Breakfast Bytes Blogs

I wrote a post a couple of years ago Numbers Everyone Should Know. It was about timing of things in microprocessor-based systems, in particular the huge disparity between the cycle-time of a modern processor and the time required to access DRAM, let alone disk or network traffic.

I recently watched Malcolm Penn's video No Man Is an Island which looks at the long-term state of the semiconductor industry. I'll summarize a few key points but you should really watch the video (26 minutes):

Malcolm has been analyzing the semiconductor industry for a long time. In fact, when I see his name, I think of Dataquest, since he analyzed semiconductors there. Well, Dataquest was acquired by Gartner in 1995, quarter of a century ago. He says he's seen all the semiconductor cycles except the very first one, around 50 years. Future Horizons (his current company) was founded over 30 years ago in 1989.

The key drivers, the four horsemen of the semiconductor apocalypse, are:

• Economy determines what users can afford to buy. A more buoyant economy, stronger demand.
  • Strong 2021 bounce back of +5.1% after 2020's -4.4% contraction
  • Recovery hampered by shortages, supply line imbalances, and inflation concerns (interest rates)
• Unit demand: what users actually buy (plus or minus inventory adjustments)
  • Shipments - real demand +/- inventory and lead-time, adjusted or double ordering
  • Shipments currently running well above normal 8% trend line...won't go on forever
• Capacity: how much demand can be met (under-supply or over-supply). Capacity is now only built with firm commitments, and not speculatively like it used to be
  • Completely sold out following years of under-investment
  • Modest 2020 capex increase on-stream late 2021 or 2022
  • More robust early 2021 capex increase (up 82% on recent trends), on-stream late 2022
• ASPs: the price users can be sold for, which is based on supply-demand plus value proposition (pricing power). ASP impact is not usually prices rising but not falling as much as they might have done.
  • Recover lats units by 20 months (this time only 8)
  • Long-term trend is 0% since pricing is hostage to Moore's Law

The industry is inherently cyclical due to mismatch between supply and demand. Production cycles are long (four months to manufacture a chip) and new capacity lead time is about one year if you already have an empty building and two to three years if you don't. On the other hand, demand is highly volatile and there is often poor visibility of end markets, which leads into poor visibility of semiconductor demand and so "atrociously unreliable forecasts".

In July 2020, things fell apart (and this is pre-Covid) due to under-investment in Q4-2019, worsened by widespread disregard for supply-demand fragility, despite clear warnings. The strong 2H-2020 set the stage for a double-digit 2021. But still nobody paid attention and believed, consensus forecast was for 8% growth. Malcolm's company was alone in predicting 18% growth, and it turned out they were low and it was 25%.

For 2022, "it will either be -4% or +6.5% so we forecast 4%". The issues:

• Uncertain supply-demand impact due to 2H-2020 capex increase/slowdown in shipments
• Supply surge in 1H-2022 due to 1H-2021 capex increase
• Uncertain economic outlook and risks (Covid, inflation, widespread shortages of all sorts of stuff)
• Not much upside potential

The biggest issue of all is "When does the capacity and demand come into balance and that will trigger this whole thing turning south? It is hard to predict right now. When do prices stop rising? When do buyers get all their double-ordered parts? When do orders start to be canceled? But there is no such thing as a soft-landing in this business. When it collapses, it will collapse, it will crash."

Malcolm also thinks that re-shoring is politically driven, not business driven, which is "a sure-fire recipe for disaster". There is no market pull (customers don't care) and it is not in Wall Street's interest. He thinks there will be backlash from states and countries that don't benefit. Plus politicians will lose interest in about six months when something else becomes a new vote-winning item.

Numbers for Semiconductors

One interesting part of his presentation is numbers that never change. Malcolm says:

I love these numbers because they are so constant. They are so predictable. And nobody believes in them. Everybody hates them.

Whether it is the number of IC sales, about 8% growth. The number of ICs per square inch, constant at about 3% per year growth, average selling price growth...they are slightly negative right now but the long-term trend is zero. Capex as a percent of sales, silicon per square inch, world GDP, world IC sales...these are good numbers. They are good solid trend lines and those are the benchmarks that we measure ourselves against when we want to understand whether we are above or below the trend line.

How much do you make?

• IP supplier, 10¢ per square cm, (Arm gets 20¢ because they're #1, a lot of suppliers only get 5¢)
• EDA 18¢ per square centimeter
• Semiconductors: $9 per square centimeter, it's been that way since time began. Or $4 per square centimeter for the foundries
• End-equipment manufacturer $35 per square centimeter
• Services: $450 per square centimeter, double that for someone like Apple

Malcolm again:

Selling on value has never been a proposition unless you have a monopoly. Unfortunately, we've always sold semiconductors by the square centimeter right from the beginning of times. It was perpetuated by the forward pricing policies of companies like TI, it was driven by the economies of scale that you get from the learning curve, and stuff like that. We've simply said never mind value, we'll just give you more for your buck and keep the prices up on that. So it's a cost-plus industry. It's a cost-driven sector. We find it very hard to break out of that.

Every startup needs a sugar daddy. You've got to have a sugar daddy to keep you going. Arm had Nokia, TSMC has Apple. They have 50 times the buying power that you do as a semiconductor company. That's why Apple is so good at making SoCs since for them it's peanuts. They have so much revenue base, so much profitability, for them to do chip design is very, very easy. But for a chip company to do chip design when you only have your chip revenues to fund it is very hard. So find yourself a sugar daddy, you've got to find an end-customer out there to buy from you and give you that initial market push that you need.

The audience for the video was semiconductor startups. Malcolm had advice for them.

Market pull is 1000X stronger than technology push. Who are the customers? How big is the market? Who are the competitors? What are the barriers to entry? Why should anyone bother?

Rule #1: Better (best) is not relevant...only 'first and good enough' matters

Rule #2: Most startups fail due to poor marketing not bad technology

In my experience of doing consulting to startups, and also looking at attempts by friends to start companies, I can confirm #2. People love that they know how to create some unique technology, but they tend not to do it with a teaching company, what Malcolm calls a sugar daddy. Far too often, people create technology for which there is either no market, or else (in EDA) the market will take a couple more process nodes before the problem needs to be solved.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.

.