can't they just create enough for decades and stock pile it somewhere?

Ben Rich talks about this being a rough spot for dealing with military contracts in the epilogue of his book, Skunk Works. I think it depends on the contract, for plane parts they did not allow for storage of parts :(

I guess they want the capacity to make planes indefinitely, if a really big war happens?

When I used to work in defense contracting, this is precisely what we (the contractor) did. We would buy up all available stock of any difficult-to-replace parts (often specific SBCs) when the manufacturer announced end of life.

There's a market in conforming interface and ABI spec meeting hardware to emulate the boot and upgrade devices for German tanks, or some other hardware. Sd card or USB behind, giant milspec plug to the fore.

Doesn’t it rot?

Given that it's definitionally military-grade hardware, and very high value-density and therefore easy to store very securely, I doubt it.

It is not the first choice of revenue of course, they and everyone would like to be Nvidia and pick and choose the customers standing patiently in line outside the door.

However in tough times, having the culture to change and deal with customers or use cases you wouldn't prefer to survive is a good sign. AMD during the bulldozer era, had many deals they didn't like or barely made money just to survive, it paid off with Ryzen, it can happen to Intel as well. Good to see Intel branching out and trying new things to turn the ship around.

This is nothing more than Intel joining the ranks of the other tbtf government contractors. That's how they plan to survive.

Intel has had the government as a customer for a long time. They even have a division for it (Intel Federal)

Did they manufacture military designs though? Or just sell Intel designs (with some modifications) to military?

Intel was making 386 chips for the F-22 for a looking time after they became obsolete.

386 in the F22?! Why ? Such an old chip even considering the plane was in design for years no ?

I thought the F-22 initially used the i960, not the 386, but then switched to PowerPC.

[0] http://www.righto.com/2023/07/the-complex-history-of-intel-i...

It also increases the likelihood of a secure supply chain for mil components. Seems like a win-win.

In the sense that any voluntary trade is a win-win for both parties, yes. But it is inglorious compared to where Intel was in, say, the early 2000s. It is a big change to look at market cap and see traders treating AMD as a big player, then Intel somewhere off in the also-ran territory.

So much for their monopoly.

I did not know intel was so short of cash it needs military money to stay afloat.

They are not immediately short of actual cash per se, but they have been doing a lot of layoffs and pulling back on big ticket investments like the foundry they had planned in Germany, financial health is a top concern .

they are also loosing revenue hard in multiple segments including their core high margin enterprise server chips and also not gaining foothold in others (mobile/lower power device or GPUs)

They need fresh revenue really quickly to keep markets happy on share price which has really tanked this year, also to keep the supply chain healthy and talent motivated

Any foundary business is always in need of cash, a leading edge foundary is $20-$30B outlay minimum these days , not many companies in the world are so flush they can easily spend $30B without sweating it

Seems like that facility they were supposed to build in Germany is going to be built in Poland.

Their long term plan is to make money via foundry/fab orders, so it makes sense to see this. But otherwise their share price is currently rock bottom because they're pouring money into infrastructure development (plus late EUV adoption) and a lot of people are sceptical it'll pan out.

They're not short of cash, they're short of business throughput for their fab because it's uncompetitive next to TSMC, but a lot of their chip design business relies on integration with it so they'd lose a ton of money if they went fabless like AMD and switched to TSMC.

Intel is arguably on the cusp of death as the company that it was between, say, 2000-2013.

https://www.economist.com/business/2024/09/12/intel-is-on-li...

It’s pretty nuts reading The Atlantic’s piece from 10 years ago and then seeing where the company is now.

https://www.theatlantic.com/technology/archive/2013/05/paul-...

losing Apple's desktop and laptop business ended up being quite a loss for them

Intel could't have done it profitably and chose to lose the business. Intel has been a cash cow business for a while, coasting on its early wins and doomed to be gradually surpassed. For shareholders seeking short and medium term value, this works out quite well.

It's not about the small time military money it's about the big ticket government bailout. If they have some hooks into the military then they have, what's the term, a "moral hazard" card to freely take risk, military dependency puts them at the top of the bailout list (which they kinda were already, now even moreso)

This is not a good sign, it only means that Intel will survive from government handouts. If I ware a shareholder I would look for an exit.

Wise

The military is probably using a lot more tech now than they did in 1997?

Procurement is probably just as screwed up though.

Chips are used in a lot more than expensive, low-quantity planes.

Increasingly, every "round of ammo" has a chip in it. Every missile, bomb, single-use drone, and artillery shell has sensors and guidance. These are expended in use.

Instead of maintaining these "rounds", the military simply asks for a new, improved, and more expensive version of them, and destroys / gives away the old versions.

It's not just one country buying these, it's the entirety of NATO.

What Are You Talking about. Ordnance gets expended, goes boom.

> Moreover, the military wants parts for decades

nothing wrong with this. Government pays for the product. The product should be serviceable whether the company becomes defunct or business shifts. At this point, military or consumer should be able to give blueprints of the parts to different manufacturer or manufacturer it themselves.

Wafers come in batch sizes of a FOUP, these are the little sealed boxes you see the robots on the tracks moving around a semiconductor factory. One might only need one wafer of chips or less for a project, those other unused wafers go in storage after some testing. When the military needs spare parts, those wafers can be brought out of storage and diced and packaged (if we will even still do that in the future).

This technique is already done for things with long operational lifetimes. It would be nice if we made a distributed (geographically) wafer bank so that we will have a long supply of semiconductors, esp after a civilization scale catastrophe.

https://en.wikipedia.org/wiki/FOUP

The analogy of a wafer bank to a seed bank is new to me. I would imagine the wafer bank is a bad idea, because they would just be playbooks that would lead to the collapse. So kinda pointless.

Interesting. But why store the somewhat unwieldy wafer instead the packaged chip? I’d imagine that doesn’t add much cost but makes storage significantly simpler.

Storing the wafers avoids paying the costs for packaging and testing in the case the chips will not be needed.

Moreover, the packaged chips typically require a much larger volume for their storage than the wafers. A wafer may contain many thousands of chips.

The wafers are normally stored in dry nitrogen, to avoid their chemical degradation during long term storage.

There are three things that already exist and address these kinds of concerns (I've removed some, but not all, color commentary, this is somewhat effective but definitely not 100%):

Data Rights - Every acquisitions contract includes data rights. The specific data varies by contract. For things like an LRU, the data rights may include schematics. This is, in theory, enough information to recreate the device but may leave out certain key proprietary pieces. Like if a 1980s era LRU had an M68k, no schematics from Motorola will be included. But the architecture is known so recreation is technically feasible. The schematics also offer a foundation for producing a like-product replacing the obsolete components, though a project like that can take years.

ICD - Interface Control Document. The device itself becomes a blackbox. Instead a description is provided, along with other requirements and spec documents, on how it behaves. The good ICDs are really enough to start a clean room project without ever needing to crack open the to-be-replaced devices. Unfortunately the good ones are rare, they often stop getting updated at some point and modern ICDs are shit compared to the documentation from last century.

COTS - DOD (and the US gov't in general) has had a major 30+ year push to go COTS as much as possible. Obviously this doesn't work for everything, but go back to that M68k example. There's no reason to ask for a custom chip when a COTS one will do. Same for other parts of major systems. Computer motherboards can be COTS (or very near) even if the chassis is bespoke to make it form and fit suitable for its intended environment. COTS, in theory, also makes it possible to do incremental refreshes more easily. Like replace that computer hardware in the custom chassis every 5 years, it's not trivial but it's a small jump and updating software components in such "short" (by DOD standards) increments is hardly onerous. In practice, updates may not happen for 20+ years which is a more substantial undertaking.

These, and other things that are supposed to be done in acquisitions, largely resolve the "What will we do in 20 years when the supplier has gone under" questions.

It's so sad to see how miserable ICDs have gotten. Some of them are just REST APIs and you are on your own to figure out what everything means, it's pathetic. The old ones, you can (and I have) interfaced with stuff from the 80s without any customer input.

(alright, it has happened though that they needed to wheel out the old-timer who was around when the hardware was originally delivered, come integration time - since the way you hook up to the thing isn't always clearly described. But the software was fine!)

Yep. The older the ICD, the better it seems to be. I miss the era of dedicated technical writers, too. We received an "ICD" from a vendor and it's just garbage. It's incomplete, for one, and kind of just ends in the middle of things and leaves out crucial details so we can't use it without reverse engineering their system. A major problem is that it appears their developers are their document writers now, and they just don't do it, either out of laziness or insufficient time (higher priority dev tasks in the queue). They've been publishing the same incomplete document for years.

Modern military uses more cots and less bespoke designs because developing cutting edge custom asics is so horrifically expensive.

But also custom ASICS are no longer necessary because things are so fast and military acquisitions are much slower than commercial development: the state of the art commercial tech state of the art will go through several iterations before a single military product is released so why chase the bleeding edge?

Unless they have really classified tech like a productively useful quantum computer or something, it’s just a better idea all around to use off the shelf or very slightly modified off the shelf parts. (Feature flags, extra testing, expanded environmental margins, etc, but not total redesigns)

That's an horrible business if you plan on being nimble and competitive and blow away customers with your innovations.

My read is Intel isn't going for that trajectory anymore and stabilized for markets where politics can play a larger role.

"a"

Intel hasn't been nimble for a while. It's a huge organisation. Makes sense to attach themselves to the military, both sides need a stable, reliable partner for a long-term relationship.

> Moreover, the military wants parts for decades, so you're stuck building obsolete parts on an obsolete process.

Intel recently built expertise in this area, so I think they are a good match.

> so you're stuck building obsolete parts on an obsolete process.

Isn't this what every corporate wants? Steady supply of income with no R&D, marketing, etc.

Keeping the obsolete process alive can be very expensive.

Could you elaborte? Maybe with illustrative example? (Intuitively it would just seem like a good thing)

Keeping a factory alive is not free. If you have committed to delivering a certain product for the next 30-50 years, you need to be able to deliver that product even in year 49. Even if it's only three parts and the factory is only profitable when producing thousands a year.

Some costs that will crop up even without investing in R&D and maintenance: - Factory maintenance. The lights need to stay on, the floors need to be swept and the bearings need to remain greased. Things like ISO and security certifications also need to be kept up to date. - The longer a product runs, the more likely it is that the original employees on the production line retire or leave for another job. This means you'll need to be able to find and train new people for a job that uses tools and methodologies no longer used anywhere else. It will probably be more expensive to hire for those jobs than for jobs where people learn transferable skills that they could use in the rest of their careers. - After 40 years, many of the components in the production machinery will be difficult to come by. A CNC machine from that time might use the (then brand new) 286 processor. If it breaks, where would you source extra 286 processors? Alternatively you can redesign the process to use up-to-date components, but that costs a lot of extra money. - Usually the demand for components drops off over time as the world moves on to something more modern. For example, demand for components of older fighter jets will slowly drop off as new airframes are no longer being built and the existing ones slowly get taken out of service. This means you'll need to spread the fixed costs of the above points over fewer and fewer components over time. - Finally, the need for R&D and marketing doesn't actually go away. If you only focus on producing (say) targeting processors for the F16, your company will go out of business at the latest when the last F16 leaves service. Your shareholders will probably not be happy about that, so it's still important to invest in gaining new contracts too.

> Even if it's only three parts and the factory is only profitable when producing thousands a year.

If that is the case you manufacture more than enough and put them on a shelf.

> Some costs that will crop up even without investing in R&D and maintenance

Obviously. You put those costs in the contract and make the costumer pay for them.

If they want to fab 3 chips a year that is going to cost them dearly. This is not a property of the old processes. This is a property of low volumes.

You have to manufacture enough parts for if wwiii breaks out even though odds are it won't and even if it does the military will probable notewant those parts - but they might.

Ok. Built in (reasonable) assumption which I was missing - demand decreases very significantly.

If you are a company facing bankruptcy because your process technology is not competitive with world leaders like TSMC and Samsung, you resort to military work. Intel is like the Boeing of the semiconductor industry except it might be less crucial to the industry.

What normally happens is they will store a bunch of chips for the future and then shut the line down, when they need more they'd have to use a newer chip in it's place.

> you're stuck building obsolete parts on an obsolete process.

TSMC's oldest plant still in operation (Fab 2) started production in 1990. So far, the only fab they have closed is Fab 1.

Reference: https://en.wikipedia.org/wiki/List_of_semiconductor_fabricat...

Admittedly, leading-edge process is where all the excitement it, but the old fabs are fully depreciated and can remain profitable for decades.

As another example, NAND is manufactured on 40nm and for the foreseeable future it'll be. https://thememoryguy.com/why-3d-nand-is-stuck-at-40nm/

I’d love to be able to fab my own designs, 40nm would be awesome. Yes there are projects where you can. But they allow small area/transistor counts. Pcbway for chips would be amazing!

won't happen for such a complicated process like 40nm. The design tooling itself would cost 100s of k$.

Honestly I love that old processes are still running and old parts are still being made.

Modern SMD stuff is really rubbish for learning electronics. It's nigh impossible to hand solder because all of it is tiny af, and since it's newer it's expensive and you don't get as many attempts. Luckily there's still an immense supply of arguably completely obsolete thruhole parts that are super easy to work with, can be breadboarded and bought for basically nothing.

I respectfully disagree with what you say about SMD parts :)

SMD parts are much cheaper compared to THT in my experience. Which parts are you thinking of specifically?

I majored in CompSci but got my first job in embedded at a hardware manufacturer. IME it didn’t require many hours of practice to learn to solder e.g. 0603-parts. Around 4-5 hours of deliberate practice for me, although smaller can still be annoying.

I have slighty worse than average vision so personally need a magnifying glass or a microscope for 0402 and smaller, but I can do it and it looks pretty afterwards (says the HW techs).

I was about 5-6x slower than the HW techs at solder work and repair, so I didn’t solder something up every week/month. I can still mount a board easily with 0805s and 0603s even though it’s been a few years now.

Well there are lots of breakout/dev boards these days that do make it possible, but for something like adding an indicator LED I'd rather reach for an ol' 1/4W resistor than those ridiculously tiny SMD specks. But in general modules like the CD4051 mux, SS49E hall sensor, 74AHCT125 level shifter, etc. Cheap and super reliable stuff.

I admit I'm really crap at accurate soldering so it might be more of a me issue, but I doubt I'm entirely alone in this.

I can see what you mean. For many, it looks scary in the beginning (using tweezers etc.), but it comes quickly with practice.

It is easier than it looks IMO, though I did get some tips from our techs which likely accellerated my learning (e.g. compared to learning on your own).

With smaller (and thus lighter) parts, the surface tension helps pull the parts into place etc., so you get help on “pad-alignment”.

Hot air soldering gets more difficult though, as you risk blowing parts offboard :D

I don’t know the parts, but my experience is that everything SMD is cheaper than THT. Almost always. The hall effect sensor can likely be found cheaper in Sot23?

The only reason I ever saw for using THT parts, was the (typically) much higher power ratings.

At my previous job, a typical board with e.g. 6 layers and 500 components would have perhaps 5-10 THT parts and the rest in SMD (0805 and 0603 mostly).

YMMV and for hobbyists it doesn’t make a big difference.

SMD-soldering skills can be handy though. E.g. I’ve repaired a few of my friends’ TVs with broken backlights for pennies.

I find SMD so much easier to work with. You just drop the parts with fine tweezers after some solder paste and apply hot air. It's also much cheaper. You can also rework without trouble in my experience. Apply hot air, lift the chip with tweezers. Remember to use lot of flux.

> Remember to use lot of flux.

Agree with your post and this part especially. More flux means easier work, but more mess to clean up when done. Flux pens are practical.

I only used hot air for removing parts, but the techs who controlled it well absolutely loved it and were insanely productive.

I solder QFP by dragging a “saturated” tip across the leads and then wick up any excess solder if needed (e.g. if I bridged/shorted a few legs).

well, true, but this doesn't guarantee a node to stay in production. fabs sometimes shut down a node to use the equipment and space for another node with more demand. also, sone of these machines from 80s and 90s are simply old and it's not financially viable to buy a new machine for such a process. they typically cannibalize some of the machines for parts, or even buy old machines for parts from other fabs closing down. I believe a fan requies refreshing and investment every 15-20 years.

Between ASIC and FPGA it's cheap enough to do for any (large competent) company, but choosing an enterprise is a guarantee they'll be around to ship too. Or a separate fab can be set up (or purchased from another company) to run batches of "obsolete" chips that then can be resold for washing machines and ink printer carts etc.

The article makes it sound like a sort of special deal, but Intel does run a foundry service now, so in theory that means that they aren't picking and choosing who their customers are -- anyone with money can buy manufacturing capacity.

It wasn't clear from what was said before the paywall whether Intel would be manufacturing products designed by the government or another third party, or if these are going to be Intel-designed products. (It's not all too out-of-the-ordinary for Intel or any other chip designer to make a variant of a product with certain application-specific tweaks for certain customers.) If Intel is designing the product, then that's a lot more than ordinary foundry services.

> The volume is small

That is changing with the drone warfare becoming large part of the future wars. US military (and pretty much everyone else) will make drones major focus of advancement and there is definitely lot of money to be made by supplying chips for those.

I can't read the article, and so it probably mentions this but I think this is a big case of "the past was prologue".

DOD is making a big bet that anything chip and silicon is going to be really hard to acquire past 2027.

This may be a great big gravy train for Intel.

>so you're stuck building obsolete parts on an obsolete process.

"Stuck"? Generally speaking, they can charge an arm and a leg for those older parts due to small volume. When I previously worked at a company that supplied hardware for a military application, they were still buying decades old hardware at about a 10x markup from when it was still in production for the general commercial market.

Heck, in 2019 Global Foundries sold a 20 year old fab that IBM built in Fishkill to ON Semiconductor for $430m. IBM originally built it for ~$2.5B. You don't think they got their money's worth and then some? ON didn't buy it as an act of charity, they've clearly got a plan to continue printing money building chips out of that fab.

https://en.wikipedia.org/wiki/GlobalFoundries

https://www.semiconductor-technology.com/projects/ibm_fishki...

sounds right up intels ally.

“Stuck making obsolete parts” sounds great. Guaranteed revenue without needing to stay competitive.