Capex Signal: April 18, 2026 — TSMC & ASML Prove the Supercycle Is Intact

This week was the one the bears were waiting for. TSMC and ASML both reported — the two companies that sit at the absolute chokepoints of the AI hardware stack. One makes the wafers. One makes the only machines that can pattern the leading-edge nodes those wafers run at. If either cracked, the supercycle narrative would need serious revision.

They didn't crack. In fact, TSMC raised its full-year revenue growth outlook from “close to 30%” to “above 30%” in USD terms, guided Q2 to $39–$40.2 billion, and CC Wei said the CAGR for AI accelerator revenue through 2029 is now expected at 54–56%— up from the 45% guide given a year ago. That's not a rounding error. That's a step-change in conviction.

ASML beat on revenue and profit, raised 2026 guidance, but the stock dropped 6% because China revenue fell from 36% of sales to 19% in a single quarter. That's the bifurcation story in one data point: the AI capex wave is accelerating everywhere except China, where the export control vice is tightening.

Let me walk through everything that mattered this week.

Earnings Scorecard — Week of April 14–18

Next up: LRCX reports April 22. AMAT, KLAC — May. AMD reports May 5. NVDA late May.

TSMC Q1 2026: The Most Important Earnings Print of the Quarter

I say that without hyperbole. TSMC is the physical substrate of the AI supercycle — every Blackwell, every TPU, every XPU, every Instinct MI-series chip runs through their fabs. When TSMC tells you demand is “extremely robust,” that's not marketing language. That's CC Wei reading his order book.

The margin improvement story is critical here. TSMC's gross margin reached 66.2%— a 390 basis point improvement — driven by higher capacity utilization and the relentless mix shift toward advanced nodes where pricing power is highest. More 3nm. More CoWoS. More advanced packaging. Every dollar of incremental AI accelerator revenue carries significantly better margin than the mature-node baseline.

Q2 2026 Guidance — Even Better

TSMC guided Q2 revenue to $39.0–$40.2 billion, a sequential increase of roughly 10% from Q1's $35.9B. That implies roughly 40%+ YoY growthin Q2. The full-year 2026 outlook is now “above 30% YoY in USD terms” — and with Q1 at 35% and Q2 guided even higher, the full-year bar feels conservative.

The 54–56% AI CAGR Call — Stop and Read This

CC Wei updated TSMC's AI accelerator revenue CAGR estimate for the period 2024–2029 from 45% to 54–56%. That is not noise. An incremental 9–11 percentage points on a five-year CAGR, compounded, translates to an enormous difference in cumulative revenue. At 45% CAGR, you roughly 6.4x the base in five years. At 55% CAGR, you nearly 8.6x it. TSMC's foundry volumes are the most direct read on where actual silicon demand is going, because they manufacture what gets ordered. This is order-book-driven guidance, not analyst extrapolation.

Wei also said TSMC's overall revenue CAGR through 2029 is now expected at ~25%, up from 20% previously. The semiconductor cycle has structurally re-rated, and TSMC is saying so in the most explicit terms management can.

CoWoS & Advanced Packaging — Still the Constraint

TSMC is targeting CoWoS capacity of roughly 130,000 wafers/month by late 2026— up from ~70,000 at end-2025 and ~35,000 at end-2024. That's nearly a 4x expansion in two years. And it's still not enough: NVIDIA is booking well over 50% of that capacity through 2027. OSATs like Amkor (AMKR) and ASE (ASX) are being pulled in as overflow CoWoS-S partners.

The new announcement this quarter: TSMC confirmed a pilot line for CoPoS— the next-generation packaging platform designed to handle even larger chip-on-wafer configurations that CoWoS can't address. Mass production is still years away, but the fact that they're piloting it now tells you where compute cluster sizes are headed.

Tariffs — The Stated Risk, the Real Hedge

Wei acknowledged tariff policy as a risk factor, as you'd expect. The more interesting story is TSMC's Arizona expansion, which has grown to a committed $165 billionwith plans for up to 12 fabs. The U.S.–Taiwan tariff deal framework gives TSMC effective exemption from semiconductor import tariffs for chips made domestically — which means Arizona production is simultaneously geopolitical insurance, customer relationship management (all the major U.S. hyperscalers want local fab options), and a tariff moat. Overseas fab margins are lower than Taiwan's right now, but the customer lock-in and tariff protection more than offset that.

ASML Q1 2026: Strong Beat, China Compression, Stock Down 6%

ASML reported on April 15th. The headline numbers were good — better than expected on both revenue and profit. The 2026 guidance went up. But the stock fell 6%, which tells you the market was already pricing in a lot, and the China revenue collapse (36% → 19% of sales in one quarter) was worse than many had modeled.

The China compression is happening fast. A bipartisan U.S. bill targeting DUV tool exports — the machines China can still buy — would represent roughly a 5% revenue impact if enacted, on top of the existing restrictions. ASML's CFO guided that China would account for ~20% of full-year demand in 2026, down from ~45% in 2023. That's a structural loss of the revenue base they built over the past decade of Chinese fab expansion.

The offset: TSMC now represents 23% of ASML's Q1 sales, up from 13% in Q4. The AI-driven capacity expansion at TSMC (and Samsung, and Intel in the U.S.) is more than absorbing the China shortfall for now. ASML guided it can deliver up to 80 EUV systems in 2027 if customer demand supports it.

One oddity: ASML quietly stopped disclosing quarterly order numbers in Q1. They say orders “continue to be very strong.” Whether that's a timing decision or a deliberate move to reduce noise in the stock, I don't know. I'd prefer transparency, but the 2026 guidance raise implies the backlog is healthy.

Supply Chain Ripple Effects

Memory: Micron HBM Sold Out Through 2026

No earnings this week, but memory remains a critical thread. Micron raised its FY2026 capex budget to $20 billion (from $18B) specifically to fund HBM capacity expansion. Their 2026 HBM supply is already fully committed. HBM4 ramp is underway in Q2. Morgan Stanley upgraded their WFE forecast: DRAM WFE projected at $34.9B (+18% YoY), NAND WFE at $13.8B (+35%). This flows directly to AMAT, LRCX, KLAC — all of whom report in the next three weeks.

WFE Outlook: +10% to $128B in 2026

Morgan Stanley revised their 2026 WFE market estimate upward from +5% to +10%, with 2027 looking like +14% to ~$150B. The drivers are familiar: leading-edge logic (3nm/2nm), HBM/DRAM, and advanced packaging tooling. LRCX reports April 22 with consensus at $5.76B (+22% YoY) for their March quarter. That will be the first equipment print of the cycle.

Advanced Packaging: Bottleneck Is Structural Through 2028

Global CoWoS demand is on track to hit 1 million wafers in 2026, up from 670K in 2025. NVIDIA is booking over half of TSMC's CoWoS allocation through 2027. The overflow is going to OSATs: Amkor (AMKR) is doubling capex to $2.5–3B in 2026, with the Arizona campus build-out absorbing most of that. ASE (ASX) is similarly scaling.

The bottleneck isn't going away fast. Advanced packaging — CoWoS, SoIC, EMIB — requires specialized equipment and yield development that can't be brute-forced. The constraint persists well into 2028, which means AMKR and ASX have multi-year demand visibility at above-normal utilization.

AMD Rallies 6% on Strong Pre-Earnings Sentiment

AMD was up 6%mid-week, rising from ~$258 to ~$274, ahead of its May 5 Q1 earnings. Data Center revenue grew 39% YoY to $5.38B in Q4. The MI350 and MI400 ramp timing is what the market will focus on. The French government collaboration announcement — a Letter of Intent on AI ecosystem development — is diplomatic noise relative to what matters: whether AMD's MI-series can actually take meaningful share from Blackwell at scale. I'm watching MI350 yield data and hyperscaler qualification timelines more than AI-cooperation press releases.

Marvell Up 7% — Barclays Upgrade, Custom Silicon Thesis Intact

MRVL climbed 7%to above $106, driven by a Barclays upgrade. The context matters: the custom silicon / XPU thesis continues to accelerate, and Marvell is one of the primary design partners for hyperscaler custom AI chips (alongside Broadcom). Their FY2027 $11B revenue target requires continued XPU program ramp. The Barclays upgrade is confirming what the TSMC earnings implied — the custom silicon order book is expanding.

Intel Up 8% — Riding the Sector Wave

INTC caught an 8% rally mid-week in line with the broad semiconductor sector move. Intel's own story — the foundry transition, the $8.9B federal investment, the 18A process node — is separate from the AI demand wave driving everyone else. I don't want to overread a sector rally in Intel. The foundry execution story remains the watch item, not near-term AI chip demand.

The Macro Frame: Constraint → Intensity → Durability

Running the Capex Signal framework this week:

Export Controls: The Policy Risk That Won't Settle

The most important policy overhang this week: the U.S. BIS tightened AI chip export rules in January, putting H200-equivalent and MI325X-equivalent chips on case-by-case review for China. More significantly, a bipartisan congressional bill now proposes restricting DUV (not just EUV) tool exports to China — which would directly impact ASML's remaining China business and, through supply chain effects, any Chinese fab expansion plans that depend on mature-node DUV tools.

The ASML China revenue drop this quarter — from 36% to 19% of sales — reflects the cumulative effect of existing restrictions. If DUV tools are further restricted, that 19% shrinks further. ASML has said their guidance assumes this scenario with “enough margin.” That's management signaling they can absorb it. The non-China AI capex wave is big enough to offset, for now.

Separately: Cadence Design Systems (CDNS) paid a $95 million administrative penaltyin a BIS enforcement action related to export controls. EDA tools — which enable chip design — are increasingly in the crosshairs. SNPS and CDNS both need to watch their China exposure carefully; the enforcement environment is tightening.

What to Watch Next Week

Lam Research (LRCX) — April 22: First equipment print of the cycle. Consensus at $5.76B (+22% YoY) for their March quarter, with guidance of $5.7B +/- $300M from the company. A beat here confirms the WFE demand thesis. Watch the China mix and any commentary on etch/deposition demand for HBM and advanced logic nodes.

AMAT, KLAC — May: Applied Materials reported a soft Q1 ($7.01B, -2% YoY) back in February, citing timing of tool deliveries. The Q2 and Q3 prints will be more telling as the WFE ramp accelerates into H2. KLAC will be a bellwether for process control intensity as N3/N2 yields mature.

AMD — May 5: Q1 data center GPU revenue and MI-series traction. This is the print that tests whether AMD is a real AI GPU competitor or a distant second.

Hyperscaler Earnings — Late April/May:Google, Meta, Microsoft, Amazon all report in the coming weeks. Their capex commentary and data center build pace will be the next major catalyst for the supply chain. After TSMC's order book confirmation, I expect all four to reiterate or raise their 2026 capex guidance.

NVDA — Late May:The most anticipated print of the season. Q1 FY2027 guidance will test whether the B200/B300 ramp is tracking or slipping. Given TSMC's blowout, the setup looks favorable.

Export control legislation: Watch the DUV bill in Congress. If it advances through committee, ASML stock will have another leg down, and the broader China fab build-out thesis gets further impaired.

Bottom Line

The two chokepoint companies in the AI hardware stack — the fab and the lithography monopoly — just told you the cycle is accelerating, not decelerating. TSMC raised its 5-year AI CAGR estimate by 10 percentage points and guided Q2 to $39–40B. ASML raised its 2026 guidance even as China revenue fell by half. WFE is heading to $128B+ in 2026. CoWoS is sold out through 2027. HBM is sold out through 2026.

The supercycle isn't stress-tested — it's accelerating. The risks are real (export controls, tariff volatility, packaging bottlenecks, execution on overseas fab ramps), but they're execution risks on a known trajectory, not fundamental demand risks.

The bet hasn't changed. More compute, more packaging, more interconnect, more power. Everything else is noise.

— J