Canada’s biggest AI trade is a contract manufacturer, and that is a stranger position than it looks

Ask which Canadian listed company gives the most direct exposure to artificial intelligence and the intuitive answers are wrong. It is not Shopify, which is a commerce platform. It is not Constellation Software, which acquires vertical market software businesses. It is Celestica, a Toronto-headquartered electronics manufacturing services company that builds hardware to other companies' designs.

The numbers behind that claim have moved quickly. Celestica raised its full-year outlook after quarterly revenue expanded sharply on demand for artificial intelligence data centre hardware and advanced networking, and has guided to a rise of roughly 190 percent year over year in enterprise revenue for the third quarter of 2026. Full-year adjusted earnings guidance moved from $10.15 to $11.30 a share, with free cash flow guidance lifted from $500 million to $600 million.

Those are not small revisions. A company raising full-year earnings guidance by more than eleven percent partway through a year is telling you its order book changed faster than its own planning assumptions, and in this case the reason is specific: the customers building large artificial intelligence facilities need hardware assembled, and there are not many firms able to do it at the required scale and quality.

It is worth being precise about what the business actually is, because electronics manufacturing services has an unglamorous reputation that no longer fits what this segment does. The traditional model is contract assembly at thin margins, competing on cost, with the value captured by whoever owns the design. What Celestica and its peers have moved toward is closer to joint design and manufacture, particularly in networking, where the switches and interconnect that link thousands of accelerators together are a genuinely difficult engineering problem rather than a bill of materials to be assembled.

That distinction shows up in the economics. A pure assembler earns a few percent and has no pricing power. A partner that co-designs the hardware, holds the qualification, understands the thermal and power envelope of a specific rack architecture and can be trusted with volume production is much harder to replace mid-programme, and earns accordingly. The revenue growth is the visible number; the shift in what kind of supplier this is matters more for whether it persists.

Now the position that deserves scrutiny, and it is not a knock on the company. Celestica does not own the demand. Its revenue is a function of how much its customers choose to build, and those customers are a small number of very large buyers whose capital plans are set annually and revised without consultation. This is exposure to the artificial intelligence buildout with essentially no control over its pace.

Set that against the deceleration now visible in the spending itself. Hyperscaler capital expenditure growth is projected to fall from around 90 percent this year to roughly 28 percent next, on a much larger base. A supplier can post excellent absolute numbers into that and still be re-rated downward, because the multiple attached to a hardware supplier reflects expected growth rather than current revenue. The risk is not that Celestica stops growing. It is that it grows at a rate the market decides to pay less for.

Customer concentration is the specific vulnerability worth watching in the disclosures. Electronics manufacturing services businesses in this segment typically derive a large share of revenue from a handful of relationships, and losing or winning one programme moves the numbers materially. That cuts both ways and has plainly been cutting favourably. It also means the reported growth rate contains less diversification than a similar rate at a company selling to thousands of customers.

For Canada, the strategic reading is more interesting than the trading one. The country has three well-regarded artificial intelligence research institutes, a genuine claim to the foundational research, and its largest listed exposure to the resulting boom is in manufacturing rather than in models or software. There is a version of that observation which is a complaint about value capture. There is another version, and I think it is the better one, which is that manufacturing at this level is a real capability that is difficult to build and difficult to displace, and Canada has one of the few firms in the world doing it at scale.

The comparison with Cohere is instructive precisely because the two occupy opposite positions. Cohere is a private company building foundation models, which is the layer everyone agrees is strategically important and where economics remain unproven at almost every company attempting it. Celestica is a listed company building the physical layer, where the economics are proven, currently excellent, and dependent on someone else's investment cycle. A country would rather have both, and Canada does, with one of them accessible to public market investors and one of them not.

The other thing this exposes is how narrow Canadian technology indices are. A sector where a handful of names carry the aggregate is a sector where the index tells you about those companies rather than about the domestic technology economy. A great deal of what is actually being built in Canada, in Calgary and Edmonton as much as in Toronto and Montreal, is private, and none of it appears on this page.

What to watch is concrete rather than atmospheric. Whether the enterprise growth rate holds as comparisons get harder, since 190 percent against a small base is a different achievement from 190 percent against a large one. Whether customer concentration disclosures move in either direction. Whether margin expands as the mix shifts toward co-designed networking, which would be evidence the business has genuinely moved up the value chain rather than simply catching a wave. And whether hyperscaler capital guidance holds at the next round of reporting, because that is the demand this entire position sits on.