news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

editor's note: intel has a glorious past. no company is more important than intel, and no company has defined silicon valley in every way like intel. but despite strong support from the united states, this old technology company that missed the technological wave of the past decade is still acting like a hopeless loser. this article suggests that splitting up intel may be the only chance to save the chip giant. the article is compiled.

this is indeed a valley:

with the santa cruz mountains to the west and the diablo mountains to the east, the once sleepy city of mountain view sits in the middle of the santa clara valley. in 1955, when william shockley, one of the inventors of the transistor at bell labs, returned to neighboring palo alto to care for his ailing mother, the land in mountain view was the u.s. navy’s moffett field.

shockley was convinced that silicon was the superior material for making transistors (bell labs focused on germanium), but because of the distance from new jersey and the rough management style, shockley could not hire many colleagues from bell labs, so he founded shockley semiconductor laboratory with a group of young scientists in mountain view in 1956. just one year later, eight scientists led by robert noyce and gordon moore left shockley (he was really a bad manager) and founded fairchild semiconductor in neighboring sunnyvale, a new division of fairchild camera and instrument company.

it was fairchild semiconductor that gave the birthplace of the technology industry the other half of the name silicon valley: yes, we say "the valley", but at least when talking about technology,we mean silicon valley.from techcrunch in 2014:

as fairchild grew, employees began to leave the company and start new spin-offs. many of these companies grew just as quickly, inspiring other employees who remained at the company... the growth of these new companies began to reshape the region. in just 12 years, fairchild co-founders and former employees founded more than 30 spin-offs and provided financial support for many more. by 1970, chip companies in the san francisco area employed 12,000 people...

the success of these companies eventually caught people’s attention. in 1971, a reporter named don hoefler wrote an article about the success of computer chip companies in the bay area. the companies he described all used silicon to make chips, and they were all located in a valley south of san francisco. so hoefler combined these two points and created a new name for the area: silicon valley.

hofler’s articles and the names he coined are well-known, but there’s a key aspect of his analysis that’s often overlooked: almost all of the silicon chip companies he profiles can ultimately be traced back to fairchild and its co-founders.

still, despite fairchild semiconductor’s enormous success, no company has been more important than intel, and no company has come to define silicon valley in as many ways as intel. arthur rock, who had helped the so-called “traitorous eight” found fairchild camera and instrument, bankrolled intel, and in the process established the compensation structure that would come to define silicon valley. gordon moore developed a roadmap for intel (often called moore’s law)—it “predicted” that the number of transistors would double at a certain rate, which would both increase the speed of computing and make it cheaper; “predicted” is in quotes because moore’s law is not a law of physics but a law of economics, the inevitable result of intel’s relentless improvement. in broad terms,that means intel sets the pace for innovation in all things technology, from making processors for personal computers to the much-underestimated cloud computing disruption of the early 2000s and defining the expectations of every software engineer around the world.

intel's long decline

unfortunately, the need to support intel has always been obvious:in may 2013, i wrote that intel needed to build a foundry business because its mobile business was failing and the economics of its idm business (integrated device manufacturer model) faced long-term challenges.

unfortunately, intel not only didn’t listen, but its business also suffered: intel struggled when trying to move to 10nm in the late 2010s, in part because they were unwilling to adopt the much more expensive euv lithography process, ceding the performance crown to tsmc. meanwhile, intel’s chip design team grew increasingly bloated and lazy, thanks to intel’s once industry-leading processes, and began to fall behind amd; now amd not only has better designs, but also better processes because they have tsmc making their chips. meanwhile, the rise of the hyperscalers meant that some entities had both the scale to prove they could wipe out intel’s software advantage, and the resources to do so; the result is that amd has been eating into datacenter share for years and is about to pass 50%:

[note: these two paragraphs are technically incorrect, as amd’s data center revenue includes ai chips; the trends still hold, but the author regrets the error]

this chart actually underestimates the problem because it only includes x86 processors; in fact, the same capabilities that allow hyperscalers to take advantage of amd's increasingly superior total cost of ownership are also being used to develop arm-based server chips. amazon in particular has invested heavily in the graviton series of chips, taking advantage of arm's theoretically higher efficiency and lower licensing fees (compared to intel's profit margins).

beyond that, what’s particularly problematic (and why intel’s data center revenue actually declined year over year) is that more and more data center spending is going toward ai, and intel is missing out on yet another wave.

the story that intel (or at least past management) would like you to believe about mobile is that they foolishly passed on the opportunity to supply apple with the iphone, not realizing that the sales would more than make up for the lost profits. in fact, tony fadell once told me that while steve jobs wanted to use intel (apple had just switched to intel chips for mac computers at the time), intel chips were not competitive:

it seems to me that in the mid-2000s, intel's thinking was,“we can take something from a desktop and repackage it for a laptop, and then repackage it for an embedded device.”this reminds me of the windows saying, "i'm going to do windows, then i'm going to do windows mobile, then i'm going to do windows embedded." use the same kernel and core, just streamline it...

intel's mindset was never "which one are we going to be?" when they went through the cisc-risc route, they chose cisc, which was the right decision at the time. if you fast forward, you will find that they also made another decision. they abandoned architecture and moved to more manufacturing. at that time, their idea was "don't worry about all these different product lines meeting all these architectural needs.just let moore’s law take over,” so to some extent it locks you into one path, which is why intel, not until the pat era, but before the pat era, was driven by manufacturing capabilities and laws.intel is not driven by architectural decisions, it's driven by "we have these things now, we're going to propagate these things, and we're going to continue to reuse these things."

in fact, this goes back to the days of pat, when ceo pat gelsinger-gannan first started at intel. he was the one who pushed for cisc over risc, arguing that intel's cisc software advantage combined with the company's superior manufacturing capabilities would ensure the company's dominance in microprocessors. and, as fadell points out, it worked, at least in pcs and servers.

intel has not succeeded in mobile: intel has not been able to leverage its manufacturing capabilities to make x86 compete with arm, especially since the latter has a head start in software; intel has also not succeeded in gpus, where intel spent years trying to build x86-based gaming chips that were originally built on intel's manufacturing capabilities. of course, gpus are the foundation of today's wave of ai enthusiasm.although intel acquired gaudi to provide ai chips, they have not made any progress in the market—by the way, the gaudi chip is manufactured by tsmc.

IDM 2.0

nothing about this story is new; i told it in my 2021 article, “the problem at intel.” my solution then—shortly after kissinger returned to intel and 15 years after he was ousted as ceo—was that the company needed to be broken up.

for decades, intel has been committed to integrating design and manufacturing, but this integration has become a straitjacket between the two businesses. intel's design is hampered by the company's manufacturing capabilities, and its manufacturing has a lack of momentum.

the key to understanding the chip industry is to understand that chip design is much more profitable. for example, nvidia's gross margins are between 60% and 65%, while tsmc, which manufactures chips for nvidia, has a gross margin closer to 50%. as mentioned above, intel's margins have traditionally been close to nvidia's due to its design and manufacturing foundation, which is why intel's own chips will always be the priority of its own manufacturing department. this will mean that potential customers will only receive less good service, and there will be less willingness to change its own manufacturing methods to accommodate customers, and to absorb best-in-class suppliers (because it will further reduce profits). there is also a trust issue here: will companies competing with intel be willing to share their designs with competitors? especially considering that the competitor is incentivized to prioritize its own business.

the only way to solve this incentive problem is to spin off intel's manufacturing business.yes, it will take some time to develop the customer service components required to work with third parties, not to mention the vast library of ip development modules that would make working with a company like tsmc (relatively) easier. but independent manufacturers will have one of the most powerful incentives to make this transition: the need to survive.

two months later, gelsinger announced his turnaround plan: idm 2.0. intel would spin off its manufacturing business into an independent unit that would provide services to third parties, but the unit would remain under intel. gelsinger told me in an interview that this was the only way intel could remain competitive in the chip field while continuing to invest in cutting-edge technology; after all, amd got into trouble after spinning off globalfoundries, and only got out of trouble after breaking its purchase agreement with globalfoundries and turning to tsmc, while globalfoundries gave up its leading edge.

gelsinger is persuasive and optimistic, and i have had faith in him for the past three years. suddenly, though, talk of a spinoff is back on the table; from bloomberg:

intel is working with investment bankers to help it through the most difficult period in its 56-year history, according to people familiar with the matter.the company is discussing options, including splitting its product design and manufacturing businesses, and which factory projects might be canceled, said the people, who asked not to be identified because the discussions are private.

intel could spin off or sell the foundry unit, which would represent a radical shift for chief executive pat gelsinger, given that the unit is designed to make chips for outside customers. gelsinger sees the business as key to restoring intel's own position as a chipmaker and hopes it can eventually compete with companies such as tsmc, a pioneer in the foundry industry.

as the article points out, intel is likely to consider some less radical measures first; reuters reported that these ideas included selling businesses such as altera's programmable chip business and reducing capital spending, including canceling a proposed foundry in germany. the company also ended up canceling its dividend and laying off 15,000 people, which, to be honest, wasn't enough; i pointed out in last week's update:

intel had 124,800 employees last year; tsmc had 76,478 and amd had 26,000, meaning both companies have fewer employees than intel but make better x86 chips, truly competitive gpus, and oh, yes, they also make chips for all the companies in the world, including apple and nvidia. cutting 15,000 jobs is too little, too late.

this passage summarizes the fundamental problem facing this company:

intel doesn't have the best manufacturing capabilities

intel doesn't design the best chips

intel is out of the game in artificial intelligence

furthermore, the future doesn't look bright; there were three issues with intel's recent earnings call:

technically, intel is on track to achieve the five nodes promised by kissinger in four years (two of which are iterations, in fact), but they haven’t scaled any of them; the first attempt was intel 3, and the profits were wiped out. this is not surprising: it’s hard to skip these steps not just because the technology is advancing, but because you have to really learn how to implement new technologies at scale and get sustainable benefits. back to intel’s 10nm failure: the company was technically capable of making 10nm chips, but it couldn’t do it economically; now the problem with intel 3 has not been solved, not to mention the 18a promised next year.

intel is pushing its lunar lake architecture hard because it's the only design the company can compete with qualcomm's arm architecture, which microsoft's copilot+ pc plans rely on;the problem is that the lunar lake architecture (including its cpu) is manufactured by tsmc, which is both embarrassing and has a huge impact on profit margins.

the third problem is that the goal that kissinger has been working hard to achieve is the aforementioned 18a, but intel has not yet released a real partner for scale. yes, the company is still negotiating with many companies and claims to have reached some secret agreements, but the current foundry strategy needs real evidence; unfortunately, even if intel has lost control of costs, it is also increasing its inclination towards tsmc, and it is difficult to convince any third party that it should leave its fate to intel.

to sum up, my initial reaction to intel's earnings decline was to defend kissinger; intel's current situation is the result of mistakes kissinger made in the first few years of his return to the company. this is still true, but kissinger does have a fatal flaw: he still believes in intel, but i no longer do.

market reality

intel, and by extension, the fundamental problem facing the us dream of controlling advanced production capacity is here: intel foundries have no reason to exist. apple, nvidia, amd, and other leading fabless chip companies all rely on tsmc, so why doesn't intel go to tsmc? tsmc has invested in euv, surpassed intel, and continues to spend tens of billions of dollars each year to advance 2nm and higher processes. yes, tsmc's 3nm pricing is too low, but even if the company raises future process node prices as i predict, its relative cost is far less important than tsmc's excellent customer service and reliability.

the point is that the smartest decision for intel's own chip division (as they did with lunar lake) is to hand over manufacturing to tsmc. intel still has an advantage in pcs and still dominates local and government data centers, but the best way to exploit these remaining areas of advantage is to have tsmc make their chips for them.

that’s why gelsinger does have a reason to keep the company intact; intel’s foundries need volume, and the easiest way to get it is from intel itself. however, by definition, this decision was not made out of what was best for intel’s theoretical fabless business, but to restore intel’s manufacturing capabilities, even if that manufacturing capacity was largely driven by satisfying intel’s chip business at the expense of external customers.

kissinger’s trump card is tsmc’s geographical location.in fact, kissinger has been very clear on this point; excerpted from cna english news in 2021:

intel ceo pat gelsinger said wednesday at fortune’s brainstorm tech summit in california that the u.s. government should support the creation of a sustainable semiconductor supply chain in the united states, in part because “that place is not stable”…

kissinger also expressed the same view in that interview when he explained why intel could not be broken up:

in my opinion, we have to look at this almost from a global perspective because i believe deeply that the west needs a world-class technology provider, and i don't think intel can survive for many years by splitting it in two, and it may be difficult for intel to survive long enough to become that world-class technology provider. remember, i am determined to solve this problem, and i think we are on the right path to solve this problem, given the cash flow, the r&d flow, and the products that allow us to drive this. so for these three different reasons, we chose the idm 2.0 path, but it's not because we didn't consider alternatives, but partly because we considered alternatives.

it is at this point that all those investing in american manufacturing must speak out.if the u.s. government and u.s. technology companies do not want to rely on tsmc, they will have to pay the price directly.yes, the chips act passed, but while intel received a ton of money, it needed a lot more—and that money would come at the cost of a smarter incentive structure that would allow intel to excel.

i suggested back in 2021 that pre-order commitments rather than subsidies be the way to go, and i am even more convinced now that this is the only viable path.

that’s why the federal subsidy program should function as a purchase guarantee: the us will buy a quantity of us-produced 5nm processors at price b; c quantity of us-produced 3nm processors at price d; e quantity of us-produced 2nm processors at price f; etc. not only will this give new intel manufacturing spinoffs something to strive for, it will also incentivize others to invest; maybe globalfoundries will get back in the game, or tsmc will build more us fabs, and, in a world of nearly free capital, maybe finally a startup will be willing to take the leap.

the world of free capital no longer exists, and it’s probably not realistic for a startup to figure out how to make the most complex device ever created by humanity; the best idea right now is a new company with the expertise and starting point of an intel foundry. but crucially, it shouldn’t be beholden to x86 chips, shouldn’t have hundreds of thousands of employees, and shouldn’t have the cultural advantages that once led the computing world. the best we can do is buy guarantees — hundreds of billions of dollars over the next decade — and pray that someone will let such an entity grow independently.

in summary, there is no market reason for intel to become a foundry; this is not a market failure in a purely economic sense, but the extent to which the us national security agencies believe that intel's failure as a foundry is detrimental to national security, the united states must pay a corresponding price to make this happen.and if the u.s. is willing to pay, that means giving the foundry the best chance to become self-reliant in the long run. that means effectively taking orders from apple, nvidia, amd, and fabless intel. the tech world is no longer reliant on intel; the only chance the u.s. has to have to lead in cutting-edge manufacturing is if it does too.

translator: boxi.