Watch Tech Series · Semiconductor Watch | September 29, 2026
The CODEW Semiconductor Watch tracks the companies, technologies, investments, and supply-chain shifts reshaping AI chips, advanced manufacturing, memory, packaging, and computing.
The semiconductor industry's binding constraint has shifted from chip design to everything around the chip. TSMC says AI accelerator wafer demand will grow 11-fold from 2022 to 2026, but the more consequential bottleneck this year isn't wafer starts — it's advanced packaging (CoWoS capacity is racing to double from roughly 130,000 to 260,000 wafers per month by 2028) and HBM (high-bandwidth memory), where a genuine supply crunch has reportedly pushed Google and Microsoft to escalate their own procurement efforts directly.
That shift — from "can we design a fast enough chip" to "can we package, memory-attach, and power it fast enough" — is the throughline connecting every development in this edition, from TSMC's capex increase to ASML's back-to-back guidance raises to the scramble among Samsung, SK Hynix, and Micron for next-generation HBM4 supply.
Semiconductor Developments to Watch
1. TSMC raises 2026 capex to $60-64 billion, races to double CoWoS capacity
What happened: TSMC raised its 2026 capital spending guidance from $52-56 billion to $60-64 billion, with 70-80% allocated to advanced process technologies. The company plans to build nine phases of wafer fabs and advanced packaging facilities in 2026 alone, following a record Q2 with net profit up 77% year-over-year to roughly $22 billion.
Financial/capacity implications: CoWoS advanced packaging capacity — the technology used to package NVIDIA's and other AI accelerators — is targeted to roughly double from about 130,000 wafers per month at end-2026 to 260,000 by end-2028, with a compound annual growth rate above 80% from 2022 to 2027.
Why it matters: Advanced packaging, not raw wafer capacity, has become the primary constraint on how many AI accelerators the industry can actually ship — CoWoS now contributes about 10% of TSMC's revenue and is expected to keep rising as a profit driver.
Competitive implications: TSMC's dominance in advanced packaging is significant enough that Intel's EMIB-T packaging technology is drawing renewed interest from Google and Amazon as a potential alternative path for their custom TPU and Trainium silicon.
2. HBM4 supply race intensifies as memory shortage bites
What happened: SK Hynix, Samsung, and Micron are racing to ramp sixth-generation high-bandwidth memory (HBM4) for NVIDIA and other AI accelerator customers. SK Hynix led global HBM revenue with 56.4% share in Q1 2026 (IDC), while Samsung and Micron have both shipped HBM4 samples and are contesting SK Hynix's lead for 2026-2027 supply share.
Financial/capacity implications: Samsung and SK Hynix reportedly raised HBM3E prices nearly 20% for 2026 deliveries as demand continues to outstrip supply, and reports indicate Google and Microsoft have both escalated internal procurement efforts to secure adequate memory allocation.
Why it matters: HBM4 chips now include a customer-specific logic "base die," meaning memory suppliers can no longer be swapped as easily as in prior generations — a structural change that deepens each hyperscaler's dependency on its chosen memory partner.
Competitive implications: A three-way HBM4 contest, rather than the two-supplier structure many expected, gives NVIDIA and other accelerator buyers more negotiating leverage even as overall supply remains tight.
3. NVIDIA's $5 billion Intel stake tests Intel Foundry's 18A ambitions
What happened: NVIDIA completed a $5 billion investment in Intel in January 2026, purchasing roughly 214.7 million shares at $23.28 each for a stake of approximately 4%, tied to securing 18A manufacturing capacity at Intel's Fab 52 in Arizona. NVIDIA has reportedly tested the 18A process but has not placed production orders.
Financial/capacity implications: Intel's foundry business posted a $2.5 billion operating loss on $4.5 billion of revenue in Q4 2025, with the company acknowledging it has not yet secured meaningful external foundry customers at any node.
Why it matters: The investment is a strategic hedge for NVIDIA — geographic and supplier diversification away from total dependence on TSMC — as much as it is a financial bet, and a genuine vote of confidence Intel badly needs to attract other customers.
Competitive implications: Microsoft and Amazon are reported to already be Intel 18A customers; whether NVIDIA and Broadcom (also reported to be testing 18A) convert from evaluation to production orders in 2026-2027 will determine whether Intel Foundry becomes a credible third option alongside TSMC and Samsung.
4. ASML raises 2026 guidance twice as AI demand outpaces equipment supply
What happened: ASML raised its 2026 revenue forecast twice this year — first from €34-39 billion to €36-40 billion, then to €43-45 billion after a Q2 beat — citing accelerating customer capacity expansion plans. Q2 2026 net sales reached €9.33 billion with a 54% gross margin, both ahead of consensus.
Financial/capacity implications: ASML plans to increase its EUV lithography shipment capacity by roughly 30% for 2027 (with another 30% increase under evaluation for 2028) and is expanding deep ultraviolet immersion capacity by a similar margin — expansion plans CEO Christophe Fouquet said exceeded market expectations.
Why it matters: As the sole global supplier of EUV lithography tools — machines that can cost over $300 million each — ASML's capacity decisions function as a hard ceiling on how fast every advanced chipmaker, including TSMC, Samsung, and Intel, can expand leading-edge output.
Competitive implications: ASML's upgraded outlook is itself a signal that its major customers — TSMC, Samsung, Intel, SK Hynix — are all accelerating capacity plans simultaneously, reinforcing that the current investment cycle is industry-wide rather than concentrated in one company's roadmap.
5. Qualcomm returns to the data center with a multi-generation AWS silicon deal
What happened: Qualcomm announced on September 8, 2026, a multi-generational agreement with Amazon Web Services to supply customized AI data center silicon, alongside optical connectivity technology extending up to 1.6T. The deal follows Qualcomm's June 2026 Investor Day unveiling of its Dragonfly C1000 CPU, Dragonfly AI300 inference accelerator, and a broader custom-ASIC capability built partly through its Alphawave acquisition.
Financial/capacity implications: Qualcomm's AI200, AI250, and AI300 accelerators form a now-annual product cadence; the Dragonfly C1000 CPU, a chiplet design with more than 250 custom Oryon cores, is slated for commercial availability in 2028.
Why it matters: Qualcomm exited the data center CPU market in 2018 after its Centriq Arm server chip failed to gain traction; the Arm-server ecosystem hyperscalers now rely on for custom silicon has matured enough that Qualcomm's second attempt starts from a fundamentally different competitive position.
Competitive implications: Qualcomm becomes a new custom-silicon option for hyperscalers alongside Broadcom and Marvell, intensifying competition in the ASIC design market that NVIDIA's largest customers use to reduce dependence on merchant GPUs.
6. Advanced packaging capacity race spreads beyond TSMC
What happened: Amkor raised planned investment for its second-phase Arizona advanced packaging expansion from $7 billion to $12 billion on stronger customer demand, and signed a 10-year agreement with TSMC to provide packaging and testing services there. ASE's subsidiary SPIL is reportedly finalizing its own Arizona facility, and separately raised its 2026 capex by $2 billion to $10.5 billion.
Financial/capacity implications: Institutional investors estimate non-TSMC advanced-packaging players (ASE, Amkor) will expand capacity meaningfully as customers actively qualify alternative suppliers to reduce reliance on TSMC's constrained CoWoS lines.
Why it matters: This is the clearest evidence yet that the advanced-packaging bottleneck is severe enough to pull outside suppliers, and billions of dollars of new U.S. investment, into a category TSMC has effectively owned for years.
Competitive implications: A credible multi-supplier advanced packaging market would reduce AI accelerator makers' single-point dependency on TSMC — a meaningful structural shift if it materializes at scale by the 2028 timelines these companies have set.
7. Marvell and Broadcom deepen custom silicon and optical interconnect bets
What happened: Marvell Technology acquired Celestial AI (photonic fabric technology for scale-up optical interconnects) and completed its acquisition of XConn Technologies (PCIe/CXL switching silicon), while Broadcom continues expanding its custom XPU design program for hyperscalers alongside its Tomahawk 6 merchant switch silicon.
Financial/capacity implications: Both companies are positioning custom silicon design services — not just merchant chips — as a primary growth vector, competing directly for the hyperscaler dollars that might otherwise go to NVIDIA GPUs.
Why it matters: Custom ASIC design has become one of the most strategically contested categories in semiconductors precisely because it's how NVIDIA's largest customers reduce dependence on NVIDIA — Broadcom and Marvell are the primary beneficiaries of that shift regardless of which specific hyperscaler silicon program wins.
Competitive implications: Qualcomm's entry (development 5, above) makes this a genuinely three-way contest for custom silicon design wins, rather than the Broadcom-Marvell duopoly it has largely been to date.
Table 1 · Developments to Watch, Summary
| Company/Tech | Category | Key Figure |
|---|---|---|
| TSMC | Foundry, advanced packaging | $60-64B 2026 capex |
| SK Hynix / Samsung / Micron | HBM4 memory | SK Hynix 56.4% HBM share (Q1 26) |
| NVIDIA – Intel | Foundry investment | $5B stake, ~4% ownership |
| ASML | Lithography equipment | €43-45B 2026 revenue guide |
| Qualcomm – AWS | Custom AI silicon | Multi-generation deal, Sept 8, 2026 |
| Amkor / ASE-SPIL | Advanced packaging | Amkor Arizona raised to $12B |
| Marvell / Broadcom | Custom silicon, optical interconnect | Celestial AI, XConn acquisitions |
The AI Chip Race
AI accelerators and GPUs remain led by NVIDIA, with AMD's Instinct line narrowing the specification gap and hyperscaler custom silicon (Google TPU, Amazon Trainium) expanding as an internal alternative. Custom silicon is now a genuinely three-way contest among Broadcom, Marvell, and Qualcomm's newly announced entry, each competing to design workload-specific chips for hyperscalers that want to reduce GPU dependence without sacrificing performance.
Inference chips are becoming a distinct product category rather than a smaller version of training silicon — Qualcomm's AI200/AI250/AI300 line and its emphasis on "tokens per dollar and tokens per watt" reflect an industry-wide recognition that inference economics, not training performance, will determine long-term AI infrastructure costs as usage scales into the billions of daily queries.
Memory has become the single tightest constraint in the entire AI chip race — the HBM4 base-die customization now locking suppliers and customers together more tightly than in any prior memory generation. Networking silicon (covered in depth in CODEW's Networking Watch) and advanced packaging round out the constraint picture: even a fully available GPU or custom accelerator is only as useful as the memory, packaging, and interconnect capacity available to pair with it.
The CODEW Lens: The AI chip race has quietly become an AI supply chain race — the accelerator itself is no longer the scarcest input; memory, packaging, and power increasingly are.
Manufacturing & Supply Chain
Fabs and process technology: TSMC continues leading at the process edge, with 3nm capacity around 150,000 wafers per month and 2nm ramping to 50,000-60,000 per month in the first half of 2026, alongside 2nm/A16 capacity guided to grow at a 70% CAGR from 2026 to 2028. Intel's 18A is production-ready but still searching for meaningful external customers beyond Microsoft and Amazon's reported engagements.
Equipment: ASML's twice-raised 2026 guidance is the clearest signal that leading-edge capacity expansion is accelerating industry-wide rather than concentrated at any single chipmaker, with EUV shipment capacity guided up roughly 30% for both 2027 and potentially 2028.
Geographic diversification: TSMC's Arizona campus is scaling toward a fourth fab and its first U.S. advanced packaging facility; Amkor and ASE/SPIL are both making multi-billion-dollar Arizona commitments of their own. The pattern across foundry, memory, and packaging is the same: U.S.-based capacity is expanding meaningfully for the first time in decades, driven by both customer demand and policy incentive, even as Taiwan and South Korea remain the technological center of gravity.
Critical supply-chain developments: The HBM shortage and the advanced-packaging bottleneck are the two most consequential near-term constraints on how many AI accelerators can actually reach customers, regardless of how much raw wafer capacity exists — a reminder that semiconductor supply chains have multiple potential choke points, not just the leading-edge fab everyone watches most closely.
The CODEW Lens: Supply-chain diversification is happening in real time, in dollars committed and fabs under construction — but the Taiwan-South Korea concentration at the leading edge won't meaningfully change before the end of this decade.
Competitive Landscape
NVIDIA is positioning across the stack — GPUs, networking, and now a strategic manufacturing hedge through its Intel stake — rather than defending GPU share alone. AMD continues narrowing the accelerator specification gap while building its own systems and networking capabilities. Intel is fighting simultaneously to modernize its own product roadmap and prove Intel Foundry can serve external customers at scale — the more urgent and less certain of the two fights.
Broadcom and Marvell are both betting that custom silicon design services, not merchant chips, capture the largest share of hyperscaler spending growth over the next several years, and Qualcomm has now entered that same contest from outside the traditional data-center silicon world. TSMC remains the indispensable manufacturing layer beneath nearly all of the above, giving it leverage that shows up in pricing power and allocation decisions as much as in its own product roadmap.
Samsung is pursuing an unusually broad strategy — foundry, memory, and HBM all at once — betting that vertical integration across those three businesses is a durable advantage even where it trails category leaders in each individually. SK Hynix has built the strongest single-category position in this entire landscape as HBM revenue leader, while Micron is fighting to prove it belongs in the HBM4 conversation at all after facing early doubts about meeting NVIDIA's performance bar. Arm's architecture underpins a growing share of custom data-center silicon (Qualcomm's Oryon cores included), positioning it as an infrastructure layer beneath the custom-silicon wave rather than a direct competitor within it. ASML, Applied Materials, and Lam Research sit underneath all of it as the equipment layer whose capacity decisions gate how fast every other company on this list can actually expand.
The CODEW Lens: The companies gaining the most ground right now aren't necessarily building the best chip — they're the ones controlling a layer (memory, packaging, or equipment) that every chip, regardless of who designed it, has to pass through.
What to Watch Next
Technology transitions: Watch whether SK Hynix, Samsung, or Micron pulls decisively ahead in HBM4 mass production quality and yield through 2027, and whether the 2nm-to-A16 TSMC roadmap holds its guided 70% CAGR through 2028.
Capacity expansions: TSMC's Arizona fourth fab and first U.S. advanced packaging facility, Amkor's and ASE/SPIL's Arizona builds, and ASML's evaluated further EUV capacity increase for 2028 are all multi-year capacity commitments worth tracking for on-schedule delivery.
Product launches: Qualcomm's Dragonfly C1000 CPU (commercial availability guided for 2028) and continued AI200/250/300 accelerator cadence are the clearest near-term test of whether Qualcomm's data-center re-entry converts into real design wins beyond the announced AWS relationship.
Partnerships and investment decisions: Whether NVIDIA or Broadcom convert their Intel 18A evaluations into production orders is the single clearest signal available on whether Intel Foundry's turnaround is real — an evaluation-to-order conversion would matter more than any other Intel news this year.
Potential strategic moves: Continued M&A in custom silicon IP and optical interconnect, following Marvell's Celestial AI and XConn acquisitions, is likely — watch Broadcom and Qualcomm for similar moves to fill roadmap gaps as the custom-silicon contest intensifies.
The CODEW Lens: What is changing in semiconductors, and how will it reshape the next generation of computing infrastructure? The industry is being re-architected around scarcity in memory and packaging rather than around chip design alone — which means the winners of this cycle may be decided as much by who locks up HBM and CoWoS capacity as by whose chip benchmarks best.
The CODEW Stat
11x AI wafer demand growth (2022-2026) · CoWoS capacity doubling by 2028 TSMC's own numbers capture the scale of this cycle: AI accelerator wafer demand growing elevenfold in four years, met by an advanced-packaging buildout racing to keep pace — with equipment supplier ASML, memory makers SK Hynix and Samsung, and packaging specialists Amkor and ASE all raising capacity and spending guidance in the same twelve-month window.
Reviewed by Erwin Castro
on
Tuesday, September 29, 2026
Rating:
No comments: