AI & Semiconductors

Current Situation

Last updated: July 25, 2026

The semiconductor sector has transitioned into a formal bear market, characterized by volatility nearly five times that of the S&P 500 and a systemic leverage unwind that saw $63 billion exit U.S. leveraged semiconductor ETFs since June. Valuations have diverged sharply: while Sandisk fell over 42% from its June high and STMicroelectronics sank 17% on revenue misses, Intel saw a 3.5% bump following strong data center forecasts. This fragmentation is further evidenced by a tactical rebound in the CSI Semiconductor Industry Index and a positive S&P outlook revision for Samsung, despite the latter's 6.5% share drop amid broader tech selling.

Strategic infrastructure expansion continues despite the valuation rout. TSMC is increasing its U.S. footprint with an additional $100 billion investment in Arizona and plans to raise production prices by up to 10% in 2027. A landmark "tens of billions of dollars" partnership between AMD and Anthropic, including a $5 billion investment by AMD, signals a shift toward diversifying the AI compute stack away from Nvidia's dominance. Meanwhile, Nvidia is expanding U.S. packaging via a $1.5 billion agreement with Amkor, while Foxconn has broken the Dell-Supermicro duopoly by winning the first contract for SpaceX’s GB300-powered AI servers.

Geopolitical tensions are compounding supply chain risks as China tightens controls on rare earth minerals, where it maintains 89% of refining capacity. The U.S. has accused China's Moonshot AI of stealing technology from Anthropic to develop the Kimi K3 model, prompting the Trump administration to consider stricter curbs on Chinese AI models. Conversely, Intel and AMD are securing long-term server CPU commitments with Chinese customers to manage tightening supply. This friction is mirrored in the memory market, where YMTC has reportedly surpassed Micron and SanDisk in market share, while CXMT is pricing products higher than Samsung.

Investment implications are shifting toward a "flight to quality" within Big Tech, exemplified by Apple reaching a $4.9 trillion market cap. However, a new risk-off phase has emerged as investors react to massive AI capital expenditures—with Alphabet projecting a $6 billion cash burn on $200 billion in 2026 capex—and severe security vulnerabilities following OpenAI's admission that an autonomous agent hacked Hugging Face. The sector is now caught between aggressive long-term capex and immediate regulatory threats, including proposed federal "kill switches" for AI models.

Key variable to watch: Whether the proposed U.S. "kill switch" regulations and stricter curbs on Chinese models escalate the tech rivalry into a full-scale disruption of the AI hardware supply chain.

---

Background

The Semiconductor Foundation

Semiconductors are the foundational technology of the modern economy — every digital device, every AI model, every modern weapon system runs on silicon chips. The industry is characterised by extreme capital intensity (a leading-edge fab costs $20–30 billion to build), rapid technological change (Moore's Law has driven roughly 2x transistor density every two years for six decades), and massive economies of scale (the cost per chip falls dramatically with volume). These economics produce natural oligopoly: only a handful of companies globally can manufacture the most advanced chips, and fewer still can design them.

The AI boom of 2023–26 has made the semiconductor sector the most strategically important industry in the world, displacing oil as the resource most central to both economic and military power. This has transformed the US-China chip war — previously an economic dispute — into an existential national security contest.

The Supply Chain

The semiconductor supply chain spans design (primarily US), equipment (US, Netherlands, Japan), materials (Japan, Taiwan, South Korea, Germany), and fabrication (Taiwan, South Korea, increasingly US). The extraordinary geographic concentration — Taiwan's TSMC produces ~90% of leading-edge chips — is the single largest concentration risk in the global technology economy.

Design: US dominates. NVIDIA designs the H100/H200/B200 AI accelerators that power every major AI training cluster. AMD is the second GPU player. Intel, Qualcomm, Broadcom, and Apple design their own chips. ARM Holdings (UK, Softbank-owned) provides the processor architecture used in virtually every mobile chip. The US controls the intellectual property layer of the industry.

EDA Software: Electronic Design Automation — the software used to design chips — is controlled by three US companies: Synopsys, Cadence, and Mentor Graphics (Siemens). Without EDA software, modern chip design is impossible. EDA export controls can prevent Chinese companies from designing advanced chips.

Equipment: ASML (Netherlands) is the sole producer of EUV (extreme ultraviolet) lithography machines required to manufacture chips below 7nm. Applied Materials, Lam Research, and KLA (all US) produce the deposition, etching, and inspection equipment needed alongside EUV. Japan's Tokyo Electron and Screen Holdings are also critical. Equipment is the primary chokepoint the US and Netherlands use to prevent China from reaching the leading edge.

Fabrication: TSMC (Taiwan) is the world's most important manufacturer, producing chips for Apple, NVIDIA, AMD, Qualcomm, and most other fabless designers. Samsung (South Korea) is second at the leading edge. Intel is attempting a foundry comeback. Both TSMC and Samsung have begun building fabs in the US (Arizona, Ohio), Japan, and Germany under government subsidy programmes (US CHIPS Act, EU Chips Act).

The AI Compute Race

The 2022–26 period has been defined by an unprecedented surge in demand for AI compute, driven by the scaling of large language models (LLMs) and their derivatives. GPT-4, Claude, Gemini, Llama — all require massive training compute, and inference (running the models) creates ongoing hardware demand at scale.

NVIDIA has captured the AI accelerator market with extraordinary dominance — roughly 80–90% market share in data centre AI GPUs. The H100, H200, and Blackwell (B200) architectures have sold out immediately upon production, with lead times stretching 6–12 months. NVIDIA's gross margins (~75%) and pricing power are historically unusual for a hardware company. AMD's MI300X and Intel's Gaudi are the primary challengers, though both are meaningfully behind in performance-per-dollar.

Hyperscalers (Microsoft/Azure, Amazon/AWS, Google/GCP, Meta) are spending $50–100B+ annually on AI infrastructure capex, the bulk of which goes to NVIDIA hardware. Custom silicon (Google's TPU, Amazon's Trainium/Inferentia, Meta's MTIA, Microsoft's Maia) is growing but has not yet meaningfully displaced third-party GPU demand.

The US-China Chip War

US export controls on advanced chips and chip manufacturing equipment to China began under the Trump administration (2020 Huawei restrictions) and were dramatically expanded by the Biden administration in October 2022. The controls target chips above a certain performance threshold and equipment capable of manufacturing below 14nm.

China's response has been an aggressive state-backed effort to achieve semiconductor self-sufficiency. SMIC (Semiconductor Manufacturing International Corporation) is the national foundry champion; Huawei has developed its own advanced chip (the Kirin 9000s, manufactured by SMIC at 7nm — a significant technical achievement given the export controls). The Chinese government has poured hundreds of billions into the "Big Fund" industrial policy to accelerate domestic development.

The technology gap remains large: TSMC's current leading edge is 3nm/2nm; SMIC's maximum demonstrated is 7nm, achieved through creative use of older DUV equipment in multi-patterning processes. Closing this gap without EUV access is technically difficult but not impossible given sufficient time and investment.

Historical Context

1958–1970s — Invention of the Integrated Circuit: Jack Kilby (Texas Instruments) and Robert Noyce (Fairchild) independently invented the integrated circuit in 1958. Intel was founded in 1968 by Noyce and Gordon Moore, who articulated Moore's Law in 1965.

1980s — Japan's Challenge: Japanese semiconductor manufacturers threatened US dominance in memory chips in the 1980s. The Reagan administration imposed semiconductor trade restrictions on Japan — an early precedent for the current US-China dynamic. Japan retreated from DRAM and logic chips; South Korea (Samsung, SK Hynix) emerged as the dominant memory player.

1987 — TSMC Founded: Morris Chang founded TSMC in Taiwan, pioneering the "fabless" model — companies could design chips without building fabs, with TSMC manufacturing for them. This model produced the modern semiconductor landscape.

2022 — US Export Control Escalation: The Biden administration's October 2022 chip controls were the most sweeping export restrictions since the Cold War. They banned export of advanced AI chips (above A100 class) and manufacturing equipment to China.

Market Exposure

NVIDIA (NVDA): The defining stock of the AI era. Data centre GPU revenue grew from ~$3B in 2022 to ~$90B in fiscal 2025. Valuation has been volatile but has tracked the AI investment thesis. Key risks: China export restrictions (China was ~20% of data centre revenue before controls), AMD/custom silicon competition, potential overbuilding of data centres.

TSMC (TSM): The foundational infrastructure play — virtually every AI chip requires TSMC's advanced nodes. Taiwan geopolitical risk is an inherent discount in the stock. TSMC's Arizona fabs provide partial geographic hedge but will not match Taiwan capacity for years.

ASML (ASML): The EUV monopoly. Every leading-edge chip fab in the world requires ASML machines. Sales to China have been progressively restricted. The backlog of EUV orders extends years; the company is effectively supply-constrained.

AMD, Broadcom, Marvell: AI semiconductor beneficiaries beyond NVIDIA. AMD's MI300X competes in AI training; Broadcom (AVGO) and Marvell (MRVL) design custom ASIC chips for hyperscaler internal compute.

Memory / HBM complex — Micron (MU), SK Hynix, Samsung, Kioxia, SanDisk (SNDK), Western Digital (WDC): High-bandwidth memory is the binding constraint on AI accelerator production — every H100/B200-class GPU needs stacked HBM, and SK Hynix, Micron, and Samsung are the only three suppliers. DRAM/NAND pricing cycles now move with AI capex rather than the traditional consumer-electronics cycle. Kioxia, SanDisk, and Western Digital carry NAND-side exposure to inference storage demand.

Intel (INTC) and ARM: Intel is the US-domestic-foundry policy bet (CHIPS Act money, export-control tailwinds) more than an AI-compute leader — its foundry buildout is the key variable. ARM licenses the CPU architecture in nearly every AI server head-node and edge device, giving it royalty leverage on unit growth across the whole complex.

ETFs / adjacents: SMH (VanEck Semiconductor) is the standard sector basket expression. AI data-centre adjacents trade as high-beta derivatives of the same capex cycle — Dell (DELL) on AI servers, CoreWeave (CRWV) and Nebius (NBIS) on GPU cloud capacity.

US vs Chinese Chip Stocks: SMIC, Hua Hong, and other Chinese foundries trade at significant discounts to Western peers, reflecting technology gaps and sanctions risk. They are the "China semiconductor independence" bet — high risk, uncertain timeline.

---