The next PCB demand cycle: AI, advanced boards and Thailand

What the numbers say

MPBxChange's finding: the next PCB cycle combines computing and networking demand, greater board complexity, automotive electronics and geographically diversified supply. Public evidence does not establish a predominantly humanoid-robot-driven investment cycle in Thailand. This study uses Chinese, Japanese, Korean, German, US and international sources. Its data cutoff is 30 September 2026.

The June 2026 Prismark series reproduced in Chinese issuer filings puts annual global PCB value at $85.836bn in 2025 and $144.610bn in 2030, an increase of 68.5%. The scope includes IC package substrates and excludes mounted components and assembly revenue. The first endpoint is an estimate; the second is a forecast, not booked orders.^[1]

Shipment area grows from approximately 483 million to 628 million m², or 30.0%. Area is neither individual board count nor copper-layer area. Its underlying public methodology is incomplete. Dividing value growth by area growth implies a 29.6% increase in pooled value per reported square metre. Product mix, complexity, pricing and possibly exchange rates all contribute; this is not the price increase of an unchanged board.^[2]

Where the additional value goes

These endpoints imply 73.2% of incremental value comes from 18+ layer boards, HDI and substrates. This is a technology breakdown, not AI's market share. Excluding substrates, the published totals imply $70.945bn in 2025 and $115.000bn in 2030, or 62.1% growth. The 2030 categories reconcile exactly. The rounded 2026 category sum is $0.001bn below its published total.^[1]

Prismark's separate end-market view gives server and data-storage PCB a $16.031bn baseline and 24.8% annual growth to 2030. Applying that rounded rate gives approximately $48.53bn in 2030, or 55.3% of total projected incremental value. This includes conventional servers and storage. Wired infrastructure has 19.0% projected annual growth; phones 4.5%, computers 2.9% and consumer electronics 3.7%. Never add end-market totals to technology totals: they classify the same market differently.^[3]

Why factories choose Thailand

Our inference is a combination of global demand growth and geographic redistribution of manufacturing. Company disclosures establish stated reasons and operating results, not every private motive or a precise percentage caused by AI.

• Shengyi's 2023 decision cites overseas orders, trade-policy exposure, comparative production costs and the supplier base. These are the company's assessments.^[4]
• Shennan's original RMB1.27418bn project budget explicitly addresses international customers and global supply capability. A budget is not realized sales.^[5]
• CMK's September 2022 plan allocated approximately JPY25bn to phase one and targeted twice its baseline HDI capacity, with up to four times if phase two were fully equipped. Its stated demand rationale included automotive CASE. These are historical plans, not a current capacity measurement.^[6]
• WUS's H1 2026 report describes over 90% of overseas customers certified in its Thai data-communications division, original capacity essentially full, and volume production of boards up to 30 layers for 400G switches and AI servers. This is stronger evidence of actual qualification and production than an investment announcement.^[2]

Do company results support the demand story?

Sources: Samsung's comparable quarterly releases,^[11][12] Meiko's May briefing,^[13] CMK's business report,^[7] and Schweizer's interim release.^[18] Meiko's FY2025 ended March 2026. Its management says its plan uses customer forecasts; these are not firm orders. Its Vietnam expansion also demonstrates competition between manufacturing locations.^[14]

Shennan discloses depreciation, limited initial output and longer overseas qualification during the Thai factory ramp. Demand forecasts need a qualification and cash-flow schedule before they become an investment case.^[17]

What about robots?

The German-based IFR reports 603,000 industrial robot installations in 2025 and forecasts 806,000 in 2029.^[9][8] Separate coverage counts approximately 7,000 full-size humanoids above 140cm sold in 2025, a limited population.^[10]

The missing coefficient is verified bare-PCB content per robot, excluding chips, motors and assembly charges. At 806,000 units, a $1bn annual PCB market would require about $1,241 of bare boards per unit. At 7,000 humanoids, the threshold is about $142,857. These are divisions, not bill-of-materials estimates. We add no speculative robotics uplift to the PCB forecast.

How to build an independent forecast

1. Demand: annual hardware shipments × boards per system × finished board area, calculated separately for servers, switches, accelerators, storage and power boards. Avoid counting modules twice.
2. Value: good finished area × realized price per good square metre, by technology. Keep IC substrates in a compatible, separate series.
3. Factory workload: finished area ÷ (panel nesting efficiency × final yield). Check laser drilling, lamination, plating and testing constraints; area alone is insufficient.
4. Thailand share: change in output = initial share × global growth + initial global market × share change + global growth × share change. Use consistent units and product definitions.
5. Validation: fit quarterly models using only information published at each forecast date. Test on rolling future periods against seasonal-naive and trend baselines. Report out-of-sample errors and interval coverage.

The IEA's 485TWh to 950TWh data-centre electricity projection is a consistency check, not a PCB-volume conversion. Hardware efficiency, utilization and replacement rates can change.^[16]

Model status: this article reconciles external forecasts and computes transparent derived quantities. It does not present a fitted or backtested predictive model. No Thailand-wide output estimate is issued without a comparable factory capacity and sales census.

Check the source disagreements

Prismark's public summary gives $123.3bn for 2030, while the dated June 2026 version quoted in issuer filings gives $144.610bn. The public page does not date its forecast. The 17.3% difference is a version discrepancy, not a statistical confidence band; we do not average the two.^[15]

Shengyi's 18+ layer row prints 72.8% growth for 2026, although its $5.254bn and $9.406bn endpoints calculate to 79.0%. WUS's adjacent area statements mention both 5.4% and 5.8% annual growth; the rounded 483 and 628 endpoints imply 5.39%. We retain endpoints and recalculate. Multiple issuers quoting Prismark do not constitute independent forecasts.