Celestica (NYSE: CLS) Deep Research

Publication date: 2026-06-30 Method: supply-chain bottleneck research framework Nature: research support only, not a trading instruction

Bottom Line First

Celestica should no longer be understood as a traditional EMS contractor. What it is really monetizing is the design-and-manufacturing bottleneck created as AI data-center networking moves from 400G/800G toward 1.6T, and from standalone racks toward rack-scale, scale-out, and scale-up architectures. This bottleneck is not “assembling boxes for customers.” It is the ability to integrate Broadcom/NVIDIA/AMD/hyperscaler switch silicon, optical modules, power, liquid cooling, OCP rack form factors, supply chains, and production yield into high-performance switching platforms that can be deployed at massive scale.

FY2026 Q1 already put that change into the financial statements: revenue was $4.05B, up 53% YoY; CCS revenue was $3.24B, up 76% YoY and equal to 80% of total revenue; HPS revenue was about $1.7B, up 63% YoY and equal to 42% of total revenue. Management raised 2026 revenue guidance from $17.0B to $19.0B, and adjusted EPS guidance from $8.75 to $10.15.

More importantly, Celestica’s growth has moved beyond “800G leadership” into three new tracks:

1. 1.6T switches: DS6000/DS6001 based on Broadcom Tomahawk 6, with 64 1.6TbE ports, 102.4Tbps switching capacity, 224G SerDes, LPO/LRO optics, and hybrid cooling support.
2. CPO switches: in FY2026 Q1, Celestica won a hyperscaler design-and-manufacturing project for a CPO Ethernet switch using 1.6T switch silicon, co-packaged optical interconnects, and liquid cooling, with ramp expected in 2027.
3. AMD Helios scale-up networking: Celestica is working with AMD on scale-up networking switches for the Helios rack-scale AI platform, supporting MI450 GPUs and UALoE architecture, with the platform planned for customer availability by late 2026.

But this is not a risk-free “AI autopilot” story. Celestica’s bottleneck is design manufacturing and supply-chain integration, not GPUs, switch ASICs, optical modules, HBM, or the network standards themselves. It is pulled by customer roadmaps and locked into customer concentration: in FY2026 Q1, the top three customers represented 35%, 15%, and 15% of revenue, while the top ten represented 84%. The most accurate research frame is therefore: AI networking ODM bottlenecks are being re-rated, but the bottleneck rent will be squeezed by hyperscalers, switch-chip vendors, and competing ODMs.

As of the latest U.S. trading session on 2026-06-29, CLS traded around $343.25, with market cap around $39.7B and P/E around 41.6x. This is no longer a “cheap EMS” valuation. The market is pricing whether Celestica can turn the 2026 revenue baseline of $19B into another step-up in 2027.

1. Market Narrative to System Change

Market narrative: Celestica is an AI data-center infrastructure supplier, driven by 800G/1.6T switching, AMD Helios, Google TPU production, and hyperscaler capex.

What actually changed in the system:

1. The bottleneck in AI clusters has expanded from GPUs to networking. The larger training and inference clusters become, the more GPU utilization is constrained by latency, congestion, bandwidth, topology, and thermal design. Switches are no longer data-center accessories; they are the throughput base of the AI factory.
2. Ethernet is entering AI back-end and scale-out networks. InfiniBand remains strong, but hyperscalers keep pushing Ethernet AI fabrics for cost, vendor diversity, and open ecosystems. Celestica’s 800G/1.6T HPS platforms sit directly in this layer.
3. ODM has moved from manufacturing to platform design plus supply chain plus volume production. 800G/1.6T switches require high-speed PCBs, 224G SerDes signal integrity, thermal design, SONiC software, OCP form factors, optical-module adaptation, and global manufacturing coordination. Pure EMS capability is not enough; HPS platform capability matters.
4. Rack-scale AI pulls customer demand from individual devices into full rack and network platforms. Architectures like AMD Helios show that AI systems are moving from server procurement toward rack-scale platform procurement. Celestica’s position extends from switch ODM toward scale-up networking design partner.

In one sentence: AI server competition is becoming a combined contest of chips, networks, racks, cooling, and supply chains. Celestica’s value is that it is one of the few companies able to engineer, manufacture, and deliver high-performance networking hardware quickly for hyperscalers.

2. Required Components, Supply-Chain Layers, and Bottlenecks

The four points that truly constrain Celestica’s revenue elasticity:

1. 1.6T switch platform volume production. DS6000/DS6001 have moved from product announcement into order and introduction stages. The 2026 H2 to 2027 ramp will decide whether HPS growth can continue.
2. The stickiness of hyperscaler custom design wins. 800G/1.6T switches, CPO switches, AI/ML compute, TPU systems, and Helios scale-up switches can bind lifecycle and capacity more strongly than ordinary EMS programs once introduced.
3. The speed of manufacturing-network expansion. 2026 capex is planned around $1B, including Dallas-Fort Worth, multiple new Thailand buildings, Mexico/Japan, and U.S. and Asia HPS design centers. If expansion is slow, revenue recognition is slow; if expansion is too fast, depreciation pressure rises when demand moves.
4. Capacity allocation under customer concentration. The largest customer represents 35% of revenue, and the top three together represent 65%. Customers can provide long-cycle programs, but they can also demand pricing, capacity relocation, and second-source structures.

3. Company Evidence

Strong Evidence

• FY2026 Q1 results: revenue of $4.047B, up 53% YoY; GAAP operating margin of 6.7%; adjusted operating margin of 8.0%, a new milestone; GAAP EPS of $1.83; adjusted EPS of $2.16.
• Raised 2026 guidance: Q2 FY2026 revenue guidance of $4.15B-$4.45B; full-year 2026 revenue guidance raised from $17.0B to $19.0B; adjusted EPS raised from $8.75 to $10.15; adjusted operating margin guidance of 8.1%.
• CCS/HPS is the main growth engine: FY2026 Q1 CCS revenue of $3.241B, up 76% YoY; Communications revenue of $2.411B, up 69% YoY; Enterprise revenue of $830M, up 101% YoY; HPS revenue around $1.7B, up 63% YoY and equal to 42% of total revenue.
• Customer concentration is very high: in FY2026 Q1, the top ten customers represented 84% of revenue; the three largest customers represented 35%, 15%, and 15%, all in CCS.
• 2025 full-year baseline: 2025 revenue of $12.39B, up 28% YoY; GAAP EPS of $7.16; adjusted EPS of $6.05; free cash flow of $458.3M.
• 1.6T product landing: in October 2025, Celestica launched DS6000/DS6001, 64 x 1.6TbE, Broadcom Tomahawk 6, 102.4Tbps switching capacity, 224G SerDes, LPO/LRO optics, with DS6001 supporting OCP ORv3 hybrid cooling.
• CPO hyperscaler design win: FY2026 Q1 disclosed a CPO Ethernet switch design-and-manufacturing win with a hyperscaler, targeting AI scale-out networks and using 1.6T switch silicon, co-packaged optical interconnects, and liquid cooling, with expected 2027 ramp.
• AMD Helios collaboration: on 2026-03-16, AMD and Celestica announced collaboration on the Helios rack-scale AI platform. Celestica is responsible for R&D, design, and manufacturing of scale-up networking switches, based on OCP ORW, supporting MI450 GPUs and UALoE, with Helios planned for customer availability by late 2026.
• Google TPU production expansion: alongside 2025 annual results, the company said it would expand in the U.S. to support next-generation AI infrastructure, including Google TPU systems, while continuing to support Google TPU and leading networking technologies in Southeast Asia.

Medium Evidence

• Dell’Oro share validation: Celestica cited Dell’Oro data showing more than 1.6M 800G ports shipped in 2025 Q1 and market-share leadership awards for 2024 Ethernet Switch - AI Networks and High-Speed Networks of at least 800G.
• 1.6T industry deployment cadence: 2026 is expected to be the first year of 1.6Tbps switch volume deployment, with the 1.6T ramp potentially faster than 800G and exceeding 5M ports within one to two years.
• Third 1.6T hyperscaler project: 2025 Q4 results disclosed a third hyperscaler customer design-and-manufacturing program for a 1.6T switching platform, expected to ramp in 2027.
• Liquidity supports expansion: in FY2026 Q1, the senior secured credit facility was expanded to about $2.5B, with the revolver increased from $750M to $1.75B, matching the larger business scale.

Weak Evidence and Items to Verify

• Public evidence on the exact identity of the largest customer is incomplete. The market generally points to hyperscalers, but company disclosures provide customer concentration, not a full list. No single-customer revenue share should be treated as confirmed unless disclosed.
• 2027 revenue potential is still mostly based on management language such as “outlook for 2027 continues to strengthen,” “new program wins,” and “improved visibility.” That is a real signal, but not formal numerical guidance.
• Whether the CPO switch project becomes a large revenue line depends on customer architecture choices and CPO deployment timing. The confirmed facts are the design win and expected 2027 ramp, not high-margin exclusivity.

4. What the Market May Be Missing

1. The core change at CLS is ODM-ization, not an EMS cycle recovery. From 2025 to 2026, growth is mainly driven by CCS/HPS. HPS design, platform, and supply-chain capabilities give it more negotiating power than traditional contract manufacturing. But that also means competitors are not ordinary EMS peers; they include Accton, NVIDIA’s own systems, Arista’s ecosystem, and large manufacturing platforms such as Foxconn, Flex, and Jabil.
2. An 8% operating margin is high for EMS but thin for an AI hardware narrative. The market likes revenue growth, but CLS is not NVIDIA, Broadcom, or an optical-chip vendor. FY2026 Q1 adjusted operating margin was 8.0%, and CCS segment margin was 8.6%. If customers pressure price or mix worsens, valuation sensitivity will be high.
3. Customer concentration cuts both ways. The largest customer is 35%, and the top three are 65%. That explains why revenue can jump quickly, but it also explains why one customer’s inventory, technology transition, in-house platform move, or second-source strategy can create large quarterly volatility.
4. 1.6T is not a simple replacement for 800G; it revalidates the supply chain. 224G SerDes, LPO/LRO optics, thermal design, liquid cooling/OCP racks, AI routing, and SONiC support all increase yield and delivery complexity. Celestica’s opportunity is here, and so is its execution risk.
5. Helios is an important entry into the anti-NVIDIA rack-scale ecosystem. Celestica is not a direct AMD GPU beneficiary, but it sits in the AMD scale-up networking switch layer. If Helios becomes an important hyperscaler route for diversifying away from NVIDIA, CLS gains a higher strategic position. If AMD rack-scale adoption disappoints, this option moves out in time.
6. Cash flow and capex must be read together. 2025 free cash flow was $458M, while 2026 capex is planned around $1B. Management says operating cash flow can support it, but in a fast-expansion phase, inventory, receivables, customer advances, and tariff recoveries all affect real cash quality.
7. TRS fair-value adjustments can disturb GAAP profit. 2025 GAAP EPS included a positive total return swap impact, while FY2026 Q1 had a negative one. Operating quality should be separated from financial/accounting items.

5. Bear Case: What Would Prove This View Wrong

Ranked by what is easiest to observe first:

1. Q2 FY2026 revenue below $4.15B, or adjusted operating margin below 8%. That would suggest Q1’s beat-and-raise may have included short-term pull-forward or mix effects.
2. 1.6T ramp does not show up in HPS/Communications growth in 2026 H2. The 1.6T bottleneck logic would weaken.
3. The largest customer rises above 40% of revenue. Near-term revenue could look stronger, but bargaining power and volatility risk would become more concentrated.
4. CCS segment margin falls from 8.6% to below 7.5%. That would show AI networking ODM rent being competed or negotiated away.
5. OCS/CPO/liquid-cooling projects slip beyond 2028. The high-end networking upgrade cadence would be below market expectations.
6. AMD Helios or the open UALoE ecosystem advances slowly, while cloud customers continue to concentrate on NVIDIA’s closed ecosystem. Celestica’s non-NVIDIA rack-scale option would be discounted.
7. Capex is executed too quickly while demand is revised down. 2026-2027 expansion would create depreciation and utilization pressure, and the EMS model’s operating leverage would reverse.
8. Competing ODMs win more second-source share in 1.6T or CPO switches. Celestica’s market share and margin would be repriced.

On the other side, the bullish view strengthens if Q2 beats the high end of guidance, full-year 2026 guidance is raised again, HPS quarterly revenue exceeds $2B, 1.6T projects enter production in 2026 H2 with stronger 2027 visibility, the CPO switch customer count expands, Helios ships on schedule in late 2026 and wins hyperscaler deployment, and CCS margin stays above 8.5%-9%.

6. Next Checks