Key takeaways
- Search traffic still piles onto “NVIDIA.” The cash that builds an AI cluster also hits TSMC packaging, Broadcom XPUs and networking, and optical vendors shipping 800G/1.6T and coherent DCI. Those are different income statements.
- NVIDIA’s last official print (Q1 FY2027, quarter ended April 26, 2026): $81.6 billion revenue, $75.2 billion data center, networking inside that at $14.8 billion (+199% year over year). Guide for the July quarter: $91.0 billion ±2%, no China data-center compute assumed. Call is Aug. 26.
- Coherent reported $2.05 billion on Aug. 12 and still sold off. Ciena’s prior quarter had cloud at 46% of sales, up 70% year over year. Same AI story. Different mix, cash conversion, and what was already in the price.
AI semiconductor stocks are not a synonym for one ticker. The query is what Google sees. The bill of materials is what hyperscalers pay: a GPU or a custom accelerator, HBM, a CoWoS (or equivalent) package, a switch, copper that is dying at rack scale, then optics — pluggables now, co-packaged and near-packaged optics as density climbs. TrendForce put the AI-focused optical transceiver market at about $16.5 billion in 2025 and $26 billion in 2026. That is not a rounding error on NVIDIA’s data-center line. It is a second industry riding the same capex.
This week already showed the split. Optical names did not move as a bloc after earnings. If your model is “AI = NVDA,” you will misread every print between now and Aug. 26.
Three bills, one search term
Map the stack before the tickers:
- Compute — merchant GPUs (NVIDIA, AMD) and custom XPUs/ASICs that Broadcom (and others) design with Google, Meta, Amazon, Microsoft. Software demand (Azure, Copilot seats, frontier-model token burn) shows up here as orders, not as a SaaS multiple.
- Wafer and package — TSMC process plus CoWoS / SoIC. Lead times live here. A sold-out 2 nm line does not help if the interposer queue is the gate.
- Move the bits — NVIDIA’s own networking (the $14.8 billion line), Broadcom switch/SerDes, then Coherent, Ciena, Lumentum-class photonics, and the systems houses. Inside the hall: 800G and 1.6T transceivers. Between halls: coherent DCI and multi-rail line systems.
Power and construction are a fourth bill. They do not show up in a semiconductor ETF the same way. Ignore them and you will not understand why a “beat” still gaps down: the customer ran out of megawatts, not GPUs.
The GPU print
NVIDIA’s last 8-K-quality number is Q1 FY2027. Company release, May 20, 2026:
- Revenue $81.6 billion, +85% year over year.
- Data Center $75.2 billion, +92%.
- Inside Data Center: compute $60.4 billion (+77%); networking $14.8 billion (+199% year over year, +35% sequential).
- Q2 FY2027 outlook: $91.0 billion ±2% revenue; GAAP / non-GAAP gross margin 74.9% / 75.0% ±50 bps. No Data Center compute from China in that outlook.
- Extra $80 billion repurchase authorization; quarterly dividend lifted from $0.01 to $0.25.
Networking growing faster than compute is the tell. Scale-out is eating copper. That is why optical vendors and Broadcom’s AI networking mix are not a sideshow. The $80 billion authorization is a ceiling, not cash already spent — we already unpacked that distinction: stock buybacks.
The Aug. 26 call covers May–July. Consensus will argue over $91 billion versus a few billion of China H200 hope. Gross margin at ~75% is the other tripwire. Mix toward systems and memory cost can move that number without “AI demand dying.”
Foundry and CoWoS
TSMC is the factory the other logos rent. July 2026 consolidated revenue: about NT$467.58 billion, +5.6% month over month, +44.7% versus July 2025. January–July 2026: NT$2,872.06 billion, +37.0% versus the same stretch in 2025. Source: TSMC July 2026 revenue report (SEC).
Second-quarter 2026 (ended June 30): NT$1,270.38 billion revenue, NT$27.25 diluted EPS (US$4.31 per ADR). Third-quarter guide on that release: US$44.6–45.8 billion revenue, 65–67% gross margin. That is foundry math, not a GPU ASP.
CoWoS capacity is the constraint everyone cites in field notes (tens of thousands of wafers per month scaling through 2026, NVIDIA and Broadcom taking the bulk). Treat sell-side wafer counts as estimates. Treat TSMC’s monthly revenue as the hard print. If July keeps running hot into August while NVIDIA guides “only” in-line, the bottleneck thesis is packaging and HBM, not a demand cliff.
Custom silicon and the software pull
Hyperscalers do not want a single queue at NVIDIA. Custom ASICs are how they dual-source compute. Broadcom is the name that actually reports an “AI semiconductor” line.
Q2 FY2026 (Broadcom): AI semiconductor revenue $10.8 billion, +143% year over year. Networking was almost 40% of that AI bucket. Bookings for AI semiconductors in the quarter: over $30 billion against $10.8 billion shipped. Q3 guide: AI semiconductor revenue $16.0 billion (+200%+ year over year); company revenue about $29.4 billion. Full-year FY2026 AI semiconductor: about $56 billion. FY2027: still talking more than $100 billion. Source: Broadcom Q2 FY2026 release.
That is software demand with a two-year lead time. Token consumption at Microsoft, Google, Meta, Amazon, and the frontier labs is the order. Azure crossing $100 billion annual run-rate on Microsoft’s FY2026 print (year ended June 30, 2026) is the pull. Capex in that Q4 was reported around $41 billion, roughly two-thirds short-lived assets (GPUs/CPUs). A later “cut” of calendar-2026 reported capex toward ~$175 billion from ~$190 billion was framed as lease reclassification after stretching building lives — not a cancellation of halls. Read the 10-K footnotes, not the headline that Microsoft “slowed AI.”
Infrastructure software inside Broadcom (VMware estate) is a different multiple. Do not mash it into the XPU CAGR.
The optics split this week
Copper dies as rack power and reach blow past what DACs can do. 800G is the volume workhorse. 2026 is the year 1.6T pluggables leave sampling. CPO/NPO is the next architecture argument; TrendForce’s June 2026 note had that niche tiny in 2025 and a tens-of-billions 2030 story. Do not price 2030 CPO into 2026 EPS.
Coherent (NYSE: COHR) — fiscal Q4 ended June 30, 2026, reported Aug. 12: revenue $2.05 billion, +34% year over year (about +42% pro forma). Datacenter & Communications $1.62 billion of that. GAAP gross margin 38.5%; non-GAAP EPS $1.74. FY2026 revenue $7.12 billion versus $5.81 billion. Q1 FY2027 guide: $2.2–2.4 billion revenue, non-GAAP EPS $1.85–$2.05. CEO line: copper-to-optical inside AI datacenters. Source: Coherent Q4/FY2026 release (PDF). The stock still dropped. Beat-and-fade is what happens when the print was in the price and investors rotate to cash conversion and industrial mix. Photonics capex to build indium phosphide and 1.6T capacity is real cash leaving before the transceiver ASP shows up.
Ciena (NYSE: CIEN) — fiscal Q2 ended May 2, 2026: about $1.57 billion revenue, ~40% year over year. Company slides: cloud provider revenue 46% of total, +70% year over year; RLS and Waveserver each +50%+; first hyperscaler multi-rail (RLS Hyper-Rail) award. That is DCI and line systems between buildings, not the same SKU as a 1.6T DR8 pluggable. Ciena Q2 FY2026 earnings slides (SEC).
Cisco’s FY2026 close put AI infrastructure orders in the billions and still sold off on gross-margin mix (more iron, memory inflation). Nokia’s AI/cloud order book is the other systems tell. Same demand. Different gross-margin math. Optical “AI exposure” is not a single beta.
What Aug. 26 actually tests
Three questions, not a religion:
- Does NVIDIA clear ~$91 billion with ~75% gross margin without needing China compute to paper it?
- Does networking commentary still outrun compute — which keeps Broadcom’s mix story and the 1.6T ramp intact?
- Does management talk supply (CoWoS, HBM, optics, power) or demand? Supply talk is a full book. Demand talk that sounds like “customers digesting” is the multiple-compression script.
A green NVIDIA and a red Coherent can be consistent. One is a 75% margin software-defined chip company. One is a photonics manufacturer spending to catch a ramp. TSMC’s July print already told you wafers are still shipping. Ciena already told you cloud is paying for glass between sites.
If you only own the search-term ticker, you are not in “semiconductors, AI software, and optical.” You are in one income statement that happens to sit on top of the other two.
Education, not a recommendation to buy or sell NVDA, TSM, AVGO, COHR, CIEN, or anything else. Figures are from company releases and SEC filings dated through Aug. 14, 2026. Prints move. Read the 8-K.
