GlobalFoundries' AI Rally Is a Supply Problem, Not a Chip Problem

GlobalFoundries has spent most of 2026 telling investors that its growth is "shifting to AI and photonics." At its Goldman Sachs Communacopia appearance on September 8 it said the same thing, alongside a newly finalized $375 million award from the U.S. Department of Commerce to advance quantum leadership. The stock has already voted on the story: it nearly tripled from a 52-week low of $31.59 to a high of $92.55, then gave most of it back to trade around $45 — up 29% for the year, but down roughly 10% over the last month. You are not discovering this theme; the market already paid up for it once and has since de-rated it. The live question is whether the underlying economics can defend the multiple the stock is being asked to earn again.

Here is the part most summaries get backwards. GlobalFoundries does not make the AI chip.

Not the compute chip. The interconnect.

When you scale an AI data center, the headline component is the compute chip — the GPU or the custom AI ASIC. Those are built on leading-edge nodes, below 10 nanometers, by a small group of foundries: TSMC, Samsung, Intel. GlobalFoundries left that race years ago. It deliberately stopped chasing the leading edge and refocused on specialty and mature nodes (28 nanometers and above, plus specialized processes like silicon germanium and silicon photonics). It does not compete for a single AI compute socket.

What it does make is the rest of the machine. Every GPU in a rack is paired with a wall of optical transceivers that move data between servers and across the building, and those transceivers run on silicon photonics and silicon germanium. Each one also draws enough power that the power-delivery chips feeding the accelerator are themselves a constraint. As AI clusters scale, the bottleneck in the data center has migrated — from the compute chip to the optical interconnect and the power that feeds it. That is precisely the layer GlobalFoundries owns, and it is one of the few pure-play, Western-aligned foundries that can qualify both of those on specialty nodes.

That reframes "GlobalFoundries is an AI beneficiary" from a broad demand claim into a specific supply one. The AI tailwind for the company is not that it sells chips into the compute market. It is that the AI data center's bottleneck has landed on the node it makes, and it cannot currently keep up with the orders.

The split, in one table

The table below is the company's own second-quarter 2026 end-market breakdown, pulled from its earnings call. It is the whole story in four rows.

End market (Q2 2026)	Revenue	Year over year	Share of revenue	Full-year 2026 guide
Communications infrastructure & data center	$286M	+62%	~16%	+50–60% (raised)
Smart mobile	$643M	−6%	~36%	low-teens decline
Automotive	$339M	−10%	~19%	low double-digit growth
Home & industrial IoT	$339M	+10%	~19%	+10–15%

Read it as two markets, not one. The communications infrastructure and data center corridor — the one tied to optical networking and power — is $286 million, up 62% year over year, the fastest quarterly growth in that segment since 2022, and management raised its full-year guide for it to 50–60% growth from the high 30s. Everything else is flat to down: smart mobile, the single largest piece at about 36% of revenue, fell 6% and is guided to decline in the low teens for the year; automotive fell 10% in the quarter on customer timing rather than lost demand; industrial IoT is a modest 10% up.

The implication is that the company's headline revenue grew only 6% while a 16%-of-revenue segment grew 62%. The growth is real but highly concentrated, and it is being used to re-mix the portfolio away from the declining handset business and toward the sold-out interconnect corridors. That mix shift, plus pricing increases that management says will begin showing in revenue in 2027, is what pushed non-IFRS gross margin up 470 basis points year over year to 29.9%.

Sold out, and priced for it

The constraint is not a claim; the company states it plainly. Its silicon germanium capacity at the Vermont fab is oversubscribed through well into 2027. It describes itself as "capacity constrained" across the interconnect corridors and, in CEO Timothy Breen's words, "hard at work producing more wafers every day for these oversold corridors." Silicon photonics revenue is expected to more than double in 2026; the company works with four of the top five optical transceiver makers; it booked seven new optical networking design wins in the quarter alone; and its fabs are running at utilization in the high 80s. That is a supply-constraint signature, not a demand-hype one.

But a constraint is also an execution and capex problem, and the multiple already reflects it. At roughly $24.7 billion of market value, GlobalFoundries trades around 34 times trailing (reported) earnings and 16 times trailing EBITDA, on about $1.8 billion of quarterly revenue. For a specialty foundry, that is growth pricing. The drawdown from $92.55 is the market stress-testing exactly what the table shows: the engine is one 16% corridor, and the drag — mobile — is bigger than the engine and still shrinking.

The risks are structural, not cyclical. First, concentration: the re-rating depends on the interconnect corridor compounding fast enough to out-earn the mobile decline — a mix race, not a revenue-growth story. Second, capex: the company is a heavy spender. Q2 capital spending was 23% of revenue, and it is mid-way through a roughly $16 billion U.S. build-out backed by about $1.5 billion in CHIPS funding and a $300 million silicon photonics award. "Sold out" is only as good as landing new capacity on time without overbuilding. Third, timing: the pricing increases that support the margin runway do not hit revenue until 2027, so part of the story is already front-run. Quantum — eight engagements, including PsiQuantum and Quantinuum — is real optionality, but management puts manufacturing revenue toward the end of the decade; it is not in this year's numbers.

The question is not whether an AI data center needs optical interconnect and power delivery — it does, and GlobalFoundries is among the few Western foundries that make both. The question is whether it lands the Vermont silicon germanium and photonics capacity before demand outruns it, and whether that 16% segment keeps compounding at 50–60% fast enough to outrun a shrinking handset base. If the constraint holds and the mix keeps shifting, the re-rating has a floor. If capacity lands ahead of demand, or mobile bites harder, the stock is paying growth prices for a specialty foundry that cannot show the growth on the compute side — which is the exact spot the market already checked, on the way down from $92.55, and found wanting.