Semiconductors, advanced hardware & quantum

Semiconductors, advanced hardware & quantum

Industry ID: 03 | Slug: semiconductors-hardware | Researched: 2026-09-15 | Analyst: agent

1. Definition and boundaries

In scope. The design, manufacture, packaging, test and sale of integrated circuits and the capital equipment, materials and design software required to make them: logic (CPU, GPU, AI accelerators, custom ASICs), memory (DRAM, HBM, NAND), foundry services, advanced packaging and OSAT, electronic design automation, semiconductor capital equipment, networking and optical silicon, and the consumer devices whose economics are currently dominated by component cost. Quantum computing hardware is included as a compute-roadmap subindustry (see §2 and the boxed treatment below).

Explicitly out of scope, and who owns it.

• AI models, training runs, cloud compute sold as a service → sector 01 (AI foundation models & compute infrastructure). The boundary is the sale of silicon versus the sale of FLOPs.
• Data-centre electricity generation, grid interconnection, nuclear and gas → sector 05 (energy & power). This dossier treats power as a constraint on semiconductors (T-03-07), not as a market.
• Data-centre buildings, land and REIT economics → sector 18 (real estate & construction).
• Post-quantum cryptography, Q-Day risk, cryptographic migration → sector 04 (cybersecurity). This is the most consequential boundary dispute in this dossier: PQC is quantum computing's single most monetisable near-term externality and it lives in a different sector from the hardware that causes it.
• Critical minerals and wafer-grade silicon upstream → sector 09 (industrial & supply chain).
• 6G and telecom network equipment → no standalone sector exists; network silicon routes here and the roster names telecom as the first candidate for a Phase 2 addition.
• Automotive semiconductors → demand side sits in sector 10; the silicon itself sits here.

Named boundary disputes.

1. Hyperscaler custom silicon. Google TPU and AWS Trainium are semiconductor products designed by companies whose primary sector is cloud. Their revenue is invisible - it appears as cloud revenue, not chip revenue - which systematically understates this sector's measured size.
2. Consumer devices. Phones and PCs were folded into 03 by the roster. In 2026 that call looks right for the wrong reason: the dominant story in consumer devices is not device innovation but component cost pass-through (T-03-08, T-03-18), which is a semiconductor story.
3. Quantum. Treated at length below.

2. Subcategories

Box: the quantum merge — justified, but lossy

The roster folded quantum into 03 rather than giving it a standalone slot. The evidence gathered here supports the merge on 2026 economics and challenges it on analytical grounds.

The merge is right because: quantum's binding constraints have become semiconductor constraints. IonQ acquired SkyWater Technology, a US foundry, for $1.8bn in January 2026. PsiQuantum's photonic approach is a 300mm CMOS process problem. Intel's spin qubits run on conventional lines. Rigetti's manufacturing partner is Quanta, a Taiwanese ODM. Cryo-CMOS control electronics, packaging, metrology and test are the same toolchain. And the scale argument is decisive: the entire quantum industry's annual revenue is smaller than roughly one week of NVIDIA data-centre revenue. A standalone sector slot for it would distort a ranked list built on economic weight.

The merge is lossy because: (a) quantum's capital formation, talent pool and government funding are structurally separate from semiconductors; (b) its most monetisable near-term consequence, post-quantum cryptography, has been routed to sector 04, so the hardware and its commercial externality now sit in different dossiers; (c) its analytic clock is 2029+ while this sector's is twelve months, which means blended scoring pushes every quantum trend's adoption, revenue and persistence toward zero and renders them invisible in ranking; (d) company-tracking density is high relative to revenue - six modalities, dozens of firms - which is an intelligence-demand argument the economic-weight argument cannot see.

Recommendation: keep the merge for Phase 1 ranking, carve quantum out as an explicitly tagged subindustry with its own trend records (T-03-13, T-03-14, T-03-19 here), and revisit it in Phase 2 alongside telecom. Also flag to sector 04 that PQC urgency claims must be checked against T-03-14 and T-03-19 rather than against vendor announcements.

3. Market structure

Concentration: winner-take-most at every chokepoint, and the chokepoints are stacked.

• Foundry: one supplier at the leading edge. Top-10 foundry revenue was US$53.49bn in Q2 2026, up 11.5% QoQ (TrendForce, 2026-09-09), with TSMC taking the overwhelming majority of leading-edge value. Samsung is a distant second and improving; SMIC is narrowing the gap with Samsung on total revenue while remaining far behind on node.
• Lithography: ASML is the sole EUV supplier. Q2 2026 net sales EUR 9.3bn; 86 new systems sold.
• Memory: three suppliers, all currently earning utility-monopoly-like returns.
• Advanced packaging: TSMC dominant, with Amkor and ASE/SPIL as capacity relief valves.
• EDA: effectively three firms.

Where margin actually sits. In 2026, margin has moved up the stack toward the scarcest physical asset, and that is no longer logic. The observable ranking of gross margin among firms that disclose it:

SK hynix reported a 76% operating margin in Q2 2026 and Samsung's Device Solutions division turned KRW 127.5tn of revenue into KRW 89.2tn of operating profit. A memory manufacturer out-earning the leading-edge foundry on margin has no precedent in this industry's history.

Barriers to entry. Capital intensity (TSMC 2026 capex $60-64bn; Micron FY2026 capex ~$27bn), process know-how that does not transfer, EUV tool allocation, a two-to-three-year fab construction lead time, and - newly - packaging capacity that is allocated years ahead. Rapidus and Intel Foundry are both proof that state capital alone does not clear these barriers quickly.

Who has pricing power. Memory makers, unambiguously and for the first time in a decade. TSMC, structurally. ASML, absolutely. Everyone else is a price taker, including NVIDIA on its HBM inputs and Apple on its DRAM inputs.

A note on market size. This dossier deliberately quotes no total-addressable-market figure. Search results for this sector are saturated with "$XXX billion by 2032 at NN.N% CAGR" content-farm pages with no stated methodology. The reference series for semiconductor market size is WSTS, whose detailed data is member-restricted; the equipment reference series is SEMI WWSEMS, whose regional detail is subscription-only. Where a size figure is genuinely needed, it should be built bottom-up from company-disclosed revenue, which is what the table above does.

4. Who matters

Leading companies. TSMC (https://www.tsmc.com) · Samsung Electronics (https://www.samsung.com/semiconductor/) · SK hynix (https://www.skhynix.com) · Micron (https://www.micron.com) · NVIDIA (https://www.nvidia.com) · Broadcom (https://www.broadcom.com) · Marvell (https://www.marvell.com) · Intel (https://www.intel.com) · ASML (https://www.asml.com) · Applied Materials · Lam Research · Tokyo Electron · KLA · Advantest · Kioxia (https://www.kioxia.com) · Amkor (https://www.amkor.com) · ASE Technology Holding (https://www.aseglobal.com) · Sony Semiconductor Solutions (https://www.sony-semicon.com) · SMIC (https://www.smics.com) · Huawei/HiSilicon (https://www.huawei.com) · CXMT (https://www.cxmt.com) · Arm · Qualcomm · MediaTek · Synopsys · Cadence.

Notable startups and challengers. Rapidus (https://www.rapidus.inc) - state-backed 2nm foundry, mass production targeted 2027 · SiCarrier - Chinese equipment · Ayar Labs, Celestial AI, Lightmatter - optical I/O · Astera Labs, Credo - connectivity · Tenstorrent, Groq, Cerebras - alternative accelerators. In quantum: PsiQuantum (https://www.psiquantum.com), IonQ (https://ionq.com), Quantinuum (https://www.quantinuum.com), Rigetti (https://www.rigetti.com), QuEra (https://www.quera.com), Pasqal, Atom Computing, Infleqtion, Xanadu, D-Wave (https://www.dwavesys.com).

Active investors. China National IC Industry Investment Fund Phase III ("Big Fund III") - the largest single state vehicle, reported at $2bn into CXMT in 2024 · SoftBank · Quanta Computer (strategic, $100m+ into Rigetti) · Churchill Capital Corp X (Infleqtion SPAC, ~$550m) · public equity markets, which are now the dominant channel: SK hynix raised $26.5bn in a US offering in July 2026 and Kioxia is reported to be weighing a $10bn+ ADR.

Platforms and standards bodies. SEMI (https://www.semi.org) - standards and the WWSEMS equipment dataset · WSTS (https://www.wsts.org) - the market-size reference series · JEDEC - memory standards including HBM4 · SPIE/BACUS - the lithography and photomask venue where the TSMC-ASML 12-inch mask initiative was staged · UCIe - chiplet interconnect · OCP - rack and power standards.

Regulators. US Bureau of Industry and Security (https://www.bis.doc.gov) · USTR and Commerce under Section 232 (https://ustr.gov) · US Customs and Border Protection (tariff administration) · Japan METI (https://www.meti.go.jp/english/) · Korea MOTIE · Taiwan MOEA · Netherlands Ministry of Foreign Affairs (ASML export licensing) · European Commission DG GROW (EU Chips Act).

Research institutions. imec (https://www.imec-int.com) · IBM Research (https://www.ibm.com/quantum) · Google Quantum AI (https://quantumai.google) · ITRI (Taiwan) · AIST and LSTC (Japan) · Fraunhofer (Germany) · NSTC (US National Semiconductor Technology Center).

Trade organisations. SEMI · SIA (US) · SEMI Japan · TSIA (Taiwan Semiconductor Industry Association) · KSIA (Korea) · ESIA (Europe) · CSIA (China).

Consumer and civil-society groups. Thin and this is a real gap. The Right to Repair coalition and consumer bodies engage with device pricing but not with component markets; no significant civil-society organisation systematically tracks memory pricing as a consumer-welfare issue, despite DRAM reportedly being ~60% of a budget handset's bill of materials.

5. Products, business models, technologies, customers

Major products. Leading-edge logic wafers (N2, N3, 18A); HBM3E/HBM4/HBM4E stacks; DDR5 and LPDDR; NAND and enterprise SSD; merchant accelerators (NVIDIA Vera Rubin, AMD Instinct); custom XPUs (Google TPU v7 Ironwood, AWS Trainium2/3, Meta MTIA, Microsoft Maia); switch and optical silicon; EUV and DUV scanners; CoWoS/SoIC packaging services; EDA tool suites; and the consumer devices at the end of the chain.

How money is actually made, and how that is changing.

• Traditional: wafer sales at negotiated prices, memory at spot-and-contract, equipment as capital sales plus a large installed-base service annuity (ASML: EUR 2.762bn of installed-base management revenue in one quarter).
• What changed in 2026: memory moved from spot pricing to administered allocation. SK hynix reports long-term agreements with around ten customers; Micron has signed multi-year deals; Korean reporting suggests roughly half of total capacity may go to priority customers on contract from 2027. If most output is sold forward, the classic memory cycle - spot collapse forcing discipline - may not operate as it has since the 1990s. This is T-03-09 and it is the most under-observed structural change in the sector.
• Also changing: the custom-ASIC model moves accelerator margin from a merchant vendor to a design partner plus the buyer. Broadcom's AI semiconductor revenue reached $16.7bn in FQ3 2026, +221% YoY, guided to $21.7bn in FQ4.

Technologies that matter now. Gate-all-around transistors and backside power delivery (N2, 18A); HBM4 and hybrid bonding; CoWoS-L and the transition to panel-level CoPoS; co-packaged optics; 800VDC power distribution and direct liquid cooling; chiplet disaggregation and UCIe; and, on a decade horizon, High NA EUV with 12-inch photomasks.

Customers and what they buy on. Hyperscalers and frontier labs buy on total cost of ownership per unit of useful compute, subject to availability - availability is currently the binding term, which is why Anthropic runs a million TPUs, a million Trainium and NVIDIA GPUs simultaneously. Device OEMs buy on bill-of-materials cost, which is why memory pricing is now their dominant variable. Automotive and industrial buy on qualification and longevity, and have been largely crowded out of capacity. Governments buy on sovereignty.

6. Geography

Production is concentrated in East Asia to a degree that has no parallel in any other sector. Leading-edge logic is Taiwan; HBM and DRAM are South Korea with Micron in the US, Japan and Singapore; NAND is Korea and Japan; advanced packaging is Taiwan; materials and chemicals are overwhelmingly Japan. Lithography is a single Dutch company.

Capital is increasingly American even where production is Asian. SK hynix raised $26.5bn in a US offering in July 2026; Kioxia is reported to be considering a $10bn+ US ADR. US CHIPS Act money, Japanese METI subsidy for Rapidus and Kumamoto, the EU Chips Act and China's Big Fund III all push in the same direction - state capital subsidising duplicate capacity - while private capital concentrates in US-listed vehicles.

Demand is concentrated in the US: NVIDIA's guidance now assumes zero China data-centre compute revenue, and Americas took 66% of Q2 2026 global VC per the macro brief. This is a genuine single-market dependency for a globally-produced industry.

Regulation is concentrated in Washington, with a veto shared by The Hague and Tokyo on equipment, and an increasingly decisive counter-veto in Beijing.

Non-US market, with regional sources: Taiwan and South Korea

Taiwan. TSMC's own IR disclosure (investor.tsmc.com) is the sector's primary document. Taiwan's mandatory monthly revenue filing regime via the TWSE Market Observation Post System gives higher-frequency hard data on foundry, OSAT, ODM and test than any other jurisdiction on earth - TSMC published its August 2026 revenue on 10 September 2026. Taiwanese research house TrendForce and Taiwan News carried the ASML-TSMC 12-inch photomask initiative (https://www.taiwannews.com.tw/news/6435631, 2026-09-08) and the Q2 2026 foundry ranking. TSMC has simultaneously committed a further $100bn to Arizona, taking that site to $265bn - a genuine geographic hedge, but one that does not move packaging out of Taiwan at comparable scale.

South Korea. SK hynix Newsroom and Samsung Global Newsroom are both Tier A Korean-origin primaries and both were used directly here. Seoul Economic Daily (https://en.sedaily.com/finance/2026/07/12/hbm4-prices-to-double-next-year-as-samsung-sk-hynix-keep, 2026-07-12) reported HBM4 pricing moving from ~$2/Gb in H2 2026 to $4-5+/Gb in 2027, and that HBM would take ~30% of DRAM capacity - weeks ahead of equivalent English-language trade coverage. The Korea Times (2026-07-29) supplied the consensus-versus-actual comparison that explains why SK hynix shares fell 9.61% on a record result. Korea's MOTIE monthly export statistics, published within days of month-end, are the fastest public read on global memory demand that exists.

Japan is the sector's least-covered critical node. Materials (Shin-Etsu, SUMCO, JSR), photoresist, mask blanks (Hoya) and much of the equipment base are Japanese, and English-language coverage outside Nikkei Asia is thin. Rapidus is targeting 2nm mass production in 2027 from Hokkaido. See §13.

7. Historical trend patterns

Ten to twenty-five years of this sector's cycles, and the false positives that matter:

The memory cycle, four times over. 2007-09, 2011-12, 2015-16 and 2018-19 all followed the same script: shortage → record margins → capacity commitment at the peak → oversupply → margins to zero or below. Elpida went bankrupt in 2012 after exactly this sequence. The single most important question in this dossier is whether 2026 is the fifth instance or a genuine structural break (see §12). The bull case rests on long-term agreements changing the mechanism (T-03-09); the bear case is that every prior peak also produced a "this time is different" argument.

False positive: 450mm wafers. The industry spent roughly a decade and substantial capital on a 450mm transition consortium (G450C) that was quietly abandoned around 2017. Relevant because the TSMC-ASML 12-inch photomask initiative (T-03-11) is a structurally identical bet: a decade-long, industry-wide format transition requiring coordinated investment across scanner makers, mask shops and fabs. It may work. 450mm did not.

False positive: the "AI PC" of 2024-25. A well-funded industry consensus held that on-device AI would drive a broad replacement cycle. In 2026 the mechanism inverted - the memory those features require is what makes the device unaffordable (T-03-18). The general lesson: feature-led upgrade-supercycle theses in consumer hardware have a poor track record in this sector, and the failure mode is usually cost, not capability.

False positive: Moore's Law is dead, repeatedly. Declared dead around 2016 at 10nm, again around 2019 at 7nm. Scaling continued via new transistor architectures (FinFET → GAA), backside power and packaging. The correct read was never "scaling stops" but "scaling moves" - from the transistor to the interconnect to the package. That is exactly what T-03-02 and T-03-11 describe.

Partial false positive: quantum, twice. A 2017-19 wave (D-Wave, IBM Q, Google supremacy) and a 2021-22 SPAC wave (IonQ, Rigetti, D-Wave all going public) both produced valuations far ahead of capability, and both retraced. The 2025-26 wave has the same structural signature - SPACs (Infleqtion, Xanadu), large private rounds, milestone announcements without comparable benchmarks. What is genuinely different this time is qLDPC overhead reduction (T-03-14), which is a real technical result rather than a scaling announcement. That does not make the valuations right.

Genuine structural break: the foundry model. The 2000s consensus that integrated device manufacturers would remain competitive at the leading edge was wrong. TI, IBM, GlobalFoundries and eventually Intel all exited or fell behind leading-edge logic. This one really was permanent, and it is why Intel Foundry's difficulty (T-03-20) should be read against a twenty-year base rate of failure rather than against its own roadmap.

Cycles in equipment. Equipment billings have historically led chip revenue by four to six quarters and have overshot in both directions. SEMI's own December 2025 forecast of $145bn for 2026 is already being exceeded by its own quarterly actuals - a forecast error in the upward direction, which is the rarer and more interesting failure.

8. What is changing now (as of 2026-09-15)

Five things have changed since the start of 2026, and the macro brief frames all of them.

1. Memory, not logic, is the story. The FOMC named "the surge in demand related to the AI buildout" as an inflation driver; this sector is where that surge lands first. Three independent Tier A disclosures - Micron (84.9% gross margin), SK hynix (76% operating margin) and Samsung DS (KRW 89.2tn operating profit on KRW 127.5tn revenue) - describe a supercycle without precedent.

2. The January 2026 export-control shift did not do what it appeared to do. The macro brief correctly records that BIS moved to case-by-case review. What the shipment data shows is that the loosening was inert: licences were granted from February 2026, China restricted purchases, NVIDIA wrote down $0.4bn of H200 inventory, and its guidance now assumes zero China data-centre compute revenue. Export policy loosened on the US side and tightened on the Chinese side, netting to closure. Any forecast built on the January liberalisation as a demand event is wrong (T-03-05, T-03-17).

3. The 25% Section 232 tariff is far narrower than its headline. It applies to three HTS headings and, within them, only to two narrow performance bands - TPP 14,000-17,500 with DRAM bandwidth 4,500-5,000 GB/s, or TPP 20,800-21,100 with 5,800-6,200 GB/s - with exemptions for US data centres, US R&D, US startups, non-data-centre consumer electronics, non-data-centre industrial use, repairs and the US public sector. Those bands map precisely onto H200/MI325X-class parts imported for third-party lab certification before re-export to China. It is functionally a toll on the export-control compliance pathway, not a broad semiconductor import tariff. Its direct cost impact on the sector is small; its signalling value is large; and Phase 2 - which the proclamation left open for additional semiconductors and manufacturing equipment - has not been imposed as of this research date despite an April 14, 2026 negotiation status checkpoint.

4. The bottleneck has moved twice. From wafers to advanced packaging (T-03-02), and now increasingly from packaging to power and thermal (T-03-07), where racks exceed 250kW and data centres represent ~90% of ERCOT's large-load interconnection pipeline. The rate-limiting step on AI compute is drifting out of this sector entirely.

5. Capital cost is a live constraint and the sector is raising equity into it. Against the macro brief's rising-rate-risk backdrop, SK hynix raised $26.5bn and Kioxia is reported to be considering $10bn+, both in US markets, both to fund capacity that arrives after the current pricing peak. That is the classic memory-cycle error executed at unprecedented scale (T-03-12).

9. The five lists

Five most important current trends

1. Memory supercycle: HBM-led DRAM/NAND pricing at unprecedented margins (T-03-01)
2. Advanced packaging, not wafers, is the binding constraint on AI compute (T-03-02)
3. Custom accelerators take measurable share from merchant GPUs (T-03-03)
4. Leading-edge foundry concentrates further as TSMC ramps N2 and capex (T-03-04)
5. Export-control loosening on paper, tightening in practice (T-03-05)

Five fastest-growing signals

1. Memory supercycle (T-03-01) — overlaps with list 1
2. Custom accelerators (T-03-03) — overlaps with list 1; Broadcom AI revenue +221% YoY
3. Semicap equipment outrunning its own forecasts (T-03-06)
4. Power and thermal as the primary design constraint (T-03-07)
5. Scale-up fabric and optical interconnect as a separate profit pool (T-03-15)

Five trends most likely to affect businesses

1. Memory supercycle (T-03-01) — overlaps with lists 1 and 2; it is now a line item in every hardware budget
2. Long-term agreements re-engineering the memory cycle (T-03-09) — determines who can buy at all
3. Advanced packaging constraint (T-03-02) — overlaps with list 1; determines accelerator delivery dates
4. Power and thermal limits (T-03-07) — overlaps with list 2; determines deployment dates
5. Export control and tariff volatility (T-03-05) — determines supply-chain design

Five trends most likely to affect consumers

1. Memory cost pass-through: the disappearing entry-tier device (T-03-08)
2. The AI-PC and on-device-AI upgrade supercycle stalling (T-03-18) — closely coupled to T-03-08
3. Memory supercycle (T-03-01) — overlaps with lists 1, 2 and 3; it is the cause of both of the above
4. Power and thermal limits (T-03-07) — reaches consumers as local electricity and water politics
5. Quantum advantage claims (T-03-19) — reaches consumers indirectly via cryptographic migration and security marketing

Overlaps are substantial and stated deliberately: T-03-01 appears in four of the four lists because a single input-cost shock is currently propagating through the entire sector.

10. Overhyped / overlooked / cooling / reversing

Most overhyped

1. "Quantum advantage" and logical-qubit counts as proxies for usefulness (T-03-19). The evidence that hype outruns substance is specific and quantitative: the best publicly reported logical-qubit results in 2026 range from roughly 12 (Infleqtion) to 96 (academic neutral-atom architectures), while IBM's own roadmap puts the first large-scale fault-tolerant machine at 200 logical qubits and 100 million gates in 2029. That is a two-orders-of-magnitude gap in gate count between what exists and what is needed, and "logical qubit" is not a standardised term - it is used for memory-only encoded qubits, for non-universal gate sets, and for full fault-tolerant computation, which differ enormously in difficulty. Meanwhile over $4bn of disclosed capital has flowed in during 2025-26 against essentially no commercial compute revenue, much of it through SPAC vehicles with weak disclosure.

2. Intel Foundry's 14A turnaround treated as settled (T-03-20). 18A reached high-volume manufacturing, which is real. But yields remain below profitable levels and are not expected to reach target cost thresholds until end-2026 at the earliest, 14A rests on two unnamed prospective customers whose decisions run from H2 2026 into H1 2027, and Intel's own CEO has said 14A may be slowed or cancelled without external foundry revenue. Two prospects are not a customer base, and the twenty-year base rate for IDMs defending the leading edge is poor.

3. (Honourable mention) Near-term High NA EUV. Covered under cooling below, but the volume of coverage relative to a 2030 production date qualifies it as hype by any reasonable test.

Most overlooked

1. The re-engineering of the memory cycle through long-term agreements (T-03-09). If a majority of memory output is contracted years forward, the mechanism that has produced four boom-bust cycles since 2007 may not operate. This changes the volatility profile of a $100bn+ market. It is almost entirely uncovered because it is invisible - contract terms, volumes and prices are undisclosed, and there is no announcement to write about. Attention has missed it because trend detection is biased toward events, and this is the absence of an event.

2. Taiwan's mandatory monthly revenue disclosure regime. Every TWSE-listed company - TSMC, UMC, ASE, MediaTek, Quanta, KYEC - must publish monthly revenue. This is the highest-frequency hard data in the global semiconductor industry and it is systematically underused outside Taiwan because the MOPS interface is Chinese-language. Anyone building a monitoring pipeline should start here.

3. Scale-up fabric and optical interconnect as a distinct profit pool (T-03-15). Bandwidth per accelerator scales super-linearly as coherent compute domains grow beyond the rack. Broadcom already reports networking inside its AI line; Intel's Jaguar Shores is specified as a rack-level design with silicon photonics. This is the clearest attention-versus-substance gap in the sector - the accelerator gets the coverage, the fabric gets the attach-rate growth.

4. The Japanese materials layer. Photoresist, mask blanks, CMP slurry, wafer substrates - a set of near-monopolies concentrated in a handful of Japanese firms, essentially uncovered in English. A supply disruption here would be as consequential as a foundry outage and would be visible to almost nobody in advance.

Trends that appear to be cooling

1. Near-term High NA EUV adoption (T-03-16). The indicator that turned: TSMC stated on 2026-09-08 that it will deploy High NA in high-volume manufacturing only from 2030, with the 12-inch mask pilot line in 2031 and system readiness in 2033. Corroborating indicator: ASML's own 2027 capacity expansion is low-NA EUV (+30% from ~65 units) and DUV immersion (+30% from ~130 units) — not High NA.

2. China as a merchant-GPU revenue line (T-03-17). The indicator that turned: NVIDIA's guidance language. A company does not remove a revenue line from its outlook while it still believes in it.

3. The AI-PC upgrade supercycle (T-03-18). The indicator that turned: Samsung's own Q2 2026 commentary acknowledging "partial demand moderation" in mobile and PC — from the company selling the memory, in the same document as record memory profits.

Trends that may reverse, and the mechanism

• The memory supercycle (T-03-01) reverses if bit demand growth falls below supply additions. Mechanism: capacity committed in 2026-27 lands in 2028-29; if hyperscaler capex digests at the same time, pricing gaps down hard. Watch Micron's quarterly bit-growth guidance and Korean monthly export data — both turn before contract prices do.
• The China closure (T-03-17) reverses in either direction on a political decision. Mechanism: a US-China settlement reopens licences, or Beijing concludes domestic supply is inadequate and lifts its purchase restriction. This is the least persistent trend in the sector.
• Custom-ASIC share gains (T-03-03) reverse if NVIDIA's rack-scale systems integration and CUDA moat make the total cost of a custom programme uneconomic between generations. Mechanism: a hyperscaler quietly cancelling a next-generation XPU would be the first visible sign.
• Intel Foundry scepticism (T-03-20) reverses on a single named 14A customer commitment.

11. Risks and major uncertainties

Sector-specific risks.

1. Geographic concentration. Leading-edge logic and advanced packaging are concentrated in Taiwan to a degree that no amount of Arizona capex materially changes in this decade. A Taiwan Strait disruption is the sector's dominant tail risk and it is not diversifiable.
2. Capex at the cycle peak. TSMC at $60-64bn, Micron at ~$27bn, SK hynix raising $26.5bn — all committed while margins are at record highs and the Fed is leaning hawkish. The macro brief's warning that "any trend thesis that assumes cheap capital must be stress-tested" applies here more than anywhere.
3. Single points of failure. ASML in EUV; a handful of Japanese firms in materials; TSMC in packaging. Each is a genuine monopoly or near-monopoly.
4. Policy volatility in both directions. Section 232 Phase 2 on equipment remains open; BIS licensing can tighten or loosen at short notice; China's informal purchase restrictions are undocumented and therefore unpredictable.
5. Demand concentration. A handful of hyperscalers and two frontier labs drive the marginal order. The macro brief notes OpenAI and Anthropic alone took 43% of H1 2026 global VC; that concentration is mirrored in this sector's order book.
6. Power as an external gate. Interconnection queues and cooling supply chains can defer revenue recognition for chip vendors regardless of silicon availability.

Genuine unknowns — and the distinction matters.

Things we don't know but could find out: CXMT's actual HBM yield and output (unlisted, no disclosure — obtainable via equipment shipment tracing and teardown); TSMC's real CoWoS capacity (not published; every circulating figure is a channel estimate); BIS licence approval and denial counts (not published, but a matter of record inside government); memory long-term agreement terms (commercially confidential but known to counterparties); Intel's 14A customer pipeline (known to Intel).

Things nobody can know: whether 2026-27 memory capacity commitments will meet 2028-29 demand — this depends on AI model economics that do not yet exist; whether quantum error correction scales as qLDPC theory predicts when c-coupler yield meets manufacturing reality; whether the Taiwan Strait status quo holds; whether the current AI capex wave is investment or a bubble, which the FOMC itself flagged as a financial-stability concern without resolving.

12. Scenarios to 2030

Base case — "the bottleneck keeps moving." Memory pricing peaks in 2027 and normalises rather than collapsing, cushioned by long-term agreements. TSMC holds leading-edge share; N2 and A16 ramp on schedule; CoWoS roughly doubles by 2028 and packaging ceases to be the binding constraint, which passes the constraint to power. Custom ASICs reach perhaps a quarter to a third of accelerator value without displacing NVIDIA. Intel 14A secures one external customer and survives at subscale. Quantum remains pre-commercial. Falsifiable early indicator: Micron's quarterly DRAM bit-growth guidance staying in the low-to-mid 20s percent through 2027 while contract prices flatten rather than fall.

Upside — "structural break in memory." Long-term agreements genuinely change the cycle; AI-driven bit demand grows faster than capacity through 2028; memory makers sustain 50%+ operating margins for multiple years and fund capacity without a bust. Equipment billings hold above $170bn. Falsifiable early indicator: a full year in which DRAM contract prices do not decline quarter-on-quarter — something that has not happened in this industry since the 1990s.

Downside — "the fifth memory cycle." 2026-27 capacity commitments land into a hyperscaler capex digestion in 2028. DRAM prices fall by more than half; Korean operating margins go from 76% to single digits; the firms that raised US equity at the peak carry the impairment. Consumer devices get cheaper, which is the one genuinely good outcome in this branch. Falsifiable early indicator: Korean monthly semiconductor export values turning negative year-on-year for two consecutive months — historically the earliest reliable turn signal in this sector.

Disruption — "the package eats the chip." Panel-level CoPoS and chiplet standardisation arrive faster than expected, commoditising interposers and letting second-tier designers assemble competitive parts from purchased chiplets. Leading-edge wafer share becomes less decisive than packaging and integration skill. Optical I/O moves on-package and the fabric becomes the differentiator. Falsifiable early indicator: CoPoS reaching mass production on schedule in late 2027 with a non-TSMC supplier qualified.

Regulatory — "Phase 2 lands." Section 232 expands to semiconductor manufacturing equipment and a broader set of chips, which the January 2026 proclamation explicitly reserved. Fab construction costs in the US rise materially; equipment makers face a tariff on their own inputs; retaliation touches materials, where Japan and China both have leverage. Falsifiable early indicator: new HTS subheadings under 9903.79 covering equipment, or a CBP CSMS guidance bulletin extending coverage.

Failure — "the AI capex thesis breaks." The FOMC's flagged financial-stability risk from "high AI-firm equity valuations and increased leveraged financing of infrastructure buildout" materialises. Orders are cancelled from the equipment layer inward, in the reverse order they were placed: tools first, then packaging, then wafers, then memory. Given that equipment leads chip revenue by four to six quarters, this sector would see it before the AI sector does. Falsifiable early indicator: SEMI quarterly equipment billings posting a sequential decline of more than 10% while TSMC's monthly revenue is still growing — the divergence is the tell.

13. Data gaps and limitations

Could not verify.

1. TSMC's 2026 capex is not on the primary document. TSMC's IR quarterly results page publishes revenue and margin guidance but not capex; the $60-64bn figure (raised from $52-56bn) comes from earnings-call reporting corroborated across MacroMicro and Bits&Chips. Well corroborated, but the primary artefact is the call transcript, not a filed document. The same applies to the $100bn Arizona increment and the $265bn site total.
2. CoWoS capacity is unpublished. TSMC does not disclose it. Circulating end-2026 figures range from 120-130k to 140k wafers/month and are all channel estimates. Recorded as a contradiction in T-03-02.
3. SEMI's regional equipment breakdown is subscription-only. The Q2 2026 press release gives the $40.53bn global total but the per-region table — which is where China's equipment trajectory is visible — requires the Equipment Market Data Subscription. This is where Phase 2 licensing budget must go.
4. WSTS detailed billings are member-restricted. Only headline forecasts are public, which is precisely why content-farm "market size" pages fill the search results.
5. BIS publishes no licence approval/denial statistics. The practical effect of the January 2026 policy change is therefore only observable through company disclosure — NVIDIA's inventory write-down and guidance language. There is no official dataset.
6. China figures are almost entirely third-party estimates. CXMT is unlisted and discloses essentially nothing; SMIC's advanced-node capacity, Huawei's Ascend volumes, and China's HBM stockpile all trace to one or two analyst houses. T-03-10 is marked single_source with confidence: low for this reason, and its evidence_quality is capped at 2.
7. The best available China analysis is a year old. The SemiAnalysis Ascend production piece used here is dated 2025-09-08. A 2026 update was not locatable within this research budget.
8. Intel's current position is assessed from March 2026 reporting. H2 2026 14A customer decisions may already have changed it; Intel's own press page was unreachable through the research proxy on 2026-09-15 (robots/rate-limit).
9. Korean company leadership is left unverified in entity records rather than guessed. The contract's instruction not to invent people was followed over the temptation to fill fields.
10. Micron's stated price increases are not unambiguously period-labelled in the extract obtained ("low-60s percent" DRAM, "mid-80s percent" NAND). Read as sequential quarterly contract price movement; the prepared-remarks PDF should be re-read in full in Phase 2.

Paywalled or access-restricted. Nikkei Asia (metered), DigiTimes (subscription — and the likely origin of much uncredited Taiwan supply-chain reporting), SemiAnalysis (subscriber-only), SEMI WWSEMS regional data, WSTS detail, Nature and Science quantum papers.

Missing for non-US markets. Japanese materials and chemicals coverage in English is thin outside Nikkei — a serious gap given that this layer contains several near-monopolies. Chinese primary disclosure is minimal by design. Taiwan's MOPS monthly filings are excellent but Chinese-language.

Where numbers conflict. Four contradictions are recorded in the trend records: CoWoS end-2026 capacity (120-130k vs 140k wafers/month); SEMI's own 2026 equipment forecast ($145bn) versus its own quarterly actuals (~$155bn+ run rate); Section 232 effective date (14 vs 15 January 2026, proclamation versus entry into force); and the definition of "logical qubit," which is the central definitional problem in quantum and makes vendor counts non-comparable.

Methodological caution applied. Search results for this sector are dominated by "AI chip wars 2026" and "$XXX billion by 2032 at NN.N% CAGR" content-farm pages. None is load-bearing in this dossier. Every quantitative claim above traces to a company IR page, an SEC filing, a regulatory document, a standards/trade body, or a named-byline trade publication, and estimates are labelled as estimates.

14. Ranking scorecard

Mean: 4.82. Scored comparatively: a sector that in 2026 is simultaneously the largest capital sink, the tightest supply constraint and the primary instrument of great-power technology policy. The single point deducted on data availability is deliberate and specific.

15. Sources

1. TSMC Second Quarter 2026 Quarterly Results — TSMC Investor Relations — https://investor.tsmc.com/english/quarterly-results/2026/q2 — 2026-07-16 — A
2. TSMC Latest News (press release index; Q2 2026 EPS NT$27.25, Aug 2026 revenue, Sony JV) — TSMC — https://pr.tsmc.com/english/latest-news — 2026-09-10 — A
3. TSMC and ASML Announce Initiative to Pioneer Industry Transition to Large-Format Photomasks for High NA EUV — ASML / TSMC — https://www.asml.com/en/news/press-releases/2026/tsmc-and-asml-announce-industry-transition-to-large-format-photomasks-for-high-na-euv — 2026-09-08 — A
4. ASML reports EUR 9.3 billion total net sales and EUR 2.9 billion net income in Q2 2026 — ASML — https://www.asml.com/en/news/press-releases/2026/q2-2026-financial-results — 2026-07-15 — A
5. SK hynix Announces 2Q26 Financial Results — SK hynix Newsroom (South Korea) — https://news.skhynix.com/en/q2-2026-business-results/ — 2026-07-29 — A
6. Samsung Electronics Announces Second Quarter 2026 Results — Samsung Global Newsroom (South Korea) — https://news.samsung.com/global/samsung-electronics-announces-second-quarter-2026-results — 2026-07-30 — A
7. Samsung Electronics Announces First Quarter 2026 Results — Samsung Global Newsroom (South Korea) — https://news.samsung.com/global/samsung-electronics-announces-first-quarter-2026-results — 2026-04-30 — A
8. Micron Technology Fiscal Q3 2026 Earnings Call Prepared Remarks — Micron Investor Relations — https://s25.q4cdn.com/621799436/files/doc_financials/2026/q3/Q3-FY26-Prepared-Remarks.pdf — 2026-06-24 — A
9. NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 (EX-99.1) — US SEC EDGAR / NVIDIA — https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm — 2026-08-26 — A
10. NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 — NVIDIA Newsroom — https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027 — 2026-08-26 — A
11. Broadcom Inc. Announces Third Quarter Fiscal Year 2026 Financial Results — Broadcom Investor Relations — https://investors.broadcom.com/news-releases/news-release-details/broadcom-inc-announces-third-quarter-fiscal-year-2026-financial — 2026-09-02 — A
12. Global Semiconductor Equipment Billings Increased 23% Year-Over-Year in Q2 2026 — SEMI — https://www.semi.org/en/semi-press-release/global-semiconductor-equipment-billings-increased-23-percent-year-over-year-in-q2-2026-semi-reports — 2026-09-03 — A
13. Global Semiconductor Equipment Sales Projected to Reach a Record of $156 Billion in 2027 — SEMI (Year-End Forecast, SEMICON Japan) — https://www.semi.org/en/semi-press-release/global-semiconductor-equipment-sales-projected-to-reach-a-record-of-156-billion-dollars-in-2027-semi-reports — 2025-12-16 — A
14. IBM lays out clear path to fault-tolerant quantum computing — IBM Quantum Blog — https://www.ibm.com/quantum/blog/large-scale-ftqc — 2025-06-03 — A
15. Rapidus Corporation — 2nm semiconductor programme — Rapidus (Japan) — https://www.rapidus.inc/en/ — 2026-09-15 — A
16. President Trump orders narrowly targeted 25% Section 232 tariff on certain advanced semiconductor articles — White & Case LLP — https://www.whitecase.com/insight-alert/president-trump-orders-narrowly-targeted-25-section-232-tariff-certain-advanced — 2026-01-21 — B
17. Initial Section 232 Actions on Semiconductors, Critical Minerals Announced — Cassidy Levy Kent — https://www.cassidylevy.com/news/initial-section-232-actions-on-semiconductors-critical-minerals-announced/ — 2026-01-15 — B
18. BIS Revises License Review Policy for Advanced Computing Commodities (AI Semiconductors) to China and Macau — Baker McKenzie Global Sanctions and Export Controls Blog — https://sanctionsnews.bakermckenzie.com/bis-revises-license-review-policy-for-advanced-computing-commodities-ai-semiconductors-to-china-and-macau-when-exported-from-the-united-states/ — 2026-01-28 — B
19. HBM4 Prices to Double Next Year as Samsung, SK hynix Keep Upper Hand — Seoul Economic Daily (South Korea) — https://en.sedaily.com/finance/2026/07/12/hbm4-prices-to-double-next-year-as-samsung-sk-hynix-keep — 2026-07-12 — B
20. SK hynix shatters Q2 earnings record amid AI peak concerns — The Korea Times (South Korea) — https://www.koreatimes.co.kr/amp/business/companies/20260729/sk-hynix-shatters-q2-earnings-record-amid-ai-peak-concerns — 2026-07-29 — B
21. Global Foundry Revenue Approaches US$53.49 Billion in 2Q26 as SMIC Narrows Market Share Gap with Samsung — TrendForce (Taiwan) — https://www.trendforce.com/presscenter/news — 2026-09-09 — B
22. TrendForce news wire (TSMC 2nm/3nm capacity boost and CoWoS doubling; Kioxia ADR; China high-purity quartz qualification) — TrendForce (Taiwan) — https://www.trendforce.com/news/ — 2026-09-15 — B
23. N2 Arrives, Capex Climbs: TSMC Doubles Down on AI Investment — MacroMicro (Taiwan) — https://en.macromicro.me/blog/n2-arrives-capex-climbs-tsmc-doubles-down-on-ai-investment — 2026-07-17 — B
24. TSMC raises 2026 capex to as much as $64B — Bits&Chips — https://bits-chips.com/article/tsmc-raises-2026-capex-to-as-much-as-64b/ — 2026-07-16 — B
25. ASML and TSMC Announce Initiative to Pioneer Industry Transition to Large-Format Photomasks for High NA EUV — Taiwan News (Taiwan) — https://www.taiwannews.com.tw/news/6435631 — 2026-09-08 — B
26. Intel's roadmaps examined — 14A, Nova Lake, Diamond Rapids & AI accelerator push (Luke James) — Tom's Hardware — https://www.tomshardware.com/tech-industry/semiconductors/intel-chip-roadmap-2026-2028 — 2026-03-16 — B
27. Huawei's Ascend AI chip ecosystem scales up as China pushes for semiconductor independence — Tom's Hardware — https://www.tomshardware.com/tech-industry/semiconductors/huaweis-ascend-ai-chip-ecosystem-scales — 2025-11-21 — B
28. Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck — SemiAnalysis — https://newsletter.semianalysis.com/p/huawei-ascend-production-ramp — 2025-09-08 — B
29. The Companies Building Quantum Computing Chips in 2026 — The Quantum Insider — https://thequantuminsider.com/2026/06/05/how-many-quantum-chip-companies-are-there/ — 2026-06-05 — B
30. Inside Anthropic's Multi-Cloud AI Factory: How AWS Trainium and Google TPUs Shape Its Next Phase (David Chernicoff) — Data Center Frontier — https://www.datacenterfrontier.com/machine-learning/article/55335703/inside-anthropics-multi-cloud-ai-factory-how-aws-trainium-and-google-tpus-shape-its-next-phase — 2025-12-03 — B
31. Installing Cooling that Scales / Rack Densities Are on the Rise / ERCOT Puts Texas AI Megawatts to the Test — Data Center Frontier — https://www.datacenterfrontier.com/ — 2026-09-09 — B
32. DRAM prices are killing the cheap smartphone — The Register — https://www.theregister.com/personal-tech/2026/07/07/dram-prices-are-killing-the-cheap-smartphone/5267643 — 2026-07-07 — B
33. Apple passes RAMpocalypse costs on to consumers — The Register — https://www.theregister.com/personal-tech/2026/06/25/apple-passes-rampocalypse-costs-on-to-consumers/5262266 — 2026-06-25 — B
34. Expect more of those DRAM price hikes as memory shortage continues to bite — The Register — https://www.theregister.com/storage/2026/06/02/expect-more-of-those-dram-price-hikes-as-memory-shortage-continues-to-bite/5250049 — 2026-06-02 — B
35. SK Hynix to boost memory production 3x — by 2034 — The Register — https://www.theregister.com/systems/2026/06/12/sk-hynix-to-triple-wafer-capacity-by-2034/5255151 — 2026-06-12 — B
36. US Export Controls on Nvidia AI Chips to China: A100 to H20 to H200 — Timeline — whatledto.com — https://whatledto.com/events/nvidia-china-export-controls — 2026-09-03 — C
37. TSMC's CoWoS Capacity: Scaling Up, Outsourcing, and Shifting to Next-Gen Packaging — Global Semi Research (Substack; sourced to an investor-shared expert transcript) — https://globalsemiresearch.substack.com/p/tsmcs-cowos-capacity-scaling-up-outsourcing — 2025-12-28 — C
38. TSMC Reportedly Plans to Double CoWoS Capacity by 2028 as AI Packaging Demand Keeps Rising — Digital Citizen — https://www.digitalcitizen.life/tsmc-reportedly-plans-to-double-cowos-capacity-by-2028-as-ai-packaging-demand-keeps-rising/ — 2026-09-12 — C

Tier C sources (36-38) are used only for corroboration or for explicitly-labelled estimates and are never the sole support for a claim.

More research in this sector · Explore trend records