Snapshot

Contract electronics manufacturers (EMS — Electronics Manufacturing Services) design, assemble, and test complex hardware on behalf of companies that don't run their own factories. For AI and data centers, that means building GPU server racks, networking switches, power distribution units, and custom AI accelerator boards. The four US-listed names with meaningful AI/DC exposure are Celestica (CLS), Flex (FLEX), Jabil (JBL), and Plexus (PLXS). Combined, these four generated roughly $80B in total revenue in their most recent fiscal years, with AI/data-center-linked revenue estimated at $15–20B and growing 40–75% year-over-year depending on the company. The global EMS market was $648B in 2025; the IT & telecom slice (which includes data center hardware) is the fastest-growing segment at an 8.7% CAGR through 2034.

Hyperscaler capex is running at $660–690B in 2026, nearly double 2025 est.. Every dollar of that capex going to servers, switches, and racks flows through contract manufacturers — hyperscalers design chips but outsource physical assembly. CLS has the most concentrated AI exposure (~80% of revenue in CCS, growing 76% YoY); FLEX is spinning off its Cloud & Power Infrastructure segment ($6.6B, +38%) as a standalone company; JBL's Intelligent Infrastructure segment hit $4.1B in Q2 FY2026 alone (+52% YoY); PLXS is the smallest and least AI-exposed, entering DC through power management and thermal solutions adjacencies.

The product & how money is made

An EMS company takes a customer's design — say, an 800G network switch layout from a hyperscaler, or an AI server rack architecture from NVIDIA — and handles the physical work: sourcing components, assembling printed circuit boards, integrating systems (compute, power, cooling), testing, and shipping finished units. The customer owns the design; the EMS company owns the factory, the workforce, and the supply chain relationships with component vendors.

Cash comes in through three channels. First, hardware assembly revenue — the largest piece — paid per unit built, typically on a cost-plus or fixed-price-per-unit basis. Gross margins on assembly run 6–10%, with AI/DC products trending higher because they are more complex (liquid-cooled racks, high-layer-count PCBs, precision power delivery). Second, engineering and design services — helping customers turn a concept into a manufacturable product, sometimes called "HPS" (Hardware Platform Solutions) at Celestica or "co-design" at others. This carries higher margins because it is labor-intensive and sticky. Third, aftermarket and lifecycle services — repair, refurbishment, and spare parts management. Smaller today but growing as the installed base of AI infrastructure expands.

The unit of demand is EMS content per dollar of hyperscaler capex — how many cents of every capex dollar pass through a contract manufacturer. As hyperscalers increasingly outsource full-rack integration (not just board-level assembly), that content-per-dollar figure is rising.

Demand

Contracted and observed

• Celestica raised FY2026 revenue guidance three times in six months: from $16.0B (Oct 2025) to $17.0B (Jan 2026) to $19.0B (Apr 2026). Q1 2026 CCS segment revenue was $3.24B, up 76% YoY. HPS (the AI server/rack sub-segment) hit ~$1.7B in Q1 alone, +63% YoY. contracted
• Flex reported FY2026 (ending Mar 2026) total revenue of $27.9B. The Cloud & Power Infrastructure segment generated $6.6B, up 38% YoY. Full-year FY2027 guidance is $32.3–33.8B, implying ~18% growth at midpoint. contracted
• Jabil Q2 FY2026 (ended Feb 2026) revenue was $8.3B. The Intelligent Infrastructure segment — covering AI cloud, data center, and networking — hit $4.1B in the quarter, up 52% YoY. Updated FY2026 guidance: $34B total revenue with core EPS of $12.25. contracted
• Plexus Q2 FY2026 (ended Mar 2026) revenue was $1.164B, up 19% YoY. No standalone AI/DC segment. DC exposure enters through power management, thermal/cooling, and semiconductor capital equipment wins. Record $355M in annualized new manufacturing wins in Q2. Qualified funnel: $4B, up 11% sequentially. contracted

Demand drivers (forward-looking)

• Hyperscaler capex: the five largest US hyperscalers (Amazon, Alphabet, Microsoft, Meta, Oracle) spent ~$380B in 2025 and are forecast to spend $660–690B in 2026. Roughly 75% (~$450B) is directly tied to AI infrastructure: servers, GPUs, data centers, equipment. est.
• 800G → 1.6T networking transition: Celestica cites the 1.6T switch ramp in H2 2026 as a major catalyst. Each generation increases assembly complexity and EMS content per unit.
• Custom silicon proliferation: hyperscalers designing their own chips (Google TPU, Amazon Trainium, Meta MTIA) still need someone to assemble systems around them. That assembly job goes to EMS firms.
• Jabil's $500M capacity investment (announced Jun 2025) specifically targeting cloud and AI DC infrastructure manufacturing.

Sources: CLS Q1 2026 earnings (Apr 2026), FLEX FY2026 results (May 2026), JBL Q2 FY2026 results (Mar 2026), PLXS Q2 FY2026 results (Apr 2026), Futurum AI Capex 2026 report.

Supply

Capacity

EMS capacity is a network of specialized assembly lines, clean rooms, testing stations, and engineering teams spread across dozens of sites globally. The four companies here operate 100+ manufacturing sites combined. Key constraint: AI server assembly lines are not interchangeable with consumer electronics lines. A facility building smartphones cannot be retooled overnight to assemble liquid-cooled GPU racks with precision power delivery. Specialized tooling, cleanroom requirements, and workforce training take 6–18 months to stand up. est.

• Celestica is investing ~$1B in capex for FY2026 (~6% of revenue, up from 1.6% historically), expanding in the US, Thailand, and Mexico specifically for hyperscaler orders.
• Flex grew capex to $633M in FY2026 (from $438M in FY2025). The Cloud & Power Infrastructure segment operates across 22 engineering and manufacturing centers globally. Flex plans to spin off this segment as a standalone public company, targeting close in Q1 calendar 2027.
• Jabil announced $500M in capacity expansion (Jun 2025) for cloud/AI DC infrastructure. Also investing in liquid cooling and silicon photonics integration capabilities.
• Plexus states current facilities can support "well in excess of $5B in annualized revenue" (current run rate ~$4.7B annualized), with headroom from efficiency gains rather than new builds.
• Sanmina (SANM, not covered here) completed its acquisition of ZT Systems' DC manufacturing business from AMD in October 2025, adding facilities in New Jersey, Texas, and the Netherlands with liquid cooling capabilities.

Bottlenecks

• Skilled labor: GPU server assembly at scale requires thermal/mechanical engineering talent and precision manufacturing workers. Training takes months. Multiple companies cite workforce as a binding constraint.
• Component supply: Plexus flagged tightening in "portions of semiconductor, portions of passives, memory, and raw PCB fabs" in Q2 FY2026. EMS firms depend on component availability; they cannot assemble what they cannot source.
• Floor space in key geographies: US-based manufacturing is increasingly demanded by hyperscalers (for data sovereignty and supply chain resilience), but US floor space for high-complexity electronics assembly is limited.

Sources: CLS Q1 2026 guidance call, FLEX FY2026 annual report, JBL FY2025 results, PLXS Q2 FY2026 earnings call, Sanmina press release (Oct 2025).

The gap

Hyperscaler capex is roughly doubling year-over-year (from ~$380B in 2025 to $660–690B forecast in 2026 est.). EMS capacity for AI/DC products is expanding at 20–40% est. — fast, but slower than demand. The result is pricing power shifting toward EMS firms: operating margins across the group have expanded 40–180 basis points YoY, with Celestica reaching 8.0% adjusted operating margin (up from 7.1% a year earlier) and Flex hitting a record 6.3%.

If hyperscaler capex reaches $670B in 2026 and roughly 15–25% flows through EMS firms for server/switch/rack assembly, the addressable spend is $100–170B est.. The four companies here capture $15–20B of that — roughly 10–15% share of the EMS-addressable slice est.. The rest goes to Foxconn (by far the largest EMS globally, not US-listed), Sanmina, Wistron, Quanta, and other Asia-based ODMs.

The players

Sources: CLS Q1 2026 (Apr 2026), FLEX FY2026 (May 2026), JBL Q2 FY2026 (Mar 2026), PLXS Q2 FY2026 (Apr 2026), StockAnalysis.com for market caps (Jun 2, 2026).

The price of exposure

All four trade above their 5-year median P/E multiples.

The 5-year median P/E for EMS companies was 10–18x. Every name here is 2–4x above that historical range. The multiple embeds AI/DC growth continuing at 30–50%+ for multiple years. If hyperscaler capex plateaus or contracts, these multiples compress toward historical norms. If capex continues doubling, forward earnings catch up to the multiple.

What to deep-dive next

• CLS customer concentration — one customer is 30% of revenue; top 3 are 59%. Identify which hyperscalers and what happens if any one reduces orders. The 10-K names no customer explicitly; cross-reference with CCS segment geography and product mix.
• FLEX CPI spin-off economics — SpinCo targets 65–75% revenue growth in FY2027 and 80%+ in FY2028. Standalone margin profile, capital structure, and customer base will appear in the Form 10 filing (expected H2 2026).
• JBL Intelligent Infrastructure segment margins — at $4.1B/quarter and growing 52% YoY, this is a $16B+ annualized business. Segment-level operating margins vs. legacy Regulated Industries are not yet broken out.
• Component supply tightness — PLXS flagged semiconductor, passive, memory, and PCB fab tightening. If components become the bottleneck, EMS firms cannot convert backlog into revenue. Monitor lead time data from Fusion Worldwide, IHS, and distributor commentary.
• Sanmina (SANM) — completed the ZT Systems acquisition from AMD (Oct 2025), adding liquid-cooled DC manufacturing capacity. Track the first full quarter of consolidated results.

Sources & confidence