The semiconductor industry as it is, not as it is told

Claus Aasholm has a habit that most corporate cultures would find uncomfortable. He reads the investor calls, checks the underlying numbers, and then checks whether the two things are actually saying the same thing.

The exercise that launched Semiconductor Business Intelligence began with a simple observation: Aasholm reviewed 25 companies and found that all of them had reported a great quarter, all of them were above average, and all of them had gained market share.

"I was thinking, hey, this is turning into a great industry," he told the audience at Evertiq Expo Zürich. "Everybody can win. We don't need to worry about market shares, winning or losing, up or down." 

The problem, of course, is that companies, on average, are average. "And when you start cooking the numbers, you find out that while the CEO is not lying, they're certainly presenting a version of the truth that is palatable to the company."

That gap between the narrative and the data is what he has spent years documenting. At Zürich, he shared his current read – and it was not particularly comfortable.

New territory

The semiconductor cycle has historically been predictable: roughly four years, resilient to crises from the dot-com bust to COVID, and legible enough that most participants have learnt when to invest and when to wait. What Aasholm is now tracking looks different.

WSTS data puts the industry at around 45% above its last cycle peak – territory it has never been in before. Forecasts point toward a doubling of that peak by year's end. But the more significant number is the share of total semiconductor revenue now attributable to data centres: just over 51%. Everything else – industrial, automotive, mobile, PC — remains below its last peak.

This, Aasholm argued, is not cyclical. It is structural. The rules governing investment, allocation, and competitive position are being rewritten, and not every company in the industry is positioned on the right side of that shift.

The TSMC situation nobody is talking about

One of the sharper observations in the presentation ran directly against the prevailing narrative of a sweeping semiconductor capacity build-out. When measured in actual physical output – 12-inch equivalent wafers – TSMC's overall capacity is not increasing. The company is redirecting investment toward leading-edge nodes at the expense of mature ones, which means that anyone depending on older technology is facing a quietly shrinking supplier.

"This is not the story that's out there. The story out there is: we're building American fabs, we're building 16 fabs in Taiwan, we're building fabs everywhere. But the overall capacity is not going up."

On the Arizona fabs, he was equally direct. TSMC operates them. TSMC holds the IP. Wafers produced in Arizona are currently shipped back to Taiwan for advanced packaging.

"The Americans are very proud — look at this, we now have homegrown, homemade American capacity," he said. "Yeah, I know, it's Taiwanese operating the fabs, and Taiwanese know all of the IP. But it's American. No, it's not. It's just politics."

The packaging revolution hiding in plain sight

What is changing, Aasholm argued, is not primarily where chips are made but how they are assembled. The core driver of AI performance gains is advanced packaging – memory getting closer to compute, compute getting closer to compute. TSMC formalised this under its Foundry 2.0 strategy and rapidly captured close to 30% of all OSAT investment. The companies that have understood this and invested accordingly are pulling ahead of those that have not.

"Think of it as a packaging revolution more than a semiconductor revolution as in technology."

Memory: the super-cycle begins

If the packaging argument was the structural observation of the presentation, the memory outlook was its most urgent practical warning. Memory CapEx has been restrained for years – first by brutal pricing conditions, then by NAND price collapses that spooked even well-capitalised companies. That investment void is now the industry's problem.

Memory prices are rising sharply. Micron has redirected roughly a third of its mobile and PC client capacity to the data centre. Purchasing commitments from the largest buyers have jumped from around 250 billion dollars to 600 billion in recent quarters, with the last two quarters alone accounting for the majority of that increase. One Micron customer has signed a five-year supply contract – longer than a full semiconductor cycle.

"If you need memory, you should have booked it last year."

The revenue and margin consequences are already visible. SK Hynix is on track to become the second-largest semiconductor company by revenue this year, ahead of TSMC. Gross margins at the leading memory companies are projected to exceed 80% – a level never previously recorded in the industry.

The companies being left behind

Not everyone benefits. The hybrid manufacturers — companies combining in-house production with foundry outsourcing — have now been in a downturn for more than 14 quarters. These are the companies producing the components that keep industrial equipment running, vehicles operating, and buildings heated. Their CEOs continue to signal an imminent upturn. Aasholm was not convinced.

"These companies are much more important, personal opinion, than AI — because these are the ones that keep us safe, fit, warm, mobile, employed with all of the products, the multitude of products that they pour into the world."

What is actually going on

Aasholm closed with a comparison of the industry in 2020 versus today. Semiconductor revenue has nearly doubled. Cloud CapEx is up 300%. Semiconductor CapEx has grown 69%. "Maybe these two numbers don't need to be married," he said, "but I would expect them to date at least."

The industry he described at Zürich is concentrated, structurally shifting, and running at utilisation rates that leave little room for error. The companies winning are winning at a scale and margin the industry has not seen before. The ones that are not are experiencing downturns of a kind and duration the cycle has not previously produced.

The Silicon Cowboys, as Aasholm put it, used to tell you the truth directly. The current industry prefers a more palatable version. His job, as he sees it, is to check whether the numbers agree. They often do not.

Evertiq Expo will return to Zurich on April 22 next year, and registration for the event has already opened.