AI & chip supply chains, tracked with customs data
Official trade statistics from the US, Japan and Korea show what physically ships — memory from Korean fabs, transceivers from Thailand, servers from Mexico. We map each flow to the company that ships it, test it against reported revenue, and use the flows that really move with revenue to estimate the current quarter before earnings.
Upcoming earnings — our revenue estimates
Current-quarter revenue from our data model, next 45 days. Consensus shown only as a gap.
| Company | Reports | Quarter | Our estimate | YoY | vs consensus | Backtest ±5% | Model |
|---|---|---|---|---|---|---|---|
| 298040.KS Hyosung Heavy Industries | Oct 31 | Sep 2026 | ₩1.9T ₩1.8T–₩2.1T | +19% | — | 5/10 Medium | Trade data |
| FN Fabrinet | Nov 2 | Sep 2026 | $1.45B $1.43B–$1.47B | +48% | +3.1% | 12/12 High | Guidance-based |
| ANET Arista Networks | Nov 3 | Sep 2026 | $3.44B $3.29B–$3.59B | +49% | +3.2% | 11/12 High | Guidance-based |
| COHR Coherent | Nov 4 | Sep 2026 | $2.38B $2.30B–$2.45B | +50% | +3.1% | 8/8 High | Guidance-based |
| LITE Lumentum | Nov 5 | Sep 2026 | $1.28B $1.25B–$1.30B | +139% | +2.0% | 12/12 High | Guidance-based |
| NET Cloudflare | Nov 5 | Sep 2026 | $757M $746M–$768M | +35% | +2.5% | 12/12 High | Guidance-based |
| 103590.KS Iljin Electric | Nov 13 | Sep 2026 | ₩562B ₩508B–₩616B | +25% | — | 4/8 Medium | Operating data |
| 059090.KQ MIKO | Nov 14 | Sep 2026 | ₩421B ₩341B–₩502B | +72% | — | 5/10 Medium | Operating data |
| PANW Palo Alto Networks | Nov 19 | Oct 2026 | $3.34B $3.32B–$3.36B | +35% | +1.0% | 12/12 High | Guidance-based |
| DELL Dell Technologies | Nov 24 | Oct 2026 | $51.0B $39.8B–$62.2B | +89% | +3.7% | 3/4 Medium | Guidance-based |
| ZS Zscaler | Nov 24 | Oct 2026 | $956M $938M–$975M | +21% | +1.9% | 12/12 High | Guidance-based |
| CRWD CrowdStrike | Dec 1 | Oct 2026 | $1.55B $1.53B–$1.58B | +26% | +1.6% | 12/12 High | Guidance-based |
| MRVL Marvell Technology | Dec 1 | Nov 2026 | $3.18B $3.15B–$3.21B | +53% | +0.8% | 12/12 High | Guidance-based |
| NTAP NetApp | Dec 1 | Oct 2026 | $2.13B $2.05B–$2.20B | +25% | +1.1% | 11/12 High | Guidance-based |
| OKTA Okta | Dec 1 | Oct 2026 | $831M $818M–$845M | +12% | +1.8% | 12/12 High | Guidance-based |
| HPE Hewlett Packard Enterprise | Dec 4 | Oct 2026 | $14.8B $13.9B–$15.6B | +52% | +2.6% | 7/11 Medium | Guidance-based |
| RBRK Rubrik | Dec 4 | Oct 2026 | $469M $453M–$484M | +34% | +9.0% | 7/7 High | Guidance-based |
| S SentinelOne | Dec 4 | Oct 2026 | $312M $305M–$318M | +20% | +0.4% | 12/12 High | Guidance-based |
Backtest = quarters where the model, run as of that time, landed within ±5% of reported revenue. “Guidance” = company guidance adjusted by its historical beat rate.
Scorecard — estimates locked before each report
Our last estimate before the company reported vs the actual revenue. Hits and misses both stay here.
Biggest supply-chain moves
Company-specific flows with the largest change: last 3 months vs the same 3 months a year earlier.
* Check before reading: a small base a year ago or a short history against revenue.
Coverage
Memory
DRAM, HBM and NAND made in Hiroshima and Taiwan, assembled in Malaysia. US imports of Taiwanese memory track revenue most directly (grade A).
NAND from the Kioxia JV fabs in Japan, assembled in Malaysia and China. Japan's nationwide memory exports match revenue growth best. Multi-year contracts from 2026 can delay how fast spot prices reach reported revenue.
World's largest memory maker. We track the semiconductor (DS) division only. Memory made at Pyeongtaek, Hwaseong and Giheung is over half of Korea's memory exports.
No. 1 in HBM, No. 2 in DRAM. Company-wide revenue. Icheon (DRAM, HBM) and Cheongju (NAND, HBM packaging) exports make up most of the rest of Korea's memory exports; China fabs are not in Korean data.
Optical
InP lasers (San Jose, Japan) plus transceivers built by Fabrinet in Thailand.
Optical module contract manufacturer in Thailand — the core of the Thai transceiver flow (peer comparison).
Optical transceivers, lasers and materials made in Malaysia, China and Vietnam.
Servers & networking
AI servers drive growth. Built in Mexico, Malaysia and China and shipped into the US; parts inventory leads shipments.
Servers, networking (Juniper) and storage. Network gear shipped from Mexico and Malaysia is the key flow.
All-flash storage, contract-built in Guadalajara and shipped via Laredo. Watch units and unit prices separately.
AI server racks assembled mainly in San Jose, so trade data sees the parts going in. Risk of revenue slipping on rack integration delays.
AI back-end Ethernet switches, contract-built in Thailand, Mexico and Malaysia and shipped into the US.
Semiconductors
Custom AI ASICs and optical DSPs made at TSMC and packaged in Asia — not directly visible in US imports.
Monopoly in EUV lithography. Dutch litho exports show shipments, but service revenue and EUV acceptance timing blur the link; exports to Korea and the US track revenue best.
Components & materials
Copper-clad laminate (CCL) for AI accelerator boards. CCL exports from Jeungpyeong county track the electronics division's quarterly revenue (correlation 0.97).
MLCCs. MLCC exports are filed under Suwon (HQ), not the Busan plant. Japan's MLCC exports track the cycle better.
Power & cooling
Solid-oxide fuel cells for on-site data center power, built in Delaware and California. Power-module exports from Delaware go mostly to its Korean partner; US sales are not in trade data.
UPS, power distribution and liquid cooling built in Mexico, China and India and shipped into the US. Lead times and capacity drive results.
HVAC and data center cooling (BasX), mostly made in the US, so trade data only catches some exports.
Battery storage systems, contract-built in Vietnam with Chinese cells. Revenue is lumpy around project deliveries.
Ultra-high-voltage transformers for US grid and data center demand. Korean transformer exports explain revenue only partly (US plant, domestic sales).
Transformers and breakers plus construction. Korean large transformer exports move with revenue (grade B).
Switchboards, breakers and automation. Export data does not explain revenue well (domestic sales, subsidiaries).
Wire, cable and transformers. Korean cable exports move with revenue (grade B).
Chip parts (cleaning, coating, ceramic heaters) plus heat recovery steam generators for gas power — a lagging play on data center power build-outs.
Neoclouds
Bitcoin miner turning into a GPU cloud (Microsoft, NVIDIA and other contracts). No company-specific trade flow — tracked through AI server imports as an industry proxy.
Full-stack GPU cloud with long-term hyperscaler contracts (Microsoft, Meta) and shorter AI-lab deals at higher prices per MW. Publishes its own spot GPU prices.
GPU cloud with direct OpenAI, Anthropic and Meta contracts; Microsoft was about two-thirds of 2025 revenue. Heavy secured borrowing.
AI cloud
Building OCI GPU data centers on a huge backlog, funded with bonds and leases — revenue depends on build speed, credit on borrowing pace.
Ad revenue funds AI data centers, partly through off-balance-sheet financing. Capex is a leading signal for AI infrastructure demand.
Cybersecurity
Platform security. Software has no trade flow, so we use disclosed ransomware victims as an industry proxy (weak link, grade C).
Endpoint and cloud security (Falcon). Ransomware counts have led revenue growth by about two quarters, but not in the last two years.
Backup-based data security. Listed in 2024, so the history is short.
AI-based endpoint and cloud security. Ransomware counts used as an industry proxy.
Zero-trust cloud security. Ransomware counts used as an industry proxy.
Firewalls and SASE. Hardware refresh cycles matter more than attack counts.
Identity and access management. Ransomware counts used as an industry proxy.
CDN, DDoS protection and zero trust. Security is only part of revenue.
Biotech
AI drug discovery. No meaningful trade proxy — driven by clinical and partnership news.
ARIKAYCE and brensocatib. Drug HS codes are too broad for a useful trade proxy.
Single-cell and spatial genomics instruments made in California; trade data only catches exports (mainly to China).
How it works
- Find where each company makes its products (filings, contract manufacturers) and the customs flows that carry them — by country, HS code, US state, Japanese customs office or Korean county.
- Test every flow against quarterly revenue: year-over-year correlation with a 0–2 quarter lead. Grades: A ≥0.7 direct link, B ≥0.5 linked, C ≥0.3 industry context, D weak.
- Estimate the current quarter only from A/B flows, pick the model with the smallest backtest error, and publish the range. Low-confidence cases are shown as such or not at all.
Data updates daily as sources publish (trade data is monthly). As of Oct 30 (KST).