Topic report
Custom silicon & edge AI
Who designs, builds, equips and runs the custom parts that are taking share from general-purpose accelerators.
Brief:
Workloads are moving from one general-purpose accelerator to a mix of custom parts, and inference is moving toward the device. The theme follows the designers, the foundry, the equipment makers and memory.
Not to be missed:
nothing specified
Head to head
The names with real exposure, most likely to succeed first. Everything
else the report found is in the full table below.
| Company | Odds | Why | What sets it apart | If it works | If it doesn’t |
|---|---|---|---|---|---|
|
Broadcom Inc.
Core
|
already earning | The AI semiconductor revenue line is reported every quarter. | Designs the custom parts for several of the largest cloud customers at once, which no other name on the list does. | The disclosed design wins ramp on schedule and the customer count widens. | A large customer moves a programme in-house or to a second supplier. |
|
Taiwan Semiconductor Manufactur
Core
|
already earning | Leading-edge revenue share and packaging capacity are disclosed; utilisation is the variable. | The one link every other link on the list has to pass through. | Capacity additions land on time and the leading-edge mix keeps rising. | A demand pause leaves new capacity under-utilised while the capital is already spent. |
|
ASML Holding N.V. - New York Re
Satellite
|
already earning | Orders and backlog are reported; the timing of shipments is the variable. | A monopoly link; the exposure is to the capacity build, not to any one designer. | Leading-edge capacity keeps being added and the services base grows with the installed tools. | Export-control scope widens or a capacity pause stretches the order cycle. |
|
Advanced Micro Devices, Inc.
Satellite
|
strong | The data-center segment is reported separately and its growth is visible. | A second source for accelerators; exposure is diluted by the client and gaming segments. | The accelerator roadmap wins a second large customer. | The roadmap slips a generation and the segment stalls. |
|
Arm Holdings plc
Satellite
|
strong | Royalty rates and licensing revenue are reported; the custom-silicon share is inferred. | Per-unit exposure rather than per-design exposure, so it rides volume rather than wins. | Custom designs standardise on the architecture and royalty rates step up with the newer cores. | Designers move to an open architecture for the new parts. |
|
Micron Technology, Inc.
Satellite
|
even | HBM share and pricing are discussed each quarter; the cycle decides the margin. | Rides the volume but with memory pricing on top, so the exposure swings in both directions. | HBM stays sold out while the pricing cycle holds. | Memory supply catches up and pricing turns while the capital is committed. |
|
QUALCOMM Incorporated
Satellite
|
even | Handset volumes and the AI content per device are disclosed in part. | The only name on the list whose exposure is through the device rather than the data center. | AI features lift the content per device and the PC share grows. | A large customer insources its modem and the content gain is offset. |
v1 ·
generated 2026-09-29 14:53 by Morgan ·
11 verified names
Scope
In scope: the listed companies that design custom accelerators, license the architecture they use, manufacture and package them, supply the equipment and the memory, and make the edge silicon that runs inference on the device. Out of scope: the hyperscalers themselves and the power supply chain, which are other themes.
The space
Workloads are moving from one general-purpose accelerator to a mix of custom parts, and inference is moving toward the device. The designers report the custom revenue, the foundry reports the leading-edge mix, the equipment makers report orders and backlog, and the memory makers report high-bandwidth share; the device names are where the shift is least visible in the numbers so far.
Why now
Design wins, foundry mix and equipment backlog are all disclosed, so the theme is measurable quarter by quarter; the device half is the part that still has to show up in the reported numbers.
Timeline
General-purpose accelerators carried the first wave; custom parts for the largest cloud customers followed two years behind, and inference on the device is the newest and least measured part of the shift.
How it works
A cloud customer specifies an accelerator for its own workload, a designer builds it on a licensed architecture, a foundry manufactures and packages it beside high-bandwidth memory, and the equipment makers supply the tools for each step. The device half runs smaller models locally on mobile and PC processors.
How we got here
The first wave was served by one general-purpose supplier; the largest customers then funded custom parts to lower cost per unit of work, and the foundry and packaging capacity to make them became the constraint.
Where it stands today
Several custom programmes are shipping and disclosed; packaging capacity is being added; the device makers have announced AI features whose attach rate is still being measured.
The open arguments
- Whether custom parts keep taking share or the general-purpose supplier's roadmap closes the gap.
- How much of the value accrues to the designer versus the foundry and the memory maker.
- Whether on-device inference changes the unit economics of the device makers at all.
Value chain
Each link was searched separately for public companies — which is what
surfaces the suppliers a single query omits.
Custom accelerator designers
Companies that design application-specific accelerators for the large cloud customers.
Design wins are multi-year; the revenue lands with a lag.
Architecture & IP licensing
The instruction-set and core IP the custom parts are built on.
Royalty economics; the exposure is per unit, not per design.
Foundry & advanced packaging
Leading-edge manufacturing and the packaging that joins compute to memory.
Capacity allocation decides who ships; pricing follows utilisation.
Lithography & process equipment
The tools that make the leading edge possible.
Order cycles are lumpy; backlog and services are the stable part.
High-bandwidth memory
The memory stacked beside the accelerators.
The most cyclical link; pricing swings with supply.
Edge & device silicon
Mobile and embedded processors that run inference on the device.
Volume businesses; the AI content per device is the variable.
The names
Tier and role rank each name within this theme — they are not
buy/sell/hold calls.
Core — the way to own this theme
(2)
| Ticker | Company | Segment | Role | Odds | Mkt cap | Listing | Ours | The read |
|---|---|---|---|---|---|---|---|---|
| Taiwan Semiconductor Manufactur | Foundry & advanced packaging | pure play | already earning | $2.4T | US listed | HELD |
Manufactures the leading-edge parts for every designer on the list and owns the advanced packaging they depend on.
DifferentThe one link every other link on the list has to pass through.
BasisLeading-edge revenue share and packaging capacity are disclosed; utilisation is the variable.
Works ifCapacity additions land on time and the leading-edge mix keeps rising.
Fails ifA demand pause leaves new capacity under-utilised while the capital is already spent.
|
|
| Broadcom Inc. | Custom accelerator designers | pure play | already earning | $1.8T | US listed | HELD |
The largest custom-accelerator design business; the design wins are disclosed and the revenue is already reported.
DifferentDesigns the custom parts for several of the largest cloud customers at once, which no other name on the list does.
BasisThe AI semiconductor revenue line is reported every quarter.
Works ifThe disclosed design wins ramp on schedule and the customer count widens.
Fails ifA large customer moves a programme in-house or to a second supplier.
|
Satellite — real exposure, with a qualifier
(5)
| Ticker | Company | Segment | Role | Odds | Mkt cap | Listing | Ours | The read |
|---|---|---|---|---|---|---|---|---|
| Micron Technology, Inc. | High-bandwidth memory | pure play | even | $1.2T | US listed | HELD |
High-bandwidth memory stacked beside the accelerators; the most cyclical link.
DifferentRides the volume but with memory pricing on top, so the exposure swings in both directions.
BasisHBM share and pricing are discussed each quarter; the cycle decides the margin.
Works ifHBM stays sold out while the pricing cycle holds.
Fails ifMemory supply catches up and pricing turns while the capital is committed.
|
|
| Advanced Micro Devices, Inc. | Custom accelerator designers | diversified | strong | $1.0T | US listed | HELD |
Accelerators and custom work inside a broader processor business.
DifferentA second source for accelerators; exposure is diluted by the client and gaming segments.
BasisThe data-center segment is reported separately and its growth is visible.
Works ifThe accelerator roadmap wins a second large customer.
Fails ifThe roadmap slips a generation and the segment stalls.
|
|
| ASML Holding N.V. - New York Re | Lithography & process equipment | pure play | already earning | $702.2B | US listed | HELD |
The only supplier of the lithography tools the leading edge needs.
DifferentA monopoly link; the exposure is to the capacity build, not to any one designer.
BasisOrders and backlog are reported; the timing of shipments is the variable.
Works ifLeading-edge capacity keeps being added and the services base grows with the installed tools.
Fails ifExport-control scope widens or a capacity pause stretches the order cycle.
|
|
| Arm Holdings plc | Architecture & IP licensing | pure play | strong | $300.6B | US listed | pipeline |
Licenses the architecture many custom parts are built on; royalty economics per unit.
DifferentPer-unit exposure rather than per-design exposure, so it rides volume rather than wins.
BasisRoyalty rates and licensing revenue are reported; the custom-silicon share is inferred.
Works ifCustom designs standardise on the architecture and royalty rates step up with the newer cores.
Fails ifDesigners move to an open architecture for the new parts.
|
|
| QUALCOMM Incorporated | Edge & device silicon | diversified | even | $184.9B | US listed | pipeline |
Device processors that run inference on the phone and the PC; the edge half of the theme.
DifferentThe only name on the list whose exposure is through the device rather than the data center.
BasisHandset volumes and the AI content per device are disclosed in part.
Works ifAI features lift the content per device and the PC share grows.
Fails ifA large customer insources its modem and the content gain is offset.
|
Watch — something has to happen first
(4)
| Ticker | Company | Segment | Role | Odds | Mkt cap | Listing | Ours | The read |
|---|---|---|---|---|---|---|---|---|
| Intel Corporation | Foundry & advanced packaging | diversified | long shot | $577.8B | US listed | pipeline |
A second foundry option and a packaging business inside a company rebuilding its process lead.
DifferentThe only alternative foundry on the list; the case rests on execution still ahead.
BasisProcess milestones and external foundry customers are disclosed as they land.
Works ifAn external customer commits to the new node and the packaging business fills.
Fails ifThe node slips and the external customers wait for the next one.
|
|
| Applied Materials, Inc. | Lithography & process equipment | diversified | strong | $413.2B | US listed | HELD |
Deposition and etch tools across the leading edge and the packaging step.
DifferentBroader than lithography; the exposure is to the whole capacity build.
BasisOrders and the backlog are reported; the mix by node is partly disclosed.
Works ifThe leading-edge and packaging mix keeps rising within the backlog.
Fails ifA capacity pause arrives before the backlog converts.
|
|
| Lam Research Corporation | Lithography & process equipment | diversified | strong | $411.3B | US listed | HELD |
Etch and deposition tools with a large memory exposure, which ties it to the HBM link too.
DifferentExposed to both the logic and the memory build, so it moves with both cycles.
BasisOrders are reported; the memory share of the mix is the variable.
Works ifMemory capex recovers while leading-edge logic orders hold.
Fails ifThe memory cycle turns before logic demand offsets it.
|
|
| Texas Instruments Incorporated | Edge & device silicon | diversified | even | $263.4B | US listed | pipeline |
Analog and embedded parts around the edge devices; the theme is a small share of a wide catalogue.
DifferentThe most diversified name on the list; the exposure is indirect.
BasisIndustrial and automotive demand dominate the reported numbers; the edge share is small.
Works ifEmbedded AI content lifts the attach rate per device.
Fails ifThe broader industrial cycle masks any edge-driven growth.
|
How we would play it
The desk's read is that the designers and the foundry are the direct expressions with the clearest numbers; equipment and memory ride the capacity build with their own cycles on top; the edge names are where the theme is least priced and least proven. This is a description of where the exposure sits, not a sizing instruction.
What must be true
- The largest cloud customers keep funding custom programmes through the capex cycle.
- Foundry and packaging capacity keep pace with the design wins.
- AI features on the device lift content per unit rather than merely shifting it.
What would kill it
- A capex pause at the large cloud customers that stretches every design win.
- A general-purpose roadmap that makes the custom parts uneconomic.
- Export-control changes that remove a large part of the demand.
What to watch
- Custom-silicon revenue lines at the designers.
- Leading-edge mix and packaging capacity at the foundry.
- Orders and backlog at the equipment makers.
- High-bandwidth memory pricing and the attach rate of AI features on devices.