Celestica Inc. (CLS.TO / NYSE: CLS) - Deep Dive Research Report

Research as of May 11, 2026. Four concalls used: Q2 2025 (July 29, 2025), Q3 2025 (October 27-28, 2025), Q4 2025 (January 29, 2026), Q1 2026 (April 28, 2026).

1. What the Company Does

Celestica designs, engineers, and manufactures complex electronics hardware for other companies. It does not make end products that consumers buy. Instead, it sits in the middle of the technology supply chain: a customer - typically a hyperscaler, an aerospace prime contractor, or a medical device company - brings Celestica an engineering challenge, and Celestica figures out how to build it, source the components, and deliver it at scale.

The founding story matters here because it explains everything about the company's DNA. Celestica was born in 1994 as IBM's internal manufacturing operation in Toronto. In 1996, Onex Corporation - a Canadian private equity firm - bought it from IBM for $550 million and turned it into an independent contract manufacturer. It went public in 1998. For the first fifteen years after its IPO, Celestica was essentially a commodity assembler: it took other people's designs and built them cheaply. The 2000-2002 dot-com bust nearly destroyed the company; it lost $4 billion and cut 3,000 jobs in 2001 alone. For a decade after, it ground along at sub-scale margins while competitors like Foxconn dominated by sheer volume and Jabil and Flex captured the complex-assembly tier.

The pivot began when Rob Mionis joined as CEO in 2015. Mionis's thesis was simple but difficult: move up the value chain from pure assembly into design. Rather than waiting for customers to hand over a finished blueprint, Celestica would co-develop hardware with customers, embed itself in the engineering process, and own more intellectual property. This was the origin of the Hardware Platform Solutions (HPS) model. Instead of being told "build 10,000 units of this exact thing," Celestica now says, "here is our switch platform, built on Broadcom's Tomahawk silicon, optimized for your AI cluster - we designed the PCB, we qualified the thermal management, we built the firmware stack, and we can produce it at scale from Thailand and Mexico." The customer gets a faster path to deployment; Celestica gets deeper integration, higher margins, and much stickier relationships.

The transformation became visible to the outside world when AI data center spending exploded in 2023. Celestica had spent years developing 400G and then 800G Ethernet switch platforms for hyperscalers who wanted an alternative to Cisco's proprietary networking gear. When Microsoft, Google, Meta, and Amazon began spending tens of billions building AI training clusters that required enormous, flat, low-latency networks, they found Celestica already had the engineering capability, the qualified supply chain, and the manufacturing capacity to deliver the switches they needed. The result was a business that grew revenue 28% in 2025 and is on track to grow 53% in 2026 - not because the end market was kind to an undifferentiated assembler, but because Celestica had spent a decade building a capability that nobody else had at scale.

The core value proposition, stated plainly: Celestica offers hyperscalers and other technology OEMs a way to get custom hardware faster, at high volume, with supply chain managed, at lower cost than doing it internally, and with much more design control than a standard white-box ODM provides. The "design-to-delivery" model means Celestica holds the engineering files, the component qualifications, the test and validation processes, and the manufacturing relationships. Switching to a competitor is not simply a matter of signing a new purchase order.

A concrete example: A hyperscaler needs a new spine switch for its AI training cluster. The switch must handle 1.6 terabits per second of bandwidth per port, fit in a 3U chassis, survive years of continuous operation in a hot aisle, integrate with the customer's orchestration software, and be delivered by the thousands within a defined ramp window. Celestica engineers work alongside the customer's architects from the design phase, select Broadcom's Tomahawk 6 silicon, design the PCB layout with 40+ layer stack-ups that require specialized manufacturing, develop the thermal management (air-cooled for standard racks, hybrid-cooled for OCP ORv3 environments), integrate the optical modules, write the firmware, qualify the supply chain, and then produce units in Thailand. By the time a competitor could reverse-engineer the design and qualify a supply chain, Celestica is already delivering the next generation.

2. Business Segments

Connectivity and Cloud Solutions (CCS)

CCS is the engine of Celestica's transformation and, as of Q1 2026, generates roughly 80% of total company revenue. It serves two end markets: Communications (networking switches) and Enterprise (compute systems, storage, AI/ML accelerator platforms). The CCS segment is growing at exceptional rates - 76% year-over-year in Q1 2026 - driven almost entirely by hyperscaler demand for AI infrastructure.

The core capability here is what Celestica calls Hardware Platform Solutions (HPS). HPS is not simple assembly; it is the design, development, and manufacturing of complete hardware platforms - switches, compute sleds, rack-scale systems - typically based on merchant silicon from Broadcom or custom silicon from hyperscalers. In Q1 2026, HPS generated roughly $1.7 billion in revenue and represented 42% of Celestica's total company revenue. Three years ago, HPS barely existed. The capability that made it possible was Celestica's investment in design engineering: the company scaled from a small engineering team to over 1,100 design engineers across seven global sites, with nearly 400 dedicated software engineers and a growing portfolio of proprietary IP in thermal management, signal integrity, and firmware.

The Communications end market within CCS is primarily networking switches - the DS5000 series (800G) and DS6000/DS6001 series (1.6T) sold to hyperscalers. These are Celestica-designed and Celestica-manufactured products built around Broadcom Tomahawk ASICs. Revenue from this end market grew 69% in Q1 2026, comfortably exceeding management's guidance of "low 60s" growth. The 800G programs are mature; the 1.6T programs are just beginning to ramp. The company has 10 active 1.6T programs in its pipeline with mass production beginning in H2 2026 with two hyperscalers, and additional programs ramping in 2027.

The Enterprise end market within CCS is primarily AI/ML compute platforms - this is where Celestica manufactures Google's TPU (Tensor Processing Unit) server systems, making Celestica one of the most direct plays on Google's custom silicon ambitions. Management has repeatedly characterized Celestica as Google's "preferred manufacturing partner" for TPU systems, and in Q4 2025 stated that "our partnership with Google has never been stronger or more integrated." This relationship extends beyond current TPU generations: Celestica has visibility into future TPU programs, which means the revenue from this end market is highly recurring so long as Google continues its AI infrastructure buildout. Enterprise revenue grew 101% in Q1 2026 (slightly below the "high-teens" guidance due to component constraints on memory and custom silicon).

CCS competes against Taiwanese ODMs - Quanta, Wistron, Inventec, and Foxconn - that have historically dominated server and switch manufacturing. Celestica's edge is the HPS model: these ODMs are excellent volume manufacturers but they are generally less integrated into the design process and less capable of handling the extreme complexity of frontier AI infrastructure. The tradeoff is that CCS margins (8.6% segment margin in Q1 2026) are lower than pure software businesses but higher than what Celestica would earn as a commodity assembler.

Advanced Technology Solutions (ATS)

ATS generates roughly 20% of total company revenue and covers four end markets: Aerospace and Defense (A&D), Industrial, HealthTech, and Capital Equipment. ATS is a fundamentally different business from CCS - it is characterized by lower volumes, more extreme technical complexity, long qualification cycles, and stringent regulatory requirements. It is not a growth story in the near term; ATS grew roughly flat to slightly positive through 2025 and continued flat in Q1 2026.

The aerospace and defense business within ATS is among the most technically demanding in Celestica's portfolio. Products include avionics systems, satellite components, radar assemblies, and defense electronics for prime contractors like Boeing, Raytheon, and Lockheed Martin. The qualification barriers here are extreme: components must be certified to MIL-SPEC standards, manufacturing processes are audited regularly by the customer and sometimes by government agencies, and design changes require recertification cycles that can take years. This creates powerful moats once qualified: an A&D customer does not switch contract manufacturers casually. The ATS A&D business is going through a portfolio reshaping - management has been deliberately exiting lower-margin programs to improve segment margins, which is why A&D has been a drag on growth even as underlying demand from defense customers is strong.

HealthTech within ATS covers medical device manufacturing - diagnostic equipment, surgical robots, imaging systems, monitoring equipment. This end market requires ISO 13485 certification, FDA audit readiness, and traceability documentation that goes far beyond standard electronics manufacturing. HealthTech was one of the brightest spots in ATS through 2025, with management citing "healthy traction" repeatedly across multiple concalls. The growth in robotic surgery, diagnostic automation, and connected care devices is driving demand for complex, high-reliability medical electronics.

Industrial and Capital Equipment within ATS serve smart energy infrastructure, industrial automation, semiconductor manufacturing equipment, and similar markets. These businesses showed weakness in 2024-2025 due to elevated inventory in the supply chain, particularly in industrial end markets. Capital equipment specifically - where Celestica builds components for semiconductor fab equipment used by the likes of ASML and Applied Materials - is cyclical and tied directly to the semiconductor capex cycle.

The strategic rationale for keeping ATS separate from CCS is threefold: the end markets, the technical requirements, and the economics are all materially different. ATS operates at 5-6% segment margins versus CCS at 8%+ and rising. ATS requires deep regulatory compliance infrastructure that would be wasteful to bolt onto CCS. And the customer relationships are fundamentally different - a HealthTech customer with a 10-year product lifecycle has nothing in common with a hyperscaler re-platforming its networking stack every 18 months.

3. Products and Business Detail

The Hardware Platform Solutions (HPS) Portfolio

Celestica's HPS products are the commercially differentiated heart of the business. They are not commodity assemblies; they are proprietary platforms that Celestica has designed and that it sells (or licenses to manufacture for) large hyperscalers. The product line spans three categories: networking switches, compute systems, and storage platforms, though networking is where the most investment and growth is concentrated.

DS5000 (800G Ethernet Switch): Celestica's current flagship switch platform for AI cluster networking. Built around Broadcom's Tomahawk 5 ASIC, the DS5000 provides 64 OSFP 800GbE ports in a 2U form factor, delivering 51.2 Terabits per second of total switching capacity. The Tomahawk 5 consolidates 128 x 400GbE ports on a single chip, enabling the "flat" low-latency topologies that AI training workloads require. The DS5000 delivers more than 75% power reduction versus the previous generation of 400G switches - a critical differentiator when data centers are constrained by power and cooling. This platform is in mass production and generating the bulk of HPS revenue as of 2025.

DS6000 and DS6001 (1.6T Ethernet Switches): Celestica's next-generation switch platform, announced in October 2025 and available for order as of April 2026. Built around Broadcom's Tomahawk 6 (TH6) silicon, the DS6000 series delivers 102.4 Tbps of non-blocking switching capacity with 64 ports of 1.6TbE connectivity - a full doubling of capacity versus the DS5000. Two form factors: DS6000 is air-cooled in a 3U chassis for standard 19-inch racks; DS6001 uses hybrid (liquid-assisted) cooling in a 2U OCP ORv3 form factor. Mass production is expected to ramp with two hyperscalers in H2 2026, with additional programs ramping through 2027. The company has 10 active 1.6T programs at various stages.

Co-Packaged Optics (CPO) Switch (2027): Perhaps the most significant technology announcement in Celestica's recent history. In Q1 2026, CEO Rob Mionis disclosed a "landmark program award" to design and manufacture a 1.6T CPO Ethernet switch with an unnamed hyperscaler. CPO is a cutting-edge technology that integrates optical components directly with switch ASICs in the same package, eliminating the power loss and signal degradation that comes from separate pluggable optical modules. This is genuinely frontier technology - the first major production-scale deployment of CPO at this level of integration. Mass production is expected to begin in H2 2027. This award validates years of Celestica's investment in optical integration and is the clearest signal yet that its engineering capabilities are advancing ahead of what standard ODMs can offer.

AMD Helios Scale-Up Networking Switch (2026-2027): Announced in March 2026, Celestica is collaborating with AMD to design and manufacture the scale-up networking switch for AMD's "Helios" rack-scale AI platform. Helios is AMD's response to the disaggregated AI infrastructure market - an open-standards architecture based on OCP Open-Rack-Wide (ORW) form factor, integrating AMD's next-generation Instinct MI450 GPUs with high-speed networking using the Ultra Accelerator Link over Ethernet (UALoE) protocol. Celestica is providing the networking switch element, with samples expected by year-end 2026 and customer availability in late 2026. This is significant because it diversifies Celestica's exposure beyond Broadcom-Tomahawk-based platforms and embeds it in AMD's ecosystem.

Google TPU Systems: Celestica is Google's preferred manufacturing partner for its TPU (Tensor Processing Unit) server systems. TPUs are Google's custom AI accelerators, the silicon that powers Google Gemini and other large-scale AI services. Celestica handles board-level assembly (primarily in Mexico, USMCA-compliant) and rack-level integration in the US. The relationship covers current generation TPUs and has joint commitment language that extends into future TPU generations. This is a multi-year, multi-generation program that carries significant revenue and has NCNR (Non-Cancellable, Non-Returnable) commitments.

Digital Native Rack Scale System (2027): Management has repeatedly referenced a "digital native customer" - a hyperscaler-adjacent company - for whom Celestica is designing and manufacturing a complete rack-scale AI system. This is a full system integration program, not just a switch. Mass production is planned for early 2027. Management has described this as adding another leg of diversification to the CCS revenue base beyond the established hyperscaler relationships.

ATS Products and Capabilities

ATS products are broadly composed of highly regulated, safety-critical electronics. Key manufacturing capabilities specific to ATS include:

• Precision machining through subsidiaries Atrenne and Sturgeon: metal enclosures and chassis for defense and space applications, made to thousandths-of-an-inch tolerances
• Clean room manufacturing at multiple sites for medical devices
• Conformal coating, potting, and environmental sealing for aerospace electronics that must survive extreme temperature, vibration, and humidity
• Full Box Build and Systems Integration for complex medical and defense systems
• Repair and after-market services through NCS Global: a profitable services business that handles warranty repair, end-of-life part sourcing, and depot services for installed equipment

Manufacturing Geography

Celestica's manufacturing footprint spans 50 sites across 15 countries, with major concentrations in:

• Thailand (Chonburi campus): Celestica's largest and fastest-growing site for HPS. Over 4,000 employees. Full spectrum of services including printed circuit assembly, box build, regional technology center, and failure analysis laboratory. Multiple new buildings under construction to support 2027 programs. Thailand is strategically important because it sits outside US-China tariff exposure and qualifies under US trade agreements that make its products importable to the US without punitive tariffs.
• Malaysia (Kulim and Johor): Over 1,500 employees in Kulim alone, focused on commercial aerospace (serving A&D customers). Johor serves enterprise, industrial, semiconductor, and healthcare markets.
• Mexico (Monterrey): Established in 1998. Handles board-level assembly for Google TPU systems, which can then be shipped to the US as USMCA-compliant imports, avoiding tariffs. Celestica has used its Mexico footprint as a tariff buffer - customers with US-bound products can route board assembly through Mexico rather than China or Southeast Asia.
• United States (Richardson and Fort Worth, Texas): Expanding for rack-level integration, particularly for US government and defense customers who require ITAR-compliant domestic manufacturing. New HPS Design Center planned in Austin. The Texas expansion is partly driven by customer requirements for US-origin content.
• Other: Sites in the UK, Ireland, Romania, Singapore, Japan, China (reduced footprint as programs shift), and others.

The geographic diversification is not accidental. As US-China tariff tensions escalated from 2018 onward, Celestica systematically moved programs out of China into Thailand, Malaysia, and Mexico. The company's China footprint has been deliberately reduced, and programs that required China-origin manufacturing have largely been transferred or exited. This de-risking provides a structural advantage over Taiwanese ODMs with large China manufacturing bases.

4. Customers

Who Buys and Why

Celestica's customers fall into two distinct populations.

In CCS, the customers are a small number of hyperscalers and cloud-infrastructure companies: Google (explicitly named as "preferred partner"), Microsoft, Meta, Amazon, and what management calls "digital native" companies that are building large-scale AI infrastructure. As of Q1 2026, three customers each represented 10% or more of Celestica's total revenue, at approximately 35%, 15%, and 15% respectively - meaning three companies account for roughly 65% of all revenue. The largest single customer (almost certainly Google given the TPU program disclosures) is 35% of revenue.

These customers buy from Celestica for a specific set of reasons:

1. Design capability: Hyperscalers want the best possible silicon integrated into the best possible platform, at speed. They do not want to wait 24 months for a standard ODM to catch up with their architecture choices. Celestica's 1,100-person design engineering team can co-develop hardware in real time.

2. Supply chain control: Custom silicon (Broadcom Tomahawk, hyperscaler custom ASICs) has lead times of 52+ weeks. Celestica has established allocation agreements with suppliers, NCNR commitments backed by customer purchase orders, and the financial strength to commit capital ahead of demand. A hyperscaler building out a multi-billion-dollar cluster cannot afford supply chain surprises.

3. Geographic footprint: US-China tariff dynamics create strong incentives for hyperscalers to source from non-China manufacturers. Celestica's Thailand, Malaysia, and Mexico footprint is precisely what these customers need. The SemiAnalysis tariff analysis explicitly noted that Google's TPU program with Celestica benefits from USMCA qualification that provides "minimal" tariff impact.

4. Integrated rack delivery: Hyperscalers increasingly want to receive complete, tested racks rather than components. Celestica can receive boards from Thailand, perform rack integration in Texas, and deliver fully tested AI infrastructure to a customer's data center. This end-to-end capability reduces the customer's own operational complexity.

The buying decision in HPS programs typically involves multiple stakeholders: the customer's infrastructure engineering team (who specify the technical requirements), the supply chain team (who evaluate manufacturing capability and financial stability), and increasingly, the customer's AI/ML research teams (who need the hardware to support specific workloads). Sales cycles are long - typically 12-24 months from initial engagement to production award, and programs run for 2-4 years per generation before the next platform.

In ATS, the customer base is more fragmented: aerospace prime contractors, defense agencies, HealthTech OEMs, semiconductor equipment makers. These relationships are even stickier. An A&D customer that has qualified Celestica's manufacturing process to MIL-SPEC standards and incorporated it into their AS9100-certified quality system will not switch manufacturers casually - the recertification cost alone can run into millions of dollars and years of schedule delay. HealthTech OEMs face similar constraints: an ISO 13485-qualified supply base is not interchangeable.

Switching Costs

Switching costs in CCS are substantial but not permanent. The NCNR terms on component purchases mean that if a hyperscaler wanted to move a program to a different manufacturer mid-generation, it would need to assume the component liabilities that Celestica has already committed. The design IP - the PCB layout, thermal characterization, firmware, test vectors - typically belongs to Celestica or is jointly developed. A hyperscaler that moves to a new manufacturer does not automatically inherit this IP; it starts over. The requalification of a new manufacturer for a high-volume networking switch program takes 12-18 months in ideal conditions. Given that networking generations turn over every 18-24 months, this creates effective lock-in for at least one product generation.

The CPO and AMD Helios programs, if they ramp as expected, would extend this lock-in further. Co-developed technology, where Celestica owns meaningful IP in the packaging architecture and optical integration, is materially harder to replicate than a design built entirely on merchant silicon.

In ATS, switching costs are even higher and more formal - regulatory re-qualification is a hard gate, not merely a preference.

5. Competitive Landscape

The EMS Universe

The global electronics manufacturing services industry is large but highly stratified. At the top end by volume are Foxconn (Hon Hai) and its subsidiaries Pegatron and Wistron - companies with hundreds of billions of dollars of revenue that built their dominance assembling iPhones, consumer electronics, and commodity servers at enormous scale in China. They compete on cost and volume. Celestica does not primarily compete with them.

The second tier, where Celestica operates, includes Jabil (approximately $30 billion in revenue), Flex (approximately $26 billion), and Celestica itself (approaching $19 billion in 2026). These companies all position themselves as "high-complexity" EMS providers, but their strategies have diverged meaningfully.

Jabil is the most direct scale competitor. It has deep customer relationships with Apple (which accounts for a substantial share of its revenue), exposure to healthcare, and a growing data center infrastructure business. Jabil's data center business has been growing but starts from a lower base and has not demonstrated the same HPS-design-led differentiation that Celestica has. Jabil's margins in its data center business are lower than Celestica's CCS margins - though direct comparisons are complicated by different accounting treatments and segment definitions.

Flex is Celestica's most similar strategic competitor. Flex has explicitly pivoted toward a "joint design manufacturing" (JDM) model that mirrors Celestica's HPS approach - working closely with hyperscalers on custom hardware rather than waiting for a blueprint. Flex has significant exposure to cloud infrastructure and networking. Flex is a genuine competitive threat in new program competitions. However, Flex has a larger industrial and automotive portfolio that dilutes its data center focus, and Celestica appears to have won more hyperscaler networking programs to date.

Accton Technology and its subsidiary Edgecore Networks are the most direct competitors in the networking switch space. Accton is a Taiwanese company that pioneered the open networking switch market, manufacturing white-box switches based on Broadcom and other merchant silicon for hyperscalers, cloud operators, and enterprises. Accton's strength is deep manufacturing expertise and cost competitiveness. Celestica's edge over Accton is design capability - Celestica can co-develop from the architecture level, whereas Accton is primarily a manufacturer of designs it receives. For commodity switch programs, Accton is a fierce and capable competitor.

Quanta Computer and Wistron have historically dominated hyperscaler server manufacturing and are now expanding into networking and storage. Both have enormous manufacturing scale in Taiwan and China (though China exposure is shrinking due to tariffs). They are genuine threats in the Enterprise (compute) end market within CCS - both can manufacture AI server systems at scale and have long-standing hyperscaler relationships. Celestica's edge in this space is primarily its Google TPU relationship, which is a proprietary program rather than an open competition.

ODM-Direct model: Some hyperscalers are experimenting with designing switches entirely in-house (Meta's Wedge, Microsoft's Sonic) and having ODMs manufacture against those designs. If this trend accelerates, it could commoditize the switch manufacturing business. Celestica's response is to move further up the stack into CPO, into rack-scale systems, and into co-developed technology that is not easily replicated - which is exactly what the AMD Helios and CPO switch programs represent.

Barriers to Entry

Building a credible HPS competitor to Celestica would require: (1) a 1,000+ person design engineering organization with deep expertise in high-speed board design, thermal management, and optical integration; (2) established allocation relationships with Broadcom and other silicon suppliers; (3) qualified manufacturing capacity in tariff-friendly geographies; (4) the financial strength to commit $1 billion+ in annual capex; and (5) proven program delivery history that hyperscalers will accept as evidence of capability. This combination of people, relationships, capital, and track record took Celestica a decade to build. A new entrant would face the same 10-year path.

Where Celestica is exposed: in programs that do not require deep design co-development, purely volume-driven manufacturing can shift to Taiwanese or Southeast Asian ODMs if they can meet quality and delivery standards. The risk is that the commodity portion of the CCS business - particularly if hyperscalers gain design confidence and move to purely ODM-manufactured products - could be contested.

6. Industry

The AI Infrastructure Investment Cycle

Celestica's CCS growth is fundamentally a bet on the AI infrastructure investment cycle. Hyperscalers - Amazon, Google, Microsoft, Meta, Oracle, and increasingly non-US sovereigns - are spending at a rate that was unimaginable five years ago. Annual data center capital expenditure from this group is expected to surpass $1 trillion by 2028, up from roughly $200 billion as recently as 2022. This is not primarily spending on real estate or power - it is spending on compute and networking hardware. Every GPU cluster requires networking infrastructure; every additional rack requires switches, cables, and optical modules. Celestica is in the path of this spending.

Within the broader AI infrastructure market, the networking subsegment is particularly important for Celestica. Management cited (in Q3 2025) that the high-bandwidth Ethernet TAM is projected to reach $50 billion by 2029, with 800G and above segments expected to grow at a 54% CAGR. This is the specific sub-market where Celestica's DS5000/DS6000 switches compete. The transition from 400G to 800G to 1.6T is a hardware refresh cycle that is already underway and that forces customers to re-evaluate their supply chain.

The EMS Market

The global EMS market is approximately $600-650 billion in size as of 2025, growing at a 6-7% CAGR. Asia Pacific is the largest region (~$290 billion), reflecting the concentration of manufacturing capacity in Taiwan, China, and Southeast Asia. For context, Celestica at $12.4 billion in 2025 revenue represents roughly 2% of the total EMS market - but the relevant sub-market (complex AI and cloud infrastructure) is much smaller and growing much faster.

Cyclicality and Demand Patterns

The EMS industry as a whole is cyclical. Consumer electronics demand is highly seasonal and sensitive to economic conditions. Enterprise IT spending follows multi-year cycles tied to upgrade cadences and corporate capex budgets. But Celestica's AI data center business is operating on a different cycle driver: the race between hyperscalers to secure the compute and networking capacity needed for AI model training. This is a structural buildout, not a discretionary upgrade. The hyperscalers themselves have stated publicly that they are supply-constrained for AI compute - they cannot build fast enough - which means the limiting factor on Celestica's growth is manufacturing capacity and component supply, not customer demand.

The ATS segments are cyclical in more traditional ways. Capital equipment (semiconductor fab equipment components) is directly tied to the global semiconductor capex cycle, which peaked in 2022 and is recovering. Industrial was weaker in 2024-2025 due to post-COVID inventory normalization. A&D is subject to government budget cycles and geopolitical events.

Tariff and Trade Policy

The tariff environment has been both a risk and an opportunity for Celestica. As US tariffs on Chinese goods have escalated - with effective tariff rates on many electronics components from China reaching 47%+ as of 2025 - customers with US-bound products have strong incentives to shift manufacturing out of China. Celestica's reduced China footprint and its established capacity in Thailand, Malaysia, and Mexico positions it as a beneficiary of this shift. The company's management disclosed that it has "transferred numerous customer programs, primarily located in China, to countries unaffected by these tariffs including Thailand." The USMCA qualification of its Mexico operations for Google TPU board assembly is a concrete example of tariff optimization delivering competitive advantage.

The risk is that tariff policy is unstable. If tariffs on Southeast Asian goods were to escalate materially, Celestica's Thailand advantage would be reduced. Management has acknowledged this as a risk in filings.

Regulation

Celestica's products must comply with RoHS (Restriction of Hazardous Substances), WEEE (Waste Electrical and Electronic Equipment), and country-specific environmental regulations. ATS aerospace and defense products must comply with ITAR (International Traffic in Arms Regulations), which restricts who can manufacture and handle defense components. ITAR compliance requires domestic (US) manufacturing for certain programs, which is partly why Celestica is expanding in Texas.

7. Growth Triggers

(Sourced exclusively from the four concall transcripts: Q2 2025, Q3 2025, Q4 2025, Q1 2026)

• 1.6T switch mass production ramp with two hyperscalers beginning H2 2026. Ten active 1.6T programs in the pipeline; management guided that two would reach mass production in the second half of 2026, with additional programs in 2027. First disclosed in Q3 2025, reaffirmed in Q4 2025, updated with specific customer count in Q1 2026. (Q3 2025, Q4 2025, Q1 2026 concalls)

"We have 10 active 1.6T programs, and we are expecting mass production to begin in the second half of 2026 with two hyperscalers, with additional programs ramping in 2027."

• Digital native rack scale system beginning mass production early 2027. An unnamed "digital native" customer has awarded Celestica a program to design and manufacture a complete rack-scale AI system. This is distinct from the switch programs - it is a full system integration. Mass production planned for early 2027 and expected to contribute meaningfully to CCS revenue growth beyond the networking switch base. (Q4 2025, Q1 2026 concalls)

• AMD Helios scale-up networking switch, samples by year-end 2026. Announced in March 2026, this collaboration with AMD is for the networking switch component of AMD's Helios rack-scale AI platform. Development is underway, with customer samples expected by year-end 2026 and commercial availability in late 2026. This diversifies Celestica's switch business beyond Broadcom-only platforms. (Q1 2026 concall)

• Co-Packaged Optics (CPO) 1.6T Ethernet switch program, mass production H2 2027. Described by CEO Mionis as a "landmark" first major production-scale CPO deployment. Awarded by an unnamed hyperscaler, this program validates years of optical integration investment and is expected to begin mass production in H2 2027. Management called CPO the "next phase" of switch architecture and positioned this award as evidence of Celestica's capability to lead the industry transition. (Q1 2026 concall)

"We have been awarded a program for the design and manufacturing of a CPO Ethernet switch with a hyperscaler customer. The switch, optimized for AI scale out networks, will utilize 1.6 Terabit switch silicon, leveraging co-packaged optical interconnects, and liquid cooling technology."

• Texas and Thailand capacity expansion coming online for 2027-2028 programs. The $1 billion 2026 capex plan (rising to $1.5 billion in 2027) includes construction of new buildings at the Thailand campus, expansion of the Dallas-Fort Worth manufacturing footprint, and a new HPS Design Center in Austin, Texas. This capacity is specifically tied to program-level business cases with customer demand commitments. (Q3 2025, Q4 2025, Q1 2026 concalls)

• New HPS Design Centers in Austin, Texas and Taiwan. Design engineering is the foundation of the HPS competitive moat. The Austin and Taiwan design centers will increase Celestica's engineering headcount and geographic coverage, supporting faster co-development with US-based hyperscalers and Taiwan-based silicon partners. (Q4 2025 concall)

• R&D spend increasing 50%+ in 2025 and 2026. Management flagged over 50% increase in R&D spending in 2025, with similar or greater increases planned in 2026. This is the investment underlying the CPO, AMD Helios, and liquid cooling capabilities that will define the next generation of programs. (Q3 2025 concall)

• ATS HealthTech acceleration. Within ATS, the HealthTech sub-segment was described as the primary bright spot, with "healthy traction" repeatedly cited across multiple calls. The ramp of robotic surgery, diagnostic automation, and connected monitoring devices is driving demand for complex medical electronics. (Q2 2025, Q3 2025 concalls)

• Supply chain capacity alignment providing revenue floor. Management stated in Q1 2026 that customer demand exceeds stated guidance, and that the primary growth constraint is supply chain capacity and component availability, not customer willingness to buy. As silicon and component suppliers add capacity, Celestica expects incremental revenue unlock above its stated outlook. (Q1 2026 concall)

"Awarded backlog and opportunity pipeline with both existing and new customers are the strongest they have ever been during my tenure."

8. Key Risks

Customer Concentration

The most significant structural risk in Celestica's business is that three customers represent approximately 65% of total revenue, with one customer alone at 35%. This is not a generic statement about diversification - it is a specific exposure to the capital allocation decisions of a small number of technology companies. If the largest customer were to significantly reduce its AI infrastructure spend, shift a major program to a different manufacturer, or go through an internal reorganization that disrupts procurement, Celestica could face a revenue decline of potentially catastrophic proportions.

Management has attempted to mitigate this through the "preferred partner" language on the Google TPU program and by pointing to NCNR commitments that extend visibility through 2028. But NCNR terms protect against sudden program cancellation, not against a customer shifting future programs away after current programs complete. The Google TPU program is generation-by-generation: today's commitment covers today's TPU. Tomorrow's commitment is dependent on tomorrow's decision. There have been press reports suggesting Google is reconsidering how it allocates AI server assembly work across suppliers, and while management has firmly denied any weakening of the relationship, the risk is real and structural.

Supply Chain Constraints Limiting Growth

As of Q1 2026, management explicitly stated that "we are experiencing more component shortages now than 90 days ago." Custom silicon (Broadcom Tomahawk series, hyperscaler custom ASICs), memory, 40-layer-plus PCBs, optical modules, and power components are all on allocation with lead times extending beyond a year for some items. Celestica's stated revenue guidance assumes specific supply deliveries on specific schedules. If those deliveries slip - whether due to Broadcom capacity constraints, memory supply disruptions, or optical module production issues - revenue guidance cannot be met even if customer demand remains intact.

The mechanism: Celestica commits to customer delivery schedules based on expected component availability. Customers commit to NCNR orders based on those schedules. If components arrive late, Celestica misses customer commitments, and the downstream demand does not simply shift to the next quarter - some customers may turn to alternative suppliers for portions of their demand.

Execution Risk on Hyper-Growth Capex

Celestica is spending $1 billion on capital expenditures in 2026 and guiding $1.5 billion in 2027. This is 6% and approximately 7-8% of expected revenue respectively - extraordinarily high for an EMS company. These investments are tied to specific customer programs that are "expected" to ramp in 2027 and 2028. If program ramps are delayed, if customers decide to insource, or if a technology shift makes a planned facility's capability obsolete, Celestica could find itself with expensive physical capacity and no revenue to fill it. A $1.5 billion misallocated capex commitment is not a minor error for a company generating $500 million in free cash flow.

Technology Transition Risk

The switch from pluggable optical modules to co-packaged optics (CPO) is a genuine architectural transition with uncertain timing and winners. If CPO adoption accelerates faster than expected, equipment built on current pluggable architectures may become obsolete. Celestica is investing in CPO (the 2027 program award is evidence of this), but the risk is that competitors who move earlier or more aggressively could establish CPO leadership before Celestica scales. Conversely, if CPO adoption is delayed due to manufacturing challenges, the investment Celestica has made in CPO capability sits idle.

Tariff Policy Volatility

Celestica's competitive advantage in non-China manufacturing is real, but it is partly a function of current US tariff policy. Any significant change - a US-China tariff deal that dramatically reduces China tariffs, or new tariffs on Southeast Asian goods - could change the relative economics. The Thailand facility is Celestica's largest and fastest-growing site; if tariffs were to apply to Thailand-origin goods at rates comparable to current China tariffs, the competitive advantage disappears. Management has disclosed this as a risk in filings, and it is one that the company cannot control.

Margin Compression from Input Cost Inflation

Management noted in Q1 2026 that "input costs that are going up materially, whether it be memory or silicon" represent a margin headwind. Celestica's contract structures with hyperscalers typically allow cost pass-throughs for major component cost changes, but there is a lag between when costs increase and when they are recovered in customer pricing. In a period of aggressive growth with significant component allocation commitments, Celestica is exposed to short-term margin compression from component price inflation.

9. Walk the Talk

Concall dates used: Q2 2025 (July 29, 2025), Q3 2025 (October 27-28, 2025), Q4 2025 (January 29, 2026), Q1 2026 (April 28, 2026).

The most recent concall (Q1 2026, April 28, 2026) is 13 days ago - well within the 90-day window.

Q2 2025: The Cautious Raise

On July 29, 2025, Celestica reported Q2 results that beat guidance and raised its full-year outlook to $11.55 billion in revenue and $5.50 in adjusted EPS. The raise was notable for what was in the commentary around it: CFO Mandeep Chawla explicitly said, "Our demand outlook is higher than the $11,550, but we're taking into account challenges." This was a signal that management was deliberately guiding below what they believed the demand environment warranted, cushioning for supply chain and tariff uncertainties.

At this point, management was flagging 800G at parity with 400G volumes - meaning the technology generation transition was accelerating - and guiding for a recovery in enterprise AIML compute starting in Q3.

Q3 2025: Confident Acceleration

By October 27-28, 2025, the confidence in management's tone had shifted noticeably. CEO Rob Mionis stated: "We are navigating the most rapid period of change in our company's history." The company raised full-year guidance to $12.2 billion and $5.90 EPS, announced it would hold an Investor and Analyst Day, and revealed that HPS revenue was running at 79% growth and represented 44% of total company revenue. The Q3 Analyst Day was used to introduce long-term guidance for 2026: $16 billion revenue and $8.20 EPS.

Management's Q2 guidance for Q3 2025 called for revenue of $2.875-$3.125 billion and EPS of $1.37-$1.53. Actual Q3 results: revenue of $3.19 billion (at the top of the range), EPS of $1.58 (above the top of the range). Management had promised acceleration and delivered it.

Q4 2025: Record Margins and the $1 Billion Capex Commitment

On January 29, 2026, Celestica reported Q4 results that again beat guidance: revenue of $3.65 billion (well above the $3.325-$3.575 billion range guided), adjusted EPS of $1.89 (above the $1.65-$1.81 range). Full-year 2025 free cash flow came in at $458 million, above the $425 million updated guidance from Q3.

The more important disclosure was the 2026 outlook: $17 billion revenue, $8.75 EPS, and $1 billion in capex. Management described the 7.5% operating margin as a "floor" rather than a ceiling. The Google partnership language was notably strong: CEO Mionis said, "our partnership with Google has never been stronger or more integrated." A 1.6T switch award with a third hyperscaler was announced. A new HPS Design Center in Austin was disclosed.

Q4 2025 demonstrated a pattern that had been visible since at least Q2: management guides conservatively, beats, and raises. In Q4, they beat the high end of revenue guidance by $75 million and beat the high end of EPS guidance by $0.08.

Q1 2026: The Full-Year Revision That Nobody Expected This Early

The Q1 2026 results on April 28, 2026 were, by any measure, remarkable. Revenue of $4.05 billion exceeded the guidance midpoint of $4.0 billion and beat the high end of the $3.85-$4.15 billion range. Adjusted EPS of $2.16 beat the high end of the $1.95-$2.15 range. Management raised full-year 2026 guidance from $17 billion to $19 billion - a $2 billion raise in a single quarter - and raised full-year EPS guidance from $8.75 to $10.15. The CPO switch program and the AMD Helios collaboration were both disclosed for the first time. Management stated that 2027 revenue would grow "significantly more than $6.5 billion" over 2026.

"Awarded backlog and opportunity pipeline with both existing and new customers are the strongest they have ever been during my tenure."

This is management at peak credibility: four consecutive quarters of beating the high end of guidance, multiple upward revisions, and new program disclosures that expand the long-term opportunity.

Assessment: Celestica's management team has, across these four quarters, demonstrated a consistent pattern of conservative guidance followed by beats and raises. This is not accidental - CFO Chawla explicitly said in Q2 that demand was running ahead of guidance. The conservatism appears to be deliberate, driven by genuine supply chain uncertainty (they genuinely do not know exactly when components will arrive) and a preference for positive surprise over missed guidance. Management has delivered on every major operational promise made across these four concalls: 800G at parity with 400G, enterprise AIML recovery, HPS revenue growth, margin expansion, and free cash flow generation. New program wins (third hyperscaler 1.6T, CPO, AMD Helios) have each been delivered and announced when management said they would be.

This is management that does what they say. The only mild exception is that some supply-related quarterly guidance has bumped against component availability (Enterprise grew 101% in Q1 2026 but was "slightly below" high-teens expectations due to component constraints) - which reflects genuine supply chain complexity, not management forecasting dishonesty.

10. Shareholder Friendliness Index

Dividends: Celestica pays no dividend. It never has as a public company. Given the current phase of hypergrowth requiring $1-$1.5 billion in annual capex, a dividend initiation is not expected or appropriate. The free cash flow the business generates is being reinvested into capacity and technology at extraordinarily high returns on invested capital (adjusted ROIC reached 50% in Q1 2026). For now, the most rational use of capital is reinvestment, not distribution.

Buybacks and dilution: Celestica has maintained a meaningful share repurchase program. The company repurchased 1.36 million shares in 2025 at a total cost of approximately $151 million. This is modest relative to the $458 million in free cash flow generated in 2025, reflecting management's decision to prioritize balance sheet strength and capex investment. Shares outstanding have been declining modestly - the repurchases are more than offsetting executive equity compensation dilution, so the share count is shrinking. Net debt leverage is low at 0.6-0.7x trailing EBITDA, and the credit facility was recently expanded to $1.75 billion, providing liquidity for the capex program without equity dilution.

Verdict: Neutral - actively shrinking the share count via buybacks but reinvesting the majority of cash flow into a high-return growth cycle; no dividend is warranted at this stage.

11. Insider Activities

Source: SEC Form 4 filings (Celestica is NYSE-listed; SEDI is the Canadian system for TSX, but the SEC Form 4 system is the primary disclosure mechanism for US-listed securities). Data from MarketBeat and SEC filings, supplemented by Stocktitan SEC filing reports.

Note: SEDI.ca access was attempted but the system does not provide easily searchable results for TSX-listed companies. Given Celestica's dual-listing on NYSE as the primary trading venue, SEC Form 4 is the appropriate and used primary source. Canadian SEDI filings may contain additional data not captured here.

Buys - reading the signal: The single buy in this period was by Director Laurette Koellner, who purchased 6,000 shares on October 30, 2025, in four separate transactions totaling $2.05 million. This is a meaningful purchase for a board director - not merely a token gesture. The timing is notable: the purchase was made the same week as the Q3 2025 earnings call and Investor and Analyst Day, suggesting Koellner had heard the long-term 2027 outlook and chose to put personal capital behind it. Koellner has since been appointed Lead Independent Director, giving her purchase extra significance as a signal of board-level conviction. This is a genuinely positive signal, though it is from a single director and is one data point.

Sells - working out the why: The concentrated selling in early February 2026 across three senior executives (Jason Phillips, Yann L'Etienvre, Todd Cooper) requires careful analysis. The aggregate value is approximately $87.7 million across roughly three days - a large sum. However, context is essential:

• The timing (early February 2026, just after Q4 2025 results on January 29) is consistent with a pre-scheduled trading window that opens after earnings blackout periods lift.
• Yann L'Etienvre and Todd Cooper's sales include tax-related disposals consistent with RSU vesting events (the same-day $300K-level sales on February 4 are textbook tax-withholding sales following a larger RSU vest).
• The larger sales by L'Etienvre (~86K shares) and Cooper (~89K shares) on February 2 are also likely tied to RSU vesting - this is the standard pattern for executive compensation where RSUs vest annually and executives sell shares to diversify. The scale of these sales is meaningful but not disproportionate given that the stock had appreciated enormously in the prior 12 months.
• Jason Phillips's sale of 120,000 shares on February 5-6 is the largest by dollar value ($36.9 million). This has been disclosed as a planned program (Form 144 pre-filing noted). At $37 million, this is substantial - roughly three to four times an average executive's annual compensation package - and is categorized as planned diversification. There is a late-filed amended Form 3 on behalf of Phillips from February 2026, suggesting an administrative filing correction.

Net assessment: Celestica insiders are net sellers by a wide margin over the past 12 months - approximately $87.7 million in sales versus $2.05 million in purchases. The selling is concentrated in a single two-day window, is largely tied to RSU vesting events (standard executive compensation mechanics), and followed an earnings period where the stock had significantly appreciated. None of the sales show unusual timing (ahead of negative news), and subsequent to the selling, Celestica beat Q1 2026 guidance significantly and raised full-year guidance dramatically. The Koellner open-market purchase is a mild bullish signal. The overall insider picture is neutral to mildly cautious - the magnitude of executive selling relative to buying is large, but the mechanics are consistent with planned, schedule-driven diversification rather than conviction selling. The critical distinguishing factor is that these sales happened before the stock continued to appreciate; insiders who sold in February 2026 have missed the Q1 2026 earnings beat and guidance raise that followed two months later.

12. Scenarios

Bull Case

The bull case is essentially that every growth trigger fires on schedule and the technology transitions Celestica has bet on prove transformative. In this scenario, the 1.6T switch programs ramp as planned in H2 2026 with two hyperscalers, and the additional programs add up in 2027 to create a networking business that dwarfs the 800G generation at its peak. The digital native rack scale system reaches mass production in early 2027 and becomes a third major revenue leg alongside the switch business and Google TPU systems. The AMD Helios collaboration succeeds commercially, opening the door to AMD's entire customer ecosystem and reducing dependence on any single hyperscaler. The CPO switch program, entering mass production in H2 2027, establishes Celestica as the first at-scale CPO manufacturer - a capability advantage that puts it years ahead of any competitor trying to catch up.

Simultaneously, ATS recovers as the semiconductor capex cycle turns and industrial inventory destocking ends. The capital equipment sub-segment rebounds as TSMC, Samsung, and Intel ramp new fabs, creating demand for the precision semiconductor equipment components that ATS manufactures. HealthTech continues its steady growth as robotic surgery and connected diagnostics proliferate. ATS margins, which management has been deliberately improving by exiting low-margin programs, approach 7% - meaningfully higher than today's 5-6%.

The balance sheet stays clean despite aggressive capex because the NCNR-backed program revenue is so high-confidence that free cash flow more than offsets the capital spend. Celestica becomes the unambiguous infrastructure manufacturing partner of choice for the AI supercycle - a company whose revenue doubles in three years and whose margin structure continues to expand as fixed costs are spread across a rapidly growing revenue base.

Base Case

The base case assumes management delivers roughly what it has guided - which, given the consistent track record of beating guidance, may itself be conservative. The 1.6T switch ramp happens roughly on schedule in H2 2026, driving material CCS acceleration. The digital native rack scale system reaches production in early 2027 with some schedule slippage (new product introductions routinely slip 1-2 quarters). The AMD Helios collaboration produces samples on time but commercial volume builds gradually into 2027-2028 rather than immediately. The CPO program develops on the disclosed 2027 schedule and proves Celestica's technical capability, but large-scale CPO volumes are a 2028-2029 story.

ATS grows low-to-mid single digits as HealthTech strength is partially offset by ongoing A&D portfolio reshaping and modest recovery in industrial end markets. The $1 billion capex in 2026 and $1.5 billion in 2027 are fully absorbed by the revenue programs they are tied to, with no major stranded assets. Free cash flow grows as revenue scales and margin expands modestly. Customer concentration remains elevated but the third hyperscaler 1.6T award and the digital native program provide meaningful diversification relative to today's 35%/15%/15% structure. Management continues to grow the business while maintaining conservative guidance and delivering positive surprises.

Bear Case

The bear case is grounded in three specific risks materializing simultaneously. First, Google - the 35% customer - decides after the current TPU generation to bring more assembly in-house or shift a portion of the work to Foxconn or Quanta. A 20-point reduction in Google-related revenue at current scale would be a multi-billion-dollar impact that no amount of 1.6T switch growth could immediately offset. The press reports of Google reconsidering assembly allocation are the early signal; the bear case is those reports proving prescient.

Second, the component shortage that is already intensifying as of Q1 2026 worsens. Custom silicon lead times stretch to 18 months, optical components remain constrained, and the 1.6T ramp schedules slip from H2 2026 to 2027. Customer demand does not disappear, but Celestica cannot fulfill it. The revenue that looked locked in through NCNR terms gets deferred, and the aggressive capex committed for 2026-2027 becomes a cash burn with no near-term return.

Third, a tariff shock hits Thailand. US policymakers, in an attempt to force semiconductor manufacturing back to domestic soil, impose significant tariffs on electronics manufactured in Southeast Asia - a scenario that would have seemed implausible two years ago but is entirely consistent with the current geopolitical trajectory. Celestica's Thailand campus, which is the anchor of its 2027 manufacturing plan, would suddenly be subject to tariffs that make its products more expensive for US-bound customers than equivalent products from tariff-exempt geographies. Repricing to pass tariffs through to customers takes time; in the meantime, margins compress and some business moves to competitors with domestic US capacity.

In the bear case, Celestica is still a real business with genuine capabilities - but it is a business that overbuilt capacity, faces a temporarily weakened customer base, and is burning through free cash flow servicing its capex program. The bear case is a painful multi-year reset rather than an existential scenario.

13. Further Reading

• Tariff Armageddon? GPU Loopholes, Mexico Supply Chain Shift, Wafer Fab Equipment Vulnerabilities, Optical Module Pricing Surge, Datacenter Equipment - SemiAnalysis [paid] - Celestica mentioned as Google's tariff-advantaged TPU assembly partner in Mexico; article is primarily paywalled.

• 100,000 H100 Clusters: Power, Network Topology, Ethernet vs InfiniBand, Reliability, Failures, Checkpointing - SemiAnalysis [paid] - Celestica named as a leading ODM for hyperscale networking switches based on Broadcom ASICs; article is primarily paywalled.

No coverage found from Stratechery or MBI Deep Dives at the time this report was generated.

Sources:

• Celestica Q1 2026 Earnings Transcript - The Motley Fool
• Celestica Q1 2026 Financial Results - Corporate IR
• Celestica Q4 2025 Earnings Transcript - The Motley Fool
• Celestica Q3 2025 Earnings - Investing.com Transcript
• Celestica Q2 2025 Earnings - Investing.com Transcript
• Celestica Q3 2025 Financial Results - Celestica Corporate
• Celestica Q4 2025 Financial Results - Celestica Corporate
• AMD and Celestica Helios Rack-Scale AI Platform - AMD Newsroom
• Celestica DS6000 1.6T Switch Launch
• Celestica DS6000-Series Available for Order
• Celestica Insider Trading - MarketBeat
• Celestica Director Koellner Buys $2M in Shares - Investing.com
• Celestica Wikipedia
• SemiAnalysis - Tariff Armageddon
• SemiAnalysis - 100K H100 Clusters
• Celestica Annual Reports - IR Page
• Celestica Products and Services
• EMS Market Size - Fortune Business Insights