[AnthonyMouse]

> But at same time, extend IRBM range by 1000km, and replace refineries with hyperscalers, or whatever targets that worth deterrent value (energy at top of list).

Hyperscalers are probably not a great example because a) they don't really benefit that much from being physically centralized (especially at the building level rather than the regional level) and b) data is one of the easiest things to keep redundant, and then even if you destroy a large facility, backups get restored to another facility or distributed set of facilities with no downtime at all if the target is well-prepared and only a short period of time if they've done even minimum diligence.

The critical ones can also do the "build it on the inside of a mountain" thing and then your capacity to take down grandpa's WordPress is mainly useful to the target for rallying opposition against you.

> whatever targets that worth deterrent value (energy at top of list).

If "energy" turns into solar panels on the roof of every house and widely distributed low density wind farms etc., that's pretty hard to target.

In general centralization is often done because it has economies of scale, but those same economies of scale have diminishing returns. One huge facility reduces certain fixed costs by a million to one (i.e. 99.9999%) over having a million small facilities, but a thousand medium facilities are much harder to target while still reducing them by 99.9% and the remaining 0.0999% is negligible because you're long since already dominated by unit costs. The target can also choose where to take the trade off based on how likely they expect to be targeted. And that's a broadly applicable principle rather than something that only applies to any specific industry.

[maxglute]

Hyperscale/data just one example, f35 manufacturing, specialty feed stock production, transformers, gas compression etc, the list of currently centralized (as in have large target profiles) that will remain soft for decades is long with varying degree of disruption/dislocation, i.e. you don't restore hardware with multi year lead times.

Those are ridiculous / absurd economies of scale numbers, splitting piles up 20-50% per duplication inefficiency, especially in US context (expensive regulatory/physical buildout), splitting 1 hyper to 1000 medium is not marginal more cost, it's magnitudes / 1000%s more cost - costs private or public will not go for, and is prematurely self defeating because others can always build cheaper missiles than US can build infra (hence goldendome theatrics).

In principle, US can preempt CONUS physical vulnerabilities, where 100+ years of built up over assumption of CONUS not being vulnerable. In practice the chance of that happening approaches 0. Didn't even harden CENTCOM air shelters and planners have been noting vulnerability for years. Not just economies scale, but JIT and all other aggregate downstream optimizations US likes to make in name of efficiency. US simply not culturally PRC who does not mind (and is optimized for) some extra concrete for physical security. Not that PRC does not have huge vulnerabilities, just development has been made with mainland strikes in mind.

[AnthonyMouse]

> splitting 1 hyper to 1000 medium is not marginal more cost, it's magnitudes / 1000%s more cost

It isn't. The primary costs of both the medium and enormous facility are the same: Server hardware and electricity, and server equipment and electricity don't have significantly lower unit costs when you're buying a million instead of a thousand. Also, you can still buy a million servers and then put them in a thousand different buildings.

It's only when you get down to very small facilities that things like staffing start to become significantly different, because amortizing tens of thousands of employees over millions of servers results in a similar unit cost as amortizing tens of employees over thousands of servers. It's only when you get to the point that you have only tens or hundreds of physical servers that you get scale problems, because it's hard to hire one tenth of one employee and on top of that you want to have more than one so the one person doesn't have to be on call 24/7/365. Although even there you could split the facilities up and then have multiple employees who spend different days in different locations.

> especially in US context (expensive regulatory/physical buildout)

This is another reason that "hyper economies of scale" don't actually do you any good. Which costs less, having dozens or hundreds of suppliers for the various parts of an aircraft, or one single Lockheed that should nominally capture all of these great economies of scale from being a single company?

It's the first one, because then it's a competitive market and the competitive pressure is dramatically more effective at keeping costs under control than a single hyper-scale monopolist that should be able to do it more efficiently on paper until the reality arrives that they then have no incentive to, because a monopoly is the only one who can actually bid on the contract and a duopoly or similarly concentrated market can too easily explicitly or implicitly coordinate to divide up the market. At which point they can be as inefficient as they like with no consequences.

This does mean you have to address the regulatory environment that tends to produce concentrated markets, but we need to fix that anyway because it's a huge problem even outside of this context.

> where 100+ years of built up over assumption of CONUS not being vulnerable

That's not true, there was a significant push during the Cold War to decentralize things to make them less vulnerable to nuclear strikes. The government pushed people into the suburbs on purpose:

https://www.wagingpeace.org/nuclear-weapons-and-american-urb...

There are obviously significant costs to that but Americans were willing pay them when there was a reason to and much of the landscape is still shaped by those decisions even now.

You also see this in the design of the internet, which came out of the same era and has a design that facilitates the elimination of single points of failure, and that sort of thing is as close as we've seen to an unmitigated good.

[maxglute]

When I say hyper, I'm referring to hyper size vs distributed, not limited to data centers. It generalized reply to your insinuation of economies of scale is broadly applicable when it is absolutely not, i.e. 99.9999,99.9%,0.0999% which fantasy figures. The general economics of economies of scale is you split 1 facility in 2 you add 20-50% overhead due to duplication. The immediate cost of redundancy/resiliency is adding double digit overhead. The point is duplication doesn't happen when "only when you get down to very small facilities", it happens when you go from 1 to 2, incremental distribution increase cost disproportionately. Breaking economies of scale of 1 hyper facility int to 2,5,10,100 smaller facilities is possible on paper, but no one doing it in practice.

>don't actually do you any good.

Sure, economy of scale good for consolidator being net bad is valid, but this wasn't discussion on optimal macroeconomics, this discussion on what US politically able to do. There are things US should do, but systemically can't.

> Cold War to decentralize

Cold war dispersion for nuclear math and precise conventional strike math is different. Spreading 2 factories apart so they draw 2 nukes vs 2 factories get 2 conventional packages regardless of spatial separation.Circle back to feasibility, what is required for distributed / dispersed survivability. Is US going to dismantle gulf oil infra and move it inland. Most physical infra processes are not fragmentable or self healing like internet. How much are Americans willing to pay, coldwar was eating 15% of GDP. All this ultimately secondary to the point that doing all this costs US more (because everything in US costs more) vs adversaries simply getting more missiles, it's economically/strategically self defeating. Let's not forget Soviet answer to US disbursement was building more missiles while US still pays inefficiency tax on suburbs.

[AnthonyMouse]

> The general economics of economies of scale is you split 1 facility in 2 you add 20-50% overhead due to duplication.

There is no general figure for this, it depends both on the type of facility and the size of it to begin with. At hyper-scale the incremental efficiency gains are much smaller because a medium-sized facility is large enough to have captured most of them already. Where you get 20-50% is where you have a single small facility and try to split it in two.

The general premise of economies of scale is that there are costs that don't depend much on scale, so at increased scale, the amortization of those costs over all units contributes less to the cost per unit.

However, at very large scales, three things happen.

First, many of those costs look like "you need this piece of equipment that can handle 1000 units per year" and then the facility that produces 1000 units can amortize it over 1000 units but the facility that only produces 100 units has to amortize it over 100 units. However, the facility that produces 100,000 units then has no advantage over the one that produces 1000 because they then need a hundred of those pieces of equipment and have no advantage in unit cost.

Second, some costs only have to be paid once regardless of the number of units. If you sell each unit for $1000 and have a fixed overhead of $10,000, at 100 units the over per-unit overhead is $100 (i.e. 10%), but at 1000 units it's already down to 1%, at 10,000 units it's 0.1%, etc. The incremental advantage of doing millions of units instead of thousands is thereby negligible because it was already under 1% of the unit cost by the time you were doing thousands. The number of industries where you need to be producing some double digit percentage of the entire national capacity of some product before those costs get down to a manageable percentage of the unit cost is extremely uncommon, to the point that it may well not even exist for a country the size of the US. Especially when you're talking about costs that specifically can't be amortized over more than one facility.

And third, some costs actually increase with scale, e.g. coordination costs. So once you pass the point that those costs exceed the diminishing incremental benefit from amortizing costs of the second type over more units, increased scale reduces efficiency even before you consider the consequences of reduced competitive pressure on incentives.

> Sure, economy of scale good for consolidator being net bad is valid, but this wasn't discussion on optimal macroeconomics, this discussion on what US politically able to do. There are things US should do, but systemically can't.

"Systemic" means that in order fix problem A, you first have to fix problem B. That is not a formal proof that A is permanently unsolvable, it's a just a dependency graph for the order in which they have to happen.

> Cold war dispersion for nuclear math and precise conventional strike math is different. Spreading 2 factories apart so they draw 2 nukes vs 2 factories get 2 conventional packages regardless of spatial separation.

Sure, but the first is the stricter requirement because "2" is an insufficient number in both cases. The USSR definitely had more than two nukes. And if you need 100+ facilities, in the second case it's fine to have multiple facilities in each of a handful of cities, whereas in the first case you need them to be in 100+ different cities, which is harder to do but effective against both.

> Is US going to dismantle gulf oil infra and move it inland.

The original premise was it would rely less on petroleum, so in that sense, yes.

> Most physical infra processes are not fragmentable or self healing like internet.

That's not actually that uncommon. Transportation networks, power transmission, etc. map to the same sorts of designs where in the common case the multiple independent paths increase capacity and efficiency and in the damage case they keep the system running for critical infrastructure by redirecting critical uses from the damaged route to the operating one.

Meanwhile most infrastructure is inherently fragmented. There is no single water treatment plant in DC that runs the whole country because you need them to be closer to the point of use.

> How much are Americans willing to pay, coldwar was eating 15% of GDP.

Starting from the status quo, getting to the scenario where there are a larger number of competing suppliers for various things would lower the costs people are paying.

> All this ultimately secondary to the point that doing all this costs US more (because everything in US costs more)

Which brings us back to, that's the real problem we need to solve. If your problem is that it's now easier for someone to blow stuff up and you've made it excessively expensive to build another one, the solution is to focus on lowering the cost of building things in the US, which would benefit people independently of this anyway.