A little more than a week ago, Intel announced that Pat Gelsinger, its former chief technology officer and former manager of the predecessor of its Data Center Group as well as the current chief executive officer at server virtualization juggernaut VMware, would be returning to the world’s biggest chip maker to be its CEO and to take on the task of rescuing the company from itself and its competition.
Intel reports its earnings tomorrow, on January 21, and is expected to give an update on its delayed 7 nanometer chip making processes for future Xeon server CPUs and future Xe GPUs and its beleaguered but now in production 10 nanometer technologies for its current Xeon server CPUs. There is a possibility that outgoing CEO Bob Swan will talk about Intel’s plans to farm out chip making to rivals Taiwan Semiconductor Manufacturing Corp and Samsung Electronics, the other two foundries that can actually deliver 7 nanometer and 5 nanometer processes for those who design chips and do not have the wherewithal to have their own foundries. That would be everybody but Intel and Samsung in the entire world right now, by the way. And maybe not either in the long run.
Gelsinger takes the helm at Intel on February 15, and that does not give us very much time to play armchair quarterback with some ideas about what Gelsinger should and should not do once he takes control of Intel. We think that Peak Intel happened a few quarters ago, but that the company will be throwing off enough cash to have plenty of operating room to make some big changes in the coming weeks and months. It remains to be seen how daring Gelsinger can be – or will be, given his three decades of experience as an engineer and top executive at the chip maker.
Before we get into all of that, it is probably helpful to remember that Intel has faced some pretty tough times in the past, and came through them and ended up stronger because of it. The best example is Intel getting out of the DRAM memory business in the mid-1980s. Intel had been a DRAM memory maker since 1970 – it was the first supplier of this technology in fact – and went public in 1971 in part due to its strength in this business, which was as important to Intel as its microprocessor business. Intel had dominant market share in DRAM in the 1970s, but by the 1980s, the Japanese memory makers ramped up DRAM production and lowered the pricing boom. Many companies were driven out of business between 1984 and 1986. It just kept getting worse, but Intel had effectively decided to exit the DRAM business already because its fabs invested in a manufacturing technology aimed more at microprocessors than memory. So by the time Andy Grove, one of Intel’s founders, took over as CEO in 1987, DRAM didn’t really matter to Intel so much anyway. It was an easy decision based on the numbers, but not easy emotionally for Intel, which thought of itself as a memory maker that also did logic.
The market made the decision for Intel, based on its performance against what was perhaps unfair competition – certainly overwhelming competition – but it was Grove who helped give Intel its new identity as a CPU designer and manufacturer. The turnaround was quick, all things considered, but that had more to do with the burgeoning PC market in the late 1980s than it did anything else. It didn’t hurt either Intel or Microsoft that IBM didn’t take its own PC business very seriously and gave control to these two companies before it really understood what it was doing. (Remember that idea for the analysis below.)
The key thing for Gelsinger to do – and that Intel’s founders and managers did successfully many decades ago – is to honestly and brutally assess what is key to Intel’s future and what is not. Grove may have said that “only the paranoid survive” as a way to motivate Intel to focus on competition, but as Intel’s might in the datacenter has grown, its paranoia has kicked into overdrive and it has been a bit acquisitive without being as focused as it needs to be. This happens to all big companies that create mini monopolies, as IBM and Microsoft have done before Intel.
Such an assessment as we are suggesting that Gelsinger do would start with the obvious: Intel is not the top notch chip fab that it once was.
Since its inception, Intel has been as much a manufacturer as it has been a chip designer and marketing organization. Serious lags in Intel’s chip manufacturing processes – the 10 nanometer process is four years late if you want to be generous to Intel, and its 7 nanometer process now looks like it will be at least a year and a half to two years behind TSMC and Samsung. This foundry business eats an enormous amount of money, but ironically, Intel is not spending enough on it. In the September quarter last year, Intel’s top brass said that it expected to spend somewhere between $14.2 billion and $14.5 billion in 2020 on capital expenses, the vast majority of which is for its foundries. Through September, in an effort to prop up its share price and therefore the wealth of its biggest shareholders and its key executives who get big chunks of shares as part of their compensation, Intel shelled out $14.2 billion in cash for share buybacks against $10.4 billion in capital expenses.
Throwing money at problems does not always solve them, but not throwing money at problems, or even good running businesses, creates problems in the long run. By contrast, TSMC recently announced that it would be spending $28 billion to build out factories to support 7 nanometer, 5 nanometer, and 3 nanometer chip manufacturing in 2021. That’s a little less than twice what Intel is spending, and the good news is that Intel has been able to rake in the cash all those years because what it can charge for a chip to an end user in a device is a lot higher than what TSMC can charge to a customer who in turn will sell it to device makers. Intel will make around $70 billion selling chips in 2020 against that $14.5 billion in foundry investments, not all of which will be used in the manufacturing for that year of course. TSMC will make maybe somewhere around $52 billion against $28 billion in foundry investments in 2021 and again, most of TSMC’s chip sales are not against its most advanced nodes. But its most profitable ones probably are. But the run rates of chip revenues versus capital outlay are illustrative.
Intel’s strategy of the past decade or so works – right up to the moment that it doesn’t. Now Intel has promised to keep Wall Street happy with share buybacks at a time when it should have spent years perfecting its chip etching processes and building the foundries to compete with TSMC and Samsung. We have no idea what Intel did wrong to get into its 10 nanometer bind, but the cost will be coming home to roost now, and we shall see if Intel can pull the rabbit out of the hat with 7 nanometer and 5 nanometer processes.
Maybe what Intel really needs to do is spin off the foundry completely. Maybe that is what the United States government and what the chip industry in the United States needs, too. Or, maybe it is too late for that.
Back when IBM went up on the rocks in the early 1990s and Lou Gerstner came in from American Express to be Big Blue’s first non-IBM CEO, he must have looked around and saw a company that was more messed up than either Wall Street or Main Street thought, and the only thing he could do was to hold it all together and pivot. The IBM pivot was to stop protecting its server businesses and open itself up as a supplier of systems software and services for all platforms, IBM or otherwise, and through the magic of an outsourcing boom and then the dot-com boom that drove a slew of services into large enterprises, Gerstner got lucky. IBM ended up financially engineering itself into a hole, much as Intel has done in past years, and has sold off too many businesses to list and is trying to re-invent itself with its Red Hat acquisition.
A foundry spinoff could work, and a lot depends on the pricing that Intel would get from the spinoff to make Intel’s chips. If it is too high, then the remaining Intel would not show huge profits, and make it too low and the foundry could not, either. But in the current economic and political climate, we believe that there is some interest in having a truly indigenous and open foundry in the United States as a counter balance to TSMC and Samsung. It really depends on how much competition we want, how much control we want, and how much risk we are willing to take with only two advanced foundries if Intel decides to drop back and punt like GlobalFoundries did in August 2018 when it spiked its 7 nanometer development efforts.
This is one very real possible pivot for Gelsinger to steer. The US government is plenty nervous about not having advanced chip making inside the country, and it might even be willing to help fund the creation of an indigenous foundry operation based on Intel’s fabs as the foundation. Intel and others have been courting a $25 billion investment from Uncle Sam since last year, but this would be more than providing matching funds for companies who build a fab in the United States. The SEMATECH consortium was formed in the late 1980s and funded for over five years with $500 million, which is about $1.14 billion in 2020 dollars, to do the fundamental research to create the equipment that makes chips. What we are talking about would be a much bigger investment for an indigenous foundry, and we think that TSMC and Samsung could also get help in putting their foundries here, too. Ditto for Europe. There is no good reason that Europe should also not have some capability for advanced chip manufacturing. Having said all of that, it is not really the job of the US government to explicitly help corporations fix their problems. This has to be a big enough national security risk to warrant the investment, and we are not qualified to make that call. We don’t know who is, either.
The other option is more of a nuclear one, and it is to do what AMD and IBM did: Sell off the foundries directly and walk away. There are only two potential buyers, TSMC and Samsung, unless the investors in GlobalFoundries have a lot of money sitting around and they think they can use Intel’s foundry operations as a means to get back into the advanced process game. None of these options seem likely to us.
Here is another crazy pivot for Gelsinger to consider. There are rumors that Intel is going back into the Arm client and server chip market. Why not go all the way and try to outbid Nvidia for Arm Holdings? (We told you it was crazy.) We think the odds for this pivot are very, very low indeed. Intel would probably have more antitrust issues buying Arm Holdings than Nvidia is now having in the United Kingdom and China, and a bid by Intel might actually improve Nvidia’s chances of closing the deal.
Intel could go in the exact opposite direction, and that would be to build its own public cloud. Look at it from Intel’s point of view. The top eight buyers of its chips are all tails wagging the CPU manufacturing dog, and each and every one of them has some sort of plan to introduce Arm servers into its infrastructure. If the hyperscalers and public clouds are thinking vertically, and at some point they are going to represent maybe 65 percent of shipments, and if the OEMs are going to be under increasing pressure because of the static or dwindling compute capacity needs of their customers, then maybe Intel should be thinking vertically, too. That would mean becoming a server manufacturer for realsies, not just for special bid deals, and using those servers in its own infrastructure cloud.
The trouble with this Intel pivot is that the investment to build a public cloud on the order of that put together by the Super 8 would be absolutely ginormous. And for the X86 workloads the big clouds need to support for their customers, those big clouds might be driven right into the waiting arms of AMD.
Building a public cloud from scratch would take tens and tens and tens of billions of dollars. Intel can get its CPUs and network ASICs on the cheap (provided its fab spinoff doesn’t gouge it on the prices) but it can’t bend metal, or buy memory, or buy flash (now that it has sold off its flash business) any cheaper than an ODM operating at volume. Building a public cloud is a very tough business in 2021. And people now expect more than just the raw infrastructure services that Intel could probably provide well with an acquisition or two. Intel could not build the vertical systems and application software that AWS, Microsoft, and Google have done and that Alibaba, Baidu, and Tencent are doing to varying degrees.
There is a higher probability that Facebook will start offering cloud services based on its technology stack, and that seems like pretty miniscule chance at that. If we were running Facebook’s infrastructure, we might consider it as a way to lower costs for IT for the Facebook app. But that would be a ten year plan and it might come to nothing because AWS is creating a modern, complete platform and Google and Microsoft are trying to do the same.
The lesson of these thought experiments above is that sometimes, you cannot buy or sell your way out of your problems. You have to dig in, figure out who you are and who you want to be, and get on with that. This is how IBM, Microsoft, and Apple have all successfully done their pivots, and to a lesser extent, how Hewlett Packard Enterprise and Dell have stayed relevant if not precisely hugely profitable. (Someone has to build systems and sell them, and we thank you for that.) The trick is to not emulate the past in the strategy. HP (when it didn’t have the E but it did have PCs and printers) and Dell had Big Blue envy after Gerstner did his pivot for IBM to software and services. They bought all kinds of businesses that they then sold off. Dell did it again buying EMC and is now apparently trying to spinout VMware to unlock its potential value on Wall Street.
If Gelsinger does intend to stay in both the compute engine design and manufacturing businesses, we have some suggestions. First, decide if Intel is going to be one or the other or both as it has been historically. If both, then stop blowing cash on share buybacks and actually invest heavily in the fab business. As we come to the end of Moore’s Law as we have known it, this is a harder pivot but one that can be done with advanced packaging. Just like someone has to make the servers, someone has to make the CPUs, GPUs, FPGAs, and so on. Offload some manufacturing in the short term to TSMC and Samsung to get things back in order, but get a process and packaging roadmap together and drive it.
Also, Intel needs to stop buying its way out of problems and also rationalize its compute engines. Intel has a lot of competing and overlapping products that can be used as accelerators that offload work from CPUs as well as CPUs that onload this work. Pick a single strategy that makes sense: CPUs only do X, Y, and Z serial work and all acceleration gets done on A, B, and C devices always and linked to the CPUs in absolutely consistent ways. Make this easy. Or at least easier. Nvidia is going to do that, particularly if it launches its own Arm CPUs, as we expect it will. IBM confused customers for decades because it refused to say which of its servers should be used when, and when asked by customers which one they should buy, the attitude was, “Well, how much money you got?”
Don’t just explain that Intel has a lot of compute engines in terms of capability and style. Show system architects how to use them best, and when and where. And also show customers what engines not to use for specific workloads and situations. This information is just as valuable. And if some of these engines are not really panning out, spike ‘em. Write ‘em off. No one will remember it any more than they do Intel’s DRAM from the 1970s and 1980s. Well, unless they are history buffs like us.
One last thing. If nothing seems to be working, change the conditions of the game entirely. Open source the X86 instruction set and the Core and Xeon chip designs and let all hell break loose. . . . That’s not really a business plan, though. That is the true nuclear option for Intel, and one that only makes sense if it is committed to having the world class manufacturing operation it once had. Remember: Someone has to make the chips, and it might as well be Intel one third of the time.