How do semiconductor business models actually work?

Why can two semiconductor companies both sell “chips” and have completely different businesses?

Semiconductor companies can sit in the same industry while making money in radically different ways because the chip itself passes through several businesses before it reaches a customer.

Nvidia and AMD design processors but outsource most advanced manufacturing. TSMC owns the fabs that manufacture chips for other companies. Texas Instruments designs and manufactures its own analog chips. Arm licenses processor technology and collects royalties. Cadence and Synopsys sell the software engineers use to design chips. ASML sells lithography machines to the fabs. Amkor packages and tests chips after wafer fabrication.

Those differences decide who pays for factories, who carries unsold inventory, who suffers when utilization falls and who can keep earning from the same technology years after the original engineering work.

That contrast is especially visible now because semiconductor demand is booming while the money is flowing very unevenly. The Semiconductor Industry Association said global chip sales reached $403.3 billion in the second quarter of 2026 alone, up 35.1% from the previous quarter. Yet the companies benefiting most range from a factory operator such as TSMC to an IP licensor such as Arm and an equipment company such as ASML.

So it is more useful to look beyond “who makes chips?” The better question is which scarce part of the semiconductor chain each company controls and which expensive parts it leaves to somebody else.

Business model	What the company gets paid for	Examples	Main risk
Fabless	Chips designed internally	Nvidia, AMD, Qualcomm	Product misses, supplier dependence
Foundry	Manufacturing other companies’ chips	TSMC, GlobalFoundries, UMC	Capex, yields, utilization
IDM	Designing and manufacturing chips	Texas Instruments, Intel, Micron	Product risk plus factory risk
Semiconductor IP	Licensing chip technology	Arm	Losing architectural relevance
EDA	Chip-design software	Cadence, Synopsys	R&D and platform competition
Equipment	Machines used inside fabs	ASML, Applied Materials	Customer capex cycles
OSAT	Packaging and testing	Amkor, ASE	Lower margins, utilization

What changes when a semiconductor company owns its own fabs?

Owning semiconductor fabs gives a company much more control over production, but it also means carrying one of the heaviest fixed-cost structures in modern industry.

A leading semiconductor factory costs billions before it produces a single sellable wafer. Once the equipment is installed, depreciation, engineers, maintenance and cleanroom costs keep running whether demand is excellent or weak.

TSMC shows the scale involved. Its second-quarter revenue reached $40.2 billion, with a 67.7% gross margin and a 60.3% operating margin. Those figures are spectacular, yet they come from a business that keeps pouring enormous amounts of capital into new nodes, packaging capacity and factories across several countries.

That makes utilization crucial. Imagine two fabs with almost identical operating costs. One runs close to full capacity while the other has a large amount of idle equipment. The first can spread depreciation and labor across far more wafers, making every good chip cheaper.

Yield has a similar effect. If a wafer produces 90 usable dies instead of 70, much of the manufacturing cost has barely changed, but there are 20 extra chips available to sell. Small improvements in yield can move billions of dollars once production reaches huge volumes.

A fabless company such as Nvidia avoids most of those factory economics. Nvidia pays manufacturing partners and focuses its own spending on chip design, systems, networking and software. That makes the company more flexible, although modern fabless companies still take on a surprising amount of supply risk.

Why did the fabless semiconductor model become so powerful?

The fabless semiconductor model became so powerful because designing the best chip and running the best advanced factory have each become difficult enough to support specialist companies worth hundreds of billions of dollars.

Nvidia is the extreme example. In its latest reported quarter, the company generated $81.6 billion of revenue and a 74.9% GAAP gross margin without owning the leading-edge fabs producing its GPUs. AMD operates with the same broad structure and recently reported $11.5 billion of quarterly revenue with a 54% GAAP gross margin.

TSMC takes care of the manufacturing problem. A fabless designer can move a new architecture onto an advanced TSMC process without financing the whole fab transition itself. TSMC, meanwhile, can justify huge investments because the same manufacturing platform serves many customers.

The split also improves risk sharing. Nvidia may misjudge a GPU generation while Apple succeeds with an iPhone processor and AMD succeeds with a server CPU. TSMC can manufacture across several of those product families rather than depending entirely on one architecture.

But outsourcing production never guarantees a good business. AMD's current gross margin is roughly 20 percentage points below Nvidia's despite both companies being fabless. Nvidia earns exceptional economics because customers are paying for much more than outsourced silicon: GPU performance, CUDA, networking, systems and an ecosystem that has taken years to build.

Fabless works best when the company outsources the factory while keeping control of the part customers really care about.

If you want more recent data on this point, please see our latest semiconductor industry report.

Is Nvidia really an asset-light semiconductor company?

Nvidia is fabless, but calling Nvidia “asset-light” now hides a huge part of how the business actually works.

Nvidia's latest quarterly filing showed $119 billion of manufacturing, supply and capacity commitments. Around $95 billion of that is due during the remainder of its current fiscal year. A year earlier, the equivalent commitments were far smaller.

Nvidia also carried $25.8 billion of inventory at the end of the latest reported quarter, up from $21.4 billion only three months earlier.

So Nvidia has avoided owning TSMC's fabs while still making enormous bets on future physical supply. It has to secure advanced wafers, HBM memory, packaging, substrates, networking components and assembly capacity before it knows exactly how much of every future product customers will buy.

The risk is real rather than theoretical. Nvidia previously took a $4.5 billion charge related to excess H20 inventory and purchase obligations after U.S. export restrictions changed what it could sell into China. AMD faced a similar problem and booked an $800 million charge related to Instinct MI308 inventory and commitments after export controls hit that product.

We can see how the fabless model has changed at the AI frontier. The factory stays on somebody else's balance sheet, but the chip designer increasingly has to reserve a large part of that factory's future output.

That is still financially preferable to building every fab yourself. It is simply much less “asset-light” than the word suggests.

Why is TSMC so hard for other semiconductor foundries to copy?

TSMC is hard to copy because customers are buying a manufacturing system that combines leading process technology, high yields, enormous capacity and years of accumulated trust.

TrendForce estimated that TSMC had 72% of global foundry revenue in the first quarter of 2026. Samsung, the second-largest player, had only 6.5%. The gap between first and second place was therefore more than 65 percentage points.

TSMC then reported another strong quarter. Revenue reached $40.2 billion, up 33.7% from a year earlier. Seventy-seven percent of wafer revenue came from 7-nanometer technology and below, including 3% from its newly ramping 2-nanometer process.

The latest monthly numbers suggest demand has kept climbing. TSMC's July revenue reached NT$467.6 billion, a company record, and revenue for the first seven months of the year was 37% above the same period in 2025.

The advantages compound. Large customers bring high volumes to TSMC. Those volumes help fill expensive factories and generate cash for the next process node. Running many designs also creates more manufacturing data, which helps improve yields. Strong yields then make TSMC safer for the next large customer.

Neutrality adds another advantage. Apple, Nvidia, AMD, Qualcomm and other chip designers can send extremely valuable designs to TSMC without worrying that TSMC will launch a competing smartphone, CPU or GPU.

Foundry	Q1 2026 market share
TSMC	72.0%
Samsung Foundry	6.5%
SMIC	5.1%
UMC	3.9%
GlobalFoundries	3.3%

Can Intel actually build a TSMC-style foundry business?

Intel can become a much better semiconductor manufacturer, but its external foundry business is still tiny compared with the scale implied by the Intel Foundry name.

Intel Foundry reported $5.77 billion of second-quarter segment revenue, yet $5.47 billion came from transactions with other parts of Intel. External revenue was only $293 million.

That means roughly 95% of Intel Foundry's revenue still came from inside Intel itself.

The segment also lost $2.09 billion during the quarter. There has been genuine progress: the loss was $3.17 billion a year earlier, Intel said yields and cycle times improved, and 18A output rose more than 50% sequentially. Manufacturing performance is clearly getting better.

Commercial foundry success requires another step. Independent chip designers have to trust Intel with products worth billions of dollars. They need stable process-design kits, proven IP libraries, packaging options, high yields and confidence that Intel will keep investing in the roadmap for many years.

External scale would also change the factory economics. TSMC can fill capacity using demand from Apple, Nvidia, AMD, Broadcom, Qualcomm and many others. Intel still depends heavily on its own product roadmap to keep its factories busy.

For now, Intel Foundry is mainly a manufacturing turnaround with an external foundry option attached to it. It starts looking genuinely TSMC-like when outside customers become a meaningful share of revenue rather than a few hundred million dollars per quarter.

If you want more recent data on this point, please see our latest semiconductor industry report.

When does owning semiconductor factories still make sense?

Owning semiconductor factories can still work extremely well when manufacturing lowers product cost or improves reliability without forcing the company into the most expensive leading-edge node race.

Texas Instruments is a good example. TI recently generated $5.46 billion of quarterly revenue, up 23% from a year earlier. Analog chips contributed $4.37 billion, around 80% of the total.

Many of those chips regulate power, convert signals, manage motors or perform other specific jobs inside cars, industrial machines and electronics. They can stay in production for years, sometimes decades.

That long life changes the factory equation. TI does not need every analog product to jump onto TSMC's newest process every two years. It can build large 300-millimeter fabs, move huge volumes through them and keep driving manufacturing cost lower over a long period.

The cash generation is already showing up. TI produced $8.7 billion of operating cash flow over the latest twelve months. Capital expenditures were $3.3 billion, down sharply from the previous phase of its manufacturing buildout, while TI's reported free cash flow rose to $6.5 billion under its definition.

The IDM model deserves more nuance than “old semiconductor companies own fabs, new ones outsource them.” Owning fabs is very attractive when the process technology itself gives you a cost or supply advantage and the assets can stay useful for a long time.

Problems appear when a company combines huge fab spending with fast obsolescence, weak utilization or products that customers no longer want.

Why does the memory-chip business make money so differently?

Memory is currently behaving like an extreme scarcity business, which is why Micron's financials barely resemble those of the same company a year earlier.

Micron's latest quarterly revenue was $41.46 billion, up from just $9.30 billion in the same quarter of the previous year. GAAP gross margin jumped from 37.7% to 84.6%.

The striking part is what caused that increase.

Sequentially, Micron's DRAM bit shipments rose only in the low-single-digit percentage range. Average selling prices rose in the low-60% range. NAND bit shipments increased in the mid-single digits while prices jumped in the mid-80% range.

So most of the revenue explosion came from price rather than a similar explosion in physical volume.

This happens because memory factories cannot add supply overnight. If server builders suddenly require huge amounts of DRAM and HBM, a relatively small shortage of available bits can produce enormous price increases. Once a fab is already running, much of that extra selling price drops through to profit.

The same mechanism works brutally in reverse. Memory companies have repeatedly gone from huge profits to losses when producers add too much capacity and prices collapse.

HBM is making today's cycle more interesting because memory for AI accelerators requires advanced stacking, qualification and tight integration with GPU platforms. Micron, SK hynix and Samsung therefore compete on technical capability as well as manufacturing volume.

Still, we should be careful about treating Micron's current 80%-plus margins as a normal memory baseline. They tell us how powerful scarcity can become when supply is tight.

Micron fiscal Q3	2025	2026
Revenue	$9.3B	$41.5B
GAAP gross margin	37.7%	84.6%
DRAM sequential bit growth	—	Low single digits
DRAM sequential price change	—	Low-60% increase
NAND sequential bit growth	—	Mid-single digits
NAND sequential price change	—	Mid-80% increase

How does Arm make money from chips it never manufactures?

Arm makes money by selling access to processor technology and then collecting royalties when customers ship chips based on that technology.

In Arm's latest quarter, revenue reached $1.29 billion, up 22% from a year earlier. Royalty revenue was $715 million and license revenue was $574 million.

The distinction is useful. A semiconductor company may first pay Arm for access to an architecture, processor core or larger compute design. Once the resulting chip enters production, Arm can collect royalties tied to shipments under the relevant agreement.

One engineering relationship can therefore generate revenue at several stages.

The physical economics are unusually attractive. Arm reported a 97.2% GAAP gross margin in its latest quarter because another Arm-based processor can be manufactured without Arm purchasing the wafer, operating the fab or carrying the finished inventory.

Royalty growth is also becoming more valuable as newer Arm designs command higher rates. Arm said the latest increase was helped by adoption of Armv9, Compute Subsystems and more Arm-based chips in data centers.

The trade-off sits in R&D. Arm spent $838 million on research and development in that same quarter, up 29% year over year. The company has to keep creating technology that customers still want several years later, because semiconductor design cycles are long.

Arm also faces a strategic balancing act as it moves closer to providing larger chunks of the finished compute system. Capturing more value per customer can lift revenue, but every step toward complete silicon also moves Arm closer to businesses historically occupied by its own licensees.

How can Qualcomm get paid twice from the same smartphone?

Qualcomm can make money from both the physical chip inside a device and the cellular patents the device needs to use, which gives Qualcomm two very different profit engines.

The chip business sits inside QCT. Qualcomm sells Snapdragon processors, modems, radio-frequency components, automotive chips and other silicon.

QTL is the licensing business. Smartphone manufacturers pay for access to Qualcomm's large portfolio of wireless intellectual property, including patents tied to cellular standards.

The margin difference is huge. In Qualcomm's latest reported quarter, total company revenue was $9.9 billion. QTL remains much smaller than the chip operation, yet licensing has historically produced operating economics that the hardware side cannot match because Qualcomm does not need to manufacture another chip to collect another patent royalty.

This becomes especially useful when a customer starts designing more silicon itself.

Apple has gradually moved modem development in-house, which reduces the amount of merchant chip revenue Qualcomm can expect from Apple over time. But designing an internal modem does not automatically remove the need to license patents required by cellular standards.

Qualcomm has been trying to reduce its dependence on smartphones at the same time. Automotive and IoT revenue grew strongly again in the latest quarter, and the company now wants non-handset businesses to become a much larger part of revenue over the next few years.

The Qualcomm model combines a relatively conventional fabless semiconductor company with an unusually profitable standards-IP business sitting beside it.

Why do Cadence and Synopsys make such good money from semiconductor design software?

Cadence and Synopsys make unusually attractive semiconductor businesses because chip designers cannot realistically build today's most complex processors without their software.

Cadence's latest quarter produced $1.58 billion of revenue, up 24% year over year. Its non-GAAP operating margin reached 45.5%, and backlog hit a record $8.1 billion. The company expects $4.2 billion of its remaining contractual obligations to turn into revenue over the following twelve months.

Synopsys recently reported $2.28 billion of quarterly revenue and an $11 billion backlog. Its business is now broader following the Ansys acquisition, but electronic design automation remains central to what the company sells.

The reason customers keep paying is pretty simple: the cost of failure is enormous. A modern high-end chip can contain tens of billions of transistors. Discovering a serious design problem after tape-out can waste millions of dollars and delay a product by months.

Chip engineers therefore use EDA software to design circuits, verify them, place billions of components, model power and timing, prepare layouts and check whether the final design can actually be manufactured.

Switching tools is painful. Engineering teams build workflows around them, foundries certify design flows for particular process nodes, and third-party IP gets integrated into the same environment.

Complexity is also moving in Cadence and Synopsys's favor. Chiplets, 3D packaging, custom AI accelerators and larger systems create more design and verification work even when the number of semiconductor companies does not grow.

Every new layer of chip complexity gives the companies selling the design tools something else to monetize.

If you want more recent data on this point, please see our latest semiconductor industry report.

How does ASML make money without ever selling a semiconductor?

ASML makes money by selling lithography equipment that advanced fabs cannot operate without, then keeps earning from those machines through service and upgrades.

ASML recently reported €9.33 billion of quarterly revenue with a 54% gross margin. Installed-base management contributed €2.76 billion.

That means almost 30% of quarterly revenue came from servicing and upgrading machines already sitting inside customer factories.

This part of the model is easy to underestimate. A foundry can postpone buying some new equipment when expansion slows. Maintaining equipment already responsible for billions of dollars of wafer output is much harder to postpone.

The installed base therefore builds a recurring revenue stream underneath the more cyclical machine business.

ASML's stronger advantage comes from EUV lithography. The technology required decades of development and a highly specialized supplier network. Leading-edge chip manufacturers have no equivalent alternative they can buy from another vendor tomorrow.

Demand currently remains very strong. ASML lifted its full-year sales outlook after its latest quarter and said AI-related investments were pushing customers to accelerate capacity plans. It is preparing to expand production capacity for both EUV and DUV systems over the next two years.

ASML still feels semiconductor cycles because TSMC, Samsung, Intel and memory producers can change the pace of fab investment. Yet ASML does not have to predict whether Nvidia, AMD or a hyperscaler's internal accelerator wins the next AI battle. Several possible winners may still need wafers manufactured on equipment that depends on ASML technology.

Why do semiconductor packaging companies make much lower margins?

Semiconductor packaging companies usually make lower margins because they carry factories, equipment and material costs while controlling less of the valuable chip IP.

Amkor's latest quarter is a clean example. Revenue reached $1.90 billion and gross margin was 16.8%. Materials alone consumed 52.6% of revenue.

Compare that with businesses elsewhere in the chain. Arm's latest gross margin exceeded 97%. Nvidia was around 75% in its latest reported quarter. TSMC reached 67.7%. ASML was at 54%.

Amkor also expects around $2.5 billion to $3.0 billion of capital expenditure this year, so low gross margins do not come with an especially light asset base.

Customer concentration adds another constraint. Amkor's ten largest customers generated 66% of quarterly sales. Those customers tend to be sophisticated semiconductor companies with enough scale to negotiate aggressively.

Advanced packaging is improving the strategic importance of the business, however. AI accelerators increasingly combine multiple compute dies, HBM stacks and complex interconnects. Packaging affects bandwidth, power use, thermals and ultimately the performance of the full system.

Amkor already generated $1.56 billion, about 82% of quarterly sales, from what it classifies as advanced products. TSMC has also invested heavily in its own advanced packaging capacity, which tells us how valuable this part of manufacturing has become.

Higher technical difficulty should allow advanced packaging to capture more profit than old-fashioned assembly. The economics will still carry plenty of manufacturing cost, so we should not expect packaging companies to suddenly look like Arm or Cadence.

Why is AI making custom semiconductor design a much bigger business?

AI is making custom silicon much bigger because Google, Amazon, Meta and other hyperscalers now spend enough on computing to justify processors designed around their own workloads.

Broadcom shows how large this model has already become. Its latest reported quarter included $10.8 billion of AI semiconductor revenue, up 143% year over year. Broadcom expected the following quarter to reach around $16 billion, driven by custom AI accelerators and AI networking.

The economics differ from Nvidia's merchant-GPU model. Nvidia builds a platform and sells it across many customers. Custom-silicon companies work closely with a handful of huge customers to turn their architecture into a manufacturable chip.

That can create deep, multi-year relationships because processors are designed years before deployment. Once a custom accelerator reaches large-scale production, the volumes can be enormous.

The customer also gains more control. Google has TPUs, Amazon has Trainium and Inferentia, Microsoft has Maia, and other large cloud companies are pushing further into internal silicon.

A hyperscaler designing its own accelerator does not mean the semiconductor supply chain loses the revenue. The money moves around.

The customer may buy fewer merchant GPUs while spending more with Broadcom or another design partner. It may still license Arm technology, design through Cadence or Synopsys, manufacture at TSMC, use HBM from Micron or SK hynix and depend on advanced packaging.

Marvell is chasing the same opportunity. Data centers now account for roughly three-quarters of its revenue, with custom compute, networking and optical products playing increasingly important roles.

Hyperscaler chip design is reshuffling who gets paid. It favors companies that can remain useful even when the customer owns more of the architecture.

Does the chip company with the highest gross margin actually have the best business?

The semiconductor company with the highest gross margin does not automatically have the best business, although gross margin tells us a lot about where pricing power sits.

Arm currently sits at the extreme, with a GAAP gross margin above 97%. Very little physical cost comes with another dollar of royalty revenue.

Micron recently reached 84.6%, but that number comes during an extraordinary memory shortage and could move dramatically when pricing changes.

Nvidia's latest reported margin was 74.9%, showing how much value the company captures even after paying external manufacturers. TSMC reached 67.7% while running one of the most capital-intensive industrial operations in the world. AMD was at 54%, ASML was also around 54%, and Amkor came in at 16.8%.

Gross margin alone misses what happens below that line.

Arm spends heavily on R&D. TSMC continuously funds fabs and process development. Nvidia has $119 billion of manufacturing and capacity commitments. ASML supports a huge engineering organization and supplier network. Amkor has to finance packaging facilities despite its thinner margins.

TSMC makes the point particularly well. A 67.7% gross margin might look weaker than Nvidia's 74.9%, yet TSMC also produced a 60.3% operating margin in the same quarter. Very few industrial companies anywhere operate at that level.

If you want more recent data on this point, please see our latest semiconductor industry report.

Company	Model	Latest reported GAAP gross margin
Arm	Semiconductor IP	97.2%
Micron	Memory IDM	84.6%
Nvidia	Fabless compute	74.9%
TSMC	Foundry	67.7%
AMD	Fabless compute	54.0%
ASML	Semiconductor equipment	54.0%
Amkor	Packaging and test	16.8%

Which semiconductor business models get hurt most when the chip cycle turns?

Memory producers and factory-heavy semiconductor companies usually feel downturns fastest because their costs remain high even when prices and utilization fall.

Memory is the clearest case. Micron currently shows what the positive side of operating leverage looks like: huge price increases turned modest changes in bit shipments into an extraordinary jump in revenue and margins. A memory glut can unwind that effect quickly.

Foundries face utilization risk. An underused fab still depreciates. Workers still have to operate it, equipment still needs maintenance and technology development continues.

Packaging companies face a similar issue at lower margins. Amkor's gross margin recently improved from 12.0% a year earlier to 16.8% as revenue and factory economics strengthened. A drop in volume can push the leverage the other way.

Equipment companies experience the cycle through customer capex. ASML may have strong service revenue from its installed base, but new system sales can slow if chipmakers decide they have built enough capacity.

Fabless companies avoid factory utilization directly, yet they still suffer product and inventory cycles. Nvidia's H20 charge showed how quickly an external event can turn reserved supply into unwanted inventory obligations.

IP and EDA generally have the cleanest protection. Arm can keep collecting royalties from chips already shipping, while Cadence and Synopsys have multi-year contracts, large backlogs and tools that engineering teams continue using during long design programs.

The current semiconductor boom makes these differences easy to forget. The tougher test comes when customers stop ordering ahead, inventories rise and scarce capacity becomes ordinary capacity again.

Is selling the finished chip still the best place to make money in semiconductors?

Selling the finished chip can produce the biggest upside when the company owns a platform customers desperately want, but several of today's strongest semiconductor businesses make money one or two layers away from the finished processor.

Nvidia demonstrates the upside. Its GPUs sit at the center of AI infrastructure, while CUDA, networking and systems make the product harder to replace. The company can therefore capture economics that a normal fabless chip designer cannot.

AMD shows the difference. It has advanced products, strong growth and the same broad fabless structure, yet its margins remain materially below Nvidia's because its pricing power and ecosystem are different.

Move one layer deeper and TSMC can make money whether the successful processor is designed by Nvidia, AMD, Apple, Broadcom or a cloud provider. ASML goes deeper again: several competing foundries need lithography equipment regardless of which customer-facing chip eventually wins.

Cadence and Synopsys can earn money before the chip exists. Arm can keep collecting royalties after the design enters production. ASML can earn service revenue for years after its machine was installed.

These businesses are attractive because they can sometimes monetize several competing winners at once.

There is still concentration risk. TSMC has large customers. ASML sells into a relatively small group of giant chip manufacturers. Arm needs chip designers to keep choosing its architecture. Cadence and Synopsys have to stay deeply embedded in rapidly changing engineering workflows.

The current industry gives us a clear lesson: the company whose logo appears on the finished processor does not necessarily control the most valuable position in the chain.

So which semiconductor business model is actually the best?

There is no universal best semiconductor business model, but the strongest ones today share a clear trait: they control something difficult to replace while leaving as much unrelated risk as possible somewhere else.

For pure capital efficiency, semiconductor IP and EDA are difficult to beat. Arm can earn licensing and royalty revenue without manufacturing chips, while Cadence and Synopsys sell deeply embedded tools with high margins and large future contracted revenue.

For manufacturing, TSMC is the standout. The company accepts extraordinary capital requirements but turns its scale, yields and technology lead into margins that most manufacturers could never approach. Its roughly 72% foundry share tells us how far the model has pulled away from competitors.

Fabless can create even greater upside when the product itself becomes a platform. Nvidia is the current proof. Outsourcing the fab lets Nvidia concentrate resources on architecture, software, networking and systems, although its $119 billion of supply commitments show how much manufacturing risk now follows successful AI chip designers anyway.

The IDM model still works very well in the right market. Texas Instruments can use long-lived analog products and 300-millimeter manufacturing to lower costs for years. Micron currently shows how profitable memory manufacturing becomes when supply is scarce, while the industry's history reminds us how quickly those profits can reverse.

Custom silicon is becoming a much larger model as hyperscalers design more of their own computing infrastructure. Broadcom is already generating more than $10 billion a quarter from AI semiconductors, with custom accelerators and networking driving much of the growth.

ASML occupies perhaps the cleanest bottleneck of all. Every leading chip designer can fight over customers while ASML sells crucial equipment to the factories they depend on. Packaging companies such as Amkor sit in a tougher financial position, although advanced packaging is becoming far more valuable as AI systems move toward chiplets, HBM and 3D integration.

So when we ask how semiconductor business models actually work, the answer comes down to where a company chooses to sit in the chain.

Some companies take product risk. Some take factory risk. Some sell intellectual property. Some collect royalties. Some sell the tools everyone else needs.

The best businesses have managed to concentrate the risk they understand while getting paid for a bottleneck their customers cannot easily go around. That explains why Arm, Nvidia, TSMC, Cadence, Synopsys and ASML can all have exceptional economics despite barely resembling one another.

If you want more recent data on this point, please see our latest semiconductor industry report.