Earnings Call Highlights: Celestica (CLS) | Q3 FY2026 — Revenue Up 83%, a $40B Call for 2027, 1.6T Overtakes 800G Next Year, First Rack-Scale CPO Ramps in Late 2027

1 day ago

10 min read

1. Executive Takeaway

First, Celestica is no longer a “contract manufacturer”; it is a hyperscaler design partner for networking and racks. Its in-house HPS (Hardware Platform Solutions) business did about $2.3B this quarter, up 70%, or 40% of company revenue and roughly half of CCS. Position matters even more: in data center Ethernet switches at 800G and above, Celestica holds 31% revenue share and 44% port share; in custom switches it holds 58% revenue share, more than double the next competitor.

Second, growth is shifting from a single networking engine to two engines: networking plus custom ASIC compute. Compute and storage grew 242% this quarter and is projected at $5.6B for 2026, up 170%. In 2027, the multi-generation Google TPU program and the production ramp of OpenAI’s in-house Jalapeño accelerator will push compute’s share of CCS higher still.

Third, CPO has moved from showcase to real orders with a real timeline. At OCP, Celestica launched its first CPO switch, the DS6005 (Broadcom Tomahawk 6 Davisson), and it has won a hyperscaler co-design and manufacturing program for a 1.6T rack-scale CPO system, ramping in late 2027. Its engineering lead was equally clear, though: the 1.6T generation will still be mostly pluggable optics.

2. Revenue and Financials

Core numbers for the quarter (GAAP basis unless marked as adjusted):

Metric	Q3 FY2026	Q3 FY2025	Change
Revenue	$5.83B	$3.19B	+82.5% YoY, +24.1% QoQ
Gross margin (GAAP)	11.0%	13.0%	−2.0pp YoY
Adjusted gross margin	11.2%	11.7%	−0.5pp YoY
Operating income (GAAP)	$460.0M (7.9% margin)	$325.0M	+41.5% YoY
Adjusted operating margin	8.5%	7.6%	+0.9pp YoY, a company record
Net income (GAAP)	$397.2M	$267.8M	+48.3% YoY
Diluted EPS (GAAP)	$3.26	$2.31	+41.1% YoY
Adjusted EPS	$3.30	$1.58	+109% YoY

Segments and end markets:

Item	Q3 FY2026 revenue	YoY	Segment margin
CCS (Connectivity & Cloud Solutions)	$4.87B	102%	8.9% (+60bp YoY)
of which HPS	~$2.3B	70%	—
of which Networking (formerly Communications)	—	68%	—
of which Compute and Storage (formerly Enterprise)	—	242%	—
ATS (Advanced Technology Solutions)	$957M	23%	6.6% (+110bp YoY)

▲ Celestica FY2026 Q3 key financials (GAAP): revenue $5.83B, up 82.5% YoY; gross margin 11.0%; operating margin 7.9%; net income $397.2M; diluted EPS $3.26; Q4 revenue guidance $6.275–6.575B (Image source: Simple Tech Trend original; data from the company’s SEC 8-K and Q3 2026 earnings release)

What the numbers mean:

Revenue beat, and margins were not diluted. The gross margin decline is a mix effect, as lower-margin compute racks grow fast inside CCS, yet operating margin hit a record. Operating leverage from scale is outweighing the mix drag, which is rare in EMS, where surging revenue usually squeezes margins.
Mind the gap between GAAP and adjusted figures. A year ago GAAP gross margin was 13.0% versus 11.7% adjusted, so most of the 2.0pp GAAP decline comes from adjustment items in the base period rather than a weaker core business. For the underlying trend, the adjusted −0.5pp is the better read.
Both end markets beat the company’s own outlook by a wide margin: networking was guided to about 60% growth and delivered 68%; compute and storage was guided to about 190% and delivered 242%.
The balance sheet is gearing up for 2027. Inventory reached $4.5B, up $1.1B in one quarter; capex was $274M (4.7% of revenue, versus 1.2% a year ago). The August offering of about 11.1M shares raised $3.45B, leaving about $4B in cash and $3.2B in net cash at quarter end.

3. Networking: 1.6T Overtakes 800G Next Year, and Scale-Up Ethernet Is the New Battleground

Networking revenue is projected at $11.8B for 2026, up 66%. This year’s momentum comes from the 800G switch ramp plus the first 1.6T programs entering mass production. CCS President Steven Dorwart gave a clear marker: 1.6T revenue will exceed 800G in 2027, and 800G will keep growing rather than being displaced. In past networking transitions, demand from the old and new generations has stacked rather than replaced each other.

The mix of 1.6T programs shows where the market is going: Celestica has won 10 1.6T programs, of which 4 are scale-up and 6 are scale-out. Scale-up interconnect inside GPU racks used to be locked into closed architectures such as NVLink; Ethernet is now moving into scale-up. Citing third-party forecasts, the company expects scale-up Ethernet switch spending to grow from almost zero in 2026 to $55B by 2030, about a third of the data center Ethernet market.

Scale-up fabrics have 10x the port attach rate of scale-out and up to 100x the bandwidth needs of front-end networks. This is the slice Celestica wants most.

The flagship is the DS6010: a scale-up switch co-designed for AMD’s Helios rack architecture. It packs two Broadcom Tomahawk 6 chips into a 1OU chassis as the industry’s first 204.8Tb switch, connecting 72 AMD Instinct MI455X GPUs in a single rack over UALink over Ethernet, with direct-to-chip liquid cooling and a cabled backplane, and is expected to begin ramping by year-end. We broke down the Helios rack architecture and its scale-up design in Inside AMD Helios: Engineering AI at Rack Scale; for why Broadcom is betting on Ethernet scale-up, see Broadcom: Ethernet Scale-up, Copper and OCI.

On competition, management’s message is “fewer players.” Group President Mandeep Chawla said Celestica faces fewer competitors on 800G and 1.6T programs than in earlier generations, because signal integrity, thermal and optical integration get harder with every generation. Dorwart also acknowledged that success invites competition and that share is negotiated with customers continuously; a customer may move an older product away so Celestica’s capacity stays focused on the next generation. Switch-to-system partnerships such as Jabil with Arista are chasing the same market.

4. CPO: From the DS6005 to Rack-Scale Production in Late 2027

For optical readers, the most valuable part of the call was how concretely Celestica laid out CPO timing and division of labor.

Product: The DS6005, launched at OCP in October, is Celestica’s first co-packaged optics (CPO) design. Built on Broadcom Tomahawk 6 Davisson, it integrates high-density silicon photonics engines, liquid cooling and field-replaceable external laser modules in a 1.5OU form factor. The company says it is among the first teams in the world to bring up a functional Tomahawk 6 Davisson CPO platform, which is now in testing.

Orders: Celestica has won a hyperscaler co-design and manufacturing program for a 1.6T rack-scale CPO networking solution, with production ramping in late 2027, and other customers are discussing proofs of concept.

Manufacturing is the real challenge. Building CPO switches is nothing like building copper-connected ones: it takes expanded clean rooms, robotic assembly and inspection, fiber insertion with sub-micron tolerances in places, and multi-stage engineering validation testing. Celestica is automating its L6–L10 tray processes and L11 rack assembly, including highly repetitive work such as cabling and fiber insertion.

But do not read this as “CPO replaces pluggables next year.” Asked directly whether 1.6T would still be mostly pluggable, CCS engineering lead Gavin Cato said a lot of pluggables will remain at 1.6T, alongside efforts to prove out CPO and NPO capabilities; CPO will play a more prominent role at 3.2T and 400G SerDes. The company’s long-term view is that pluggable and integrated optics will coexist, each in the use cases that suit it, with new pluggable formats such as XPO advancing in parallel.

1.6T is the pluggable generation; 3.2T is the CPO generation. That call comes from a system builder doing both, which makes it more neutral than any single optics vendor’s view.

5. Compute: Google TPU and OpenAI Jalapeño, the Rack Business Behind Custom ASICs

Compute and storage revenue is projected at $5.6B for 2026, up 170%. The two 2027 drivers have been named: the multi-generation TPU partnership with Google and the production ramp of OpenAI’s Jalapeño platform, both weighted to the second half.

Jalapeño is Celestica’s most comprehensive and complex project to date, more than two years in development. It is OpenAI’s first-generation in-house inference accelerator, co-developed with Broadcom. The dual-rack system hosts 128 Jalapeño accelerators per unit on a cabled backplane; its scale-up network uses 1.6T Tomahawk 6 switches, connecting up to 2,048 accelerators across multiple racks and tiers in the scale-up domain, and runs a custom SONiC co-developed with Celestica. OpenAI’s published tests show higher inference throughput per watt than the comparison systems across three open models. We covered the network design of this architecture in Celestica: Designing the Next AI Fabric.

Why it matters: citing forecasts, the company expects the custom accelerator market to exceed $430B by 2030, growing more than 7x, over twice the growth rate of merchant GPUs and about a third of the AI accelerator market. Every new custom chip needs its own rack architecture, and often its own network fabric, so the more customized the silicon, the more valuable the vendor that can design from networking all the way up to the full rack.

6. Outlook: Q4 Guidance, a 2026 Raise and $40B for 2027

Q4 FY2026 guidance:

Revenue of $6.275–6.575B, up about 76% at the midpoint
Adjusted EPS of $3.37–3.57, up about 84% at the midpoint
Adjusted operating margin of about 8.7% (+100bp YoY)
End markets: networking up about 65%, compute and storage up about 160%, ATS up in the low 20s percent

Full-year 2026 (raised): revenue of $21.0B (from $20.5B, up 69%), adjusted EPS of $11.55 (from $11.30, up 91%), and adjusted operating margin unchanged at 8.4%; free cash flow cut to $460M (from $600M) to fund working capital for the 2027 ramp.

Full-year 2027 (first outlook):

Item	2027 outlook
Revenue	$40B (+90% YoY)
CCS / ATS	About +110% / about +10%
Adjusted operating margin	9.0% (+60bp)
Adjusted EPS	$21.50 (+86% YoY)
Capex	$1.5–2.0B
Adjusted tax rate	About 21%
Gross margin	Low 11s, depending on the networking/compute mix
Free cash flow	Net outflow; targeting positive in 2028

Reading the language: what is confirmed and what is expected

The firmest line: “2028 demand forecasts based solely on awarded programs are meaningfully higher than the 2027 revenue outlook.” This is not a target but a customer demand forecast built on signed programs, the highest level of certainty on the call.
On the $40B, the CFO said “high confidence”, noting that ramp risk, component availability and execution risk are already factored in, with “opportunities to exceed.” He also stressed that the year is clearly second-half weighted.
A return to positive free cash flow in 2028 is described as a “target”, a lower level of certainty that depends on the pace of the revenue ramp.
Mandeep said Celestica is already discussing programs for 2029 and even 2030, and some programs have been awarded before the silicon even exists. AI infrastructure supply chains have moved from placing orders to reserving capacity.

R&D: annual R&D spend has more than tripled since 2023 and is expected to exceed $300M in 2027. HPS engineering headcount grows from 1,600 to over 2,000 next year, mainly in Austin, Taiwan and Silicon Valley, and the team is engaged on 60–70 programs.

7. Supply Chain and Customer Signals

Customer concentration is high: three customers each exceeded 10% of revenue this quarter, at 35%, 17% and 14%, for 66% combined. Dorwart’s analogy: it is fine to keep your eggs in one basket if the basket is very strong and you watch it closely.
The bottleneck is not demand but land, power and geopolitics. Management said constraints are most likely to come from land and power. The company is adding hundreds of MW across its network: expanding its flagship Thailand campus for the major 2027–2028 ramps, adding capacity at multiple Texas sites for US-based L11 integration of liquid-cooled AI racks, and investing in Monterrey, Mexico, Japan and Malaysia (the last two for networking programs). New capacity in Asia can come online in as little as 12 months; the US takes longer.
What it means for the optical module supply chain: with 1.6T switches ramping next year and 1.6T staying mostly pluggable, demand for 1.6T optical modules and their upstream parts (EML, silicon photonics chips, DSPs) will scale with switch shipments. This is the other end of the same demand chain behind the doubling of 1.6T per-module content discussed in Earnings Call Highlights: MaxLinear (MXL) | Q3 FY2026.
What it means for CPO lasers and fiber: the DS6005 uses field-replaceable external laser modules, so once rack-scale CPO production starts in late 2027, external laser sources and high-precision fiber connectivity are the parts certain to be pulled. For the latest on CPO back-end capacity and prepayments, see CPO Capacity Goes 15x, but the Buyers Are the Most Cash-Strapped | STT Weekly 2026 W43.
Signal integrity above 800G is near its limit: moving to 400G SerDes tightens electrical loss budgets and shortens copper reach, requiring simultaneous upgrades in PCB laminates, surface-mount packaging and direct-attach copper. Rack power is heading to 500kW and beyond with a shift to 800V DC distribution, and liquid cooling will extend from GPUs and switch silicon to optics and memory.

8. Risks and Counterarguments

The cash flow pressure is real. Adding nearly $20B of revenue in 2027 means working capital and $1.5–2.0B of capex will turn the company to net cash outflow. If customer deployments slip, inventory and capacity become a burden.
Customer concentration. The top three customers make up two-thirds of revenue, so any hyperscaler capex adjustment or reallocation is amplified. Analysts also asked about the financial capacity of emerging AI labs; management said it has no concerns at this time and looks at third-party funding or contractual backstops when needed.
Second-half weighting. Both the OpenAI and TPU programs ramp in 2H27; any component or yield issue in the first half compresses the full-year schedule.
Gross margin depends on mix. A rising share of compute racks pressures gross margin; the “low 11s” for 2027 assumes networking keeps a high share.
CPO production yields are unproven. The industry still has no real volume-yield data for sub-micron fiber alignment and high-density silicon photonics engines.

9. Summary

This quarter proved one thing: value in AI infrastructure is moving from the single chip to the entire rack, and the vendor that can integrate networking, compute, liquid cooling and optics into a manufacturable rack earns three to five years of visibility. The $40B for 2027 is not built on optimism but on programs already won, some signed before the silicon was even built.

The most useful signal for the optical industry is timing: 1.6T is the pluggable generation, late 2027 is when the first rack-scale CPO goes into production, and 3.2T is when CPO truly scales.

Three metrics to track next quarter:

1.6T share of networking revenue: whether it passes 800G in 2027 as management expects.
DS6010 (AMD Helios) and Jalapeño ramp progress: whether scale-up Ethernet ramps from zero on schedule.
Working capital and free cash flow: whether inventory growth matches the revenue ramp and the path to positive FCF in 2028 is clear.

The question for 2027 is not whether Celestica can win orders, but whether it can deliver the ones it has already won under three bottlenecks: power, land and cash.

This article is for technology and industry trend analysis only and does not constitute investment advice.