iiinova — where your digital self keeps evolving

If the big labs get squeezed they will squeeze the compute providers that then squeeze the hardware vendors and so on. I have a hard time seeing how in the long run the value add by the big commercial labs over open source models is substantial enough to maintain margins.

000 0

This feels like the “big problem”. Tokens are a commodity and it’s a race to the bottom. Margins are squeezed as token cost converges toward the underlying compute cost. epoch.ai/publications...

100 0

Side note: At this point trying to figure out the current AI infrastructure build out is a hobby, why do I do this to myself?

100 0

As a percentage of GDP corporate debt in Q2 was less than it was in 2000. It’s probably accelerated since then but has been generally trending down since 2020. I suspect debt trouble is more of a case-by-case thing (cough Oracle). www.federalreserve.gov/releases/z1/...

100 0

joe @joewilliams.bsky.social · 30/09/2026

People are talking about the hyperscalers taking out debt to pay for the infra build out. Being wildly cash flow positive they’ve never done this before so it’s a surprise to be sure. It’s unclear to me that this debt is a big deal to the broader economy. robinjbrooks.substack.com/p/is-the-ai-...

100 0

Food $200 Data $150 Rent $800 Compute $31B Utility $150 someone who is good at the economy please help me budget this. my family is dying www.implicator.ai/anthropic-ip...

000 0

joe @joewilliams.bsky.social · 30/09/2026

knightcolumbia.org/content/ai-a...

According to the superintelligence view, deceptive alignment is a ticking time bomb—being superintelligent, the system will easily be able to defeat any human attempts to detect if it is actually aligned and will bide its time. But, in the normal technology view, deception is a mere engineering problem, albeit an important one, to be addressed during development and throughout deployment. Indeed, it is already a standard part of the safety evaluation of powerful AI models.

121 0

joe @joewilliams.bsky.social · 28/09/2026

Old number 6 cost $7500 in 1927, around 144k in today’s dollars or about three Rubin GPUs.

000 0

joe @joewilliams.bsky.social · 28/09/2026

. @bcantrill.bsky.social is on the right track here. I’d argue making absolute claims far outside one’s domain of expertise isn’t called out enough. People do it because it costs them little if they are wrong and makes them feel and look good if they are right bcantrill.dtrace.org/2026/09/27/f...

bcantrill.dtrace.org

Fool's Expertise | The Observation Deck

3172 0

Is the capex sustainable?

This is to say the numbers are big and increasing but higher prices across and more expensive AI focused facilities and components (liquid cooling, etc) means there is just less HW getting deployed than it would otherwise seem www.marketsentiment.co/p/is-the-cap... www.currence.ai/blog/q3-2026...

marketsentiment.co

No paywall on this one.

100 0

The increase in capex is unambiguously a lot. 2025 vs 2026, somewhere around 25% of the increase seems to be simply component price inflation. Any other price increases would be on top of that so like M$’s $190B number is like 165B in 2025 dollars at best.

100 0

AI Data Center Cost per MW: 2026 Benchmarks by Tier

AI focused infrastructure is more expensive than previous normal cloud infrastructure axis-intelligence.com/ai-data-cent... All this is to say it seems like less capacity is being built at higher costs. So all these doublings of capex aren’t making much of a dent in creating more compute capacity?

axis-intelligence.com

$11.3M/MW for standard builds. AI-optimized facilities hit $15M–$25M. Full 2026 breakdown: construction tiers, hyperscaler CapEx, and the proprietary CEIR™ metric. AI Data Center Cost per MW 2026 Upda...

100 0

Someone ran the numbers epoch.ai/data-insight... so at least a quarter of the capex increase is just component price increases rather than more deployed hardware. They did cost distribution too epoch.ai/data-insight...

100 0

Hyperscaler Capex Is Up Because of Component Prices, Not Compute Ambitions | Currence Blog

… it would seem that a huge chunk of that doubling is just paying for the increase in component cost rather than increasing the rate of the actually deployed megawatts of capacity. www.currence.ai/blog/q3-2026...

currence.ai

Currence's Q3 2026 Data Center Outlook tracks the buildout project by project, from equipment lead times to the labor gap shaping who can build.

100 0

If DDR5 costs 3x (dam.stanford.edu/memory-price... ) what it did last year, in addition to other components increasing in cost due to inflation and higher demand, and the hyperscalers have doubled their data center investment from last year …

100 0

joe @joewilliams.bsky.social · 16/09/2026

It seems like no one agrees on how much AI infrastructure is contributing to GDP growth. Here’s three different views I’ve seen so far: * Goldman Sachs: 0.1% (pwm.gs.com/nam/en-us/in...) * St Louis Fed: 0.48% (www.stlouisfed.org/on-the-econo...) * Apollo: 1% (podcast.show/theviewfroma...)

100 0

people out here talking reaction functions like they used that phrase last year

000 0

This is definitely seeming more like FB playbook of defining how they wanted to be regulated before regulators have a clue so that by the time the regulation happens it's largely influenced by the scope, language and definition provided by the companies themselves and seems like the obvious solution

000 0

metr.org/blog/2026-08...

000 0

How many times do we need to relearn solving intrinsically human problems with technological solutions is exceedingly difficult if not impossible?

130 0

Slowing down only helps when those risks are addressed. If a month from now nothing really has changed it’s because this whole thing is to control the narrative rather than de-risk.

100 0