The AI Chip Chain on One Page

Diagram of the AI chip supply chain in eight layers, from materials and equipment through design, memory, foundry, packaging, cloud systems and finally applications such as ChatGPT, with the leading companies named in each layer
Hero image: original illustration created for this post.

People argue about artificial intelligence using the names of products. ChatGPT, Copilot, Gemini, Claude. Those names sit at the very bottom of the picture above, on the last line. Everything above them is what has to exist before a single answer can be typed back to you. The map is worth reading slowly, because it explains something the product news usually hides: the industry's limit right now is not imagination, and it is not even chip design. It is the physical middle of the chain.

Read it from the top. Each layer hands something to the layer below it, and each layer contains companies with a different answer to the same question — how do we avoid being the one who gets squeezed.

1. Materials

Silicon wafers, the light-sensitive chemicals used to print circuits, ultra-pure gases, and the substrate boards a finished chip is mounted on.

Who: Shin-Etsu and SUMCO dominate wafers. JSR and Sumitomo Chemical sit in the chemistry. Ibiden makes the high-end substrates that AI accelerators need.

What they are doing about AI: mostly capacity, quietly. There is no clever strategy available at this layer — you either have qualified material or you do not. This is also the layer most exposed to export controls, because a license can stop a shipment faster than a factory fault can. A GPU is a layered object, and some of those layers can be slowed by paperwork.

2. Equipment

Nobody manufactures leading-edge chips without buying machines first.

Who: ASML is the only company selling extreme ultraviolet lithography, the tool that prints the finest features. Applied Materials, Lam Research, KLA, and Tokyo Electron handle the deposition, etching, and inspection steps around it.

What they are doing about AI: selling into two booms at once, logic and memory. ASML raised its 2026 revenue outlook to €43–45 billion and noted that memory had grown to nearly half of system sales. That second detail matters more than the first. It is the equipment layer confirming that HBM factories — not just GPU factories — are still ordering tools, which is an upstream confirmation of a downstream shortage.

3. Design and IP

This is the layer people recognise, and the one where brand names explode.

Who: Nvidia and AMD design merchant accelerators anyone can buy. Broadcom and Marvell co-design the custom chips that hyperscalers put their own badge on. Qualcomm and MediaTek are arriving from the phone world. Arm licenses the CPU and NPU blocks that ride along in the rack, and Synopsys and Cadence sell the software used to draw all of it.

Nvidia no longer sells a GPU; it sells a rack. Vera Rubin, NVLink Fusion so that even non-Nvidia silicon can plug into its fabric, and a $3.5 billion investment in MediaTek so custom ASICs still live in its neighbourhood. Its defining move this quarter was financial rather than architectural: supply and capacity commitments more than doubled from roughly $119 billion to about $279 billion, mostly to lock down memory. Revenue was $96.2 billion last quarter with about $108 billion guided, and gross margin guidance slid from 75% toward a 71–72% trough because memory prices beat the company's own forecasts. The most powerful firm in AI is paying a tax to its suppliers.

AMD is the second merchant supplier and deserves more than a footnote, because a credible alternative is exactly what cloud buyers want. Instinct MI355X ships with 288GB of HBM3E — Micron and Samsung memory — on TSMC CoWoS-S packaging, aimed at customers who want more memory per chip and a less closed software stack through ROCm. Lisa Su's larger bet is Helios, a 2026 rack combining EPYC CPUs, MI400-class GPUs, and Pensando networking. The constraint is the same as everyone else's: a competitive die that cannot get packaged never becomes a cloud instance.

Broadcom and Marvell are how cloud companies pretend they made the chip themselves. Broadcom co-designed OpenAI's Jalapeño inference chip and sits inside Google's TPU programme, which is why the report that its CoWoS allocation was trimmed toward 200,000 wafers was not trivia. Marvell appears in Amazon's Trainium and Microsoft's Maia, plus a broader attach deal around Google's racks. Qualcomm announced a multi-generation collaboration on September 8 to build custom AI-inference silicon and optical interconnects for Amazon's data centres — a company escaping a phone-only future. Arm does not run the chat, but it collects a royalty when the blocks beside the accelerator are used.

4. Memory

An accelerator is useless if it cannot be fed. High-bandwidth memory is DRAM stacked into a tower and placed millimetres from the processor, because the bottleneck is not only how fast the math units multiply — it is how fast they can be supplied with numbers.

Who: only SK hynix, Samsung, and Micron make it at volume.

What they are doing about AI: collecting. SK hynix is the incumbent, and its pricing power is now visible inside Nvidia's margin guidance. Micron is sold out of 2026 HBM and is qualifying HBM4. Samsung is trying to close the gap while using memory as the door into a wider relationship with OpenAI. Three companies deciding how fast every assistant on earth can run is the single most concentrated point of leverage on the map.

5. Foundry

Nvidia, AMD, Broadcom, and Apple do not own factories.

Who: TSMC makes almost all leading-edge AI silicon. Samsung Foundry wants to be the credible second source. Intel Foundry is trying to re-enter.

What they are doing about AI: TSMC's August revenue of NT$514.8 billion, up 53.3% year over year, is the most honest demand gauge published this month — chips shipping, not chips promised. It also treats advanced packaging as a product rather than a courtesy, which is how it stays the factory nobody can route around. Samsung is running three bets at once: catch up in HBM, sell foundry as a second source, and stay in OpenAI's conversation. OpenAI's Korea office said this month that joint research and production on next-generation chips with Samsung was among its areas of most progress, though OpenAI added that nothing new had been announced — direction, not a signed contract. Intel is attacking from a different door, discussed next.

6. Packaging and test — the actual ceiling

Here is the layer the map exists to highlight. A fabricated wafer is not a sellable product. The logic die and the memory stacks have to be bonded onto an interposer, attached to a substrate, and tested.

Who: TSMC's version is called CoWoS. ASE with its SPIL subsidiary, Amkor, and JCET take the overflow. Intel is selling an alternative.

What they are doing about AI: this step decides how many AI accelerators reach the world. Nvidia is estimated to hold around 60% of CoWoS capacity. TSMC is pushing toward roughly 120,000–130,000 CoWoS wafers a month by late 2026 and still outsources a large share of the on-substrate work — which is why ASE raised 2026 capital spending to a record $10.5 billion and expects leading-edge packaging revenue to double again in 2027, and why Amkor is building U.S. capacity on multi-year ties to TSMC and Nvidia. The overflow valve has become a business. Intel is pitching EMIB-T as a way to skip the CoWoS queue entirely, with high volume targeted for 2027; it is being discussed as a possible path for a future Google TPU generation, which would make it a packaging decision as much as a chip decision. The next fight here is optical: co-packaged optics, putting light sources next to the switch so data-centre power does not vanish into copper. TSMC expects silicon photonics to enter mass production in the second half of this year, with ASE, Nvidia, and Foxconn already in that chain.

7. Systems and cloud

Packaged chips become rentable capacity.

Who: Microsoft, Google, and Amazon. Foxconn, Dell, HPE, and Supermicro bolt the boards into racks.

What they are doing about AI: every hyperscaler is buying Nvidia in volume while designing its own silicon, so that it is not purely a customer. Microsoft hosts much of ChatGPT on Azure and is preparing Maia so Copilot is not a single-supplier product. Google is furthest down this road: TPUs on a faster cadence, and TPU systems now sold outside its own cloud, with Broadcom, MediaTek, and Marvell around the stack. Amazon keeps Trainium and Inferentia for AWS and has now added Qualcomm as a custom-silicon partner. The shared goal is to make answering a query cheaper than renting someone else's GPU.

8. Applications

The last line, and the only layer most people ever touch.

Who: ChatGPT, Copilot, Gemini, Claude, and the coding agents built on them. OpenAI launched GPT-6 Astra on September 4, pitched at computer use, professional work, science, coding, and cybersecurity.

What they are doing about AI: moving upstream. OpenAI is no longer only a model vendor. Jalapeño, its inference chip designed with Broadcom and made at TSMC, was unveiled in June; Samsung and SK hynix already sit in the Stargate memory story; and the Seoul comments show the next worry is volume too large for one foundry's queue. The company that taught the public to chat with a model is now shopping for wafers, memory, and a second manufacturing door. This layer cannot grow faster than layers four and six permit.

The same map, priced

Stock prices, measured from the end of the latest full quarter — June 30, 2025 to June 30, 2026 — say the same thing more bluntly than any keynote. This is not investment advice. It is a picture of where the market believed the bottlenecks were.

Company Layer on the map Q2 stock YoY
Micron4 — Memory+837%
SK hynix4 — Memory+808%
Intel5 / 6 — Foundry and EMIB packaging+523%
Samsung4 / 5 — Memory and foundry+440%
ASE6 — Advanced packaging+337%
Amkor6 — Advanced packaging+311%
AMD3 — Design+309%
Applied Materials2 — Equipment+295%
Marvell3 — Custom silicon+285%
ASML2 — Equipment+148%
Arm3 — Chip IP+119%
TSMC5 / 6 — Foundry and CoWoS+111%
Broadcom3 — Custom ASICs+37%
Nvidia3 — Design and systems+27%
Qualcomm3 — New data-centre entrant+16%

Read the table against the diagram rather than as a scoreboard. The top of the list is layers four and six — memory and packaging — which is precisely where the map says the ceiling is. AMD's +309% is the market paying for a second GPU supplier, which is what every cloud buyer wants if they live in fear of Nvidia's queue. Nvidia's modest +27% sits on top of an already enormous valuation: extraordinary revenue, but the stock had long since assumed the ChatGPT era, and this year it is handing margin to the memory layer. Qualcomm's small move is the opposite tell — the Amazon deal is a beginning, not a run rate. Broadcom's middling year, despite the custom-ASIC boom, fits a week in which its CoWoS allocation was the thing people argued about.

What the map is really saying

If you had drawn this diagram two years ago, the tight spot would have been layer five: could TSMC make enough advanced dies. Today the pressure has moved one step down into packaging and one step sideways into memory. That is why a company can announce a brilliant accelerator and still not ship it, and why the loudest brand in the chain is absorbing a margin squeeze while its suppliers collect it.

It also explains every strategy above. Nvidia committing $279 billion in advance is a company purchasing the bottleneck rather than complaining about it. OpenAI courting Samsung is a customer buying a second door. Intel pitching EMIB-T is a challenger attacking the queue instead of the transistor. ASE and Amkor spending records is the overflow valve turning into a business. None of it makes sense if you only look at the last line of the map.

Three unglamorous numbers will tell you whether the ceiling is lifting over the next year: whether TSMC's CoWoS output really reaches about 130,000 wafers a month, whether Micron's and Samsung's HBM4 qualifications broaden the memory supplier base, and whether co-packaged optics yields hold up in volume. They will shape the next model launch more than any keynote. The interface that changed search and work is, in the end, downstream publicity for an upstream shortage.

References