AI demand is tightening semiconductor capacity and broadening growth across equipment, packaging, testing, memory and design software

Overview

This conference-takeaway report synthesizes management commentary from TSMC, Amkor, NXP, Cadence, SanDisk, IBM, Teradyne, Applied Materials and SiTime. Its central conclusion is that AI is extending demand across the semiconductor value chain, while capacity constraints, increased design complexity and a gradually improving analog cycle support a constructive sector backdrop.

Core views

Goldman Sachs' first broad takeaway is that agentic AI is creating demand tailwinds from systems through components, services and equipment. Companies described using AI internally to accelerate product development, shorten innovation cycles and improve costs, while also commercializing AI-enabled offerings. Cadence expects chip complexity to increase 48x over the next five years, supporting agentic-AI EDA adoption and a consumption-plus-subscription model for AI agents. IBM sees lower-cost open-source tools, including smaller language models, RHEL AI and watsonx products, as a way to scale enterprise agentic AI and create consulting value. Applied Materials sees AI-enabled product development, yield improvement and service innovation as sources of return on investment for itself and customers. The second sector-wide conclusion is that structural AI demand is driving capacity additions but supply remains tight across leading-edge wafers, advanced packaging, test and memory. TSMC cited tight advanced-packaging and wafer capacity, Amkor cited strong advanced-packaging demand, and Teradyne pointed to growing test demand. SanDisk said NAND supply-demand conditions remain tight and could persist through 2027 and potentially 2028. The report frames these constraints as a consequence of longer customer visibility and the time required to build and ramp fabs, rather than simply a short-lived utilization cycle. TSMC's presentation is the report's most detailed discussion of the AI cycle. Management views AI as a multi-year megatrend still at an early stage: falling compute costs, open-source models and better efficiency should lower token costs, widen adoption and increase advanced-silicon demand. It believes the shift from generative to agentic AI broadens demand beyond accelerators to CPUs, networking and supporting silicon, meaning its disclosed AI revenue definition understates total AI-related exposure because it principally captures GPUs, custom AI accelerators and memory base die. TSMC validates demand with both top-down analysis of macroeconomic, semiconductor, fabless, system-company and foundry conditions and bottom-up feedback from customers and customers' customers. It acknowledges that semiconductors remain cyclical but considers identification of the underlying megatrend more important than forecasting every short-term fluctuation. TSMC also emphasized manufacturing execution and geographic flexibility. Arizona Phase 1, using N4, has yields comparable to Taiwan and is already profitable; Phase 2 equipment move-in for N3 is expected soon, Phase 3 construction has begun, and preparation for Phase 4 and a first US advanced-packaging facility is underway. TSMC expects about 30% of N2-and-beyond capacity to be outside Taiwan in five years, primarily in Arizona, and has acquired land that could accommodate another five to six fabs alongside its announced incremental US$100 billion investment. The company argues that technology leadership, high-volume manufacturing, capacity, customer trust and cross-node flexibility form a difficult-to-replicate moat. It can use parts of N5 capacity for N3 demand and optimize across other nodes where feasible, reducing capacity-obsolescence risk. For TSMC profitability, management reiterated a long-term gross-margin objective of 56% and above, compared with 50% five years ago and 53% previously. It expects the N2 ramp to dilute 2026 gross margin by approximately 2-3 percentage points and overseas expansion to dilute margins by 2-3 points early in the current five-year period, widening to 3 points later as more overseas fabs enter production. Management nevertheless expects disciplined capacity planning, utilization, cost reduction, productivity and value-based pricing to offset these effects. It describes advanced packaging primarily as an enabler of wafer demand rather than a standalone profit pool; CoWoS capacity remains tight, selected processes are being outsourced to OSAT partners, and its packaging roadmap targets roughly 14x reticle size in coming years from about 5.5x currently. Across packaging and equipment, Amkor expects compute revenue to grow 30% year on year in 2027 and reiterated an $11 billion 2030 revenue target. Arizona Phase 2 is expected to add 60,000 square meters of cleanroom space after the 33,000 square meters added in Phase 1; management's initial assessment implies revenue around $1 billion and a 30% gross margin at a scale similar to Phase 1. It expects 2028 and 2029 capital expenditure to be elevated, with 20%-25% of its $12 billion Arizona investment in each of those years. Applied Materials reiterated approximately 40% revenue growth in 2026 and expects strong growth to extend into 2027 and 2028. It expects leading-edge logic, advanced packaging and memory to comprise about 80% of 2026 WFE, while advanced-packaging revenue is expected to rise about 70% year on year to a $2.5-$3.0 billion run rate and Process Control to grow more than 50% in 2026. The report identifies a bullish analog backdrop, with lean inventories, content growth and steadily improving industrial and automotive demand supporting a broad recovery. NXP cited better revenue visibility from a growing backlog and longer lead times, although broad OEM restocking has not yet materialized. About 50% of its automotive business consisted of growth drivers in the second quarter, including software-defined vehicles, radar and connectivity, which are growing about 20%. NXP expects datacenter exposure to grow more than 100% in 2026 and is refocusing toward physical AI and edge intelligent systems. Its selective, staggered price increases are expected to accumulate in revenue over coming quarters, while hybrid manufacturing, joint ventures and migration from 200 mm to 300 mm wafers are intended to preserve capacity flexibility and cost competitiveness. Company-specific AI infrastructure opportunities extend into test, timing and memory. Teradyne expects test equipment to grow at least in line with semiconductor capital expenditure over the next three to five years; testing now represents about 8% of semiconductor capital expenditure versus roughly 4% historically. It expects AI-driven growth to accelerate through 2027, supported by merchant-compute share gains, hyperscale custom-silicon engagements and ASIC test momentum. Its co-packaged-optics test TAM is expected to rise from about $100 million today to $300-$700 million by 2028 and potentially about $1 billion by 2030. SiTime expects datacenter content growth from AI inference, share gains and 1.6 Tbps optical networking; it expects share at 1.6 Tbps to double to 80%, content per rack to reach $400-$600 and potentially $1,000 in some configurations. Its acquired Renesas IC-clock portfolio remains supply constrained, but management expects a potential revenue inflection in mid-2027. SanDisk described a structural shift in NAND from spot pricing toward longer-term agreements, supported by datacenter demand. Its NBMs cover 50% of planned volume in FY27 and 67% in FY28; floor pricing is intended to support approximately 80% gross margin for most of the business in a downside scenario. Management expects muted NAND supply growth, sees China competitor additions largely absorbed locally, and highlighted low capital intensity of about 5% through its Kioxia joint venture, IP and BiCS platform. It has executed roughly $4.5 billion of buybacks and continues to prioritize repurchases for excess capital. IBM, meanwhile, expects 4%-5% constant-currency revenue growth in 2026, approximately $1 billion of year-on-year free-cash-flow growth and 100 basis points of pre-tax margin expansion. It expects 6%-8% software growth in 2026, accelerating consulting growth in the second half, and sees quantum computing as a potential multi-billion-dollar opportunity before the end of the decade.

Analysis framework

The report first consolidates sector debates from conference presentations, then applies company-specific management commentary to AI demand, supply constraints, capacity investment, operating outlooks and end-market recovery. It uses supply-demand conditions, volume and content growth, manufacturing execution, margin drivers, management targets and normalized earnings multiple-based target prices for covered equities.

Methodology notes

• Industry AnalysisSupply-demand framework

  Supply-demand analysis across wafers, advanced packaging, memory and test equipment

  The report links AI-driven demand visibility and long fab construction and ramp periods to continued tightness in several semiconductor supply-chain segments.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure demand transmission through the semiconductor value chain

  The analysis traces AI demand from cloud and systems spending into foundry capacity, packaging, memory, testing, EDA software and semiconductor equipment.

• Valuation methodsP/E and PEG Valuation

  Normalized EPS multiplied by a target P/E multiple

  The valuation table derives several 12-month price targets by applying stated P/E multiples to normalized earnings estimates.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• TSMC (2330.TW; TSM)

  Leading-edge foundry and advanced-packaging beneficiary of broadening AI silicon demand.

  Strengths

  Technology roadmap, manufacturing scale, Arizona execution, capacity flexibility and customer trust.

  Weaknesses

  N2 ramp and overseas fabs initially dilute gross margin.

  Comparison

  Management argues foundry and packaging are distinct businesses and packaging wins should not meaningfully displace its leading-edge wafer position.

  Risks

  End-demand weakness, slower node migrations, lower AI investment, execution or yield issues, competition, FX and higher costs.

• Applied Materials (AMAT)

  Semiconductor-equipment beneficiary of leading-edge logic, memory and advanced-packaging investment.

  Strengths

  Exposure to AI and datacenter spending, process-control leadership and advanced-packaging growth.

  Comparison

  Process Control is expected to grow more than 50% in 2026 on demand for yield improvement and materials engineering at advanced nodes.

  Risks

  Incremental export restrictions and China-based supplier share gains.

• Amkor Technology (AMKR)

  Advanced-packaging and test provider benefiting from compute complexity and Arizona demand.

  Strengths

  Arizona campus expansion, 2.5D packaging, high-density fan-out and co-packaged-optics capacity.

  Weaknesses

  Near-term communications headwinds from memory demand destruction, a Korea-to-Vietnam capacity shift and customer build-cycle differences.

  Comparison

  Arizona proximity to TSMC is expected to improve feedback loops and cycle time.

  Risks

  2.5D roadmap execution, utilization leverage at new facilities and execution on smartphone design wins.

• NXP Semiconductors (NXPI)

  Analog and edge-intelligence supplier exposed to automotive, industrial and datacenter recovery.

  Strengths

  Automotive growth drivers, datacenter investment, physical-AI strategy and hybrid manufacturing flexibility.

  Weaknesses

  Broad OEM inventory restocking has not yet emerged.

  Comparison

  Management sees growing component-level competition in China while pursuing system-level solutions.

  Risks

  Renewed automotive end-demand weakness, uncertain recovery timing and unfavorable pricing.

• Cadence Design Systems

  EDA software provider benefiting from rising chip complexity and AI-agent adoption.

  Strengths

  Agent interoperability, broader AI-agent TAM, hardware-accelerated verification and physical-AI opportunity.

  Comparison

  Cadence sees its integrated top-layer agents and middle-layer tools as differentiated from customer-built agent layers.

  Risks

  Export restrictions, market-share losses and fewer custom chip designs.

• SanDisk (SNDK)

  NAND supplier benefiting from datacenter demand and a shift toward longer-term agreements.

  Strengths

  NBM coverage, low capital intensity, Kioxia joint venture and share-repurchase focus.

  Comparison

  Management expects competitor supply additions in China to be mainly absorbed by the local market.

  Risks

  Failure of structural NAND-pricing change, YMTC roadmap progress and failure to gain enterprise-SSD traction.

• IBM (IBM)

  Enterprise software, consulting and infrastructure participant in agentic AI adoption.

  Strengths

  Open-source AI tools, Red Hat momentum, consulting signings, recurring revenue and quantum program.

  Weaknesses

  Software shortfall affected about 15% of the portfolio after customer capital-spending reprioritization.

  Comparison

  Management believes its consulting exposure to generative-AI service areas is differentiated from peers.

  Risks

  Macro headwinds, consulting slowdown, AI-bookings deceleration and dilutive M&A.

• Teradyne (TER)

  Test-equipment and robotics provider exposed to AI infrastructure expansion.

  Strengths

  Exposure across semiconductor test, robotics and product test; hyperscale and custom-silicon opportunities.

  Weaknesses

  Large AI-infrastructure projects can create quarterly volatility.

  Comparison

  Management sees custom AI silicon as a larger growth opportunity than networking.

  Risks

  Delays at a merchant GPU customer and a slower-than-expected mobile recovery.

• SiTime (SITM)

  Timing-solutions provider benefiting from AI datacenter content growth and higher-speed optical networking.

  Strengths

  Datacenter share gains, 1.6 Tbps roadmap, acquired IC-clock portfolio and design wins across GPUs, CPUs, ASICs and switches.

  Weaknesses

  A seasonal mobile and consumer slowdown is expected in 1Q27.

  Comparison

  Management expects its 1.6 Tbps market share to double to 80%.

  Risks

  Slower AI datacenter spending and weaker design-win momentum at key datacenter or smartphone customers.

Key data

• TSMC long-term gross-margin target56% and aboveManagement expects N2 ramp dilution of about 2-3 percentage points in 2026 and overseas-expansion dilution of 2-3 points early in the five-year period.
• TSMC N2-and-beyond overseas capacity~30%Expected to be outside Taiwan in five years, primarily in Arizona.
• Amkor compute revenue growth30% YoY in 2027Supported by advanced-packaging demand.
• Applied Materials 2026 revenue growth outlook~40%Management expects strong growth to continue into 2027 and 2028.
• Applied Materials 2026 WFE mix~80%Leading-edge logic, advanced packaging and memory share of WFE.
• NXP datacenter growth expectation>100% in 2026Management identified datacenter as a key growth driver.
• SanDisk planned volume covered by NBMs50% in FY27 / 67% in FY28Longer-term agreements are intended to improve revenue visibility.
• Teradyne co-packaged-optics test TAM$300mn-$700mn by 2028; ~$1bn by 2030Compared with roughly $100mn today.
• IBM 2026 constant-currency revenue growth4%-5%Management also expects approximately $1bn of year-on-year free-cash-flow growth.

Impact & implications

The report argues that AI demand is becoming broader and more durable across the semiconductor ecosystem, benefiting capacity providers, packaging and test suppliers, memory producers and design-software vendors. It also indicates that execution, capacity availability, end-demand recovery, pricing discipline and customer adoption remain critical to translating the favorable backdrop into earnings and margins.

Risks

• A slowdown in AI investment or end demand could weaken semiconductor utilization, capacity spending and long-term content growth.
• TSMC faces node-migration, yield, execution, competition, FX and cost risks as it expands advanced-node and overseas capacity.
• NXP's recovery outlook remains exposed to automotive demand, uncertain restocking timing and pricing.
• SanDisk's thesis depends on sustained structural changes in NAND pricing and execution in enterprise SSDs.
• Teradyne remains exposed to project timing volatility and delays in major AI-related customer ramps.

What to watch

• Leading-edge wafer, CoWoS and advanced-packaging capacity additions and utilization.
• Whether lower compute costs broaden AI demand beyond accelerators into CPUs, networking, memory and supporting silicon.
• TSMC's Arizona ramp, N2 economics, overseas-capacity execution and progress toward its 56%-plus gross-margin target.
• NAND supply discipline and the volume coverage and economics of SanDisk's longer-term agreements.
• Analog demand, OEM restocking, backlog and lead-time trends in industrial and automotive markets.
• AI-related revenue ramps at Amkor, Teradyne, SiTime, Cadence, Applied Materials and IBM.