Published
AI chip stocks across 8 layers: GPUs, custom ASICs, IP, foundry, HBM, packaging, fab tools and networking. $NVDA, $AVGO, $TSM, $MU with growth and margin data.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
AI chip stocks are publicly traded companies that design, manufacture, or connect the semiconductors that train and run AI models. The obvious names are the chip designers, Nvidia ($NVDA), Broadcom ($AVGO) and AMD ($AMD). But every accelerator also passes through seven other layers: chip IP and design software, a foundry like TSMC ($TSM), high-bandwidth memory from SK Hynix ($000660.KS) or Micron ($MU), advanced packaging, fab tools from ASML ($ASML) and its peers, test, and the networking silicon that wires thousands of chips into one cluster.
Most "best AI chip stocks" lists stop at the first layer. That leaves out where a lot of the money is being made right now. In the latest reported quarter, Micron grew revenue 346% year over year, Astera Labs ($ALAB) 154% and Nvidia 106%. Micron's gross margin (85%) was higher than Nvidia's (75%). This guide maps all eight layers, names the public companies in each across US, Taiwan, Korea, Japan and Europe listings, and then uses Macroplane's own supply-chain graph and financials to show which layers are growing fastest, which keep the margin, and which suppliers depend most on a single AI customer.
👉 Track the whole stack live: the Custom Silicon basket holds the chip designers, IP and EDA vendors; the HBM / Packaging basket holds the memory-stacking, packaging and test names; the Lithography & Fab Tools basket holds the front-end equipment makers.
An AI accelerator is a team effort. Nvidia designs a Blackwell or Rubin GPU, but it does not own a fab, a memory line, or a packaging plant. Each of those steps is its own industry with its own leaders, its own bottleneck, and its own stock.
The useful way to read this map: layers one to three decide what gets built, layers four to seven decide how many can be built, and layer eight decides how big a cluster can get. When people say "AI chip shortage," they usually mean a constraint in layers four to six (CoWoS packaging capacity, HBM supply), not a shortage of GPU designs.
These companies design AI chips and sell them to anyone who will buy. This is the layer most searches for "AI chip stocks" are really about.
$NVDA (Nvidia) sells most of the world's AI training silicon. Its latest reported quarter brought in $96.2B of revenue, up 106% year over year, at a 75% gross margin and a 62% net margin. The roadmap runs Blackwell → Vera Rubin (2026) → Feynman (2028); our Nvidia Rubin roadmap breakdown covers what changes each generation. Nvidia also sells the networking (NVLink, InfiniBand, Spectrum-X) that ties GPUs into a rack, which is why it shows up again in layer eight.
$AMD is the credible second source. Revenue grew 45% to $13.4B last quarter. Its Helios rack (72 MI455X GPUs, 31 TB of HBM4) launched in July 2026 with Microsoft, Meta, OpenAI, Anthropic and Oracle committed to deployments. Our AMD Helios vs Nvidia Vera Rubin comparison goes through the rack-level fight.
$INTC (Intel) matters less as an accelerator vendor than as a foundry and packaging option (layer four). Nvidia has both bought overflow packaging capacity from Intel Foundry and taken an equity stake in the company.
$CBRS (Cerebras) is the wafer-scale outlier: one chip the size of a dinner plate, sold as inference capacity. Revenue is still small (about $180M a quarter on Macroplane's financials) and losses are large.
China has its own merchant layer, cut off from Nvidia's best parts by export rules. Cambricon ($688256.SS) is the largest listed name.
The private challengers sit here too: Etched, Groq, Tenstorrent and SambaNova have no public ticker. Our Etched AI vs Nvidia analysis compares the inference-chip startups with Nvidia and shows how public investors get indirect exposure.
👉 Browse every AI GPU maker in the AI GPUs product category and the narrower AI Training Accelerators category.
Hyperscalers want an alternative to paying Nvidia's margin, so they design their own chips: Google's TPU, Amazon's Trainium, Meta's MTIA, Microsoft's Maia. They still need a partner with high-speed SerDes IP, physical design teams and TSMC allocation. That partner is the investable piece.
$AVGO (Broadcom) is the biggest. It co-designs Google's TPUs and Meta's accelerators and also sells the Tomahawk switch chips that connect them. Revenue grew 86% to $29.6B last quarter at a 69% gross margin. On Macroplane's graph, Google's TPU program alone accounts for roughly 15.6% of Broadcom's revenue.
$MRVL (Marvell) does the same for Amazon's Trainium and Microsoft, and owns most of the optical DSP market inside 800G and 1.6T transceivers. Revenue grew 37% to $2.74B. Nvidia invested $2B in Marvell in 2026, which tells you it sees Marvell as a partner as much as a competitor.
Alchip ($3661.TW) and Global Unichip ($3443.TW) are the Taiwanese design-services houses that take a hyperscaler's architecture to a TSMC tape-out. GUC is the best performer in our Custom Silicon basket over two years, up 617%. MediaTek ($2454.TW) has moved into co-design on Google's TPU program.
Every custom chip starts in software and licensed building blocks. These companies get paid per design and per unit, regardless of who wins the GPU fight.
$ARM (Arm Holdings) licenses the CPU architecture inside Nvidia's Grace and Vera, AWS Graviton, Google Axion and Microsoft Cobalt. That model produces a 97% gross margin, the highest of any company in this guide.
$SNPS (Synopsys) and $CDNS (Cadence) are the EDA duopoly. No chip tapes out without their tools. Nvidia has an equity stake in Synopsys and a partnership to run its EDA software on GPUs. See the EDA Software category.
$RMBS (Rambus) licenses memory-interface IP and sells the DDR5 buffer chips that sit on server memory modules.
$TSM (TSMC) makes nearly every leading AI accelerator, Nvidia's, AMD's, Broadcom's and the hyperscalers' own. In TSMC's most recent quarter on file, net margin was 47.5%. Macroplane's relationship record puts Nvidia at about 11% of TSMC revenue (a 2023 estimate that is almost certainly higher now).
Samsung Electronics ($005930.KS) and $INTC are the only other companies building leading-edge logic, and both are pitching AI customers on capacity and on US-based manufacturing. See the Foundry / Fab Services category for the full producer list.
High-bandwidth memory is DRAM stacked 8 to 16 dies high and bonded right next to the GPU. Each new accelerator generation carries more of it, and three companies make it.
SK Hynix ($000660.KS) is Nvidia's primary HBM supplier and signed a multi-year co-design pact with Nvidia in June 2026. Its 2025 revenue rose 47% to ₩97.1 trillion at a 44% net margin.
$MU (Micron) is the only US-listed HBM maker and the fastest grower in this guide: revenue up 346% to $41.5B last quarter, with an 85% gross margin. Our Micron stock breakdown covers that quarter and why the stock sold off anyway.
Samsung ($005930.KS) is the third supplier and is pushing to catch up on HBM4. Our post on Samsung's HBM5 and zHBM roadmap covers what it has promised and what it hasn't.
👉 The High Bandwidth Memory trend and the HBM product category track the oligopoly plus the equipment and materials suppliers behind it.
An AI accelerator is no longer one die. It is a logic die (or two), plus six to twelve HBM stacks, on a silicon interposer, on an organic substrate. Putting that together is TSMC's CoWoS process, and CoWoS capacity has been the binding constraint on GPU supply for most of the last three years.
👉 Follow the Advanced Semiconductor Packaging trend and the CoWoS / 2.5D-3D Integration category.
Every new fab and every packaging line is a tool order. These companies sell to TSMC, Samsung, SK Hynix and Micron, so they win no matter which chip designer does.
$ASML has a monopoly on EUV lithography, the only way to print the smallest transistor features. $AMAT (Applied Materials), $LRCX (Lam Research) and Tokyo Electron ($8035.T) split deposition and etch; $KLAC (KLA) owns process control; $ASMI (ASM International) leads atomic-layer deposition.
Test is the quiet winner. Each GPU and HBM stack is screened before it ships, and test time rises with chip complexity. Advantest ($6857.T) makes the V93000 testers used on Nvidia's H100, H200 and Blackwell parts. $TER (Teradyne) grew revenue 104% last quarter.
The front-end tool makers grew more slowly: Applied +25%, Lam +24%, KLA +12%. Fab capex follows chip demand with a lag, and memory makers in particular add capacity cautiously.
A training cluster is tens of thousands of GPUs that have to behave like one computer. The chips that move data between them are a separate and fast-growing market.
The next step is co-packaged optics, which moves the laser onto the switch package itself. Our co-packaged optics deep dive covers who supplies it, and the Silicon Photonics & Optical Interconnects trend tracks the group.
| Layer | What it sells | Public leaders | Bottleneck / moat | Latest-quarter revenue growth |
|---|---|---|---|---|
| 1. Merchant GPUs | AI training and inference chips, racks | $NVDA, $AMD, $CBRS | CUDA software, rack-scale systems | NVDA +106%, AMD +45% |
| 2. Custom ASICs | Co-designed hyperscaler chips | $AVGO, $MRVL, $3661.TW, $3443.TW | SerDes IP, TSMC allocation | AVGO +86%, MRVL +37% |
| 3. IP & EDA | CPU licences, design software | $ARM, $SNPS, $CDNS, $RMBS | Per-design and per-unit royalties | ARM +27% |
| 4. Foundry | Leading-edge wafers | $TSM, $005930.KS, $INTC | 3nm/2nm capacity, yield | INTC +21% |
| 5. HBM & memory | Stacked DRAM for accelerators | $000660.KS, $MU, $005930.KS | Three-supplier oligopoly | MU +346% |
| 6. Advanced packaging | CoWoS, bonding, substrates | $ASX, $AMKR, $BESI.AS, $042700.KS, $4062.T | CoWoS and substrate capacity | AMKR +28% |
| 7. Semicap & test | Fab tools, testers | $ASML, $AMAT, $LRCX, $KLAC, $8035.T, $6857.T, $TER | EUV monopoly, test time per chip | TER +104%, AMAT +25%, KLAC +12% |
| 8. Networking & optics | Switch chips, retimers, optics | $ALAB, $CRDO, $ANET, $COHR, $LITE | Bandwidth per GPU | ALAB +154%, LITE +137%, CRDO +115% |
Growth figures compare each company's latest reported quarter with the same quarter a year earlier, from Macroplane's company financials as of November 11, 2026.
Macroplane maps supplier and customer relationships from filings, earnings calls and news, and stores quarterly financials for every company in the catalogue. Three patterns stand out when you put the AI chip stack through it.
The common assumption is that Nvidia is the AI chip trade and everyone else is a smaller, slower version of it. The quarterly numbers say otherwise.
Four companies grew faster than Nvidia's 106% in their latest quarter: Micron, Astera Labs, Lumentum and Credo, with Teradyne right behind at 104%. All of them sell parts that scale with the number of accelerators and the bandwidth between them, not with the price of the GPU. When a rack goes from 72 GPUs to 144, or from HBM3E to HBM4, these suppliers' content per rack goes up even if the GPU price doesn't.
The laggards are informative too. Front-end tool makers (Applied, Lam, KLA) and Intel grew 12–25%, because fab capacity is added over years rather than quarters. Qualcomm ($QCOM), which has an AI data-center ambition but still mostly sells phone chips, shrank 4%.
Revenue growth tells you who is selling more. Gross margin tells you who has pricing power.
The top of this chart is a list of near-monopolies: Arm's architecture, the three-company HBM oligopoly, Nvidia's CUDA ecosystem, Astera's retimers. The bottom is a list of companies that assemble other people's high-value parts. Fabrinet builds optical and electronic assemblies for Nvidia and keeps 12 cents of each dollar; Amkor packages chips and keeps 14. Both are growing, and both are useful businesses, but a dollar of AI chip revenue is worth far less to them than to the owner of the scarce input.
That gap is the practical answer to "which AI chip stock?" A stock's exposure to AI demand matters less than whether it owns a chokepoint in its layer.
Macroplane's supply-chain graph records, for many supplier relationships, how much of the supplier's revenue comes from one customer. These are the most concentrated AI-chip links in the graph:
| Supplier | Customer | Share of supplier revenue | What it sells | Source date |
|---|---|---|---|---|
| Advantest ($6857.T) | $NVDA | ~20% | V93000 GPU test systems | 2024 |
| $AVGO | Alphabet ($GOOGL) | ~15.6% | TPU co-design and networking ASICs | 2024–2025 |
| $MRVL | Amazon ($AMZN) | ~14–15% | Trainium co-design and cloud networking | 2024 |
| $TSM | $NVDA | ~11% | Foundry wafers and CoWoS | 2023 |
| $MU | $NVDA | ~8% | HBM3E for H200 and Blackwell | 2024 |
Two caveats. These shares come from filings and reports dated 2023–2025, and given Nvidia's growth since then, the Nvidia-linked figures are probably higher today. And customer concentration cuts both ways: it is the reason these suppliers grew so fast, and it is the single biggest risk to each of them if one customer slows orders or switches vendors.
Here is the full picture for Nvidia, built from the relationships in Macroplane's graph:
The graph shows something the income statement doesn't: Nvidia now holds equity on both sides of its supply chain. It has invested in Marvell ($2B), CoreWeave ($CRWV, $2B), Nebius ($NBIS, a 9.3% stake), Synopsys and Intel, and it backs laser supplier Lumentum. Those stakes help secure supply and seed demand. They also mean that a slowdown at Nvidia would travel through the stack faster than the supplier lists alone suggest.
There are three practical routes, each with a different risk profile.
Buy the designers. $NVDA, $AVGO and $AMD give the most direct exposure to AI accelerator demand. They are also the most crowded trades and the most exposed to a pause in hyperscaler spending.
Buy the chokepoints. HBM ($MU, $000660.KS), CoWoS-related packaging and test ($6857.T, $TER, $BESI.AS), EUV ($ASML) and retimers ($ALAB) sell into every accelerator regardless of who designed it. The trade-off is cyclicality: memory and equipment have historically had sharp downturns when capacity catches up with demand.
Buy a basket or ETF. Semiconductor ETFs such as the VanEck Semiconductor ETF ($SMH) and the iShares Semiconductor ETF ($SOXX) hold most of the US-listed names above, weighted heavily toward Nvidia, Broadcom and TSMC. They do not hold Korean, Japanese or most Taiwanese listings (SK Hynix, Advantest, Alchip, GUC). Macroplane's equal-weight Custom Silicon, HBM / Packaging, Memory Supercycle and Networking / Retimers baskets track those layers across exchanges, so you can see when one layer starts to diverge from the rest.
The capex cycle. Almost all AI chip revenue traces back to a few hundred billion dollars a year of hyperscaler and AI-lab spending. When investors doubted that spending on November 5, 2026, tech stocks lost about $1.3T at the open and AMD fell 12% (see why tech stocks fell). The stack moves together on that question.
Customer concentration. As the table above shows, several suppliers get 10–20% of revenue from one buyer. Nvidia itself sells most of its data-center output to a short list of hyperscalers and GPU clouds.
Custom silicon substitution. Every TPU, Trainium or MTIA chip a hyperscaler deploys is one it didn't buy from Nvidia or AMD. That is good for $AVGO, $MRVL and the Taiwanese design houses and a slow headwind for merchant GPU margins.
Taiwan and export controls. Most leading-edge logic, CoWoS capacity and substrate supply sits in Taiwan. US export rules already cut Nvidia and AMD off from much of the Chinese market and created a domestic Chinese competitor set.
Memory and equipment cyclicality. Micron's 85% gross margin is a scarcity margin. The memory makers are spending heavily on new capacity, and in past cycles that spending is what ended the boom.
Valuation. Nvidia's market value is above $5 trillion on Macroplane's latest data. Good news is widely expected across the stack, so a merely good quarter can still send a stock lower, as Micron showed after its record quarter.
AI chip stocks are public companies that design, manufacture or connect the semiconductors used to train and run AI models. That includes GPU and accelerator designers like $NVDA, $AMD and $AVGO, and the companies they depend on: TSMC ($TSM) for manufacturing, SK Hynix ($000660.KS) and Micron ($MU) for memory, packaging and test specialists, fab-tool makers like $ASML, and networking chip companies like $ALAB.
Nvidia ($NVDA) designs and sells the most AI accelerators by revenue: $96.2B in its latest reported quarter. But TSMC ($TSM) physically manufactures almost all of them, including Nvidia's, AMD's and the hyperscalers' custom chips. So the answer is Nvidia by sales and TSMC by production.
It depends on which layer you want. Broadcom ($AVGO) and Marvell ($MRVL) lead custom AI chips; AMD ($AMD) is the main GPU alternative; Micron ($MU) and SK Hynix ($000660.KS) dominate HBM memory; TSMC ($TSM) and ASML ($ASML) sit underneath everyone; Astera Labs ($ALAB) and Credo ($CRDO) are the fastest-growing networking chip names. In the latest quarter, Micron, Astera Labs, Lumentum and Credo all grew revenue faster than Nvidia.
On gross margin in the latest reported quarter: Arm ($ARM) at 97%, Micron ($MU) at 85%, Nvidia ($NVDA) at 75%, Astera Labs ($ALAB) at 72% and Broadcom ($AVGO) at 69%. Assemblers sit at the bottom: Amkor ($AMKR) at 14% and Fabrinet ($FN) at 12%.
Broad semiconductor ETFs such as $SMH and $SOXX hold most US-listed AI chip names, with heavy weights in Nvidia, Broadcom and TSMC. They largely miss the Korean, Japanese and Taiwanese listings in memory, test and design services. Macroplane's Custom Silicon and HBM / Packaging baskets track those layers across exchanges.
No. This guide is research and education, built from Macroplane's supply-chain and financial data as of November 11, 2026. It is not a recommendation to buy or sell any security. Do your own due diligence, and check the latest filings before acting on any number here.