Intel Honesty

Intel is trapped. Its debt repayments alone are massive. The poor performance in stock has encouraged 20years of wage stagnation to the point where you can literally earn over 50% more at amd or nvidia or even an startup for equivalent roles so they aren’t hiring the best. Their pay in the Bay Area is double tsmc in taiwan for equivalent roles in raw dollar terms but the ppp differences means your better off working for tsmc in Taiwan than intel in the Bay Area. That’s not a joke. Intel are literally incapable of attracting talent from Taiwan right now.

They don’t have the talent they need and the debt trap and poor performance means a lot of push back to the needed doubling of wages to attract that talent. It’s a very hard sell for any exec trying to correct this problem. They sidelined lip bu tan who was one of the advocates for even more layoffs and wage freezes but he’s one of many backwards thinkers they need to remove. It’s going to be difficult to fix their board.

Without talent intel has no hope of winning and they can’t get that talent due to poor stock performance for the past 20years leading to executives and shareholders wishing to implement the opposite of what they need right now. In fact they have ongoing layoffs right now. A true downward spiral and the only real hope is for a newcomer to step up.

> in the late 2010’s Intel got stuck trying to move to 10nm, thanks in part to their reluctance to embrace the vastly more expensive EUV lithography process

TBH, it's easy to say this with the benefit of hindsight. Throughout most of the 2010s, EUV lithography was like the "year of the Linux desktop" - i.e., this year will be the year where EUV was suitable for high volume manufacturing. I don't really blame Intel for deciding to go with self-aligned quadruple patterning for 10 nm, but combining it with cobalt interconnects was probably biting off more than they could chew.

FWIW, Intel was funding EUV R&D since 1997: https://www.intel.com/pressroom/archive/releases/1997/CN0911...

That press release had an interesting prediction:

> Intel projects that the microprocessor of the year 2011 will contain one billion transistors, operating at over 10 gigahertz and delivering 100,000 MIPS (millions of instructions per second).

They weren't that far off in estimating the transistor count and MIPS: the i7-2600k released in January 2011 had 1.16 billion transistors [0] and delivered 117k MIPS [1] @ 3.4 GHz. The clock speed prediction was way off due to the failure of Dennard scaling in the early-mid 00s.

[0] https://www.anandtech.com/show/14043/upgrading-from-an-intel...

[1] https://en.wikipedia.org/wiki/Instructions_per_second#CPU_re...

I like the proposed approach of purchase guarantees here. Directly injecting cash into Intel just creates a moral hazard where they spend the money dealing with organizational inertia than innovating in the space. Plus, the government could even turn a profit reselling its purchased semiconductors to various US companies!

The US ecosystem badly needs a serious Intel competitor -- because they just ain't it.

US chip production really needs its SpaceX moment, but it seems like it will never happen, and we're left with the crumbling empire of Intel.

By "SpaceX moment", I mean a startup entering an impossible market, where the barrier to entry is billions of dollars of research and manufacturing, a market dominated by industry giants from the 60s, and yet still coming out on top somehow.

I generally like Stratechery but this is not up to the usual standard.

To the extent that CISC means anything it means “backwards compatible”. Modern ARM cores have multi-step decode and issue: to the extent that RISC means anything it means “we developed this with twenty year’s hindsight so we don’t have to do a lot in microcode yet.” In die area, in TDP, in wafer consumption, it’s just not a big cost.

To the extent that Intel missed AI so did everyone: right now there’s one target and that’s NVIDIA, no amount of HBM/TC ratio tuning is going to change that. Like 3 of 17 Ray/Anyscale tutorials target Habana. The Gaudi stuff works fine if you’re willing to pay the “last year’s model” tax all non-CUDA does.

And bigger picture, on semiconductors it’s the geopolitics people who are watching the bright line: semiconductors are a major power arena, the markets will get adjusted along those lines.

Sidequest:

“I’m going to do Windows and then I’m going to do Windows Mobile and I’m going to do Windows embedded.” (all on intel x86)

As someone who worked on Windows Mobile and Windows Embedded, it's funny because that's what management was hoping, but it was so different. I don't think any of the phones had an x86 chip in them, although that was our debugging setup so we could use vastly more powerful desktop machines as devkits (although we still had devices to). Having to port parts of Windows to ARM and even MIPS was a giant PITA.

Windows Embedded is closer since it was based on XP, but also a completely different branch from what I remember from Win Mobile.

A lot of people think it's easy to port things, and hard to write things. With really, really complicated software, it's still really hard just to port things.

> today AMD has both better designs and, thanks to the fact they fab their chips at TSMC, better processes

On paper AMD is better, but in the scientific community, it seems that Intel has much better performance for things like NumPy and SciPy. The reason seems to be "Intel® MKL" (https://en.wikipedia.org/wiki/Math_Kernel_Library).

I'm using a lot of weasel words like "seems to" because I haven't rigorously proved it. But anecdotally, my company's AI pipelines which are NumPy/SciPy heavy run an order of magnitude faster (2 seconds vs. 20 seconds) on my laptop's Intel i7 than the do on my Ryzen 7, despite the Ryzen 7 being a newer gen than the i7.

The Altera merger pain is maybe enlightening.

Altera 10-series FPGAs were massively delayed due to Intel's 10 nm fab problems. But I speculate that there was also a big difference in toolchains, does anyone know for sure? I mean Altera was using TSMC previously, and I presume were using industry standard tools: Cadance / Synopsis. But I would guess Intel was running their fab on home-grown tools... what is the status these days? For example I know for sure that IBM's synthesis was "Booledozer".

It's interesting because maybe Intel will spin out the fab business. But are they immediately ready to become a commercial fab business instead an Intel-only business? What's the toolchain like for Intel fab?

In June 2021, Gelsinger named 22-year Intel vet Sandra Rivera to head up a new Datacenter and AI Group. Rivera, who spent two years as Intel’s chief people officer before taking on her new role, describes what Intel staffers are looking to get out of their relationship with the company, using terms that would have been unlikely to come up in its original heyday. “For sure, we are in an age now where our employees are looking for a deeper, richer connection to the purpose of the organization,” she says. “They care about diversity, equity, inclusion. They care about the planet.”

The geopolitical bit of all this is the real wild card. No one really knows what's going to happen there.

Not a trade lawyer but as I understand it, WTO treaty obligations specifically forbid the kind of purchase guarantees that Ben's arguing for here. Part of the whole point of the WTO is to to pull up the ladder and prevent countries from doing what Japan, South Korea, and Taiwan successfully did to become rich. So this kind of support for domestic manufacturers is not allowed

The OP makes a compelling case that Intel's foundry operations have fallen so far behind that they are no longer economically viable on their own, in the face of current market forces, i.e., Intel's foundry business may not be able to earn a positive return on the tens of billions of dollars of capex now required to catch up with TSMC.

If the US truly views having a domestic foundry as critical to national security, the US federal government has no choice but to pay up big time to support Intel's foundry business until -- hopefully, eventually -- that business is able to compete profitably against TSMC to manufacture chips for Apple, Nvidia, AMD, etc.

Oh, how the mighty have fallen!

I’ve heard the theory that Intel is physically incapable of doing the lithography transition they need to do without missing at least one full release cycle. Which should be bad news but shouldn’t be existential.

The theory goes that any CEO who does the necessary thing will be fired, so either they won’t do it, or the board will replace them with someone who won’t do it.

That kind of market failure can destroy arbitrary value.

Brutal. Intel processors will be the equivalent of government cheese. Not what anyone wants, but it's what you get with your government dollars.

Did I misunderstand or did the author advocate a world where the only high-end chip manufacturer in the entire world is TMSC?

Just the one single company, located in a politically unstable area?

And putting aside the politics, if they become a monopoly what do you think will happen next?

Intel was getting something like $20 billion from the CHIPS act - but it seems like by the time the fabs would have been ready TSMC will already have them beat to 1.4nm/a14 (also using High NA EUV). Intel has just consistently failed to execute whereas TSMC has been fairly close to schedule.

I wonder if Global Foundries will become relevant again on the high end. They “gave up,” but they are working on some pretty small nodes now, 12nm, obviously not what TSMC is up to, but maybe they’ll catch Intel in 10 years. I’m sure they are fine for automotive.

I'm still hopeful of the fact that Intel is, in many aspects "too big to fail", and with their cash on hand, they have enough of a burn rate to turn this ship around, even if they aren't being as intense on cutbacks as they should be.

The question remains, will Intel survive in its current form or could an activist investor stir up a hostile takeover and change the calculus?

From the article, “William Shockley, one of the inventors of the transistor at Bell Labs”.

William Shockley did not invent the transistor in 1947. He was not on the patent and was in Europe at the time of the invention. John Bardeen and Walter Brattain were on the patent. Shockley shared the Nobel Prize due to work in 1948 on the transistor effect.

Source: https://www.electronicdesign.com/technologies/analog/article...

I'm not sure if I buy the thesis entirely (that thesis most aptly being: "Gelsinger does have one fatal flaw: he still believes in Intel, and I no longer do.")

The biggest reason is geopolitics: If escalated confrontation happens with China, Intel becomes one of the most valuable companies on the planet, period. TSMC will finish western fabs eventually, but the delta-t on when they become competitive (tech, capacity & cost, remember) with even Intel's western fabs is... decades, plural? Maybe never? The CHIPS act helped, but TSMC is dragging their feet for very, very (existentially) good reason. Also, remember that TSMC is built entirely on the back of ASML; conflict with China doesn't just block the west's access to the east's fab capacity, it blocks the east's access to the west's tooling to build more and better fabs. TSMC craters in this scenario. Samsung also exists. That's the list of near-SotA fab capacity, globally.

Capacity is another good reason: TSMC is not a bottomless bucket, and every year for the past at least two years Apple has purchased 100% of their SotA fab capacity. No one is competing with Apple's margins (except Nvidia, but that's a bubble market; numbered days). His argument is that there's no market reason for Intel's fabs to exist; we don't have the data to say for certain, but I'd guess that if Intel went to TSMC and said "you're already making lunar lake, make our xeon chips too" TSMC would say "we can't" (especially on SotA nodes, but maybe even near-SotA). They're tapped out, they grow, they're instantly tapped out again, everyone wants what they're selling. Intel fabs Lunar Lake and Arc with TSMC; both very-low volume.

Also worth keeping in mind: Intel was at one time the global leader in chip fabrication; but they lost that crown. People view TSMC as this unassailable beast, but they're just as fallible; and when I hear people say "Well, Intel 18A is probably three years out and by that time we'll have TSMC N2P with backside power delivery so who will even care about Intel 18A" are just extrapolating history. That's a dangerous game when all these rankings and valuations are based on asymptotically approaching the limits of the laws of physics. And while it is unlikely that Intel will take the lead again, TSMC showing any sign of faltering will raise the relative value of the #2 companies.

Intel is not in a good spot, but they're still an interesting business, they're still designing some of the best chips on the planet, and with the right decisions could grow to be even better than interesting. IMO, this article is missing a lot of the hard analysis and data that Stratechery is usually known for; I don't feel you can have a researched discussion on Intel without talking about fabrication volume or their concerning levels of debt, but he didn't mention either of those things. Heck, more than one passing mention of China feels kinda important to the topic.

And the news just keep getting "better" [1]

[1] https://finance.yahoo.com/news/intel-just-gave-plan-laid-213...

Intel has been talking about the monumental shift that its 20A node will bring for years, but we'll never see it in action.

I'm sure all the points are correct but the assessment for ways forward being to just use TSMC as the manufacturer completely ignores the somewhat perilous position Taiwan is in with a looming Chinese invasion. I'd be very reluctant to tie the future of my business to the fates of a company that may become state property within a year or two.

Intel has $29B cash on hand. Their new fabs are being subsidized. Sure, American labor is more expensive than APAC, but I don't see that being a huge differentiator in such an automated manufacturing segment.

My personal rabbit-hole conspiracy is that AI is driving a fire under the national security apparatus, and we're going to see a restriction on AI-chip technology being exported or manufactured overseas. Intel will be in prime position to onshore that manufacturing.

What happens to the computer and tech industry if Intel fails?

The x86 licence still has value to some others in the space, but it seems like Intel could potentially be sold off for parts if it can't sell correct.

They are building a huge fab in Ohio but I don't see Intel as a going concern that will be around at the date of completion. I fully expect the project to be abandoned.

yeah intel's foundry business might not be competitive in the short term. but it's better than nothing. they can focus on other industries that don't need high performance / for low power.

because if you yield the whole thing to tsmc. you won't be able to recover the expertise, which is catastrophic for both intel / USA.

Anyone else notice how many typos are in this article? Does no one proofread anymore?

>Their pay in the Bay Area is double tsmc in taiwan for equivalent roles in raw dollar terms but the ppp differences means your better off working for tsmc in Taiwan than intel in the Bay Area. That’s not a joke. Intel are literally incapable of attracting talent from Taiwan right now.

There's some whiplash in hearing constantly how the US is a corporate-dominated oligarchy, and then also seeing trillion-dollar industries at the forefront of the global economy constantly getting their shit rocked by a few dozen NIMBY retirees at city council meetings.

Lmao @ AMD coming out on top.