its memory together on one base), and that foundry keeps operating in Taiwan without interruption.
Export controls stop short of shutting it out of markets beyond China, and a China-only product line is not the price of staying in the rest.
Its own guarantees, cloud-capacity commitments and equity stakes in customers stay small next to sales, so demand is paid for by buyers and not by NVIDIA's balance sheet; its latest quarterly filing adds a guarantee on OpenAI's data-centre leases and commitments to rent back computing from AI clouds that buy its systems, which puts this condition under strain.

What would show it failing

The site's non-NVIDIA share of AI computing shipped rising until chips from other designers make up most of the AI computing shipped, and still climbing.
NVIDIA's quarterly gross margin falling towards AMD's while the site's accelerator tightness word reads easing.
A top direct customer's share of revenue falling sharply in the same year that customer announces a larger own-chip deployment.
The largest cloud companies cutting capital spending quarter after quarter.
New export rules that go beyond the existing Gulf licensing to require licences for sales to allied countries, or a write-down of chips it cannot sell that is larger than the H20 and H200 charges already taken.
The site's reading of circular financing, chip makers funding the customers who buy their chips, growing faster than NVIDIA's revenue.
The site's chip makers' share of gross profit with the cloud and labs falling while chips stay short, which would undercut the view that whoever holds the scarcest input collects.

Drafted by claude-opus-5.5 from Broadcom annual report (10-K), 18 Dec 2025 (opens in a new tab), NVIDIA Corp annual report (10-K), 25 Feb 2026 (opens in a new tab), NVIDIA Corp quarterly report (10-Q), 26 Aug 2026 (opens in a new tab), NVIDIA Corp filing (8-K), 3 Sep 2026 (opens in a new tab), TSMC annual report (20-F), 16 Apr 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Already worth over a trillion dollars

Broadcom

Show the profile

Market: Broadcom sells two things. The first is chips, and in AI that means custom accelerators (chips it designs to each big buyer's own plan, which it calls XPUs), whole racks and systems built on them, plus the Ethernet switches, network cards and optical parts that connect them. The second is infrastructure software, mostly VMware, which lets a company run many virtual computers on its own servers. Its semiconductor segment, the chip business it reports separately, brought in most of its revenue in the year to November 2025. This site judges its AI market grows if the largest buyers shift from buying NVIDIA's chips to designing their own, because Broadcom does design and production work for such buyers. This profile judges its rent for its AI chip role, custom accelerators and networking, not for VMware.
Customers (concentrated): Hyperscalers (the handful of companies that run the largest clouds), companies building frontier AI models, system integrators, and OEMs (original equipment makers that build servers, switches and racks). The filing says distributors take a large part of its sales and a few end customers a large part of its revenue, a share that rose sharply in 2026, and it expects that concentration to continue. The annual filing does not name the custom-chip buyers, but a 2026 filing names Google, for future generations of its TPUs with supply through 2031, and Anthropic, which will take TPU computing capacity through Broadcom from 2027.
Powers: scale economies: It reuses the same high-speed connection circuits and packaging know-how across many customers' custom chips; this site judges the benefit is a lower cost per design than a buyer designing alone would bear, and the barrier is the volume across customers that a newcomer design partner lacks.
Under the site's rent rule: a fat profit, kept by the company itself. Depends on: Chip foundries (moderate), Advanced chip packaging (no reading), Memory (no reading), Networking and optics (no reading; a model judges it probably moderate), Data-centre buildings (tight), Capital (slack).

What would have to hold by 2035

Custom chips keep winning a growing share of AI computing from NVIDIA's chips, the opposite of what NVIDIA needs.
The big buyers keep paying an outside designer for that work instead of taking it in house or moving it to cheaper design partners, even as they learn to do more of it themselves.
Open-standard Ethernet built on its switch chips stays the way AI clusters are wired, rather than NVIDIA's own networking, which now includes Ethernet as well as NVLink and InfiniBand.
A few customers each keep ordering new generations, since losing a single programme would hit its AI revenue hard.
Wafers at the newest processes, advanced packaging (mounting a chip and its memory together) and high-bandwidth memory (memory stacked beside the chip so it can feed it fast) stay available to it, which today means the Taiwan foundry.
Its backstops of customers' rack leases and any notes it takes from customers stay small next to sales, so its AI revenue is paid for by buyers and not by Broadcom's balance sheet; its latest quarterly filing adds a backstop on one customer's rack leases and the possibility of taking that customer's convertible notes.

What would show it failing

A named hyperscaler announcing that its next accelerator generation will be designed in house or with another design partner.
The top five end customers' share rising while AI chip revenue falls, meaning dependence is deepening as the business shrinks.
The semiconductor segment's operating margin, the share of sales left after running costs, falling while AI chip revenue grows, a sign buyers are winning the price talks.
NVIDIA's networking taking the larger share of new AI clusters' switch spending.

Drafted by claude-opus-5.5 from Broadcom annual report (10-K), 18 Dec 2025 (opens in a new tab), Broadcom quarterly report (10-Q), 10 Sep 2026 (opens in a new tab), Broadcom filing (8-K), 6 Apr 2026 (opens in a new tab), NVIDIA Corp annual report (10-K), 25 Feb 2026 (opens in a new tab), TSMC annual report (20-F), 16 Apr 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Advanced Micro Devices

Show the profile

Market: AMD sells processors for servers (EPYC), AI accelerators (Instinct, graphics processors, or GPUs, built for AI), chips for personal computers and games consoles, and programmable chips for industry. It offers rack-scale designs for AI and its own open software platform, ROCm, to compete with NVIDIA's CUDA, the software developers write AI code against. This site judges it the main other seller of general-purpose AI accelerators besides NVIDIA, and that its AI market is whatever share buyers give to a second source, an alternative supplier they can switch orders to: large if buyers want to stop depending on one supplier, small if NVIDIA and custom chips split the market between them.
Customers (moderate): Large cloud companies, server makers and assemblers. The filing says a small number of customers account for a substantial part of revenue. In October 2025 it signed an agreement with OpenAI to deploy a very large amount of its GPUs, measured by the electricity they will draw, and at the same time issued OpenAI a warrant, a right to buy AMD shares at a nominal price, that vests as OpenAI meets GPU purchase milestones and AMD's share price hits targets.
Powers: None that passes the test of a benefit behind a barrier today.
Under the site's rent rule: a thin profit, kept by the company itself. Depends on: Chip foundries (moderate), Advanced chip packaging (no reading), Memory (no reading), AI chips (moderate), Electricity generation (tight), Data-centre buildings (tight), Capital (slack).

What would have to hold by 2035

It earns a lasting advantage it does not hold today: its open software and lower-priced racks, sold to buyers who want a second supplier, become counter-positioning (a model NVIDIA could copy only by cutting its own margins) once they show in AMD's gross margin, the share of sales left after the direct cost of making them.
Buyers want a second general-purpose accelerator supplier badly enough to give it a large, lasting share, rather than filling that role with their own custom chips.
Its open software becomes good enough that moving code off CUDA costs little, which would erode NVIDIA's switching costs.
The OpenAI agreement is delivered on time and paid for. Because the warrant hands OpenAI AMD shares as it buys GPUs, this site reads it as close to vendor financing, a seller funding its buyer's purchases, so the deal must still pay once those shares are counted as part of the price, and must not leave AMD overly dependent on one buyer whose own finances are unproven.
It keeps winning server processor share from Intel and holds off processors built on Arm designs (a rival chip blueprint licensed to many firms) by the cloud companies themselves.
It keeps enough of the Taiwan foundry's capacity at the newest processes and of its advanced packaging (mounting a chip and its memory together), where it competes for the same slots as NVIDIA and Broadcom.
Export controls leave the markets outside China open; the filing records an inventory charge after a product for China needed an export licence.

What would show it failing

Its data centre segment growing more slowly than NVIDIA's data centre revenue for years running.
The OpenAI deployment being delayed, shrunk or renegotiated.
OpenAI's GPU purchases meeting the warrant's milestones without AMD's gross margin rising, showing the volume was bought with its own shares.
Its gross margin failing to rise towards NVIDIA's as accelerators become a larger part of sales.
The site's non-NVIDIA share of AI computing rising while AMD's share within it falls, meaning custom chips took the second-source role.
Major cloud companies' own Arm server processors taking share from both AMD and Intel.

Drafted by claude-opus-5.5 from Advanced Micro Devices annual report (10-K), 4 Feb 2026 (opens in a new tab), TSMC annual report (20-F), 16 Apr 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Already worth over a trillion dollars

TSMC

Show the profile

Market: TSMC is a foundry, a company that manufactures chips other firms design, printing them on wafers (thin silicon discs later cut into chips). It also does the advanced packaging (which its filing sells as silicon stacking and packaging services, CoWoS among them: ways of mounting or stacking a processor and its memory together) that AI accelerators need. High-performance computing, which includes AI chips, made up most of its revenue in 2025, and chips at its most advanced nodes (its newest manufacturing processes, which print the smallest features) made up most of its wafer revenue. Its filing calls its scale and capacity, particularly for advanced technologies, a major competitive advantage. This site judges its market grows with every chip designer's AI demand, whoever among them wins. TSMC files a 20-F (the annual report foreign companies file with the US regulator) under IFRS, the international accounting rules, so this site's filing reader records no filed figures for it.
Customers (concentrated): Many customers, but its ten largest took most of its revenue in 2025, a share that has risen each year since 2023, while its single largest customer's share fell over the same years and its second largest customer's share rose sharply. The filing does not name them. The NVIDIA, Broadcom and AMD filings each name TSMC among the foundries that make their wafers.
Powers: scale economies: Its filing calls its scale and capacity at advanced technologies a major competitive advantage; this site judges that a factory for the newest process costs so much that only the foundry with the most such orders can fill one.
process power: This site judges it brings new processes to high yield (the share of working chips per wafer) faster than rivals, learning that takes years; its filing notes that others have invested heavily to compete.
switching costs: This site judges that a chip is designed to one foundry's rules, so moving costs a redesign and many months, and that packaging ties the chip to the same supplier.
Under the site's rent rule: a monopoly-like profit, kept by the company itself. Depends on: Chip-making equipment (no reading; a model judges it probably moderate), Electricity generation (tight), Critical minerals (no reading; a model judges it probably moderate), Advanced chip packaging (no reading), Capital (slack).

What would have to hold by 2035

No blockade, conflict or political settlement across the Taiwan Strait stops or controls its Taiwan factories, where most of its capacity for the newest processes still sits.
Its factories in Arizona, Japan and Germany reach yields and costs close enough to Taiwan's that shifting output there does not cut deep into its margins.
Intel's and Samsung's foundries stay behind at the newest process, so no second source, an alternative supplier, is good enough for AI chip designers to split their orders.
Demand for wafers at the newest processes keeps growing into the 2030s; unlike its customers, it does not need any one chip designer to win, only that the total grows.
Taiwan keeps supplying enough electricity and water for new factories, which the filing lists among its risks.
Tariffs and US policy do not force it to build or price in ways that erase the cost edge of its Taiwan base.

What would show it failing

Any interruption to shipping or factory operation in Taiwan lasting weeks.
A leading AI chip designer moving its flagship accelerator to Intel or Samsung at the newest process.
Overseas factories diluting its gross margin, the share of sales left after the direct cost of making them, year after year instead of converging.
The ten largest customers' share of revenue falling while total revenue falls, meaning big designers are leaving and not being replaced.
Power rationing or water shortages in Taiwan delaying new factories.
Its yearly gross margin falling for years running while the site's foundry tightness word reads easing.

Drafted by claude-opus-5.5 from Advanced Micro Devices annual report (10-K), 4 Feb 2026 (opens in a new tab), Broadcom annual report (10-K), 18 Dec 2025 (opens in a new tab), NVIDIA Corp annual report (10-K), 25 Feb 2026 (opens in a new tab), TSMC annual report (20-F), 16 Apr 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Already worth over a trillion dollars

Microsoft Corp

Show the profile

Market: Microsoft sells software and cloud services. Its Intelligent Cloud segment (a part of the business it reports separately) holds Azure, its public cloud: computing, storage and AI services rented from its own data centres around the world and billed mostly by how much customers use. The filing says Azure offers AI computing at very large scale, the latest models, and hardware that includes chips Microsoft designs itself alongside chips from partners. It also sells AI inside products businesses already pay for, such as its office software and GitHub. Until mid-2026 Azure sat in its Intelligent Cloud segment; from the year to June 2027 Microsoft reports it inside a new Agents and Infra segment with its office software, GitHub and licensing, and its reported Azure figure excludes GitHub and Security Copilot, which move to its office-software cloud. In the year to June 2026 Azure revenue grew strongly, and the filing says the gross margin percentage of its cloud business (the share of sales left after the direct cost of providing it) fell because of spending on AI infrastructure and rising AI use. It warns that its AI and cloud investments will keep raising costs and may lower operating margins. This site judges its market to be both renting computing for training AI models and for inference (running a trained model to answer requests), and selling AI features to the companies that already run on its software. How big that becomes depends on whether AI revenue grows into the spending.
Customers (moderate): Businesses of every size, developers and governments; the filing gives no single customer's share of total revenue. It does name OpenAI. Microsoft holds a large minority stake in OpenAI, accounted for by the equity method (it records a share of OpenAI's results rather than adding in OpenAI's whole accounts), which makes OpenAI a related party (a company close enough that dealings with it must be disclosed); it discloses revenue from commercial arrangements with OpenAI, including revenue-sharing payments, and holds rights to OpenAI's models and other intellectual property for use in its products. The filing says some partners that supply AI models are also significant Azure customers and may change how much they buy. Its commercial remaining performance obligations (revenue already contracted but not yet delivered or recognised) grew sharply; this site's records, from an earlier Microsoft quarterly filing, show a large part of that backlog came from OpenAI. This site's records also hold an Anthropic commitment to buy Azure computing. The CoreWeave and Nebius filings name Microsoft as a large buyer of their capacity, so Microsoft also rents from neoclouds (newer clouds that specialise in renting AI chips). David Teece, an economist who studied why inventors often fail to keep the returns from what they invent, found that the firm holding the complementary assets (the distribution, sales and systems a product needs to reach buyers) often keeps the profit; this site judges that Microsoft holds its own: its data centres and its customers.
Powers: switching costs: Companies run their identity, office software, servers and data on Microsoft together, and the filing names hybrid cloud (linking a customer's own data centres with Azure) as Azure's advantage; this site judges the benefit shows in an Intelligent Cloud operating margin (the share of sales left after running costs) that this site's filed figures put at or near the top of the big clouds, a segment Microsoft stops reporting from the year to June 2027, and the barrier is the cost of moving years of systems. This holds for the work firms already run there; this site judges it is not yet proven for AI work in particular.
scale economies: This site judges that a worldwide network of data centres spreads the cost of chips, power contracts and its own chip designs over the largest volumes, so a rival would need similar volume to match its cost; the filing's note that AI spending is lowering its cloud gross margin shows this benefit is under strain.
Under the site's rent rule: a fat profit, kept by the company itself. Depends on: AI chips (moderate), Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), Capital (slack), Deployment platforms (no reading; a model judges it probably easing).

What would have to hold by 2035

Revenue from AI, in Azure and inside its software, grows into its capital spending (money spent on buildings and equipment) before the servers it bought wear out; the filing depreciates servers and network gear over several years at most.
OpenAI keeps growing, pays what it has committed, and keeps much of its computing on Azure; the filing lists OpenAI as a related party and warns that partners who are big Azure customers may change their purchasing.
Its rights to OpenAI's models keep their value even as OpenAI also buys computing from other clouds, which the CoreWeave filing shows it does.
Businesses keep buying AI through Microsoft contracts they already have, rather than directly from the labs that make the models.
Its own chip designs carry a growing share of its AI work, so it depends less on buying NVIDIA's chips at NVIDIA's prices.
Power, components and the large data-centre leases it has signed but not yet started all arrive on time; the filing warns that power limits and component shortages could restrict its capacity.
Prices for AI services do not fall faster than its costs; the filing names competition and commoditisation, goods becoming interchangeable and sold on price, as a risk to AI pricing.

What would show it failing

The Agents and Infra operating margin, or the Microsoft Cloud gross margin, in this site's filed figures falling quarter after quarter while capital spending keeps rising.
Commercial remaining performance obligations shrinking against revenue, the test this site's backlog claim sets.
OpenAI moving a large block of its committed spending to other clouds, renegotiating it down, or failing to pay.
A write-down of AI servers, a shortening of their useful lives, or unwinding of data-centre leases before they start.
This site's ratio of the big platforms' capital spending to AI revenue rising for years instead of falling.

Drafted by claude-opus-5.5 from CoreWeave Inc. annual report (10-K), 2 Mar 2026 (opens in a new tab), Microsoft Corp annual report (10-K), 29 Jul 2026 (opens in a new tab), Microsoft Corp filing (8-K), 2 Sep 2026 (opens in a new tab), Nebius Group N.V. annual report (20-F), 30 Apr 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Already worth over a trillion dollars

Amazon.com Inc.

Show the profile

Market: Amazon's cloud arm, Amazon Web Services (AWS), is a segment it reports separately. It rents on-demand computing, storage, databases, analytics and AI and machine-learning services to developers and to organisations of every size, from start-ups to government agencies. The filing says its cash capital spending (money spent on buildings and equipment) rose sharply in 2025, mostly on technology infrastructure, the majority of it to support AWS growth, and that it expected that spending to increase again in 2026. In 2026 it also began borrowing heavily in several currencies and said it expects further financing that year. Having lengthened the estimated useful life of its servers in 2024, it shortened it for some servers and network equipment in 2025, citing the faster pace of AI technology, which raised its depreciation (the yearly charge for wear on equipment). This site's filed figures show the AWS operating margin (the share of sales left after running costs) holding steady through that record spending. This site judges its market to be the rented computing on which companies and AI labs train and run models, and that its size depends on how much AI work moves into the cloud rather than staying on companies' own machines or moving to the labs' own data centres.
Customers (diffuse): Developers and enterprises of all sizes; the filing discloses no customer large enough to report separately, and it says AWS holds most of its long-term customer commitments not yet recognised as revenue. It names Anthropic, a maker of AI models: Amazon has invested in Anthropic through convertible notes (loans that can turn into shares) and non-voting preferred shares (shares that rank ahead of ordinary shares for payouts but carry no vote), recorded large gains on those stakes in 2025, bought more preferred shares in 2026, and set up a loan facility that opens up as AWS delivers computing to Anthropic, taking Anthropic notes, or shares after a listing, for what is drawn; its large gains on the stake come from Anthropic's new funding rounds. It has a commercial arrangement with Anthropic mainly for AWS cloud services, including the use of AWS's own chips. In 2026 it also bought preferred shares in OpenAI, paid the rest of its commitment soon after, and widened its AWS arrangement with OpenAI; its quarterly filing names Anthropic and OpenAI as its main private stakes. This site's records also hold an Anthropic commitment to Microsoft's cloud, so Anthropic is not AWS's alone. David Teece, an economist who studied why inventors often fail to keep the returns from what they invent, found that the firm holding the complementary assets (the distribution, sales and systems a product needs to reach buyers) often keeps the profit; this site judges that Amazon holds its own: its data centres and its customers.
Powers: switching costs: The filing lists customers' ability and willingness to change business practices among the competitive factors for its enterprise services; this site judges the benefit is an AWS operating margin that held steady in this site's filed figures while spending surged, and the barrier is the work of moving data and systems built on AWS services. This holds for the work firms already run there; this site judges it is not yet proven for AI work in particular.
scale economies: This site judges that the largest pool of rented servers, networks and data centres spreads fixed costs, including its own chip designs, over more use than a smaller cloud can; the barrier is the volume a rival would need to match that cost.
Under the site's rent rule: a fat profit, kept by the company itself. Depends on: AI chips (moderate), Memory (no reading), Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), Capital (slack).

What would have to hold by 2035

AI revenue on AWS grows into the capital spending before the servers wear out; the filing already shortened server lives because AI hardware ages faster.
Anthropic and OpenAI both keep growing, keep much of their work on AWS, and pay; Amazon now holds large stakes in both and has long cloud contracts with both, so the value of those stakes and contracts rests on the two labs.
AWS's own AI chips carry a large and growing share of the work, so Amazon is not left paying NVIDIA's prices for most of its AI capacity.
Most companies choose to run AI where their data already sits, on AWS, rather than on specialist neoclouds (newer clouds that rent out AI chips) or through the labs' own services.
Power, land and chips, including memory chips whose supply the filing calls volatile, arrive fast enough that data centres are neither short nor stranded; the filing warns that mis-predicting demand can leave excess or too little capacity and lead to impairment charges (write-downs of an asset's recorded value).

What would show it failing

The AWS operating margin in this site's filed figures falling for several quarters running while capital spending keeps rising.
Anthropic moving the bulk of its new computing to other clouds or to data centres of its own.
A write-down of Amazon's stakes in Anthropic or OpenAI, or of AI servers bought in the build-out; or Anthropic paying for much of its AWS computing with money drawn from Amazon's own facility.
Another shortening of server useful lives within a few years of the last one.
AWS's growth in rented AI computing trailing the neoclouds' for several years.

Drafted by claude-opus-5.5 from Amazon.com Inc. annual report (10-K), 6 Feb 2026 (opens in a new tab), Amazon.com Inc. quarterly report (10-Q), 31 Jul 2026 (opens in a new tab), OpenAI filing (8-K), 27 Feb 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Already worth over a trillion dollars

Alphabet Inc.

Show the profile

Market: Alphabet sits in two parts of the chain. As a cloud, Google Cloud rents computing, storage and AI services, offering both graphics processors (GPUs) and its own tensor processing units (TPUs, chips Google designs for the maths AI models run on; the filing names Ironwood as its seventh generation), plus Vertex AI, a platform for building on its Gemini models and other companies' models, and Workspace office software. As a frontier lab (a company building the most capable AI models), it develops the Gemini models, which the filing says run across all of its largest consumer products, including Search through AI Overviews and AI Mode. The filing says the cost of developing its general AI models sits in Alphabet-wide costs rather than in the Google Cloud segment, and that margins on its revenue outside advertising, Google Cloud among it, are generally lower than advertising margins. It spent heavily on capital (buildings and equipment) in 2025 and said it expected to raise that spending significantly in 2026; in 2026 it also began selling new shares and mandatory convertible preferred stock, set up a programme to sell more shares over time, and borrowed heavily in several currencies. This site's filed figures show the Google Cloud operating margin (the share of sales left after running costs) rising sharply over the past year from the lowest among the big clouds. This site judges its market is both renting AI computing and selling AI inside its most widely used consumer products, and that owning chips, models and distribution together is rare.
Customers (diffuse): Advertisers across the world pay for most of Alphabet's revenue, and the filing names no customer large enough to report separately. Google Cloud sells to businesses and developers; the filing says its remaining performance obligations (revenue contracted but not yet recognised) relate primarily to Google Cloud, and that it has a number of large, long-duration commercial agreements. It also says it has made debt and equity investments in companies with which it signed commercial arrangements at or near the same time, and that if such an investment were impaired it might have to reduce the related contracted revenue; the filing does not name those companies. Google invested again in Anthropic in April 2026 and promised more if Anthropic meets performance targets; Alphabet's quarterly filing records forward funding commitments of that kind, tied to milestones, on a strategic investment it does not name. Anthropic has also signed for Google's TPU capacity. Alphabet enters significant leases with outside data-centre operators, and has agreed to backstop payments on certain third-party data-centre leases (to cover them if the tenant defaults), with the right to take over the leases.
Powers: scale economies: One infrastructure serves Search, YouTube, Gemini and Google Cloud; this site judges the benefit is a lower cost for each unit of AI computing, and the barrier is the volume across its consumer products that a cloud-only rival does not have.
process power: Process power is a way of working that takes rivals years to learn. It has designed TPUs through generation after generation alongside its own models, and the filing says it focuses on driving efficiency in its data centres; this site judges that designing chips and models together drives part of that efficiency, that the benefit is AI computing that depends less on buying NVIDIA's chips, and that the barrier is the years of joint chip, software and model work a rival would have to repeat.
Under the site's rent rule: a monopoly-like profit, kept by the company itself. Depends on: AI chips (moderate), Chip foundries (moderate), Data-centre buildings (tight), Electricity generation (tight), Building frontier models (easing), Running models (easing), Capital (slack).

What would have to hold by 2035

Gemini stays among the most capable models, so Alphabet does not have to pay another lab for the models that run its own products.
AI answers in Search keep earning at least what the old results page earned, so advertising keeps carrying most of the build-out; in 2026 Alphabet also began selling new shares and convertible preferred stock and borrowing heavily, so the build-out now leans on outside capital too.
TPUs stay competitive with NVIDIA's chips generation after generation, and outside customers keep choosing them, turning a cost advantage into a cloud business.
The Google Cloud margin keeps rising as the contracted backlog turns into revenue, even as the filing warns that margins outside advertising, cloud among them, are generally lower than advertising's.
Companies it invested in and also sells cloud services to stay solvent and pay, and the backstops it has given on data-centre leases are not called.
Its capital spending earns back before the equipment ages out; the filing warns that expected technology advances could change the period over which it benefits from these assets.

What would show it failing

The Google Cloud operating margin in this site's filed figures turning down while capital spending keeps rising.
A payment under one of its data-centre lease backstops, or a cut to contracted revenue after an impaired investment.
Large TPU customers moving their next generation of work back to GPUs.
Search revenue falling while more answers are given by AI, showing the answers earn less than the links they replace.
Gemini falling well behind rival models for several model generations in a row.

Drafted by claude-opus-5.5 from TechCrunch article, 24 Apr 2026 (opens in a new tab), Alphabet Inc. annual report (10-K), 5 Feb 2026 (opens in a new tab), Alphabet Inc. quarterly report (10-Q), 23 Jul 2026 (opens in a new tab), Alphabet Inc. filing (8-K), 4 Jun 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Oracle Corp

Show the profile

Market: Oracle sells database and business software and, increasingly, Oracle Cloud Infrastructure (OCI): rented computing from virtual machines and GPU (graphics processor) servers to bare-metal servers (whole machines rented without sharing), used for AI training and inference (running trained models), plus storage and networking. It also runs its database inside other companies' clouds, which it calls multicloud. The filing says cloud infrastructure became the larger part of its cloud revenue in the year to May 2026. To grow OCI it is spending heavily: its capital spending (money on buildings and equipment) rose steeply, its free cash flow (cash from operations after that spending, a measure the company itself reports), slightly negative the year before, turned deeply negative, and it issued new bonds, has mandatory convertible preferred shares (which must turn into ordinary shares later) outstanding, set up a programme to sell new shares into the market over time, from which it had sold none by the end of May 2026, and signed long leases with outside data-centre developers and large commitments to chip and equipment suppliers. Its remaining performance obligations (revenue contracted but not yet recognised) grew steeply, which the filing puts down mainly to certain significant cloud contracts signed in the year. This site judges its market is renting AI computing at very large scale to a few big buyers, a business whose size depends on those buyers' ability to pay.
Customers (concentrated): The filing names no AI customer. It says that in certain OCI offerings it is more concentrated among a number of large customers, that some customers may be highly leveraged (heavily borrowed), and that it takes prepayments on some committed contracts to manage credit risk. This site's records hold a press-reported agreement for OpenAI to buy computing from Oracle, which the filing does not itemise; nor does the filing mention the Stargate project by name. Oracle's older database and application customers are many and spread out. This site judges that the racks it rents out for AI are a generic asset, which anyone with money, power and a supply of chips can assemble; the scarce inputs behind that business belong to others, NVIDIA's chips on one side and the demand of a few big buyers on the other.
Powers: None that passes the test of a benefit behind a barrier today.
Under the site's rent rule: competed away to users. Depends on: AI chips (moderate), Memory (no reading), Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), Capital (slack).

What would have to hold by 2035

It earns a lasting advantage it does not hold today in renting AI computing, for example if the switching costs (the cost to a customer of moving to a rival) of its database carry over, so that companies running Oracle's database choose OCI for their AI work.
Its few large AI customers keep growing and pay; the filing warns that if a key customer cannot pay, it could be locked into multi-year commitments for data-centre space and financing without the matching revenue.
Its data-centre leases, which the filing says do not line up in length or pricing with its customer contracts, can be re-let or reused if customers do not renew.
Lenders and shareholders keep funding the build-out on terms the contracts can carry while its free cash flow is negative.
Power, sites, GPUs and memory arrive on time; the filing warns of power shortages, limits on building data centres, and developers who fail to finance or finish construction.
The price of renting GPUs does not fall below what it costs Oracle to own, power and finance them before the contracts end.

What would show it failing

Oracle's remaining performance obligations falling, or a large contract being cut or delayed, in this site's filed figures.
A write-down (a cut in the value it records) of data-centre assets, or a charge for leases it cannot fill.
Its borrowing costs rising or a credit downgrade forcing it to slow the build-out.
Capital spending falling while contracted revenue stays unconverted, a sign that capacity is not being delivered or not wanted.
A large customer failing to pay or seeking to renegotiate its commitment.

Drafted by claude-opus-5.5 from Oracle Corp annual report (10-K), 22 Jun 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Candidate to be worth a trillion dollars

CoreWeave Inc.

Show the profile

Market: CoreWeave is a neocloud, a newer cloud built only to rent out AI computing. It rents large clusters of NVIDIA graphics processors (GPUs) with the networking, storage and management software to run them, and has bought developer-tool companies including Weights & Biases. The filing says it has been among the first to bring each new NVIDIA system to market and expects to be among the first with NVIDIA's next platform, Rubin. Customers mostly sign multi-year take-or-pay contracts (they pay for reserved capacity whether or not they use it), and it pays for its build-out mainly with debt secured on specific assets and backed by those contracts, plus company-level shares and debt. Its revenue and remaining performance obligations (revenue contracted but not yet recognised) grew steeply, and it has reported a net loss every year shown in the filing. This site judges its market is the overflow of AI computing demand that the big clouds and labs cannot or will not build themselves, and that its size depends on how long that overflow lasts.
Customers (concentrated): A few very large buyers. The filing says Microsoft, its top customer, accounted for most of its revenue in 2025 and that no other customer reached the reporting threshold. It has signed contracts under which OpenAI and Meta committed to large multi-year purchases and expects both to become significant customers; it also has a stock issuance agreement with OpenAI. It expects its dependence on a limited number of customers to continue. NVIDIA supplies all of its GPUs, since its customer contracts specify NVIDIA chips, and bought CoreWeave shares in January 2026; this site's records also hold an arrangement giving NVIDIA access to CoreWeave capacity left unsold. Its competitors include the big clouds, several of which are its customers. This site judges that the racks it rents out are a generic asset, which anyone with money, power and a supply of chips can assemble; the scarce inputs behind its business belong to others, NVIDIA's chips on one side and the demand of a few big buyers on the other.
Powers: None that passes the test of a benefit behind a barrier today.
Under the site's rent rule: competed away to users. Depends on: AI chips (moderate), Networking and optics (no reading; a model judges it probably moderate), Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), Capital (slack).

What would have to hold by 2035

It earns a lasting advantage it does not hold today: its skill at bringing new NVIDIA systems online first and running them at high utilisation becomes process power (a way of working that takes rivals years to learn) that shows in margins after GPUs stop being scarce.
Its customers widen well beyond Microsoft, OpenAI and Meta, so losing one buyer does not break the business.
Those buyers keep paying through the full terms; the filing warns that some customers are early-stage or private companies with a higher risk of not paying.
Its cost of borrowing falls as it matures; this site's reading of its facility rates is too new to show a direction.
GPUs keep earning after their first contracts end; the filing warns that failing to redeploy equipment past its contracted life could hurt results.
Its big customers do not build enough computing of their own, a risk the filing names, to stop renting from it.
The price of renting GPUs does not fall faster than the cost of owning and financing them.

What would show it failing

Microsoft not renewing or shrinking its contracts without other buyers filling the gap.
The average rate on its borrowing, in this site's readings, rising over several quarters.
A write-down of GPUs, or older GPUs re-let at prices far below their first contracts.
Delays in delivering capacity that lead customers to cancel parts of their contracts.
OpenAI or Meta moving work to data centres of their own and cutting what they buy from it.

Drafted by claude-opus-5.5 from CoreWeave Inc. annual report (10-K), 2 Mar 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Candidate to be worth a trillion dollars

Nebius Group N.V.

Show the profile

Market: Nebius is a neocloud, a newer cloud built to rent out AI computing. It rents clusters of NVIDIA graphics processors (GPUs), including the GB200 and GB300 systems, in a data centre it owns in Finland and in built-to-order or shared sites in the United States, Europe and Israel, and sells software on top, including Token Factory, which serves open AI models and charges by the token (a small chunk of text a model reads or writes) instead of by the GPU-hour. It also owns smaller businesses, among them Avride, a self-driving developer. The company was called Yandex N.V. until 2024, when it sold all its businesses in Russia, which were almost all of the group at the time, and kept the AI cloud and a few smaller units. It files a 20-F (the annual report foreign companies file with the US regulator), but prepares its accounts under US accounting rules. This site holds filed figures on its debt, including its total debt and borrowing rates, but no filed revenue or capital-spending figures for it. This site judges its market is AI computing for customers the big clouds do not serve first, plus overflow capacity for the big clouds themselves.
Customers (concentrated): Until recently mostly short-term customers, the filing says. It has since signed multi-year contracts to provide dedicated GPU capacity to Microsoft, at a new site in New Jersey, and to Meta; both contracts give the buyer service credits (discounts owed when service falls short) or the right to end tranches (separate portions of the deal) if delivery is late. One Meta order obliges Meta to buy any capacity in certain clusters that Nebius fails to sell to other customers. Its largest customer took a large part of 2025 revenue, and one customer owed most of its receivables (money customers owe it) at the end of 2025. It relies on NVIDIA for its GPUs and named a strategic partnership with NVIDIA in March 2026. It funds itself with convertible notes (loans that can turn into shares), share sales, and advance payments from these large contracts. This site judges that the racks it rents out are a generic asset, which anyone with money, power and a supply of chips can assemble; the scarce inputs behind its business belong to others, NVIDIA's chips on one side and the demand of a few big buyers on the other.
Powers: None that passes the test of a benefit behind a barrier today.
Under the site's rent rule: competed away to users. Depends on: AI chips (moderate), Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), Capital (slack), Running models (easing).

What would have to hold by 2035

It earns a lasting advantage it does not hold today, for example if Token Factory and its other software turn it from a renter of chips into a platform customers would lose work by leaving.
It delivers the Microsoft and Meta capacity on schedule; the filing says it has limited experience delivering contracts this large and that late delivery lets those buyers claim credits or end tranches (portions of the contract).
It wins enough other customers that its dependence on a few big clouds shrinks as their contracts run down.
Capital stays available on terms its contracts can carry while it spends ahead of revenue.
The price of renting GPUs does not fall faster than the cost of owning, powering and financing them.
The big clouds and labs do not build enough of their own capacity to stop needing neoclouds.

What would show it failing

Microsoft or Meta ending tranches for late delivery or not renewing.
Meta having to buy large amounts of capacity Nebius could not sell, a sign that outside demand is thin.
The borrowing rates in this site's readings of its debt rising over several quarters.
Token Factory staying a small part of revenue while GPU rental prices fall.

Drafted by claude-opus-5.5 from Nebius Group N.V. annual report (20-F), 30 Apr 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Candidate to be worth a trillion dollars

Fluidstack

Show the profile

Market: Fluidstack is a neocloud (a newer company that rents out AI computing power, competing with the large clouds of Amazon, Microsoft and Google). In November 2025 Anthropic chose it to build data centres custom-made for Anthropic's work in New York and Texas, with more sites to come and the first coming online during 2026; it says Anthropic chose it for how fast it can deliver gigawatts of power (a gigawatt is a very large amount of electric power, about what a big power station produces). Its Form D (the notice a private company files with the US regulator when it sells shares without a public offering) shows a UK company that sold part of a new offering of shares to a small group of investors in June 2026, with Morgan Stanley as its paid placement agent (the bank that finds buyers for the shares); it declined to disclose its revenue. How big the market becomes depends on whether it can become a builder for many buyers or stays tied to one.
Customers (concentrated): Its one named customer in these sources is Anthropic. It says it also works with other AI labs, governments and enterprises, but names none. This site judges its revenue is concentrated. This site judges that the racks it builds are a generic asset, which others with money, power and chips can also assemble; a site built for one customer gives that customer, not Fluidstack, the stronger hand once it is built.
Powers: None that passes the test of a benefit behind a barrier today.
Under the site's rent rule: competed away to users. Depends on: Data-centre buildings (tight), Electricity generation (tight), The grid and its hookups (no reading; a model judges it probably tight), AI chips (moderate), Networking and optics (no reading; a model judges it probably moderate), Capital (slack).

What would have to hold by 2035

Power and ready sites stay scarce long enough that speed of delivery earns it a better price than a larger cloud would charge.
Anthropic keeps paying for the sites built for it; Oliver Williamson, who studied why firms own assets tied to a single partner, found that such assets invite a fight over price once they are built, and a data centre made for one customer is that kind of asset.
It wins large customers beyond Anthropic, so the loss of one does not sink it.
It can keep raising equity and debt to build ahead of payment, since its private offering was only partly sold at filing.

What would show it failing

Anthropic moving work to the large clouds' or its own sites instead of expanding with Fluidstack.
Sites coming online late, or with power it has not secured.
No second customer of Anthropic's size named within the build-out.
Its next sites being delayed for want of equity or debt to build them.

Drafted by claude-opus-5.5 from Anthropic announcement, 28 May 2026 (opens in a new tab), Fluidstack announcement, 12 Nov 2025 (opens in a new tab), Fluidstack Form D, 30 Jun 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Etched

Show the profile

Market: Etched designs systems for inference (running a trained AI model to answer requests) on the largest models: its own chips, plus the racks (cabinets holding many chips wired together), cooling, power supply and software around them, designed together. Its chip is a kind of ASIC (an application-specific integrated circuit, a chip built for one kind of work rather than general use). It says its first chip came back from TSMC, the contract chip maker, earlier in 2026, that it is testing its first rack-scale product with customers, and that it has begun production against customer contracts. It claims design advances: running its maths circuits at low voltage so they can work harder without overheating, and a memory design that mixes HBM (high-bandwidth memory, stacked beside the chip to feed it fast) with SRAM (very fast memory built into the chip) and pools it across chips. It has opened a factory in Taiwan. How big the market becomes depends on whether buyers want a specialised inference machine once NVIDIA and the clouds' own chips also chase inference.
Customers (concentrated): It names no customers; it says it designed with leading AI companies, cloud providers and hyperscalers (the largest cloud companies). This site judges its early contracts come from a few buyers.
Powers: None that passes the test of a benefit behind a barrier today.
Under the site's rent rule: a thin profit, kept by incumbent firms. Depends on: Chip foundries (moderate), Advanced chip packaging (no reading), Memory (no reading), AI chips (moderate), Networking and optics (no reading; a model judges it probably moderate), Electricity generation (tight), Capital (slack).

What would have to hold by 2035

Its racks deliver, in customers' data centres, the throughput, speed of response and power savings it reports from early tests.
Inference becomes a large enough share of AI spending, and the leading models' designs stay stable enough, that a specialised machine beats general chips.
It gets enough capacity from TSMC and the memory makers; this site judges they serve NVIDIA and the large clouds first, though Etched says it is deepening its partnership with the leading chip manufacturer and has a strategic investment from VentureTech Alliance.
Its customers buy beyond first orders, and NVIDIA's and the clouds' own chips do not close the gap first.

What would show it failing

Delays in shipping its racks, or customer tests that do not match its claims.
A major model change, such as a new architecture (the basic design of a model), that its design runs poorly.
No named large customer after its first production run.

Drafted by claude-opus-5.5 from Etched announcement, 30 Jun 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it.

Indicators

Rent concentration in semiconductors

concentrating

75.0%as of 2026-07-26

Cloud backlog (remaining performance obligations)

concentrating

$664Bas of 2026-08-31

Semiconductors' share of stack operating income

concentrating

61.0%as of 2026-06-30

Chips' share of the stack's gross profit

concentrating

72.6%as of 2026-06-30

Cloud segment operating margins

concentrating

38.7%as of 2026-09-30

Circular financing scale

concentrating

$998Bas of 2026-09-30

Concentration among covered chip filers (HHI)

stable

0.87 indexas of 2026-06-30

Concentration among covered cloud filers (HHI)

stable

0.35 indexas of 2026-06-30

Non-Nvidia share of AI compute shipped (H100-equivalents)

emerging

33.8%as of 2026-06-30

Hyperscaler capex to AI revenue (trailing year)

dispersing

4.38×as of 2026-09-30

Hyperscaler server useful lives (10-K estimates)

emerging

5.00 yearsas of 2025-12-31

Neocloud borrowing cost (CoreWeave facility rates)

emerging

10.5%as of 2026-06-30

GPU rental price per hour

emerging

$3.85as of 2026-10-26

Training power draw growth (doubling time)

emerging

419 daysas of 2026-10-15

Hardware price-performance growth (doubling time)

consistent with normal

1331 daysas of 2026-08-02

Revenue already contracted

How much future revenue the companies at this layer have already signed, from their own filings. Context with no status of its own: a backlog reads contracts, not the demand behind them.

Revenue contracted but not yet delivered, in years of revenue, at the cloudscompany filings · each dot links to its row

CoreWeaveOracleMicrosoftAlphabet

measured

Remaining performance obligations are revenue a company has contracted but not yet delivered, from its own filings; each is divided here by its revenue in the last four fiscal quarters, so a reading of one means a year of current revenue is already sold. CoreWeave only rents out computing capacity; Oracle's, Microsoft's and Alphabet's backlogs also hold the software and services they sell, and Alphabet's is divided by revenue that comes mostly from advertising. Companies define and disclose backlog differently, so read each line against its own past; a line breaks where a quarter is missing here.

Derived (rpo_to_trailing_revenue) from: SEC EDGAR XBRL company facts. SEC EDGAR, XBRL company facts API.

The numbers

line	as of	value	from
CoreWeave	2026-06-30	13.7×	sec.crwv.revenue.fy Dec 2025 · sec.crwv.revenue.q, 5 rows · sec.crwv.rpo.q Jun 2026
CoreWeave	2026-03-31	15.9×	sec.crwv.revenue.fy Dec 2025 · sec.crwv.revenue.q, 4 rows · sec.crwv.rpo.q Mar 2026
CoreWeave	2025-12-31	11.8×	sec.crwv.revenue.fy Dec 2025 · sec.crwv.revenue.q Mar 2025 · sec.crwv.revenue.q Jun 2025 · sec.crwv.revenue.q Sep 2025 · sec.crwv.rpo.q Dec 2025
CoreWeave	2025-09-30	11.6×	sec.crwv.revenue.fy Dec 2024 · sec.crwv.revenue.q, 6 rows · sec.crwv.rpo.q Sep 2025
CoreWeave	2025-06-30	8.54×	sec.crwv.revenue.fy Dec 2024 · sec.crwv.revenue.q, 5 rows · sec.crwv.rpo.q Jun 2025
CoreWeave	2025-03-31	5.43×	sec.crwv.revenue.fy Dec 2024 · sec.crwv.revenue.q, 4 rows · sec.crwv.rpo.q Mar 2025
Oracle	2026-08-31	9.25×	sec.orcl.revenue.fy May 2026 · sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Aug 2026
Oracle	2026-05-31	9.47×	sec.orcl.revenue.fy May 2026 · sec.orcl.revenue.q Aug 2025 · sec.orcl.revenue.q Nov 2025 · sec.orcl.revenue.q Feb 2026 · sec.orcl.rpo.q May 2026
Oracle	2026-02-28	8.62×	sec.orcl.revenue.fy May 2025 · sec.orcl.revenue.q, 6 rows · sec.orcl.rpo.q Feb 2026
Oracle	2025-11-30	8.58×	sec.orcl.revenue.fy May 2025 · sec.orcl.revenue.q, 5 rows · sec.orcl.rpo.q Nov 2025
Oracle	2025-08-31	7.71×	sec.orcl.revenue.fy May 2025 · sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Aug 2025
Oracle	2025-05-31	2.40×	sec.orcl.revenue.fy May 2025 · sec.orcl.revenue.q Aug 2024 · sec.orcl.revenue.q Nov 2024 · sec.orcl.revenue.q Feb 2025 · sec.orcl.rpo.q May 2025
Oracle	2025-02-28	2.33×	sec.orcl.revenue.fy May 2024 · sec.orcl.revenue.q, 6 rows · sec.orcl.rpo.q Feb 2025
Oracle	2024-11-30	1.77×	sec.orcl.revenue.fy May 2024 · sec.orcl.revenue.q, 5 rows · sec.orcl.rpo.q Nov 2024
Oracle	2024-08-31	1.84×	sec.orcl.revenue.fy May 2024 · sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Aug 2024
Oracle	2024-05-31	1.85×	sec.orcl.revenue.fy May 2024 · sec.orcl.revenue.q Aug 2023 · sec.orcl.revenue.q Nov 2023 · sec.orcl.revenue.q Feb 2024 · sec.orcl.rpo.q May 2024
Oracle	2024-02-29	1.53×	sec.orcl.revenue.fy May 2023 · sec.orcl.revenue.q, 6 rows · sec.orcl.rpo.q Feb 2024
Oracle	2023-11-30	1.27×	sec.orcl.revenue.fy May 2023 · sec.orcl.revenue.q, 5 rows · sec.orcl.rpo.q Nov 2023
Oracle	2023-08-31	1.27×	sec.orcl.revenue.fy May 2023 · sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Aug 2023
Oracle	2023-05-31	1.36×	sec.orcl.revenue.fy May 2023 · sec.orcl.revenue.q Aug 2022 · sec.orcl.revenue.q Nov 2022 · sec.orcl.revenue.q Feb 2023 · sec.orcl.rpo.q May 2023
Oracle	2023-02-28	0.79×	sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Feb 2023
Oracle	2022-11-30	0.80×	sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Nov 2022
Oracle	2022-08-31	0.81×	sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Aug 2022
Oracle	2022-05-31	0.64×	sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q May 2022
Oracle	2022-02-28	0.54×	sec.orcl.revenue.q, 4 rows · sec.orcl.rpo.q Feb 2022
Microsoft	2026-09-30	1.97×	sec.msft.revenue.fy Jun 2026 · sec.msft.revenue.q, 4 rows · sec.msft.rpo.q Sep 2026
Microsoft	2026-06-30	2.06×	sec.msft.revenue.fy Jun 2026 · sec.msft.revenue.q Sep 2025 · sec.msft.revenue.q Dec 2025 · sec.msft.revenue.q Mar 2026 · sec.msft.rpo.q Jun 2026
Microsoft	2026-03-31	1.99×	sec.msft.revenue.fy Jun 2025 · sec.msft.revenue.q, 6 rows · sec.msft.rpo.q Mar 2026
Microsoft	2025-12-31	2.07×	sec.msft.revenue.fy Jun 2025 · sec.msft.revenue.q, 5 rows · sec.msft.rpo.q Dec 2025
Microsoft	2025-09-30	1.35×	sec.msft.revenue.fy Jun 2025 · sec.msft.revenue.q, 4 rows · sec.msft.rpo.q Sep 2025
Microsoft	2025-06-30	1.33×	sec.msft.revenue.fy Jun 2025 · sec.msft.revenue.q Sep 2024 · sec.msft.revenue.q Dec 2024 · sec.msft.revenue.q Mar 2025 · sec.msft.rpo.q Jun 2025
Microsoft	2025-03-31	1.19×	sec.msft.revenue.fy Jun 2024 · sec.msft.revenue.q, 6 rows · sec.msft.rpo.q Mar 2025
Microsoft	2024-12-31	1.16×	sec.msft.revenue.fy Jun 2024 · sec.msft.revenue.q, 5 rows · sec.msft.rpo.q Dec 2024
Microsoft	2024-09-30	1.05×	sec.msft.revenue.fy Jun 2024 · sec.msft.revenue.q, 4 rows · sec.msft.rpo.q Sep 2024
Microsoft	2024-06-30	1.12×	sec.msft.revenue.fy Jun 2024 · sec.msft.revenue.q Sep 2023 · sec.msft.revenue.q Dec 2023 · sec.msft.revenue.q Mar 2024 · sec.msft.rpo.q Jun 2024
Microsoft	2024-03-31	1.02×	sec.msft.revenue.fy Jun 2023 · sec.msft.revenue.q, 6 rows · sec.msft.rpo.q Mar 2024
Microsoft	2023-12-31	1.01×	sec.msft.revenue.fy Jun 2023 · sec.msft.revenue.q, 5 rows · sec.msft.rpo.q Dec 2023
Microsoft	2023-09-30	0.99×	sec.msft.revenue.fy Jun 2023 · sec.msft.revenue.q, 4 rows · sec.msft.rpo.q Sep 2023
Microsoft	2023-06-30	1.08×	sec.msft.revenue.fy Jun 2023 · sec.msft.revenue.q Sep 2022 · sec.msft.revenue.q Dec 2022 · sec.msft.revenue.q Mar 2023 · sec.msft.rpo.q Jun 2023
Microsoft	2023-03-31	0.97×	sec.msft.revenue.fy Jun 2022 · sec.msft.revenue.q, 6 rows · sec.msft.rpo.q Mar 2023
Microsoft	2022-12-31	0.95×	sec.msft.revenue.fy Jun 2022 · sec.msft.revenue.q, 5 rows · sec.msft.rpo.q Dec 2022
Microsoft	2022-09-30	0.90×	sec.msft.revenue.fy Jun 2022 · sec.msft.revenue.q, 4 rows · sec.msft.rpo.q Sep 2022
Microsoft	2022-06-30	0.97×	sec.msft.revenue.fy Jun 2022 · sec.msft.revenue.q Sep 2021 · sec.msft.revenue.q Dec 2021 · sec.msft.revenue.q Mar 2022 · sec.msft.rpo.q Jun 2022
Microsoft	2022-03-31	0.83×	sec.msft.revenue.fy Jun 2021 · sec.msft.revenue.q, 6 rows · sec.msft.rpo.q Mar 2022
Alphabet	2026-09-30	1.21×	sec.googl.revenue.fy Dec 2025 · sec.googl.revenue.q, 6 rows · sec.googl.rpo.q Sep 2026
Alphabet	2026-06-30	1.17×	sec.googl.revenue.fy Dec 2025 · sec.googl.revenue.q, 5 rows · sec.googl.rpo.q Jun 2026
Alphabet	2026-03-31	1.11×	sec.googl.revenue.fy Dec 2025 · sec.googl.revenue.q, 4 rows · sec.googl.rpo.q Mar 2026
Alphabet	2025-12-31	0.60×	sec.googl.revenue.fy Dec 2025 · sec.googl.revenue.q Mar 2025 · sec.googl.revenue.q Jun 2025 · sec.googl.revenue.q Sep 2025 · sec.googl.rpo.q Dec 2025
Alphabet	2025-09-30	0.41×	sec.googl.revenue.fy Dec 2024 · sec.googl.revenue.q, 6 rows · sec.googl.rpo.q Sep 2025
Alphabet	2025-06-30	0.29×	sec.googl.revenue.fy Dec 2024 · sec.googl.revenue.q, 5 rows · sec.googl.rpo.q Jun 2025
Alphabet	2025-03-31	0.26×	sec.googl.revenue.fy Dec 2024 · sec.googl.revenue.q, 4 rows · sec.googl.rpo.q Mar 2025
Alphabet	2024-12-31	0.27×	sec.googl.revenue.fy Dec 2024 · sec.googl.revenue.q Mar 2024 · sec.googl.revenue.q Jun 2024 · sec.googl.revenue.q Sep 2024 · sec.googl.rpo.q Dec 2024
Alphabet	2024-09-30	0.26×	sec.googl.revenue.fy Dec 2023 · sec.googl.revenue.q, 6 rows · sec.googl.rpo.q Sep 2024
Alphabet	2024-06-30	0.24×	sec.googl.revenue.fy Dec 2023 · sec.googl.revenue.q, 5 rows · sec.googl.rpo.q Jun 2024
Alphabet	2024-03-31	0.23×	sec.googl.revenue.fy Dec 2023 · sec.googl.revenue.q, 4 rows · sec.googl.rpo.q Mar 2024
Alphabet	2023-12-31	0.24×	sec.googl.revenue.fy Dec 2023 · sec.googl.revenue.q Mar 2023 · sec.googl.revenue.q Jun 2023 · sec.googl.revenue.q Sep 2023 · sec.googl.rpo.q Dec 2023
Alphabet	2023-09-30	0.22×	sec.googl.revenue.fy Dec 2022 · sec.googl.revenue.q, 6 rows · sec.googl.rpo.q Sep 2023
Alphabet	2023-06-30	0.21×	sec.googl.revenue.fy Dec 2022 · sec.googl.revenue.q, 5 rows · sec.googl.rpo.q Jun 2023
Alphabet	2023-03-31	0.22×	sec.googl.revenue.fy Dec 2022 · sec.googl.revenue.q, 4 rows · sec.googl.rpo.q Mar 2023
Alphabet	2022-12-31	0.23×	sec.googl.revenue.fy Dec 2022 · sec.googl.revenue.q Mar 2022 · sec.googl.revenue.q Jun 2022 · sec.googl.revenue.q Sep 2022 · sec.googl.rpo.q Dec 2022
Alphabet	2022-09-30	0.19×	sec.googl.revenue.fy Dec 2021 · sec.googl.revenue.q, 6 rows · sec.googl.rpo.q Sep 2022
Alphabet	2022-06-30	0.18×	sec.googl.revenue.fy Dec 2021 · sec.googl.revenue.q, 5 rows · sec.googl.rpo.q Jun 2022
Alphabet	2022-03-31	0.19×	sec.googl.revenue.fy Dec 2021 · sec.googl.revenue.q, 4 rows · sec.googl.rpo.q Mar 2022

Revenue contracted but not yet delivered, in years of revenue, at the chip makerscompany filings · each dot links to its row

BroadcomNVIDIAAMD

measured

The same measure for chip makers, from their filings. Broadcom reports a large backlog, which also holds its software subscriptions; NVIDIA and AMD report very little, so this measure shows how each company contracts with its buyers, not how much demand it faces.

Derived (rpo_to_trailing_revenue) from: SEC EDGAR XBRL company facts. SEC EDGAR, XBRL company facts API.

The numbers

line	as of	value	from
Broadcom	2026-08-02	2.01×	sec.broadcom.revenue.fy Nov 2025 · sec.broadcom.revenue.q, 6 rows · sec.broadcom.rpo.q Aug 2026
Broadcom	2026-05-03	2.18×	sec.broadcom.revenue.fy Nov 2025 · sec.broadcom.revenue.q, 5 rows · sec.broadcom.rpo.q May 2026
Broadcom	2026-02-01	0.66×	sec.broadcom.revenue.fy Nov 2025 · sec.broadcom.revenue.q, 4 rows · sec.broadcom.rpo.q Feb 2026
Broadcom	2025-11-02	0.52×	sec.broadcom.revenue.fy Nov 2025 · sec.broadcom.revenue.q Feb 2025 · sec.broadcom.revenue.q May 2025 · sec.broadcom.revenue.q Aug 2025 · sec.broadcom.rpo.q Nov 2025
Broadcom	2025-08-03	0.46×	sec.broadcom.revenue.fy Nov 2024 · sec.broadcom.revenue.q, 6 rows · sec.broadcom.rpo.q Aug 2025
Broadcom	2025-05-04	0.43×	sec.broadcom.revenue.fy Nov 2024 · sec.broadcom.revenue.q, 5 rows · sec.broadcom.rpo.q May 2025
Broadcom	2025-02-02	0.41×	sec.broadcom.revenue.fy Nov 2024 · sec.broadcom.revenue.q, 4 rows · sec.broadcom.rpo.q Feb 2025
Broadcom	2024-11-03	0.40×	sec.broadcom.revenue.fy Nov 2024 · sec.broadcom.revenue.q Feb 2024 · sec.broadcom.revenue.q May 2024 · sec.broadcom.revenue.q Aug 2024 · sec.broadcom.rpo.q Nov 2024
Broadcom	2024-08-04	0.45×	sec.broadcom.revenue.fy Oct 2023 · sec.broadcom.revenue.q, 6 rows · sec.broadcom.rpo.q Aug 2024
Broadcom	2024-05-05	0.54×	sec.broadcom.revenue.fy Oct 2023 · sec.broadcom.revenue.q, 5 rows · sec.broadcom.rpo.q May 2024
Broadcom	2024-02-04	0.71×	sec.broadcom.revenue.fy Oct 2023 · sec.broadcom.revenue.q, 4 rows · sec.broadcom.rpo.q Feb 2024
Broadcom	2023-10-29	0.57×	sec.broadcom.revenue.fy Oct 2023 · sec.broadcom.revenue.q Jan 2023 · sec.broadcom.revenue.q Apr 2023 · sec.broadcom.revenue.q Jul 2023 · sec.broadcom.rpo.q Oct 2023
Broadcom	2023-07-30	0.60×	sec.broadcom.revenue.fy Oct 2022 · sec.broadcom.revenue.q, 6 rows · sec.broadcom.rpo.q Jul 2023
Broadcom	2023-04-30	0.61×	sec.broadcom.revenue.fy Oct 2022 · sec.broadcom.revenue.q, 5 rows · sec.broadcom.rpo.q Apr 2023
Broadcom	2023-01-29	0.66×	sec.broadcom.revenue.fy Oct 2022 · sec.broadcom.revenue.q, 4 rows · sec.broadcom.rpo.q Jan 2023
Broadcom	2022-10-30	0.71×	sec.broadcom.revenue.fy Oct 2022 · sec.broadcom.revenue.q Jan 2022 · sec.broadcom.revenue.q May 2022 · sec.broadcom.revenue.q Jul 2022 · sec.broadcom.rpo.q Oct 2022
Broadcom	2022-07-31	0.80×	sec.broadcom.revenue.fy Oct 2021 · sec.broadcom.revenue.q, 6 rows · sec.broadcom.rpo.q Jul 2022
Broadcom	2022-05-01	0.55×	sec.broadcom.revenue.fy Oct 2021 · sec.broadcom.revenue.q, 5 rows · sec.broadcom.rpo.q May 2022
Broadcom	2022-01-30	0.54×	sec.broadcom.revenue.fy Oct 2021 · sec.broadcom.revenue.q, 4 rows · sec.broadcom.rpo.q Jan 2022
NVIDIA	2026-07-26	0.011×	sec.nvda.revenue.fy Jan 2026 · sec.nvda.revenue.q, 5 rows · sec.nvda.rpo.q Jul 2026
NVIDIA	2026-04-26	0.010×	sec.nvda.revenue.fy Jan 2026 · sec.nvda.revenue.q, 4 rows · sec.nvda.rpo.q Apr 2026
NVIDIA	2026-01-25	0.011×	sec.nvda.revenue.fy Jan 2026 · sec.nvda.revenue.q Apr 2025 · sec.nvda.revenue.q Jul 2025 · sec.nvda.revenue.q Oct 2025 · sec.nvda.rpo.q Jan 2026
NVIDIA	2025-10-26	0.013×	sec.nvda.revenue.fy Jan 2025 · sec.nvda.revenue.q, 6 rows · sec.nvda.rpo.q Oct 2025
NVIDIA	2025-07-27	0.011×	sec.nvda.revenue.fy Jan 2025 · sec.nvda.revenue.q, 5 rows · sec.nvda.rpo.q Jul 2025
NVIDIA	2025-04-27	0.012×	sec.nvda.revenue.fy Jan 2025 · sec.nvda.revenue.q, 4 rows · sec.nvda.rpo.q Apr 2025
NVIDIA	2025-01-26	0.013×	sec.nvda.revenue.fy Jan 2025 · sec.nvda.revenue.q Apr 2024 · sec.nvda.revenue.q Jul 2024 · sec.nvda.revenue.q Oct 2024 · sec.nvda.rpo.q Jan 2025
NVIDIA	2024-10-27	0.014×	sec.nvda.revenue.fy Jan 2024 · sec.nvda.revenue.q, 6 rows · sec.nvda.rpo.q Oct 2024
NVIDIA	2024-07-28	0.013×	sec.nvda.revenue.fy Jan 2024 · sec.nvda.revenue.q, 5 rows · sec.nvda.rpo.q Jul 2024
NVIDIA	2024-04-28	0.016×	sec.nvda.revenue.fy Jan 2024 · sec.nvda.revenue.q, 4 rows · sec.nvda.rpo.q Apr 2024
NVIDIA	2024-01-28	0.018×	sec.nvda.revenue.fy Jan 2024 · sec.nvda.revenue.q Apr 2023 · sec.nvda.revenue.q Jul 2023 · sec.nvda.revenue.q Oct 2023 · sec.nvda.rpo.q Jan 2024
NVIDIA	2023-10-29	0.020×	sec.nvda.revenue.fy Jan 2023 · sec.nvda.revenue.q, 6 rows · sec.nvda.rpo.q Oct 2023
NVIDIA	2023-07-30	0.022×	sec.nvda.revenue.fy Jan 2023 · sec.nvda.revenue.q, 5 rows · sec.nvda.rpo.q Jul 2023
NVIDIA	2023-04-30	0.025×	sec.nvda.revenue.fy Jan 2023 · sec.nvda.revenue.q, 4 rows · sec.nvda.rpo.q Apr 2023
NVIDIA	2023-01-29	0.024×	sec.nvda.revenue.fy Jan 2023 · sec.nvda.revenue.q May 2022 · sec.nvda.revenue.q Jul 2022 · sec.nvda.revenue.q Oct 2022 · sec.nvda.rpo.q Jan 2023
NVIDIA	2022-10-30	0.024×	sec.nvda.revenue.fy Jan 2022 · sec.nvda.revenue.q, 6 rows · sec.nvda.rpo.q Oct 2022
NVIDIA	2022-07-31	0.022×	sec.nvda.revenue.fy Jan 2022 · sec.nvda.revenue.q, 5 rows · sec.nvda.rpo.q Jul 2022
NVIDIA	2022-05-01	0.022×	sec.nvda.revenue.fy Jan 2022 · sec.nvda.revenue.q, 4 rows · sec.nvda.rpo.q May 2022
NVIDIA	2022-01-30	0.023×	sec.nvda.revenue.fy Jan 2022 · sec.nvda.revenue.q May 2021 · sec.nvda.revenue.q Aug 2021 · sec.nvda.revenue.q Oct 2021 · sec.nvda.rpo.q Jan 2022
AMD	2026-09-26	0.0054×	sec.amd.revenue.fy Dec 2025 · sec.amd.revenue.q, 6 rows · sec.amd.rpo.q Sep 2026
AMD	2026-06-27	0.0054×	sec.amd.revenue.fy Dec 2025 · sec.amd.revenue.q, 5 rows · sec.amd.rpo.q Jun 2026
AMD	2026-03-28	0.0070×	sec.amd.revenue.fy Dec 2025 · sec.amd.revenue.q, 4 rows · sec.amd.rpo.q Mar 2026
AMD	2025-12-27	0.0091×	sec.amd.revenue.fy Dec 2025 · sec.amd.revenue.q Mar 2025 · sec.amd.revenue.q Jun 2025 · sec.amd.revenue.q Sep 2025 · sec.amd.rpo.q Dec 2025
AMD	2025-09-27	0.0087×	sec.amd.revenue.fy Dec 2024 · sec.amd.revenue.q, 6 rows · sec.amd.rpo.q Sep 2025
AMD	2025-06-28	0.0074×	sec.amd.revenue.fy Dec 2024 · sec.amd.revenue.q, 5 rows · sec.amd.rpo.q Jun 2025
AMD	2025-03-29	0.0027×	sec.amd.revenue.fy Dec 2024 · sec.amd.revenue.q, 4 rows · sec.amd.rpo.q Mar 2025
AMD	2024-12-28	0.0033×	sec.amd.revenue.fy Dec 2024 · sec.amd.revenue.q Mar 2024 · sec.amd.revenue.q Jun 2024 · sec.amd.revenue.q Sep 2024 · sec.amd.rpo.q Dec 2024
AMD	2024-09-28	0.0035×	sec.amd.revenue.fy Dec 2023 · sec.amd.revenue.q, 6 rows · sec.amd.rpo.q Sep 2024
AMD	2024-06-29	0.0033×	sec.amd.revenue.fy Dec 2023 · sec.amd.revenue.q, 5 rows · sec.amd.rpo.q Jun 2024
AMD	2024-03-30	0.0050×	sec.amd.revenue.fy Dec 2023 · sec.amd.revenue.q, 4 rows · sec.amd.rpo.q Mar 2024
AMD	2023-12-30	0.0060×	sec.amd.revenue.fy Dec 2023 · sec.amd.revenue.q Apr 2023 · sec.amd.revenue.q Jul 2023 · sec.amd.revenue.q Sep 2023 · sec.amd.rpo.q Dec 2023
AMD	2023-09-30	0.0087×	sec.amd.revenue.fy Dec 2022 · sec.amd.revenue.q, 6 rows · sec.amd.rpo.q Sep 2023
AMD	2023-07-01	0.013×	sec.amd.revenue.fy Dec 2022 · sec.amd.revenue.q, 5 rows · sec.amd.rpo.q Jul 2023
AMD	2023-04-01	0.015×	sec.amd.revenue.fy Dec 2022 · sec.amd.revenue.q, 4 rows · sec.amd.rpo.q Apr 2023
AMD	2022-12-31	0.010×	sec.amd.revenue.fy Dec 2022 · sec.amd.revenue.q Mar 2022 · sec.amd.revenue.q Jun 2022 · sec.amd.revenue.q Sep 2022 · sec.amd.rpo.q Dec 2022
AMD	2022-09-24	0.013×	sec.amd.revenue.fy Dec 2021 · sec.amd.revenue.q, 6 rows · sec.amd.rpo.q Sep 2022
AMD	2022-06-25	0.0079×	sec.amd.revenue.fy Dec 2021 · sec.amd.revenue.q, 5 rows · sec.amd.rpo.q Jun 2022
AMD	2022-03-26	0.011×	sec.amd.revenue.fy Dec 2021 · sec.amd.revenue.q, 4 rows · sec.amd.rpo.q Mar 2022

Sub-layers and venture flow

1. Semiconductors & AI chip startups

Semiconductors & AI chip startups: primary-round equity by quartereach segment links to its first round's record

Form D filings

Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

The numbers

quarter	source	kind	dollars	record
Q3 2026	Form D filings	equity	$1.3B	reported
Q1 2026	Form D filings	equity	$762M	reported
Q2 2025	Form D filings	equity	$141M	reported
Q1 2025	Form D filings	equity	$148M	reported
Q4 2024	Form D filings	equity	$445M	reported
Q1 2024	Form D filings	equity	$174M	reported
Q4 2023	Form D filings	equity	$180M	reported
Q2 2023	Form D filings	equity	$95.8M	reported

2. Hyperscalers

Hyperscalers: primary-round equity by quartereach segment links to its first round's record

Form D filings

Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

The numbers

quarter	source	kind	dollars	record
Q3 2026	Form D filings	equity	$3.7B	reported
Q2 2026	Form D filings	equity	$730M	reported
Q1 2026	Form D filings	equity	$450M	reported
Q4 2025	Form D filings	equity	$1.0B	reported
Q2 2025	Form D filings	equity	$100M	reported
Q4 2024	Form D filings	equity	$686M	reported
Q2 2024	Form D filings	equity	$1.1B	reported
Q4 2023	Form D filings	debt	$192M	reported

3. Data-centre developers & power
Data-centre developers & power: primary-round equity by quartereach segment links to its first round's record
Form D filings
Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.
The numbers
quarter source kind dollars record
Q3 2026 Form D filings equity $465M reported
Q3 2024 Form D filings equity $160M reported
4. Memory/HBM
Memory/HBM: primary-round equity by quartereach segment links to its first round's record
Form D filings
Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.
The numbers
quarter source kind dollars record
Q2 2023 Form D filings equity $110M reported

quarter	source	kind	dollars	record
Q3 2026	Form D filings	equity	$465M	reported
Q3 2024	Form D filings	equity	$160M	reported

quarter	source	kind	dollars	record
Q2 2023	Form D filings	equity	$110M	reported

Primary-round dollars per calendar quarter for each sub-layer, each strip on its own scale: Form D amounts sold where the entity files, Epoch's press-compiled rounds otherwise.

Crosswalk to the diffusion lens

Methods (input to) — Compute growth and hardware perf-per-dollar feed the methods stock.
Leak (leak from) — Margin stacking, HHI, circular financing: surplus exiting the chain at semis and cloud. Interpretive spine: Teece, Nordhaus and Perez in docs/interpretation; Ding through the bottleneck list.

Compute & physical

How the economics move

The chip designer's lead is a shortage profit; the foundry's and the toolmaker's are harder to copy

High-bandwidth memory is a cyclical commodity earning a shortage's price

A site already wired to the grid is scarcer than the building or the money

The big clouds keep a rent from customers who stay; the neoclouds earn the shortage's price on borrowed money

The companies that matter here

ASML

Micron

SK Hynix

Samsung Electronics

NVIDIA Corp

Broadcom

Advanced Micro Devices

TSMC

Microsoft Corp

Amazon.com Inc.

Alphabet Inc.

Oracle Corp

CoreWeave Inc.

Nebius Group N.V.

Fluidstack

Etched

Indicators

Revenue already contracted

Sub-layers and venture flow

1. Semiconductors & AI chip startups

2. Hyperscalers

3. Data-centre developers & power

4. Memory/HBM

Crosswalk to the diffusion lens