Real Countries Have Fabs

Persistent rumblings about the sale of IBM’s semiconductor unit might have seemed absurd a couple decades ago—before IBM sold off its PC unit to Lenovo and lost the gaming chip business to AMD’s x86 chips—but no one is scoffing at the possibility these days.

The reality is that IBM will never reach the volume necessary to be the No. 1 or No. 2 player in its segment. It’s not even in the top 10. Moreover, a refocusing on energy efficiency inside of data centers and more targeted servers—the so-called microserver revolution—will likely further erode IBM’s standing. In Gartner’s most recent worldwide rankings for semiconductor makers, IBM didn’t even make the list.

And as more and more processing is split between a variety of processors on SoCs and inside the cloud—including specialized processors for search being developed by Google and Facebook—IBM is looking at market prospects that point firmly down and to the right.

From IDM to fabless…to IDM, with a twist
The challenges of getting to the next process node on the semiconductor design side are enormous, but they’re even tougher on the fab side. IBM said as much back in 2006, when it joined forces with Samsung and Chartered Semiconductor (now part of GlobalFoundries) to defray costs and jointly develop next-generation processes—along with multi-sourcing options for customers of all three.

Since then, Samsung has ramped up its semiconductor production on all fronts, AMD spun off its manufacturing operations into GlobalFoundries, and TSMC emerged as the global leader with a fabless ecosystem. But as the price of developing advanced ASICs and SoCs continues to skyrocket, the cost of equipping fabs moves into the double-digit billions of dollars, and the production volume necessary to profitably maintain these kinds of operations continues to escalate, IBM’s future increasingly looks bleak.

In fact, IBM was noticeably absent from today’s announcement by GlobalFoundries and Samsung, which established a common 14nm finFET process-sharing and second-sourcing relationship that is almost identical to the to the one the Common Platform had for 28nm, 40nm and 65nm.

“We are providing common technology with a single PDK,” said Kelvin Low, senior director of foundry marketing at Samsung Semiconductor. “There has been an unprecedented change in the supply chain. This will apply to the 14LPE process, which is available now, and the 14LPP process, which is a follow-on, enhanced finFET device that will provide additional performance or power savings later this year.”

Ana Hunter, vice president at GlobalFoundries, said the agreement is focused on customer supply and technology availability—the development cost is so high that no one company can do it all—so customers can have chips manufactured in New York, Austin, or South Korea without having to change anything. “We are continuing to work with IBM on 10nm, but they are not part of this collaboration.”

What’s interesting about this deal is that the 14nm finFET technology is Samsung’s. In fact, Dean Freeman, analyst at Gartner, said the technology that GlobalFoundries was working on with IBM was floundering, and Apple was in the background demanding a second source for chips being built by Samsung.

Where’s IBM?
While IBM may continue developing some chips for its mainframe servers, the value of doing everything itself has all but disappeared—except as part of a system, where processors increasingly are less important than the middleware and power-management capabilities, which can be added on top of a base processor platform. The company may continue to design chips for internal use, but selling them as a commercial venture is less of a viable option because of the cost. This is an interesting turn of events, considering the entire PC market and ultimately the semiconductor market were the result of IBM stepping back in the first place to comply with a U.S. Department of Justice consent order in 1956. The consent decree was a way of allowing other companies into the computer business that was being monopolized by IBM, which used a “market-basket” combination of software, hardware and services to undercut its competitors.

What no one counted on, at least back in the 1970s, was that Moore’s Law would actually work for as long as it did. This famous observation, introduced by Intel co-founder Gordon Moore in a 1965 paper, has pushed the limits on manufacturing, optics, physics, and more recently, economics. But as Moore’s Law runs out of steam from an ROI standpoint, and as the cost of equipping fabs at the latest process node escalates into the tens of billions of dollars, this could well become nation-state-company types of endeavors.

There are three distinct models emerging. One involves the classic IDM-plus-foundry model, where the IDM takes in additional business. Samsung flirted with that idea earlier in the decade before extending it with the new agreement with GlobalFoundries. Intel has adopted a similar stance.

The second model is a pure-play foundry approach fueled by fabless companies, as exemplified by TSMC and UMC on one side, and ARM on the other, and fabless chip companies in between. The challenge going forward will be new systems companies such as Apple, Google and Facebook developing their own chips to their exact specs, leaving chipmakers battling for a socket facing cost pressures and power-performance challenges.

“The foundry model was built on the premise that they could get technology from an IDM and build second-generation devices by back-engineering the technology at the same process node,” said Gartner’s Freeman. “But the model evolved to the point where companies like Nvidia, Altera and Xilinx wanted it faster. But they also screamed when wafer prices increased.”

The third model appears to be a combination of nation-state/foundry-design, where the cost of equipment for an advanced fab won’t pay returns without additional revenue from the design through packaging supply chain. This model requires government support, which is why it’s not widely discussed. Uneven playing fields reach well beyond business into the political realm. However, IBM’s semiconductor business—and particularly the IP it has created—could be a key competitive advantage in this kind of competition, just as its PC business was for Lenovo and the Chinese PC industry.

Questions remain…and invite speculation
But if IBM does spin off its semiconductor business, how many other combinations are waiting in the wings? AMD spun off GlobalFoundries. But if GlobalFoundries were to buy IBM’s semiconductor business, which is the latest speculation, could another company buy both of them?

There is a possibility that IBM’s semiconductor business also could be broken up into lucrative parts, notably its 200mm fab in Essex Junction, Vt.,its 300mm fab in East Fishkill, N.Y., which develops its PowerPC processors, and its IP, which is used for manufacturing in Albany, N.Y.

But all of these carry big price tags, and the prices are only going up. In the future, only countries may be able to afford the most advanced fabs. And what that could do to the competitive landscape is anyone’s guess. It also puts a new twist on AMD co-founder Jerry Sanders’ famous comment, “Real men have fabs.” Not anymore, apparently.