Never stop poasting

I've talked about not knowing how you would go about fixing the American infosphere before. Which is a statement that has implicitly excluded "let it burn and rebuild", but I don't know that, that is out of realm of possibility anymore.

Has anybody thought about beinging back token ring?

I mean itd make sense to plan for adopting it, or atleast building new systems with the ability to incorporate them later.

The I-pace that waymo has been using as their base vehicle ended production in December '24, so I imagine a new vehicle and new systems are well underway.

I'll admit to not having kept track of the coverage, so I added the almost as a qualifier incase it isn't complete coverage yet. Which last I checked atleast it wasn't.

Ah nice, yeah that would have been quite difficult to do back then. What they've accomplished later with jupiter is also quite impressive.

Yeah I mean, a worst case scenario as i see it is that they scale back datacenter spend slightly.

The compute is exceedingly unlikely to end up sitting idle, at worst it'll pay for itself. More likely in the case of a bust we might just see a period of slightly reduced profit

Yeah it was around the time when the access and core network capacity relationship flipped and suddenly just massively building out core cap became viable as transit costs increased.

Though I'll admit to not entirely be sure what firepower is in this context. Before my time.

I mean yeah it's problematic and regulation might be needed going forward. But even if you destroy every last copy of the weights for this(you can't). Someone else will just make a new one.

Once it's been shown possible, someone else is going to try and replicate what you did.

I mean sure at an any moment in time, there is a max given a certain price per unit of compute/storage other resources. But we aren't there.

It also grows, especially as prices drop.

Like the buildup we see right now is just a few years of planned investment compressed to a single one.

As part of the gpu buildout you are also investing in all of the supporting infrastructure around it to actually feed them. Which again, if needed is capacity that can be sold.

I mean the hyperscalers are likely to retain the pricing power, they have lower costs as a result of their scale.

An example being the majority of Googles internal compute needs were gpu, even before the AI growth spurt. And excess capacity is sold.

Like the ahead of schedule buildout likely also means they can keep a higher steady state DC spend if/when the "bubble" pops and they do have to take their foot of the gas

I mean even if you cut from the very high levels of spending, you still need to keep updating/building/expanding infrastructure.
Cloud compute is sort of a "build it and they will come" situation, and is likely to remain that way.
so the cut will probably be less severe than another industry.

Would not be surprised if you start to see social media get treated as a public service with a government supported independent non-profit with partial public ownership.

Model for how it gets done is unknown obviously.

Obviously not, but it's like we know how to make them. The trick is bringing down the cost, and that is still a work in progress.

(as with everything it's about it's cost per unit bandwidth)

I mean as it stands we are also starting to see restrictions on SoMe for the younger audiences. See the danish government stating today that a SoMe ban for those under 15(13 with parental consent) is planned.

I mean, if it were anything other cloud compute being invested in, I would tend to agree.

In this situation if suddenly AI isn't worth pursuing, you've built out a few years ahead of schedule and that's basically that . It's less of an issue than it would otherwise seem, given the level of capex.

I like the way you think.

It has its charms.

When you start seeing FMF/MC fibers get commercialized. They are a pita to manufacture(and produce amplifiers for) so still a ways off.

With a TB of vram I'm fairly certain you could do a bit more than that.

Nice! Yeah, Google is doing some seriously cool stuff that I can only appreciate from the outside.

Yeah, I mean WDM let's us extend the lifetime of physical links, but even that has its limits.

At point we need to start considering FMF/MC fibers in the cables.

And that's... Not gonna be cheap.

Every household, a TB of vram worth of GPU's

Only 8?

Also if you can't find any use for them( or you accidentally drop them) sxm2 modules make for great paper weights.

Yeah I mean fabs are moving to GAA with N2 which will probably result in some serious teething issues getting those up. But I mean the actual roadmap for logic shrink is still relatively well defined for a decade or so.

Its memory/storahe that is stubborn. More compute is "easy", feeding it is not.

It's always fun to read about Google's datacenter networks.

Just casually announcing double digit petabit/s bisect bandwidth in the DC network is one of those "hold up, what the fuck" situatione.

Same. Until then we get to enjoy how cool some of what we work with is.

Yeah it's always amazing running into/working with super competent people.

Obviously video services kinda changed the game on what AN's needed to do and support, which is likely why we saw a sudden jump in their required capacity.

With the core being left having to catch up with the now much more capable AN's. And now multigig is starting to show up.

Fair, I just mean that.
Aside from Google fiber(obviously non trivial but still)
I don't think Google actually operates the access networks in most places. Being peered with the local providers instead?