It really took the Chinese slideshow for the final piece of the puzzle to click in and become clear.
The logic is remarkably easy to follow when the pieces are all in place.
Ultimately it's all about the money.
Here's the single most important piece of information:
At 90nm there were 19 ARM foundries. At 65/55nm there were 16. At 45/40 there were 12. At 32/28nm there were 8. At 16/14nm there will be three. That number could drop to one (Samsung). Regardless the trend is incontrovertible. The fabless model is broken. There is no other reason for a decline from 19 to 3 foundries.
And why are the number of ARM fabs dropping like a rock? Money pure and simple. The fabless model doesn't provide nearly enough money for CapEx or R&D. The ARM foundries have therefore tried to get by with "just good enough." And it has resulted in disaster with no access to FinFET (insufficient R&D spending) and insufficient manufacturing capacity (insufficient CapEx availability). And a drop in the number of ARM foundries that if you extend the trend line drops to zero.
And here's your proof:
93 fabless Series A Semiconductor Fundings in 2000, 42 in 2004, 16 in 2006, 16 in 2007, and 3 in 2012.
Funding for foundries has dropped like a rock and the number of foundries has dropped in a corresponding manner. No money. No fabs. Not rocket science.
Here are the results or perhaps the symptoms of the breaking of the ARM fabless model:
1.) Fewer fabless start-ups, Even fewer rising stars, Large fabs acquiring small ones, Dropping number of foundries, deals and investments.
2.) The number of designs that the ARM foundries can support at each node is dropping. At 65nm it was 65 to 80, at 40nm, it was 40 to 50, at 28nm, it was 20 to 30, at 20nm, it will be 12-16 and at 14nm it will be 5 to 8.
3.) The time to ramp on each new node is getting longer.
4.) The complexity of each node is increasing dramatically.
5.) The costs of each new node are getting higher while the benefits are getting smaller.
6.) Foundries are forced into a bad economics move to 20nm planar which is costly at minimal benefit.
7.) Foundries are forced into attempting a half-step to 14/16nm even though they haven't yet successfully assimilated 20nm upon which the move is based.
8.) Because of extremely high cost in process technology and SoC development, the relationship between fabless companies and foundries has become sensitive.
9.) Foundries will select their customers very carefully, and the criteria for acceptance will involve financial returns as well as technical capabilities. So, the rich will get foundry services and the rest of the ARM shops will be left hanging. So much for the ARM hoardes helping to compete with Intel.
10.) There is also the complementary factor that, with the increasing cost of implementing SoC designs at most advanced technology node, only a small number of fabless companies can participate.
11.) Whenever the partnership between a fabless company and foundry established, it would not be easy to change, as the change means high cost that should be paid by both parties. In other words, when an ARM shop makes a deal with one of the few remaining ARM foundries, then they are STUCK, and can only get unstuck by the application of large sums of money. And then who will they go to? One of their other one or two choices, if they even have that many?
12.) At 20nm and 14nm, the foundry business model will probably be slightly changed. An IDM-like collaboration between fabless companies and foundries will be expected. So the ARM foundries are going to have to become virtual IDMs (like Intel) in order to compete as these levels. Expect higher costs and lots of problems on implementation.
13.) Fabless companies have to accept a slow ramp-up of wafer volumes because of the IP qualification and parametric yield challenges.
14.) Design teams will need to be manufacturing-aware and have in depth understanding of the process variables that have to be included as part of design disciplines. There will be the need for collaboration between fabless company and foundry on understanding the process variations.
15.) Fabless companies cannot assume that their designs will give high yields based on their legacy interfaces that are in place with foundries. There is the need to anticipate potential yield problems and address them at the design implementation phase, and DFM factors have to be carefully analyzed.
16.) Fabless companies will need to build stronger process and product engineering and DFM expertise. It means a extra cost of 50-100M USD per year.
The fabless fabrication model has been in decline for years. Now the absence of adequate R&D spending has shown up to finish it off. No FinFET, no state-of-the-art technology, no roadmap beyond 20nm, no future.
Along with insufficient manufacturing capacity.
Does anyone really think that something will happen to increase the number of foundries up from three rather than down to one? Nothing in the trends, dynamics or funding suggests that anything other than a continuation of the trend is possible.
Pull the curtains. Que the lights. The ARM show is over...
Sentiment: Strong Buy
"At 90nm there were 19 ARM foundries. At 65/55nm there were 16. At 45/40 there were 12. At 32/28nm there were 8. At 16/14nm there will be three. "
At 16/14nm there will only be two. Intel is not currently classified as an ARM foundry. Intel is a large ARM customer (top 10 largest) because they use ARM embedded into a couple of their controllers. IF Intel decided to take a couple of large foundry customers and work closely with them, Intel could integrate their ARM products into the Intel flow.
Well, I was trying to be generous by including Samsung, TSMC and GloFo. You may have excluded Samsung as not being a traditional foundry. And I didn't include Intel. Regardless, it's clear that the number of foundries has dropped dramatically and it seems likely that the trend will continue, especially given ARMs difficulties in attempting to transition to FinFET...
Sentiment: Strong Buy
Intel explained how the fabless industry would be transformed earlier but analysts were too busy worrying about the "death of the PC" to care. All the data suggests Intel's predictions were correct. There will be Intel and maybe two others capable of competing for top tier product. But one of the remaining two, Samsung, may be captive to their own products which leaves one foundry for everyone else. The non-captive foundry will have to behave like one to compete and that means it will have to significantly downsize the number of customers it serves, make far fewer designs and make their customers share far more of the risk and cost.
The ARM ecosystem will undergo a huge transformation, fabless companies will engage in bitter infighting for limited top tier capacity and the surviving companies will face higher costs.
Very useful data and a good reminder of Intel's earlier prediction.
Yes, absolutely great news on the fab front. I'm sure this will be evident to analysts over the next 12 to 18 months.
But in terms of Intel's stock price in the near to medium-term, I am looking for Intel to deliver some knockout punches at MWC and soon after.
1) Asus Android Tablet at $99/$149 (and Tablets/Smartphones from other OEMs)
2) Microsoft starting to adopt Intel for Smartphones
3) Data leaks on improved Haswell power/performance/graphics
4) Bay Trail - power/performance/graphics and possible pull-in to Q3 or Q4
5) Last but not least, Apple starting to adopt Intel for iOS
The URL works, I am just reviewing the presentation. Substitute period for DOT.
Some call-outs from the presentation:
1) At 16/14nm, they expect only Intel, Samsung, and TSMC.
Callout: If no other foundry joined the 16/14nm camp,
fabless companies would not be happy at all.
(That would be Qualcomm, nVidia, Apple - where are they
going to get their Snapdragons, Kraits, Tegras, and A#s made?)
2) Callout: The number of designs that foundry can support is declining as
process technology advances.
(Would foundries (only TSMC at 16/14nm) even take on companies that are
directly competing with Intel in the mobile SoC space?)
3) Callout: Only a few high-end chip makers today can even afford the exorbitant cost
of NEXT-GENERATION RESEARCH AND DESIGN, much less the fabs to
Semiconductor industry is going toward the post Moore’s law era.
Uncertainty increases and roadmap is no longer clear.
Only few companies have the capability to build beyond 20nm fab..
Moreover, few fabless companies can involve the products
development at this advanced technology node.
The relationship between fabless company and foundry become
sensitive. Foundry will only select the fabless companies with high
volume products and strong capabilities. Once fabless company
chooses a foundry, it will not change. Some kind of “binding”
between them will be established.
Design engineers should be manufacturing-aware and have indepth understanding
of the process variables. DFM and DFY are new challenges.
Software is important. Fabless company should be aware of the value of
IC vendor software.
China is the most important semiconductor market for both
domestic and global industry. Although there are many difficulties,
China will certainly welcome a new turn of fast growth in this area.
Check out the Chinese slideshow:
Sustainability Challenging Fabless Beyond 20nm
It pretty much spells out why the fabless model is not sustainable beyond 20nm with enough substantiating information to be quite credible ...
Sentiment: Strong Buy
Where do Apple, Samsung, Qualcomm and NVidia get their chips fabricated? Do they get them fabricated for free? What kind of volumes are we talking about? What's the capex budget for Samsung, TSMC, GloFo in 2013?
There a lot of money at play here and its very profitable being a foundry these days. The Chinese are talking about what they saw in the last generation before the growth in mobile chip demand. They tend to catch up with what's going on a little slower.