Will Rosecrans's Avatar

Will Rosecrans

@willrosecrans.com.bsky.social

215 Followers  |  146 Following  |  1,824 Posts  |  Joined: 15.07.2023  |  2.3322

Latest posts by willrosecrans.com on Bluesky

So, big tech has spent hundreds of billions of dollars of capital expenditures, and much of those expenditures are on GPUs. These GPUs, once installed, immediately begin decaying in value, and the more you use them, the more likely they are to crap out, and even when sitting idle, these GPUs have some power consumption — somewhere between 7% and 14% of their power in the case of the H100 — making them a consistent burden on the infrastructure of any purchaser whether they have customers or not. This becomes an even bigger problem for more expensive GPUs like the B200, which has an idle power draw of 140W per GPU. 

Extrapolate that out to thousands, tens of thousands or hundreds of thousands of them, or to the 50,000-per-building of them inside Stargate Abilene’s NVIDIA’s , and you’ve got yourself a problem. 

Even without customers, even without IT load, these servers sit burning power, and while a company could turn them off and boot them on for a period of time, these racks are often run in clusters — servers of GPUs that are synced together using high-speed networking — and while I can’t find anything firm as far as how long it takes to set up a cluster, this presentation on cluster design from the CEO of Lambda details a cluster’s sheer complexity, involving discrete networking, software layer and storage methods just to get the bloody thing working. 

In simple terms, it may be expensive to leave them on, but it’s going to take a bunch of time and energy to restart these things, meaning that they are more than likely left on, slowly burning energy.

Imagine how bad that is at a scale of hundreds of thousands of GPUs? 

Yet things might be a little wonkier, as the reality may be that while hyperscalers have customers, those customers might not be using all their GPUs at all times. Although there may be some customers that agree to a minimum spend or throughput, it’s hard to imagine that these contracts are enough to surmount the destructive margins. 

And when …

So, big tech has spent hundreds of billions of dollars of capital expenditures, and much of those expenditures are on GPUs. These GPUs, once installed, immediately begin decaying in value, and the more you use them, the more likely they are to crap out, and even when sitting idle, these GPUs have some power consumption — somewhere between 7% and 14% of their power in the case of the H100 — making them a consistent burden on the infrastructure of any purchaser whether they have customers or not. This becomes an even bigger problem for more expensive GPUs like the B200, which has an idle power draw of 140W per GPU. Extrapolate that out to thousands, tens of thousands or hundreds of thousands of them, or to the 50,000-per-building of them inside Stargate Abilene’s NVIDIA’s , and you’ve got yourself a problem. Even without customers, even without IT load, these servers sit burning power, and while a company could turn them off and boot them on for a period of time, these racks are often run in clusters — servers of GPUs that are synced together using high-speed networking — and while I can’t find anything firm as far as how long it takes to set up a cluster, this presentation on cluster design from the CEO of Lambda details a cluster’s sheer complexity, involving discrete networking, software layer and storage methods just to get the bloody thing working. In simple terms, it may be expensive to leave them on, but it’s going to take a bunch of time and energy to restart these things, meaning that they are more than likely left on, slowly burning energy. Imagine how bad that is at a scale of hundreds of thousands of GPUs? Yet things might be a little wonkier, as the reality may be that while hyperscalers have customers, those customers might not be using all their GPUs at all times. Although there may be some customers that agree to a minimum spend or throughput, it’s hard to imagine that these contracts are enough to surmount the destructive margins. And when …

And when I say “destructive margins,” I mean so much more than just the power to run and cool them. There’s the construction of the data centers to house them, the necessary maintenance, and, of course, any and all debt used to buy them. 

While I can (and will!) say that for every $50,000 GPU bought, big tech needs to make another $100,000, I believe that the hole created by the GPU’s existence is likely larger than just $100,000, with every construction delay, power outage, or moment of idle compute adding dollars to the hole for each and every GPU. 

Need a very real example? The Information reported recently that Oracle lost $100 million between June and August 2025 on rentals of NVIDIA’s Blackwell chips:

In the three months that ended in August, Oracle lost nearly $100 million from rentals of Nvidia’s Blackwell chips, which arrived this year. That’s partly because there is a period between when Oracle gets its data centers ready for customers and when customers start using and paying for them, the documents show. It’s not clear what causes the gap or how Oracle plans to shorten it.

So, while we can’t say for sure, there’s a chance a large chunk of these losses were caused by the spread-out deprecation of those Blackwell GPUs — after all, its most recent quarterly earnings included $1.35 billion of depreciation, right?

Sidenote: For context, Oracle reported depreciation of $804m in the same period last year, and depreciation of $3.867bn for the whole of FY2025. This should give you a sense of how rapidly these GPUs are becoming a millstone on its earnings. 

…is there any way to actually recover that $100 million through the margins on these chips? Oracle only has 4.5 years of depreciation on these Blackwell Chips, and The Information reports that those GB200 server racks have a negative 100% gross margin according to internal Oracle documents. If that’s the case, we have a much, much bigger problem, but even if the margins aren’t literally negative…will ther…

And when I say “destructive margins,” I mean so much more than just the power to run and cool them. There’s the construction of the data centers to house them, the necessary maintenance, and, of course, any and all debt used to buy them. While I can (and will!) say that for every $50,000 GPU bought, big tech needs to make another $100,000, I believe that the hole created by the GPU’s existence is likely larger than just $100,000, with every construction delay, power outage, or moment of idle compute adding dollars to the hole for each and every GPU. Need a very real example? The Information reported recently that Oracle lost $100 million between June and August 2025 on rentals of NVIDIA’s Blackwell chips: In the three months that ended in August, Oracle lost nearly $100 million from rentals of Nvidia’s Blackwell chips, which arrived this year. That’s partly because there is a period between when Oracle gets its data centers ready for customers and when customers start using and paying for them, the documents show. It’s not clear what causes the gap or how Oracle plans to shorten it. So, while we can’t say for sure, there’s a chance a large chunk of these losses were caused by the spread-out deprecation of those Blackwell GPUs — after all, its most recent quarterly earnings included $1.35 billion of depreciation, right? Sidenote: For context, Oracle reported depreciation of $804m in the same period last year, and depreciation of $3.867bn for the whole of FY2025. This should give you a sense of how rapidly these GPUs are becoming a millstone on its earnings. …is there any way to actually recover that $100 million through the margins on these chips? Oracle only has 4.5 years of depreciation on these Blackwell Chips, and The Information reports that those GB200 server racks have a negative 100% gross margin according to internal Oracle documents. If that’s the case, we have a much, much bigger problem, but even if the margins aren’t literally negative…will ther…

Premium Newsletter tomorrow: Meta, Google, Amazon and Microsoft must add $2 Trillion in AI revenue by 2030 or they wasted their capex.

The cost of data centers GPUs may make it impossible to make a profit from AI. Here's $10 off annual.

edzitronswheresyouredatghostio.outpost.pub/public/promo...

30.10.2025 22:01 — 👍 959    🔁 153    💬 19    📌 15

Counterpoint is that tons of people were trying to get Biden to take structural court reform seriously and pack/expand.

In a scenario where some of this was taken seriously, it's entirely plausible you also get a different court deciding on immunity.

30.10.2025 17:06 — 👍 8    🔁 0    💬 0    📌 0

No doubt. And it's working, unfortunately. "Military occupation on American streets" already doesn't sell papers, and the nightly news isn't starting every broadcast with "Day X of military occupation."

30.10.2025 05:37 — 👍 1    🔁 0    💬 0    📌 0

The error rate against white male Republicans who don't hassle them doesn't matter because those guys don't get scanned. The error's effect is gated on the bias of who they *feel* like scanning so they can get a "neutral" procedural justification for abuse -> in practice it only harms opposition.

29.10.2025 22:41 — 👍 1    🔁 0    💬 1    📌 0

Absolutely the point. "Oh, you have papers? Well, my phone says I can do whatever I want." Even if six months from now you prove in court that you are a citizen and they have to let you go, they got to punish you for being an out group and blame a contractor for the decision.

29.10.2025 22:37 — 👍 1    🔁 1    💬 1    📌 0

Outsourcing of morality. If computer says what they want to hear, they can insist no human was responsible, and therefore no humans can or should be held accountable.

Training humans to hatefully abuse is... not hard, but not trivial. Training humans to *not* trust a computer is hard.

29.10.2025 22:14 — 👍 18    🔁 3    💬 1    📌 0

"Terds."

29.10.2025 22:09 — 👍 3    🔁 1    💬 0    📌 0

There are tons of American citizens that the government just... won't have any source data to verify against. Even if you (falsely) assume that this magically works 100% perfectly, not everybody has a government photo ID sitting in a database to match against. This is fucking nuts from any angle.

29.10.2025 22:08 — 👍 1    🔁 1    💬 0    📌 0

They apparently understood the documents were from the Biden administration, but didn't consider a permit issued while Biden was president to be legitimate.

Basically, personal loyalty to Trump as a person, zero loyalty to the office of the President or the Constitution. Historically, very Nazi.

29.10.2025 20:50 — 👍 0    🔁 0    💬 0    📌 0

Can we? Of course we can do it.

Will we? What, you mean LA would actually have to *do* it instead of just doing a PR photo op treating it as a fait acompli and then wasting some money and not ever doing any of it?! Naww, probably not.

29.10.2025 20:38 — 👍 4    🔁 0    💬 1    📌 0

"Rabbit, Kat, and Straw" does sound a bit like either a 70's folkpop band or a team of assassins in a B-movie, right?

29.10.2025 19:37 — 👍 2    🔁 1    💬 0    📌 0

I doubt there's a real legal principle that DOJ making a typo means you go free rather than DOJ just being told to submit a corrected document. But at this point I can sort of imagine an annoyed judge just running with it.

29.10.2025 18:17 — 👍 1    🔁 0    💬 2    📌 0
Preview
Costa Rican Man Dies After Health Decline in U.S. Immigration Custody A man from Costa Rica, died after returning from U.S. ICE detention in critical condition, prompting family demands for investigation

🧵Here are more details on Randall Gamboa, the Costa Rican man who was snatched by ICE in Dec, detained for 8 months in facilities in Texas, then deported back to Costa Rica via air ambulance in a vegetative state with brain damage.

He is now dead.

28.10.2025 16:50 — 👍 176    🔁 127    💬 17    📌 25

That's pretty much it. There's plenty of power in the world to "make them." But the people with that power don't want the inconvenience so they also adopt "make me" rather than do anything.

29.10.2025 17:28 — 👍 8    🔁 0    💬 1    📌 0

Weird how they believe they are the majority because of what they read on Twitter, but have trouble finding all those people who agree with them in real life, but still think most everybody agrees with them.

29.10.2025 16:17 — 👍 2    🔁 0    💬 1    📌 0

The only hard part about going "48 hours with No AI" is that every fucking program tries to force it on me and there's no way to turn the shit off. I'd fucking love to go 48 hours with being assaulted by it.

29.10.2025 03:40 — 👍 0    🔁 0    💬 0    📌 0

It's impressive that you can be so bad at governmenting that you create an emergency out of the register of copyrights. That's like being so bad at cooking that you have to call 911 about pancakes.

29.10.2025 01:25 — 👍 0    🔁 0    💬 0    📌 0

A year ago, Biden's staff were claiming he was 81. Now they are claiming he's 82! WHICH IS IT? If he's really 82, why did they lie about it last year? If they lied last year, how can we trust them now? CONSPIRACY OF THE CENTURY.

28.10.2025 22:51 — 👍 1    🔁 0    💬 0    📌 0

They know if they try a fight on substance, they have to talk about his sensible and broadly popular platform. But if they derail the conversation into definition of aunt, they displace any discussion of actual policy for BothSidesy "we got called liars for calling him a liar," a net plus for them.

28.10.2025 18:26 — 👍 1    🔁 1    💬 1    📌 0

It'll be interesting to see what happens if DOJ's conviction rate falls under 50%. Imagine judges routinely giving jury instructions with stuff like "When evaluating his credibility, consider the fact he intentionally violated a court order illegally turning off his body camera and assume the worst"

28.10.2025 18:06 — 👍 4    🔁 0    💬 0    📌 0

It's stupid. But it's human stupidity as exploited by an evil megacorporation's evil AI. A Trillion dollars depends on keeping people like him emotionally invested in engagement stats, and that makes it not just a him problem.

28.10.2025 04:12 — 👍 6    🔁 1    💬 0    📌 0
Post image

A hurricane's "hiding" between the weather radar coverage. But the lightning sensors are much longer range so you can see a cluster of lightning strikes coming from the storm that's just over the horizon. Very spooky. Like almost detecting a cloaked ship in Star Trek.

www.windy.com/-Weather-rad...

28.10.2025 04:09 — 👍 1    🔁 0    💬 0    📌 0

These days it just seems to depend on whether the people being sued want to bribe Ingrassia with a legal "settlement." All of Trump's big recent victories were basically just payoffs. I don't think anybody needs $150M worth of bribe work product from Ingrassia though, so it's cheaper to win.

28.10.2025 02:37 — 👍 0    🔁 0    💬 0    📌 0

It speaks to how little they have of substance. They hate his policies, but they also seem to understand that talking about his policies would be a losing strategy for them.

28.10.2025 02:21 — 👍 1    🔁 0    💬 0    📌 0

If a life saving vaccine had actual side effect numbers like that, the current HHS secretary would already have had the makers executed.

27.10.2025 22:17 — 👍 1    🔁 0    💬 0    📌 0

A product is correlated with over a million people a week expressing suicidal intent. This is now considered fine, and it's considered very important in many circles to do whatever it takes to feed the AI bubble.

27.10.2025 21:09 — 👍 0    🔁 0    💬 0    📌 0

Yup, monarchies always accepted becoming Constitutional monarchies because the alternative to courts was torches and pitchforks.

Tip to tail, the rich and powerful have dismantled the system they didn't believe was *protecting* them from baser applications of power.

27.10.2025 19:55 — 👍 1    🔁 0    💬 0    📌 0
Post image

It would be really hard to argue that "securing oil" in Kuwait wasn't a goal, given all the iconic historical images of US troops standing in oil fields, keeping them secure, after having fought to secure Kuwait's oil...

Did they do it by accident? Against orders?

25.10.2025 23:15 — 👍 21    🔁 2    💬 1    📌 0

Completely fucking irresponsible to post AI generated misinformation as if it was real.

25.10.2025 21:22 — 👍 0    🔁 0    💬 0    📌 0

When you see the steady drumbeat of arrests of Russian recruits for sabotage etc across the UK and Europe, and no equivalent announcements in the US, you can be sure we are missing cases, and mislabeling other arrests.

25.10.2025 09:01 — 👍 387    🔁 144    💬 15    📌 4

@willrosecrans.com is following 20 prominent accounts