Industry View · Semiconductors
Semiconductors stopped being a cyclical commodity business and became the physical chokepoint of the AI economy. Global chip sales hit $772 billion in 2025 and are tracking toward roughly $975 billion in 2026, with essentially all of the upward revision driven by AI demand while non-AI end markets stayed soft.
EDA 5-10x productivity gains; Nvidia data-center revenue $47.5B to $193.7B
Reinforcement-learning and generative tools automate floorplanning, optimization and verification, cutting workflows from days to minutes.
Hyperscalers commission bespoke accelerators (XPUs/TPUs) to cut cost and dependence on Nvidia, reshaping who captures design value.
Leading-edge logic (3nm/2nm) and AI accelerator volume concentrate in one foundry, making capacity allocation a geopolitical lever.
AI accelerators are memory-bound; HBM supply is sold out through 2026 and pricing is in a structural shortage.
CoWoS interposers and EUV/High-NA tools are the gating capacity; lines stay booked even as output expands fast.
01 · The thesis
The AI cycle has concentrated value, not spread it. A single foundry (TSMC) fabricates the accelerators for all five US hyperscalers plus Broadcom; a single equipment maker (ASML) holds 100% of EUV lithography; and a memory triopoly (SK Hynix, Samsung, Micron) controls the high-bandwidth memory that every AI chip needs — with all 2026 HBM capacity sold out. The binding constraint is no longer wafer starts but advanced packaging: TSMC's CoWoS lines have been fully booked with 52-78 week lead times even as capacity roughly tripled in two years. The second-order story is that AI is now designing the chips that run AI. Synopsys and Cadence have pushed reinforcement-learning and agentic tools into place-and-route and verification, with customers reporting 5-10x productivity gains and AI methods now touching 20-40% of leading-edge designs. That compresses design cost and time-to-market — and lowers the barrier for hyperscalers building custom silicon to escape Nvidia's margin.
Reinforcement-learning and generative tools automate floorplanning, optimization and verification, cutting workflows from days to minutes.
Hyperscalers commission bespoke accelerators (XPUs/TPUs) to cut cost and dependence on Nvidia, reshaping who captures design value.
Leading-edge logic (3nm/2nm) and AI accelerator volume concentrate in one foundry, making capacity allocation a geopolitical lever.
AI accelerators are memory-bound; HBM supply is sold out through 2026 and pricing is in a structural shortage.
CoWoS interposers and EUV/High-NA tools are the gating capacity; lines stay booked even as output expands fast.
02 · The two clocks
Capacity clock: TSMC's CoWoS advanced-packaging demand roughly tripled from ~370,000 wafers in 2024 to ~670,000 in 2025 and is approaching ~1.0 million in 2026, yet lines stay booked with 52-78 week lead times — the gating constraint on every AI accelerator (Silicon Analysts). Memory clock: HBM is sold out through 2026, DRAM market revenue grew 260% YoY in Q1 2026, and analysts warn of a memory pricing shortage that could persist for years as AI data centers absorb supply (Counterpoint; Deloitte). Concentration clock: Nvidia's data-center revenue grew from $47.5B (FY2024) to $193.7B (FY2026) in two years, now ~90% of company revenue — even as hyperscaler custom ASICs, a ~$30B market growing 30%+ annually, emerge as the faster threat (Nvidia 10-K; JPM).
03 · Public players & exposure
We plot the listed players on two editorial axes — how exposed each is to AI disruption, against how ready its data, brand and position are to be the answer. The figures in the table are sourced; the placement is our read.
04 · Private flagships
The companies attacking this industry AI-first, with disclosed funding where available:
Blackwell and the coming Rubin generation anchor the accelerator market; Jensen Huang cites $500B+ in booked Blackwell/Rubin orders and partnerships to deploy 10GW for OpenAI alone.
Mass-producing 2nm and ramping a $165B US expansion while managing the geopolitical exposure of fabricating nearly all advanced AI silicon in Taiwan.
Co-designs TPUs/XPUs for Google, Meta, OpenAI and Anthropic; signed a Google supply agreement running through 2031 and a 10GW OpenAI accelerator deal.
Pursues training and inference with the largest single chip ever built; 2026 IPO was the year's biggest tech debut, with shares up 68% on day one.
Deterministic LPU architecture targets ultra-low-latency LLM serving, the highest-volume recurring AI workload as inference overtakes training in spend.
Led by chip-design veteran Jim Keller, it bets on an open, CUDA-free RISC-V plus AI-accelerator stack manufactured on Samsung 4nm.
05 · Signals
06 · The exposure read
AI rewards clean, structured advantage and punishes friction. The line runs through who owns the data, the brand and the customer — and who is merely a step the technology can route around.
Sources
AI's footprint across the chip stack, 2025-2026
Sources: Nvidia FY2026 10-K; Broadcom earnings; TSMC investor materials; WSTS/SIA. Figures are latest disclosed annual or run-rate. ($B revenue)
AI absorbs the entire forecast revision
WSTS raised the 2026 global chip outlook by ~$175B in December 2025 — driven entirely by AI demand, with weakness in non-AI markets (Deloitte).
Anthropic-Google-Broadcom mega-deal
Anthropic committed to up to 1M TPU chips, then a further multi-gigawatt TPU deal for 2027, the single largest external AI-ASIC arrangement to date (Hashrate Index).
SK Hynix overtakes Samsung on profit
SK Hynix posted 47.2T won 2025 operating profit, beating Samsung for the first time, as HBM shortages pushed memory pricing into a structural up-cycle (CNBC).
2nm enters high-volume manufacturing
TSMC began N2 production with monthly capacity scaling toward 200,000 wafers by 2028; 2nm orders surged despite ~$30,000 wafer pricing (DigiTimes/Jukan).
Big debuts reopen the chip-IPO window
Cerebras' blockbuster IPO and Groq's mega-round signaled the market is again ready to fund large semiconductor challengers (EE Times).
Blackwell and the coming Rubin generation anchor the accelerator market; Jensen Huang cites $500B+ in booked Blackwell/Rubin orders and partnerships to deploy 10GW for OpenAI alone.
Mass-producing 2nm and ramping a $165B US expansion while managing the geopolitical exposure of fabricating nearly all advanced AI silicon in Taiwan.
Co-designs TPUs/XPUs for Google, Meta, OpenAI and Anthropic; signed a Google supply agreement running through 2031 and a 10GW OpenAI accelerator deal.
Pursues training and inference with the largest single chip ever built; 2026 IPO was the year's biggest tech debut, with shares up 68% on day one.
Deterministic LPU architecture targets ultra-low-latency LLM serving, the highest-volume recurring AI workload as inference overtakes training in spend.
Led by chip-design veteran Jim Keller, it bets on an open, CUDA-free RISC-V plus AI-accelerator stack manufactured on Samsung 4nm.
| Company | Stance | The sourced fact |
|---|---|---|
| Nvidia NVDA | Dominant incumbent | FY2026 revenue hit $215.9B (+65%) with data-center revenue of $193.7B and 71.1% full-year gross margin; holds roughly 80-90% of the AI accelerator market (Nvidia FY2026 results; Counterpoint/Silicon Analysts). |
| TSMC TSM | Indispensable foundry | Q3 2025 revenue of $33.1B (+30.3% YoY) with HPC at 57% of sales and ~71% global foundry share; AI revenue ~$30B, about 24% of 2025 sales (TSMC investor materials; Counterpoint). |
| ASML ASML | EUV monopoly | Record €32.7B net sales in 2025 with 100% EUV market share and 2026 guidance raised to €36-40B on AI-driven demand (ASML Q4 2025 results). |
| Broadcom AVGO | Custom-silicon king | FY2025 AI revenue of $20B (+65%); Q1 FY2026 AI revenue $8.4B (+106%) with a $73B backlog and a stated target above $100B in AI chip revenue by 2027 (Broadcom earnings; Tom's Hardware). |
| SK Hynix 000660.KS | HBM leader | Record 2025 revenue of 97.1T won (+~50%) and 47.2T won operating profit (doubled YoY); ~57-62% HBM market share with all 2026 capacity sold out (CNBC; Counterpoint). |
| AMD AMD | Credible #2 | Q3 2025 record revenue of $9.2B (+36% YoY); Instinct MI-series generated an estimated $7-8B in 2025 for roughly 5-7% accelerator share, with an OpenAI multi-gigawatt deal signed (AMD earnings; Silicon Analysts). |
| Micron MU | Memory third | Holds ~21% of HBM revenue share, expects to sell out 2026 HBM capacity and guided toward an ~$8B HBM annualized run-rate (Counterpoint; Reuters). |
| Intel INTC | Turnaround bet | 2025 revenue of $52.9B (roughly flat); 18A ramped into volume but external foundry revenue was only $307M, and custom AI processors reached just a >$1B run-rate against a $100B TAM (Intel 10-K/annual report; Investing.com). |
| Cerebras CBRS | Wafer-scale challenger | Raised $1.1B at an $8.1B valuation in 2025, then priced a 2026 IPO at $185/share raising $5.55B at a fully diluted valuation near $56B (IntuitionLabs; Metaintro/EE Times). |
| Groq GROQ-pvt | Inference upstart | Raised a large 2025 round lifting its valuation to about $6.9B on an LPU architecture built for low-latency inference (IntuitionLabs). |