Celestica’s ($CLS) Q2 Beat Is Real. Here’s What It Actually Says About the Rest of the AI Hardware Chain.
Deep-dive, July 27, 2026. Data and market conditions as of the July 27 after-hours print and July 27 market close. Historical snapshot, not auto-updated. Educational only, not financial advice.
Celestica beat on both lines tonight, above the high end of its own guidance, and raised its full-year outlook for what sounds like the second time this year. That part is straightforward. The more useful exercise is figuring out which parts of the print are genuinely new information, which number is quietly inflated by something that has nothing to do with operations, and which other names in the AI-hardware chain this actually tells you something about.
The headline, verified against the actual release
Revenue: $4.70B, +62% YoY, above the high end of its $4.15-4.45B guidance range. [CO]
Adjusted EPS: $2.54 vs $1.39 a year ago, above the high end of $2.14-2.34 guidance. [CO]
Adjusted operating margin: 8.2%, a new company high (vs 7.4% a year ago). [CO]
FY26 guidance raised: revenue to $20.5B (from $19.0B), adjusted EPS to $11.30 (from $10.15), free cash flow to $600M (from $500M). [CO]
Management said 2027 revenue growth will accelerate beyond the 65% rate now expected for 2026, with EPS growing faster than revenue. That’s a forward commitment, not hedge language. [CO]
Where the growth actually lives
Celestica has two reportable segments, not three, a point that got muddled in the wire coverage tonight:
Connectivity & Cloud Solutions (CCS): $3.81B, +84% YoY, 81% of revenue, margin 8.7% (vs 8.3%). [CO] This splits into two end-markets: Communications (networking) and Enterprise (servers/storage, branded internally as “Hardware Platform Solutions”).
Enterprise/HPS: ~$1.9B, +58% YoY. [CO]
Communications (networking): the implied remainder, roughly $1.9B. Not separately disclosed, but since the blended segment grew 84% while the disclosed Enterprise half only grew 58%, the networking side had to grow well past 84% to pull the average there. [CALC] That’s the 800G/1.6T switch business, and it’s the disproportionate engine here, not just a strong-but-average contributor.
Advanced Technology Solutions (ATS): $0.89B, +8% YoY, margin 6.3% (vs 5.3%). [CO] The non-AI aerospace/defense/industrial/healthtech book. Smaller, slower, but margin actually expanded more in percentage points than CCS did.
The number that’s quietly inflated (and why the framework flagged it)
GAAP EPS came in at $3.17, higher than the $2.54 adjusted figure, backwards from the usual pattern where adjustments typically add back costs and land above GAAP. The cause: a $104.1M non-cash gain on Celestica’s total return swap (a derivative tied to its own share price), worth $0.90/share pre-tax, baked into the GAAP number and explicitly excluded from adjusted. [CO] It isn’t new either, it was $0.84/share a year ago too, so it’s a recurring mark-to-market swing that happens to be large right now because the stock has run hard. Use the adjusted figures for the real comparison; the GAAP headline is the wrong number to get excited about.
The capex story underneath the growth
Capital expenditure jumped roughly 8x year-over-year, $263.8M this quarter versus $32.5M a year ago. [CO] Celestica is now spending like an infrastructure builder, not the classic asset-light contract manufacturer it used to be.
Free cash flow still grew ($147.1M vs $119.9M) because operating cash flow grew even faster, but working capital ballooned in both directions: inventory up about 55% since year-end, payables up about 105%, receivables up about 27%. [CO] Normal shape for a company scaling this fast, but a lot more exposed to a demand air pocket than it was a year ago.
No share buybacks this quarter, and none guided for Q3 or the rest of 2026. [CO] Every dollar is going into the buildout right now.
Second-degree sweep of the source
Two things worth naming that point outside Celestica itself:
The compute side of the business (Enterprise/HPS, +58%) is GPU-server assembly, a distinct read-through from the networking side, pointing at GPU suppliers (NVIDIA, AMD) and memory (Micron, HBM demand), not just switch-silicon and optics names.
Management names “competition risk from evolving AI technologies, including lower-cost/open-source AI models” as an explicit risk factor in its own filing language. [CO] That’s a direct tie to the cheap-open-source-model fear that hit chip stocks earlier this month. Tonight’s guidance raise is management’s own evidence that fear hasn’t shown up in real order volume yet.
Who else this actually points to
Not everyone in the “AI hardware” bucket benefits the same way, or for the same reason.
Direct, mechanism-linked (a beat here is real evidence about them too): - Broadcom (AVGO): Celestica’s 800G switch runs on Broadcom’s Tomahawk 5 chip, a named, confirmed relationship. [3P] - Marvell (MRVL): a named partner, mechanism less precisely disclosed than Broadcom’s. - NVIDIA / AMD, and Micron: the GPU-server (Enterprise/HPS) side of tonight’s beat, a read-through I’d initially undersold by treating the whole print as networking-only. - Coherent (COHR) / Lumentum (LITE): every switch port needs a transceiver, and Celestica flagged 1.6T switch programs entering mass production with two hyperscaler customers in the second half of 2026, a real leading indicator for the 800G-to-1.6T upgrade cycle that’s the core of both companies’ own thesis.
Peer comp, not supply-chain link: - Fabrinet (FN): same contract-assembly business model, different product (optical modules instead of servers/switches). If hyperscaler hardware demand is this strong broadly, that’s a read-through for Fabrinet’s own volumes.
Sentiment tailwind, not a mechanical read-through: - Dell, HPE: these are OEMs (they design, brand, and sell their own hardware), not suppliers into or peers of Celestica’s ODM/contract-assembly model. Dell in particular already had its own, much larger, independent AI-server story months ago ($16.1B AI-server revenue in a single quarter, back in May). Tonight’s print doesn’t predict anything about either company’s numbers the way it does for Fabrinet. It’s one more data point that hyperscaler capex is real, which trims a little tail risk sitting under the whole group.
Comps, normalized
HPE is the cheapest by a wide margin (12.4x forward P/E vs. 23-29x for the rest) despite the highest operating margin (13.3%, nearly double CLS’s). That’s a real disconnect worth a second look — either the market doesn’t trust HPE’s margin durability post-Juniper, or it’s mispriced relative to peers on this specific print.
Margin ranks exactly by business model: HPE (OEM, 13.3%) > FN (ODM, 10.7%) > CLS (ODM, 8.2%) > Dell (not disclosed here, but Dell’s blended margins run thinner than HPE’s historically). That’s the OEM-vs-ODM distinction from a few messages ago, showing up numerically, not just conceptually.
FN and CLS trade at almost identical forward P/E (28.7x vs 28.0x) despite FN having the better margin and cleaner “pure-play DCI/optics” story (its own DCI revenue +90% YoY this quarter) — that’s the real, tight peer-comp read: they’re priced almost the same, and FN’s fundamentals look at least as good.
Dell’s growth number isn’t apples-to-apples with the others — the +757% is the AI-server sub-line, not total company (+total beat by $8.4B is the closest like-for-like figure I have). I don’t have Dell’s op margin from tonight’s searches; flagging that gap rather than guessing at it.
The counter-case
Valuation isn’t obviously cheap. CLS trades near 28x forward earnings for a contract manufacturer, a multiple usually reserved for higher-margin, branded businesses. If the growth rate decelerates even to a merely-good rate, that multiple compresses.
The 8x capex jump cuts both ways. It’s confidence today. It’s also fixed cost committed against demand that hasn’t been delivered yet, exactly the kind of operating leverage that turns into a margin problem fast if a customer pushes out a program.
Working capital is stretched. Inventory and payables both surged double digits in six months. Fine while growth continues; a real source of cash-flow risk if it doesn’t.
The Dell/HPE comparison is a genuine limitation, not just a footnote. Because they’re OEMs and Celestica is an ODM, there’s no clean read-through in either direction, meaning tonight’s print tells you less about the health of the branded server market than it does about the hyperscaler-direct one specifically.
The single fact that would break this thesis: a hyperscaler customer publicly pushing out or canceling a 1.6T switch or AI-server program. Nothing in tonight’s release suggests that’s happening, but it’s the one data point that would matter more than anything above.
What would change this
Tomorrow’s 8am ET call should sharpen the exact Communications-versus-Enterprise split inside CCS, and may name (or decline to name) specific hyperscaler customers. That’s the next real evidence point, not a guess about where the stock goes from here.
Educational content, not financial advice. For informational purposes only. Do your own research.