Growing up: Breaking Moore’s Law

It is 40 years since Gordon Moore spelled out his famous law and it seems to be the common view that it will continue in force for the next 10-15 years. As a matter of fact, the clock frequency of microprocessors has not increased noticeably during the past two years. The improvement of microprocessor speed, however, used to be an important pillar of the economics of Moore's Law. Since about the mid-1990s there has been plenty of debate in the industry about the possible end of Moore's Law. If anything, the integration race has accelerated since then. We want to stress that we are not with those who are trying to call the end of Moore's Law altogether. Yet we would point to the speed limit that the industry has reached, which will have far-reaching consequences.

Smaller transistors have three advantages that have driven semiconductor companies to invest in the development of new production processes: 1) faster switching speed (clock frequency), 2) lower production cost per transistor and 3) more functions per chip.

One might also add reduced power consumption to this list, although this topic is a more complex one. In principle, smaller transistors enable lower power consumption. However, with smaller transistors also unwanted effects like current leakage increase, which have a significant negative impact on power efficiency.

For about two years one of the driving forces of Moore's Law has not been active. Despite continued shrinking of the transistor size the speed (as expressed by clock frequency of microprocessors) has stopped increasing. This is an important disruption. In the past, Moore's Law used to drive cost reduction and performance improvement. Today, Moore's Law is only about cost reduction. Higher performance has to come from other innovations that require additional R&D resources.

Implications for market growth
The industry needs to keep up innovation in order to keep up replacement demand. A slowdown in innovation would have a very negative impact on revenue potential. About 70% of this year's PC demand, which we estimate at around 200m units, is generated by replacements and only 30% by new users. If the slowdown in Moore's Law were to stretch the replacement cycle by one year, PC demand would only be around 170m units, or some 15% below original expectations.

Without any doubt the industry has to keep up innovation. The transition to multi-core processors, however, which is currently being pursued by Intel and AMD, is no solution, in our view. Multi-core processors deliver higher performance but, unlike in the past, only for significantly higher costs (due to the higher chip size).

We expect companies to focus less on advanced process technologies and more than they did in the past on innovative chip design. The often inefficient circuit design still holds ample opportunity for improvement in chip performance, without adding die size or other extra costs. We already see evidence for this trend, as described in the following.
Implications for R&D Intensity
The R&D intensity of chip suppliers has already started to increase, reflecting the increasing process complexity and the need for innovative circuit design. We believe this trend is likely to continue. The R&D intensity of equipment companies is about stable or even declining.

Less efficient R&D
As long as Moore's Law was fully in force the allocation of R&D resources was clear. Every dollar spent on the development of smaller transistors was a dollar well spent, as smaller transistors resulted in cost reductions and higher speed. Higher chip performance through innovative chip design came on top. Today, higher circuit speed can almost only be achieved by new design architectures. This adds to increasing process R&D. Surging R&D costs are the consequence.

Rising R&D costs posed no problem for as long as industry revenue growth kept pace with R&D progression. From today's perspective, however, one can say that the industry started to mature in the mid-1990s, when industry growth slowed down. During the past 10 years the industry has grown at an average rate of 5% p.a. R&D costs during the same time frame have risen by more than 10% p.a., roughly unchanged from the average annual increase during the high-growth phase of the industry. R&D intensity (as a percentage of revenues) has increased from 9% in 1995 to almost 16% in 2005E.

Shift from equipment suppliers to chip manufacturers
Interestingly, the R&D intensity of equipment suppliers shows a completely different trend. Whereas the R&D of chip manufacturers has grown fairly steadily over the past two decades at an average rate of 12% p.a., R&D budgets of equipment suppliers surged between 1987 and 1995 by an average 21% p.a., but have increased by only 7% p.a. between 1995 and today. In fact, equipment companies have reduced R&D spending since 2000. Thus, industry R&D is shifting from equipment suppliers to chip companies.

In the short term, the shift of R&D activity from equipment to chip companies might appear positive for the equipment sector, as this trend has clearly supported the margins of equipment companies during the past few years. Longer term, however, we are concerned about the value position of the equipment sector if innovation is shifting to chip producers.

Who can afford rising process R&D?
Process-related R&D accounts typically for around half of the total R&D costs of a semiconductor integrated device manufacturer (IDM). During the past 10 years the development costs of new process technologies have increased on average by 28% from node to node, with the rate of increase coming down, but only very moderately. Given the two-year technology cycle described by Moore's Law, this translates into a 13% annual increase in process R&D. Thus, process R&D has increased roughly proportionally to total R&D costs.

Going forward we expect the share of application- and design-related R&D to increase more rapidly. This will be necessary, in our view, to maintain a high level of innovation, given that product-level innovation from process improvements alone is slowing down. This could weigh on the financial ability to spend on process R&D, raising the question of who can still afford rising process R&D.

Intel during its last analyst day explained at length that its investment in developing the 65nm technology was paying off. This may hold true for Intel but not necessarily for the industry as a whole. Revenue scale is one important part of the equation in determining whether it is worth carrying R&D costs, which for the 65nm technology amount to more than $1bn.

During the past 10 years the industry has introduced a new technology node about every other year. For a company the size of Intel it is economically sensible to maintain this two year technology cycle, as the savings in manufacturing costs more than make up for the investments in process R&D. For a smaller company with lower absolute manufacturing costs this looks different. Here, costs increase with shorter technology cycles. A small company has lower absolute manufacturing costs, and therefore savings in manufacturing costs are smaller and are not sufficient to make up for high R&D costs.

We estimate that the critical size required to afford a two-year technology cycle stands currently at about $7bn revenue p.a. This critical size is increasing over time with rising process R&D costs. Thus, we believe that in future leading-edge technologies may be affordable only for a very small number of companies. R&D partnerships help to delay the problem but do not solve it.

Of course, companies have responded to increasing R&D costs through partnerships. However, the figures above show that partnerships are having to span a rapidly increasing revenue volume. There are few partnerships in which CMOS revenues exceed the $7bn required to cover development costs. Partnerships help to delay the problem but do not solve it.

We expect the technology gap between a very few large chip manufacturers (we think fewer than five) and smaller followers will grow rapidly during the next few years. With a shrinking group of customers buying leading-edge equipment, though, we believe it will become increasingly difficult for equipment suppliers to generate attractive returns on R&D investment in new equipment.

What does all this mean for the capital intensity of the industry? Will it continue its decline of the past few years? In contrast to the expectations of some industry observers we think this trend will continue, as innovative chip design becomes more important than leading-edge production.

Declining capital intensity
Until the mid-1990s the chip market was growing at an average rate of more than 15% p.a. During that time semiconductor manufacturers on average invested about 22% of sales through the cycle. Since then industry revenue growth rates and capital investments (as percentage of revenues) have been declining. There currently appear to be two schools of thought about the further development of capital intensity. The pessimists of the equipment industry think the decline in capital intensity will continue, with all the negative implications that this carries for the equipment industry. The optimists expect average capex-to-sales ratios to bounce back to the above-20% levels of the high-growth phase of the industry. These people cite increased production complexity and larger die sizes (introduced by innovative designs such as multi-core microprocessors) as reasons for higher capital requirements.

We take a very cautious stance on the outlook for capital spending. While we acknowledge that the transition to multi-core processors may result in an interim higher spending level by selected customers, we think this will be outweighed longer term by the structural changes in the industry: slowing industry growth, continued transition to 300mm, the slowing of the technology cycle, and changes in the economics of semiconductor cost reduction.

Reasons to reduce capital intensity pressure to generate cash
Lower capital intensity of the industry is showing in improved cash generation. The FCF margins of chip manufacturers have been improving since about the mid-1990s. Until about 1995 chip companies tended to reinvest all operating cash flow, as is characteristic of a pure growth industry. Since 1995 cash generation has improved continuously, driven by lower capital spending. The semiconductor sector now has a typical free cash flow yield (over enterprise value) of between 5% and 8%. This appears to be a level with which investors feel comfortable at current industry growth rates. Should industry growth slow further, chip companies would have to increase FCF/sales ratios. This would put additional pressure on capital spending.

The industry transition to 300mm is far from complete yet. We estimate that around 35% of this year's equipment shipments are still for 200mm lines. This figure will probably drop to below 20% in 2006.

As 300mm equipment offers about twice the capacity for only 30-40% higher costs than 200mm equipment, the ongoing trend to 300mm should further reduce capital intensity.

As described above, we see the lack of clock frequency improvement by new process technologies and increasing process R&D costs as severe risks for the economics of Moore's Law. This might well result in a slowdown of the technology cycle for most companies, even if a small group of the largest chip manufacturers can maintain the current two-year cycle for a few more years. An extended technology cycle requires less capital investment in leading-edge equipment and, therefore, reduces capital intensity.

Economics of chip manufacturing - the bigger picture
Over the past 30 years, production costs for semiconductor chips has decreased on average by 27% p.a. Consumers and manufacturers of electronic equipment have got used to this trend and it should not change going forward if a slowdown in end-demand is to be avoided. An analysis of contributors to reduced semiconductor costs shows that, in future, a 20-25% p.a. cost reduction will be possible only if equipment costs (per wafer start) increase by only about 5% p.a., well below the 17% annual increase of the past.