Celestica

IThe business

Celestica is a Toronto company that makes things other companies design. For most of its history that meant electronics assembly: circuit boards and boxes for computer, telecom, aerospace and medical makers, at a margin of two or three cents on the dollar. In the past three years a different business grew inside it. The big cloud companies, which the filings call hyperscalers, and the AI labs began asking it to help design the network switches, servers and racks they put in their data centers, and then to build them in volume. Revenue was $5.6B in 2021 and is guided to $21.0B for 2026 R (reported).⁶⁴

Fiscal years end on 31 December. The company moved from IFRS to US GAAP and from foreign-issuer to US domestic-filer forms in 2025, so its reports for 2024 and 2025 are 10-Ks, and I use their US GAAP numbers for 2023 to 2025 together with older IFRS ones.⁵¹²¹³ Where the two differ, I say so. The third quarter reported on 26 October, just before this page, and the company held an investor day the next morning, so the numbers here are current.⁴⁸

What it sells

The company has two segments. Connectivity and Cloud Solutions, CCS, is 84% of revenue now. Advanced Technology Solutions, ATS, is the older business: parts and systems for aerospace and defense, industrial, health technology and semiconductor equipment makers.³

• Networking, $3.26B in the September quarter, up 68%: the Ethernet switches that connect AI chips inside a data center. This used to be called Communications. Management plans for $11.8B in 2026 E (estimate).³⁸
• Compute and storage, $1.61B, up 242%: servers and storage, which now means racks built around custom AI chips. This used to be called Enterprise. The plan is $5.6B in 2026, of which $4.0B is servers.³⁸
• ATS, $0.96B, up 23%: the older, steadier business, with a 6.6% segment margin. The plan is $3.6B in 2026 and about 10% growth in 2027.³⁸

A subset of CCS that management calls Hardware Platform Solutions, HPS, is where the design work sits. It is the part where Celestica designs the product with the customer, then builds it. HPS was about $2.3B in the September quarter, 40% of total revenue, up 70%. It was 21% of revenue in 2023 and 41% in 2025.³⁵ Management says it will reach about $8.5B in 2026, up from $1.7B in 2023.⁸

Who pays

Revenue from cloud and AI customers is most of the company now. At the Investor Day, management put hyperscaler revenue at $2.9B in 2023, $4.8B in 2024, $7.1B in 2025 and $14.4B for 2026, 83% of CCS.⁸ The 10-K says its CCS customers include Amazon, Meta, Google, Ciena, Dell, HPE and IBM, and its ATS customers include Applied Materials, Lam Research and Honeywell.⁵ It never says which customer is which share.

What it does say matters more. In the September quarter three customers, all in CCS, were 35%, 17% and 14% of revenue, together 66%. The ten largest were 85%.³ A year earlier the largest was 30% and the ten largest 80%. In 2023 the largest was 22% and the ten largest 64%.³⁵ The company’s growth is mostly one customer’s growth, and that customer’s share is rising. I cannot map the 35% to a name, and I will not guess.

Why AI pulled it in

Training and running AI models takes thousands of chips that must exchange data constantly. The chips are the most expensive part, so the switches that link them decide how many are working at any moment. So the biggest buyers want the switches fast, in their own specification and in large numbers. Celestica’s pitch is that it can work on the design alongside the customer, so the same company that designs the switch has already worked out how to build it in volume. Management calls this Joint Design and Manufacturing, JDM. The idea is that the manufacturer is on the program from the proof of concept, so it is hard to replace after.⁸

The products match where AI networks are going. Management shows 800 gigabit switches growing, 1.6 terabit switches ramping and the next step, 3.2 terabits, coming. It describes three kinds of network: scale-up, between chips in a rack; scale-out, between racks; and scale-across, between buildings. Celestica says most of its deployments are for back-end AI networks, the ones that link the AI chips, and that it is moving into scale-up, with a switch for AMD’s Helios racks.⁸ It says a hyperscaler has awarded it a design for a co-packaged optics switch, in which the lasers sit next to the chip, with production expected to begin in late 2027.⁸ In servers, it is the rack partner for OpenAI’s custom chip, called Jalapeño, built with Broadcom, with 128 chips per rack and a production ramp planned for 2027.⁸ Those are the programs behind a plan that nearly doubles revenue again in 2027.

The market data management cites is large. Dell’Oro forecasts data center Ethernet networking growing from $44B in 2026 to $173B in 2030, and 650 Group sees custom AI chips growing from $61B to $434B. Global data center capital spending is forecast to go from $1.2T this year to $3.1T in 2030, two thirds of it by the four largest US cloud companies and the AI labs.⁸ These are the company’s chosen third-party forecasts, not mine. I can’t judge them. They are the demand the 2027 outlook rests on.

The outlook, and how it moved

The guidance path is the story of the stock. At the end of January 2026 the outlook for 2026 revenue was $17.0B, up from $16.0B in October 2025. It was raised to $19.0B in April, $20.5B in July and $21.0B on 26 October.⁷¹⁵⁴ Adjusted earnings per share for 2026 went from $8.20 to $8.75, $10.15, $11.30 and $11.55 R (reported).⁷¹⁵⁴ Q3 revenue of $5.83B beat the top of guidance, $5.55B, by 5%, and the Q4 guide is $6.275B to $6.575B.⁴ Then came 2027: revenue of $40.0B, up 90%, a record adjusted operating margin of 9.0% and adjusted earnings of $21.50 a share, with capital spending of $1.5B to $2.0B.⁴ Management says customers’ demand forecasts for 2028 are well above the 2027 outlook, but gives no number.⁴

Two things in that release are less cheerful. The 2026 free cash flow outlook was cut, from $600M to $460M, because of working capital spent ahead of growth, and the company expects “a net use of cash in 2027”.⁴³ The stock rose 12% on 27 October, 8% on 28 October and then lost 17% in the next three sessions, to $249.22.¹ The page that follows is mostly about why a company guiding to $21.50 a share trades at 11.6 times it.

IIThe moat

A useful test of protection is what a company is paid for the thing it makes. Arista and Celestica sell into the same data centers. Arista’s gross margin in its latest quarter was 62.9%. Celestica’s was 11.0% C (calculated).³¹⁶ Arista keeps about six times as much of each sales dollar as gross profit. That gap is how much of the value Celestica leaves with its buyers.

The margin has been rising, and that matters. Gross margin was between 6.5% and 8.8% in every year from 2016 to 2022, and operating margin between 1.6% and 3.6%. By the latest twelve months they were 11.5% and 8.3% C (calculated).⁶⁵³ The reason is mix. In 2023 CCS had a 6.2% segment margin and HPS was 21% of revenue. In the September quarter CCS earned 8.9% and HPS was 40%.⁵³ Programs that Celestica designs earn more than programs it only builds. The question is whether that premium lasts, because the buyers set the price.

What the filings say about the contract

Celestica’s 10-K says its master supply agreements do not typically promise a level of business or fixed prices, that it bids program by program, that some agreements require specific price reductions over their term, and that a majority require the customer to buy parts Celestica bought to meet the customer’s forecast.⁵ The last point protects it when a forecast is cut. The first points say that revenue is renewed one program at a time, at prices the buyer pushes down. Real protection stops that. The filing describes the opposite.

Evidence it is widening

Margins are the evidence. The company’s adjusted operating margin went from 5.5% in 2023 to 6.5% in 2024, 7.5% in 2025 and 8.5% in the September quarter, and management guides to 9.0% for 2027.⁴⁷⁸ Adjusted return on invested capital went from 21% in 2023 to 27% in 2024, 37% in 2025 and 57% in 2026 to date.⁸ That is what a firm earns when its capacity is full and the work is complicated. It is also what any manufacturer earns in the year when demand beats supply.

The second evidence is selection. The company says a hyperscaler chose it for a co-packaged optics switch, and that OpenAI and Broadcom chose it for a rack-scale program.⁸ Awards like these suggest that at least some of the largest buyers see Celestica as an engineering partner. I have only the company’s own account of them.

Evidence it is narrowing

The first sign is gross margin itself. It fell from 13.0% in the September 2025 quarter to 11.0% a year later, and the company blames mix.³ Part of the change is accounting, because the 2025 quarter included a $48.5M gain on the company’s total return swap, which I explain in the Financials section.³ Adjusted gross margin, which removes that gain, was 11.7% then and 11.2% now.⁴ So the underlying margin fell, and it fell in the quarter revenue rose 83%. Growth with a falling gross margin is what a company sees when its fastest customers also bargain hardest.

The second sign is the competitor list. The 10-K names six electronics makers, three Taiwanese design makers and two networking brands as rivals.⁵ Hon Hai, the world’s largest electronics maker, earned a 6.12% gross margin and a 3.75% operating margin in its June quarter, on revenue of NT$2.53 trillion.¹⁷ Celestica’s margins are about twice those. That is partly mix and partly Hon Hai’s scale in lower-margin consumer devices. It also shows where margins in this trade go when many large builders compete for the same programs. I cover Hon Hai on my Foxconn page.¹⁸

So I would call the protection real but thin, and rented. It lasts as long as each program does, and each program is a bid. It has protected a rising margin for four years. It has not yet been tested by a year of falling volume.

IIIThe competitive advantage

Celestica competes with three kinds of company. The first is the other electronics makers, such as Jabil, Flex and Sanmina. The second is the Taiwanese design makers, such as Quanta, Wiwynn and Accton, which sell finished servers and switches to the same buyers, and Hon Hai, which does both jobs. The third is the networking brands, Arista and Cisco, which sell the same switches as products rather than as a build-to-order service.⁵ The 10-K says that in data center switching customers decide between the custom route and the off-the-shelf one on cost and other grounds.⁵

Where Celestica wins

It wins on the speed of a ramp. Hyperscaler programs go from design to production in months, and the company says it is the launch partner on the newest platforms, with production in Thailand, the United States and Mexico.⁸ Management cites 58% of custom Ethernet switch revenue, 44% of 800 gigabit-and-faster Ethernet ports and 31% of that revenue, from Dell’Oro and 650 Group data for the twelve months to June 2026.⁸ A chart on the same slide, for the top four US hyperscalers and AI labs, shows Celestica’s switch revenue at those buyers rising faster than that of a US switch brand, a Taiwanese design maker, a US chip seller and a second US brand.⁸ These are the company’s figures and charts, in its own investor presentation, and I have not seen the underlying data.

It also wins on diversification of supply. Many buyers want a second source in addition to Taiwan, and about 75% of Celestica’s 2025 revenue was produced in Asia, with about 20% in North America.⁵ The investment plan, with Texas, Mexico and Thailand expansion, adds to that.⁸

Where it loses

It loses if the design and the money stay with someone else. In many of these programs the chip designer, such as Broadcom, or the buyer, such as a hyperscaler, owns the key intellectual property, and Celestica contributes the board, the thermal design and the build. Management shows the design effort it is adding, including R&D of more than $200M for 2026 and $300M or more targeted for 2027, but even that is under 1% of the planned revenue.⁸ Arista spent 13.7% of revenue on R&D.¹⁶

It also loses to its own customers. The 10-K lists potential insourcing among the risks: some buyers evaluate making the products themselves, and in some cases Celestica’s offering competes with a customer’s own hardware.⁵ If one customer is 35% of revenue, then a decision at that customer to dual-source or bring work in-house could take a large part of a year’s revenue within a program cycle.

And it loses when the market gets crowded. Hon Hai guides to more than doubling its AI rack shipments in 2026, and to high double-digit growth in cloud and networking in the third quarter.¹⁷ The same buyers can award new programs to either company. Celestica’s smaller base gives it faster growth, and the contest stays open.

IVManagement

Mr. Mionis took over in 2015 when the company was a low-margin assembler with $6B of revenue, and led the shift into design-led work and into the hyperscaler programs. The record is good by any measure. Revenue went from $6.0B in 2016 to a guided $21.0B, and operating margin from 2.6% to 8.3% C (calculated).⁶⁴ In July 2026 the company named a new head of its largest segment, and in September a new finance chief. Both changes are internal. Both come just before the largest year in the company’s history, and the year management has said will use cash.¹⁹²⁰³

How they are paid, and what that pays them to do

Mr. Mionis was paid $17.4M for 2025: $1.15M of salary, $12.4M of stock awards, a $3.45M cash incentive and $0.38M of other pay.⁹ The proxy says 92% of his target pay was at risk.⁹ The annual cash incentive, called the CTI, pays on three corporate measures: GAAP revenue (40%), adjusted operating margin (40%) and free cash flow (20%).⁹ Revenue counts as much as margin, so a dollar of low-margin, pass-through revenue still moves the score. Only the last 20% notices cash.

The stock awards are half the story. Performance share units vest after three years on adjusted EPS against a target range, modified up or down 30% by shareholder return against a technology hardware index.⁹ The 2023 units, measured to the end of 2025, paid at 200% of target, the maximum.⁹ Adjusted EPS leaves out stock pay itself, gains on the company’s total return swap, and the amortization of acquired intangibles.⁴ So executives are paid on a measure that excludes their own stock pay. A payout at twice the target means the target was set well below what the business delivered. The board describes the target as demanding but reachable.⁹

Shareholders approved the pay. Say-on-pay passed on 19 May 2026 with 68.5M votes for and 2.5M against, 96% of those cast either way.²¹ I note the vote, and I note that most of the owners are large funds whose benchmark is the share price.

How much of their own money is in

This is the strongest part of the case for management. Mr. Mionis held 1,010,091 common shares at 27 March 2026, and his shares and share units counted for ownership rules were worth $411.8M at the 2025 year-end price of $295.61, 358 times his salary against a required five.⁹ Mr. Chawla’s were $60.9M and Mr. Cooper’s $85.2M.⁹ Together all directors and officers held 1.10% of the shares.⁹ Five of the nine director nominees held no common shares at that date, though directors take part of their fees in deferred share units.⁹ The board’s lead independent director, Laurette Koellner, bought 6,000 shares on the open market on 30 October 2025, at about $340, for $2.05M. It is the only open-market purchase in the past year C (calculated).¹⁰

They are also selling. Form 4s show open-market sales by eight officers and directors of about $254M over the 12 months to 1 October 2026 C (calculated).¹⁰ Mr. Mionis sold 345,949 shares for $130.7M between 15 June and 23 September, in 109 small sales, at between about $360 and $414 a share. The forms mark them as under a preset 10b5-1 plan.¹⁰ Mr. Phillips, who is retiring, sold $36.9M in February, Mr. Cooper $33.3M, Mr. Etienvre $28.1M and Mr. Chawla $13.1M, mostly in June and August.¹⁰ Preset plans are normal, and the sales were at prices well above today’s. The 345,949 shares are about a third of the 1,010,091 the proxy showed in March. I cannot say how many shares he holds now, because the last Form 4 reports only the shares in one account. That is a limit on my knowledge.

What they did with capital over a full cycle

From 2016 to 2025 the average diluted share count fell from 143.9M to 116.2M, 19% lower, through buybacks, in a stretch when the shares were cheap.⁶ In the first nine months of 2025 the company bought 1.2M shares for cancellation at an average of $94.05.³ In the first nine months of 2026 it bought 0.1M at $273.60.³ Then on 5 August 2026 it sold 11.1M new shares at $310.00, for gross proceeds of $3.45B, about 9.7% of the shares then outstanding.¹¹³ The stated use is working capital and capital spending, with a new bid for up to 3.1M shares starting on 3 November.¹¹²²

I read this two ways. The company is raising permanent money to fund a doubling of revenue, because the cash needed to carry inventory and receivables for $40B of sales is larger than it has ever had, and that is a sound reason. It is also true that the company sold new shares at a price 3.3 times what it paid for old ones about fifteen months before, and the buyers took the dilution. Raising equity at the top of a boom is what a careful finance team does. It is also a sign that management sees the peak as high enough to sell into. Neither reading is wrong, and I cannot tell which is more right.

VFinancials

Revenue stayed between $5.6B and $6.6B for six years from 2016, and then did not. It rose 28% in 2025, and 83% in the September quarter. Operating margin went from 2.6% in 2016 to 8.3% in the past year, and operating earnings from $0.16B to $1.5B C (calculated).⁶³ Cash from the business did not keep up with it. Operating cash flow less capital spending and stock pay was $0.15B in 2023 and is $0.46B in the past twelve months, against net earnings of $1.25B C (calculated).⁵³ Profit grows with revenue here. Cash does not, because growth needs inventory, receivables and factories.

Revenue has risen every quarter, and adjusted margin has risen with it. The GAAP gross margin has not. It went from 13.0% to 10.8% and back to 12.3%, then 11.0%, because the amount of revenue passing through at lower margins grew faster than the design-led part. The adjusted operating margin rises because operating expenses are spread over more revenue: adjusted SG&A fell from 3.0% of revenue in the September 2025 quarter to 1.8%.⁴ That is a real gain, of a kind that runs out, because there is only so much overhead left to spread.

What an owner could really take out

The company’s own free cash flow was $412.5M for the first nine months of 2026, and $568M over twelve months C (calculated).⁴⁷ That figure is flattering, for two reasons. Here is the walk.

The first reason is stock pay. The income statement charges only $84M of stock expense over the twelve months. The cash flow statement shows $442M paid to settle vested awards, $38M spent buying shares for awards, and $247M received from the swap the company uses to hedge its own share price. The net $233M is the real cost of paying people in stock this year. It is bigger than the expense because the tax is paid on the value at vesting, and the shares are worth far more than when they were granted.³⁵ The company says it holds the swap to manage its cash needs and its exposure to its own share price.³

The second reason is receivables. In the first nine months the company sold $2.18B of customer invoices to a bank, $1.30B of it in the September quarter, against $22M a year earlier, under a program it raised in September from $450M to $1.0B of capacity. What is new is the balance: $700.8M of sold invoices were outstanding at 30 September, against none at the earlier quarter ends the 10-Q shows.³ Without it, receivables on the balance sheet would be $0.70B higher and operating cash flow lower by the same amount. I treat that as borrowing against the customer, so I add it to debt in the valuation. It also means the company’s cash cycle, the figure it publishes, is flattered.

So the number I would call owner cash is about $0.3B for the past year, 75% below net earnings of $1.25B. Against the enterprise value of $29B it is a yield of 1.1%, against 4.91% on a ten-year Treasury.²³ That is the cost of growing at 83% a year, and it means the price today depends on later years, which I model in Valuation.

Quality of earnings

Reported earnings have a second quirk. The company holds a total return swap on 1.25M of its own shares, with a strike price of $288.87.³ When the share price rises, the swap is a gain, booked in the income statement. It was $253.0M of gains in 2025, a quarter of pre-tax earnings of $983M, and $83.2M in the first nine months of 2026.⁵⁴ Adjusted earnings remove it, and so does my model. At today’s price the swap is worth about −$0.05B, because $249.22 is below the strike, so the company will book a loss in the fourth quarter C (calculated).

The tax rate is the third item. The GAAP rate was 15% for the first nine months, and 13% in the third quarter, helped by a discrete benefit. Adjusted, it is 18% to 20%, with 21% guided for 2027.⁴ I use 20%.

The balance sheet

The company has $3.96B of cash, up from $0.60B at the end of 2025, because of the August share sale. It has $0.74B of term loans and a $1.75B revolving credit line, none of it drawn, which the lenders extended in April through 2031.³²⁴ Cash minus borrowings and finance leases is $3.14B. That looks like a fortress. The next figure explains why it isn’t one.

Inventory doubled in nine months, to $4.5B. Receivables grew by $1.1B to $3.8B, even with the sales program.³ And accounts payable, the credit that suppliers extend, rose from $1.9B to $5.0B.³ Days in payables went from 50 at the end of 2025 to 74. Cash cycle days, a number the company reports, went from 61 to 42.³ Suppliers are financing the build. If payables fell back to 60 days, on the third quarter’s cost of sales that is about $0.8B of cash out, and on the $35B or so of 2027 cost of sales it is about $1.3B C (calculated).³ That sits on top of the capital spending of $1.5B to $2.0B and the receivables sold to the bank.

Management says in the third quarter report that it expects to use cash in 2027, because working capital and capital spending will both rise.³ I treat the $3.4B raise as spoken for, because the company decided it needed it to grow into $40B, and so I do not treat the cash as spare: my cash flows already carry the working capital, and I deduct the sold receivables and the swap from the cash I add back.

Why the price re-rated, and has now come down

The stock rose from $11.27 at the end of 2022 to a close of $472.40 on 2 June 2026, 42 times, and now stands at $249.22, 22 times the 2022 level and 47% below the peak.¹ Three things happened. Earnings rose: reported EPS went from $1.18 to $10.60 in the past twelve months, nine times C (calculated).⁶³ The mix changed from boards to hyperscaler-designed systems, and the company began to talk about AI. And the multiple went from about 10 times reported earnings to 41 times at the end of 2025.

The market re-rated it for what it was becoming, and the move down since June is the market asking whether the new earnings are repeatable. On the company’s outlook the stock is at 11.6 times 2027 earnings, a figure I compare with 18.7 times for Broadcom, and about 36 times for Arista.²⁵¹⁶ By that measure it is cheap. By the cash it can count today it is not. Both are true, and the rest of this page is about which one matters.

It tends to fall hard. The stock has fallen by more than 30% four times in ten years on daily closes: 81% from April 2017 to March 2020, 35% in the summer of 2024, 54% from February to April 2025 and 48% since June.¹ It has made new highs after each of the first three.

VIValuation

First, how this goes badly wrong

Before any numbers, I asked what would most likely destroy this investment, and worked back. The company has no balance sheet trouble, with $3.96B of cash against $0.74B of term loans, so a loss would have to come from the customers and from the price. Three ways stand out.³

• The largest buyer pauses or moves a program, with the parts bought and the factories built. One customer was 35% of revenue in the September quarter, and the filings give the company no volume guarantee.³⁵ The company is building capacity for $40B of sales, with capital spending of $1.5B to $2.0B in 2027 and $4.5B of inventory already.⁴³ A pause in 2028 or 2029, after the capacity exists, hits revenue, margin and inventory at the same time. The customers must buy unused parts under most contracts, which helps, but that clause relies on the customer being able to pay.⁵ This sets my bear.
• The money that funds the growth stops. Days in payables rose from 50 to 74 and $0.70B of invoices were sold to a bank under a program that the bank can end on three months’ notice.³ If suppliers or the bank tighten while inventory is high, the company needs cash it has to find. It has $1.74B on the revolver and the $3.4B raised, which is real protection. But the 2027 plan already uses cash, so the cushion is for the plan, and not for a bad year on top.³
• A good business at the wrong price. The business is better than it was. The question is whether guided earnings of $21.50 a share, up 86%, are a floor or a peak. Adjusted earnings per share were $2.46 in 2023 and a guided $11.55 in 2026, nearly a fivefold rise in three years.⁸ I would not assume that rate continues for a fourth year. The stock fell 81%, 35%, 54% and 48% in four drops in ten years, and none of them needed bad results.¹

The first two set my bear. The third is what the rest of this section is about. All three are in the company’s own risk factors or its own history.⁵

What the price implies

At $249.22 on the NYSE (C$348.45 in Toronto) the market values Celestica at $31.7B on 127M diluted shares, or about $29B net of the cash I count C (calculated).¹²⁸ The company’s adjusted EPS outlook is $11.55 for 2026 and $21.50 for 2027, and Nasdaq’s consensus for both is the same numbers, from a single estimate each. That is a sign that the Street is repeating the company’s guidance, not forecasting separately.⁴²⁶ So I use the guide as the first number and do not claim a view from the Street.

The reverse DCF cannot start from a base of owner cash, the way it does on my other pages, because the past year’s is $0.3B and growing it from there would prove little. So I solve for something else. I scale my base revenue path up by one share each year, at my base margins, until value equals the price. The price needs $103B of revenue in 2036, 1.9 times my base of $55B, and above my bull of $91B C (calculated). Or it needs my base revenue and a margin that is 2.8 points higher every year, about 9.3% in 2036. That is the margin management guides to for 2027, held for ten years.

The price sits between my base and my bull. An owner today pays for the 2027 margin to last, and for revenue of about twice my base by 2036, and earns less than the 12% I require if only my base happens.

Quality and price are two questions

The first question is the quality of the business. It earns a 57% adjusted return on invested capital, has more than doubled operating margin since 2022, has a new set of programs and grew 83% in the latest quarter, though it has a thin margin, no volume guarantee and a customer that is 35% of sales.⁸³⁵ I would call it a good business that has improved, in a hard industry, with one large weakness. The second question is the price. Against the outlook the price is low. Against the cash it is high. Against my base it is 47% too high and against my bull it is 26% too low. This looks like a business of improving quality at a price that needs a good deal to go right.

The next best use of the same dollar

Ten-year Treasuries yield 4.91%, with no risk to the principal.²³ My Broadcom page shows a stock at 18.7 times FY2027 consensus earnings, with a base case return of about 7.5% a year at a 10.5% hurdle. Arista on my page is at about 36 times 2027 consensus.²⁵¹⁶ Celestica’s base return of 6.1% is lower than Broadcom’s, at a higher hurdle, with a customer that is 35% of revenue. Its bull return of 16.7% is high, and its bear return is negative. That is what the price is for: a wide range.

What I can’t judge

Three inputs matter most, and I can’t know any of them. I can’t judge how long the AI buildout lasts. The companies selling into it are sure of it, and the buyers have said that 2028 will be higher still, but no one has seen capital spending of $3.1T a year and no one knows who pays for it.⁸ I can’t judge how much of the $40B is margin and how much is parts passing through. The filings do not say how much of the revenue is components, such as chips and memory, that Celestica buys and resells at a few cents on the dollar. And I can’t judge what the largest customer will do with its next generation of products. It was 35% of revenue in the quarter, and nothing public says whether its next program is already awarded.

So I show the range, and not a point. Margin matters more than growth here. One point of operating margin moves value by about $29 a share. and two points of growth a year, every year, moves it by about $21.

The three cases

Base. 2027 comes in at $38B, 5% under the guide, because first-year ramps on new rack programs usually slip. Networking grows 60% and compute and storage 170%, then 14% and 12% in 2028. I include a one-year pause in 2029, with revenue down 6%, because I have not seen a technology build-out run ten years without a pause. After it revenue grows about 5%, then 3% to 4%. Operating margin, after stock pay, goes from 8.2% to 6.5%, as buyers push price. That is still more than twice the 2016 to 2022 average. Capital spending runs at 4.0% of revenue in 2027, 3.2% in 2028, 2.5% in 2029 and 2.2% after. Working capital is 10% of added revenue.

Bear. The largest buyer moves programs and the build stops. Revenue reaches $31B in 2027, falls to $20B by 2029, and grows slowly to $26B. Margin goes back to 4.0%, between the 3.6% and 4.2% of 2022 and 2023. A fall that size releases working capital, but I release only half of it.

Bull. The company beats its 2027 outlook at $42B and keeps growing, to $91B in 2036, as the 1.6 terabit, scale-up and custom-chip programs turn into more programs. The 2027 margin of 8.6% stays near 8.8%. This is the company’s plan, extended. I have no evidence that it cannot happen, and little that it is the usual outcome.

The discount rate

I use 12.0%. The capital asset pricing model starts with the 4.91% ten-year Treasury yield in Q4 2026 and Damodaran’s 4.20% implied equity risk premium of 1 October.²³²⁷ My beta on ten years of month-end returns is 2.23, which gives 14.3%. The five-year beta of 2.12 gives 13.8%, and the three-year beta of 3.44 gives 19.3%, which is mostly this year’s run C (calculated).¹²⁸ I take 12.0%, the rate on my Micron and Vertiv pages, which also have five-year betas above two. I sit below the beta range, because the downturn is already in my bear case and I don’t want to charge for it twice. At 14% my base is $140, against $169 at 12%.

Margin of safety

I set it at $87, so that my $65 bear value is at least three quarters of what I would pay. It is 49% below my base, and 65% below the price. The stock has been there: it closed at $66.15 on 4 April 2025 and at $92.30 on the last day of 2024, a year in which Celestica earned $3.88 adjusted per share.¹⁷ It is a long way from $249.22 now.

Where the price sits

On my weights of 25%, 50% and 25%, the three values average about $185. The price is 35% above that, 47% above my base value and 26% below my bull. At this price an owner has no margin of safety: the market pays about three quarters of my bull value, and my base earns less than the rate I require.

It is reasonable to ask why the market might be right and I might be too low. The simplest argument is this. Suppose Celestica earns its 2027 outlook of $40B at a 9.0% adjusted margin, then simply holds it, growing 3% a year after that. At my 12% rate and with the reinvestment that implies, that stream is worth about $235 a share E (estimate). That is close to the price. So the price asks for one very good year to last, and my base says it will not quite: it has 2027 at $38B, a margin that fades toward 6.5% as buyers push price, and a pause in 2029. If the margin holds and the pause does not come, I am too low.

VIIThe case for the price

What most people miss

Two things cut in Celestica’s favour. First, the inventory build is less at risk than its size suggests. Customers had put $560.7M of cash deposits against inventory at 30 September, and a majority of the supply agreements oblige the customer to buy parts bought to fit its forecast.³⁵ Second, capital is not the constraint on returns that it looks like. Adjusted return on invested capital was 57% for the year to date, because customers and suppliers carry much of the working capital.⁸

The bull case

In plain words: the biggest buyers of AI hardware will spend more every year until 2030 and want custom designs, and Celestica is the partner they trust to build them fast. Chief executive Rob Mionis said on 26 October 2026 that demand is very robust and that 2028 demand is well above the 2027 outlook.⁴ On 27 October management showed 1.6 terabit switch programs, a hyperscaler optics award and OpenAI’s Jalapeño racks.⁸

Nasdaq’s analyst summary shows 14 price targets with a mean of $462.71, between $375 and $535, all above the price.²⁹ I could not see their dates or firms, so I attribute no view to a named analyst. The proxy lists FMR at 6.4% and BlackRock at 5.5% at 27 March 2026, and I attribute no view to them.⁹

Against reality

The near term supports the case. Revenue grew 83% in the September quarter and the Q4 guide implies about 76% on a year earlier C (calculated).⁴ If Celestica ships $40B in 2027 at a 9% margin, and 2028 is higher as management says, the price is fair and my bull is within reach.

The price needs that and more. Revenue has to reach about $103B by 2036, or the 2027 margin has to hold for ten years. Today the company has a 35% customer, a margin that has risen for five years but not through a downturn, a cash flow that is negative once sold receivables are counted, and an outlook that is a plan.³ These are what the case has to survive. Multiply it out: AI spending keeps rising to 2030, 60%; Celestica wins the next generation at its largest buyers, 55%; the new programs ramp on time, 55%; the margin holds near 8.8%, 45%; the largest customer does not dual-source, 60%. Those are guesses and not independent. Multiplied they give 4.9% E (estimate), perhaps 10% to 15% allowing for overlap. I still give the bull 25%, as on my other pages, which is generous.

What a 10x would take. Ten times $249.22 is $2,492, about $317B of market value. At 25 times earnings and a 6.5% net margin, near the 6.8% GAAP and 6.9% adjusted of the September quarter, that is about $195B of revenue, 4.9 times the 2027 outlook, or growth of about 25% a year for ten years from 2026 C (calculated). The stock did rise 22 times from the end of 2022, but from a market value of about $1.4B, about 23 times smaller than today’s $32B C (calculated).¹⁶ That run belonged to the small base. From here a tenfold move is close to impossible, and I would not build a case on it.

Where I come down

The outlook moved me: it is why my 2027 is not lower and why my bull is as high as it is. It did not move me on durability, because a builder’s margin tends to fall when many builders chase the same program, and the filings give the company no volume or price promise. Proof that the largest customers have awarded the next generation of racks and switches to Celestica at this margin, or a second year of positive cash flow, would raise my values most.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• Supply and cost. The company says component supply improved in July, and that component shortages are common in its industry. Shortages can delay products and raise costs it cannot always pass on.¹⁵³
• Customer credit. CCS customers now include frontier AI labs, which are not established cloud companies.³ A customer that cannot pay leaves the company with inventory and receivables. The A/R sales program depends on customers the banks consider creditworthy.³
• Tariffs and trade. Section 301 tariffs of 10% to 12.5% on most imports from certain countries began on 23 July 2026. About 75% of revenue is produced in Asia.³⁵ The company has recovered tariffs from customers so far, and refunds of earlier tariffs have arrived. A change in either direction moves working capital.³
• Geopolitics. The 10-Q names tensions between mainland China and Taiwan among its risks.³
• Power and capacity for the customers. Data center growth needs power and water that may not arrive. The company lists these constraints as a risk to demand.³
• Execution on the build. Capital spending goes from about 2% of revenue to 4.4% to 4.8%, in Thailand, Texas and Mexico. The 10-K warns that facilities may be late or underused, and that the cost base rises.⁵⁸
• Interest and debt terms. $385.6M of term loans have unhedged variable rates, and the credit agreement limits buybacks if leverage rises.³ This is small at present.
• Management change. A new chief financial officer and a new head of CCS in a year when spending and working capital double.¹⁹²⁰
• Reported earnings. The swap, the tax rate and the choice of adjusted measures all flatter earnings in some periods, and stock settlements cost cash.³

What would change my view

Toward a higher value. A 2027 outlook raised above $40B with the margin intact. A year of positive free cash flow with payables no longer than 74 days. A named second large customer above 20% of revenue and the largest one below 30%. Proof that the HPS programs keep a 9% operating margin at twice the volume. The new share buyback used at prices below my base.

Toward a lower value. A cut to the 2027 outlook. A gross margin below 11% with revenue still growing. Payables days or sold receivables rising again without a matching rise in revenue. The largest customer moving a program to another builder. A pause in hyperscaler orders. Another equity raise.