The Moment Losses Flip to Profits: Margin Inflections at Lumentum, Credo and Innodata

TL;DR The semiconductor index has fallen into a bear market, down more than 20% from its high a month ago. Sell-offs like this do not only punish the overvalued names, they drag the strong ones down too. What I am tracking right now is the group whose operating margins just flipped from negative to positive on AI demand — Lumentum, Credo, Innodata and Tower Semiconductor. Cohu and SanDisk are on a different clock.

What actually matters in a drawdown

The SOX, the index that tracks the entire semiconductor group, recently fell into a bear market. It is down more than 20% from its high just a month ago, and at one point on Friday it was down almost 6% in a single session.

The most common mistake in this environment is treating an index move as a verdict on individual companies. Sell-offs do not surgically target the overvalued names. They drag down the strong ones with them. In my experience the opportunity always hides in that gap — I laid out the same structural point in the semiconductor pullback and narrowing breadth.

But I am not making the argument that you buy whatever fell the most.

The signal I look for is the profit flip, not the growth rate

A headline that revenue doubled tells you very little on its own. A company can double revenue while doubling its losses just as fast.

What I weight far more heavily is the point where operating margin crosses from negative to positive. Past that line, most of every incremental revenue dollar clears the fixed cost base and falls straight to earnings. That is the transition from startup mode to a real business, and it is usually where momentum builds fastest.

Several AI infrastructure names have just crossed it.

Lumentum — a genuine chokepoint in optics

Lumentum owns the EML laser chip, the tiny light source buried inside every high-end AI optical transceiver that shuttles data between racks.

Every 800 gig and 1.6 terabit module needs several of them. Lumentum makes 50 to 60% of them. More importantly, it is the only supplier shipping the 200 gig per lane version at volume that the newest links depend on. That is not a marketing advantage, it is a physical bottleneck — and it is why Nvidia designated Lumentum's optics for the next-generation Rubin platform.

For years this was a struggling telecom name bleeding money, running an operating margin as low as -25%.

Then the AI orders hit. Revenue climbed 90% year over year last quarter and the operating margin swung to positive 22%. Nvidia put $2 billion into the company as an investment, and separately Lumentum landed a multi-billion-dollar optical switching deal that already carries a backlog north of $400 million. My read is that most of that demand has not even reached the reported numbers yet.

Credo — the company that bought copper a few more meters

Credo builds active electrical cables: copper with a signal processing chip built directly into the connector.

At current speeds, plain copper dies past a meter or two. Optics solve the distance problem but cost more and burn more power. Credo's cables retransmit the signal mid-run, so copper goes several meters farther, cheaper, and with far less power draw.

Inside an AI cluster, what matters most is not speed, it is reliability. One flaky connection can stall a training run worth more than $100 million, and nobody wants to own that outcome.

Credo created this category and holds roughly 88% of it. Revenue more than tripled last year, up 206%, at a 68% gross margin — software territory for a cable maker. Operating margin went from -19% to a 33% profit. On a PEG basis it still sits well under one.

Innodata — the data supplier that survived a short report

Innodata produces the expert-labeled data that AI models are trained and tested on. It is essentially the textbook a frontier model learns from and the exam it gets graded on. It already does this for five of the seven largest tech companies.

This next part is where I am speculating, and I will flag it as such. I think the bigger wave is still ahead. As every large company starts building its own AI agents, they need the same data, plus a way to continuously test whether those agents are actually correct. That turns lumpy project work into recurring revenue, and it widens the market from a handful of labs to potentially the whole economy. The agentic slice alone is projected to grow more than five times by 2030.

A couple of years ago, short sellers filed fraud accusations against the company. Those accusations completely collapsed — both the DOJ and the SEC closed with no action. Over the same stretch, revenue grew 48%.

Revenue has nearly tripled in two years to $252 million, and the company swung from a loss to $32 million in profit. The market still treats it as an afterthought.

Tower Semiconductor — early in the mix, but the capacity is already spoken for

Tower does not design chips, it manufactures them for the companies that do. It is a specialty analog foundry, and its most important platform for AI is silicon photonics — the chips that turn electrical signals into light inside optical connections.

Only about three companies make photonics at real scale, and Tower is the open, neutral one the whole industry can share. That is why names like Marvell and Nvidia run their photonic chips through it. I covered the broader shift in photonics and the move from copper to light.

Silicon photonics revenue more than doubled in a year and operating margin climbed from about 9% to nearly 16%. That said, photonics is still only about 15% of Tower's revenue. This is a forward bet layered on a very broad foundry base, and the premium multiple only works if the AI leg keeps compounding.

Which is why customer behavior is the useful evidence here. Against roughly $1.57 billion in annual revenue, Tower has already booked $1.3 billion of silicon photonics orders for 2027, and has taken $290 million up front just to reserve capacity. With earnings growing better than 50% a year, its forward PEG sits near 0.91.

Not flipped yet: Cohu and SanDisk

By the same standard, these two have not crossed the line. That does not make them uninteresting — it makes them a different clock.

Cohu makes the machines that test semiconductor chips after they come off the line, holding each chip at a precise temperature, pressing it onto the tester, and sorting good from bad. The parts that matter for AI are its Eclipse thermal handler, which tests GPUs while controlling the intense heat they throw off, and its Neon system, which inspects the stacked memory in AI chips, with inspection revenue guided to grow about 80% this year. To be clear, AI is only about 2% of sales today. This is a deeply cyclical test business coming off the bottom. At its last peak Cohu earned $3.45 a share and it is only now climbing back out of a loss, with orders already up 57%. Behind that sits an AI test pipeline of $750 million that is not even booked yet and is larger than its entire revenue last year, with 60% of sales coming from recurring consumables. You are not paying up for peak earnings here — you are buying a business still priced for the downturn it is only beginning to exit.

SanDisk is known for NAND flash, the chips that store data permanently. The memory bolted next to an AI chip today is HBM: blazing fast, but expensive and cramped, holding only about 50 GB in a stack. SanDisk pioneered high-bandwidth flash, or HBF, which essentially matches that speed but holds eight to sixteen times more — on the order of 512 GB at a similar cost — which could roughly double the company's addressable market. It is co-writing the industry standard with SK Hynix while a giant like Samsung is still at the early concept stage. So this is a company with about 13% of the NAND market holding the pin on technology that could reset the entire business. SanDisk spun out of Western Digital last year near $38 a share, then revenue exploded 251% in a single quarter. It used that windfall to wipe out its entire debt and authorize a $6 billion buyback, and it trades under 10 times next year's earnings today. Memory is still cyclical, but the multi-year supply deals already signed and HBF still ahead give real reasons this run outlasts the ones before it. I went deeper on that in the memory supercycle and SanDisk.

Side by side

Company	What it does	Operating margin shift	Recent revenue growth	Demand already locked
Lumentum	EML lasers for AI optical transceivers	-25% → +22%	+90% year over year	$400M+ optical switching backlog
Credo	Active electrical cables	-19% → +33%	+206%	~88% category share
Innodata	Labeled data for AI training and eval	Loss → $32M profit	+48%	5 of the 7 largest tech firms
Tower	Silicon photonics foundry	~9% → ~16%	Photonics more than doubled	$1.3B booked for 2027
Cohu	Semiconductor test equipment	Still recovering	Orders +57%	$750M AI test pipeline

Where I land

What these names share is not that they are cheap. Cheap is the output, not the thesis.

What they share is that they sit on demand they do not control but which is already committed through contracts and share positions — and most of them just crossed the profitability line. Reported earnings have not caught up to that crossing yet, and the index drawdown is discounting it a second time.

The risk is equally clear, and I want to be blunt about it. Every one of these names is wired to a single variable: AI capex. They look like different companies with different products, but a large share of the downside comes from the same place. Holding five of them is not diversification. That, more than valuation, is what I think individual investors are getting wrong right now.

FAQ

Q: Why does the moment margins turn positive matter so much? A: Because the fixed cost base is already paid for. Past that point, a large share of incremental revenue drops straight to earnings, and you get a window where profit growth badly outruns revenue growth. Credo going from -19% to +33% is the textbook version.

Q: The semiconductor index is in a bear market — is this a reasonable entry? A: Rather than trying to call the index, check whether the decline happened independently of business results. When a company growing revenue 90% or 206% falls 20% alongside the index, that is usually flows, not fundamentals. The trade-off is that the condition can persist for months.

Q: Does a PEG under 1 automatically mean undervalued? A: No. PEG depends entirely on the forward growth estimate, and for these companies that estimate comes from the AI capex cycle. If the cycle slows, the denominator breaks first and the PEG re-rates upward very quickly.