I run a small open source LLM inference company, Synthetic.new. As far as I can tell, CNBC isn't reporting this accurately: the problem isn't that Oracle is building "yesterday's data centers": they're building Blackwell DCs! Those are today's DCs.

The problem appears to be that Oracle is building today's DCs... Tomorrow. And by the time they come online, Vera Rubins will be out, with 5x efficiency gains. And Oracle is unlikely to want to drop the price of Blackwells 5x, despite them being 5x less efficient.

It's a little unclear to me how bad this is. Nvidia's "rack scale" machines like GB200-NVL72s and GB300-NVL72s are basically a fully built rack you roll into a DC and plug into power and network. In that case, Oracle should probably just buy the rack-scale Vera Rubins when they come out instead of Blackwells and roll them into their new DCs. Tada! Tomorrow's DCs, tomorrow.

OTOH it's possible someone at Oracle screwed up and committed to buying Blackwells at today's prices, delivered tomorrow. Or maybe construction of the physical DCs is behind schedule, so today's Blackwells are sitting around unused, waiting for power and networking tomorrow. Then they're in a bit of trouble.

Regardless, CNBC's reporting seems pretty unclear on what actually happened and whether this is actually bad or not.

I really don't want to overrule your expertise in this regard, but an 5x efficiency gain in a single generation feels like its too much, especially considering how newer process nodes have been yielding less and less improvements.

Just to compare and contrast:

https://www.videocardbenchmark.net/power_performance.html

Here's a synthethic benchmark page listing every GPU in recent memory. True, its not AI, but if we look at the 1080 Ti, a 9 year old card at this point, and compare it with the 5090 we see the gains were 190/74=2.56x in that timespan that involved multiple die shrinks and uArch changes.

I think these numbers might not hold up on IRL workloads, and afaict older datacenter cards still hold up well and are being used in production.

Newer process nodes are not the main avenue of improvement. What those transistors are used for is more important and it’s plausible that improvements between generations can increase performance by multiples on a specific task. All of the improvements aren’t necessarily in the chip itself either.

E.g. the next gen might have hardware inference for lower bits, more memory bandwidth, etc.

Almost seems as if microchips are approaching their "B-52 age":

"Those things are still flying! Introduced in 1955!"

"But that was the B version, all those that are still flying are the H version, so many iterations between them!"

"Welcome to 1962"

> but an 5x efficiency gain in a single generation feels like its too much, especially considering how newer process nodes have been yielding less and less improvements

The efficiency is in other areas too e.g. memory, network, etc. It's TOTAL.

> Here's a synthethic benchmark page listing every GPU in recent memory

We don't have the GPU gains not because of process nodes. Nvidia and later AMD stopped investing in that direction. They started optimizing for AI not graphics.

Likely aimed at classified/defence environments. In those spaces, hardware typically takes 18–36 months after commercial deployment before it’s approved—due to firmware vetting, side-channel analysis, crypto validation, and similar processes.

Meanwhile, commercial operators have already deployed their hardware for public workloads. Existing Blackwell capacity won’t just be shifted into classified environments—governments don’t repurpose hardware from unclassified infrastructure for secret/TS systems. That deployed stock will stay in the private sector for hosted AI workloads.

For many high-security use cases, new Blackwell systems may effectively be the only viable option, especially given the slow review cycles around new firmware and GPU software stacks. Newer chipsets will also be prioritized for training due to performance gains.

Oracle likely recognizes this dynamic and is betting competitors may eventually need to deploy in their data centers. Governments haven’t historically deployed GPU capacity at this scale-beyond ASIC/FPGA crypto workloads.. and likely don’t have large pools of pristine Blackwell hardware available.

They’re also purchasing late in the cycle, which may work in their favour.

they are saying what you are saying. At least Deirde Bosa did. I think there is a lot of folks internally who don't understand the gravity of it and keep questioning it.

You are right about the building of today's DC's. There is a small part of me that feels Oracle might be a bit toxic long term with all this debt him and his kid have taken on. And this could be the first reaction to it.

But this is exactly why Oracle is the wrong partner. They don’t get it. They never will. To them, it’s just a “workload”.

> Nvidia's "rack scale" machines like GB200-NVL72s and GB300-NVL72s are basically a fully built rack you roll into a DC and plug into power and network. In that case, Oracle should probably just buy the rack-scale Vera Rubins when they come out instead of Blackwells and roll them into their new DCs.

This is what I don't understand. Why is the article making the assumption that the DC itself is tied to a particular GPU generation? AWS doesn't knock down a building and start over every time Intel releases a new Xeon.

Xeons have a much longer shelf life and diverse workloads. If you order hardware specifically for LLM inference and then some new hardware/model combination is much better at that (which it will be, because a lot of people are working on that), you might be in trouble.

It's like setting up a warehouse of GPUs to mine bitcoin while others are switching to ASICs.

Training you mean. Doing inference on last year's chip is probably ok, but training a frontier model on it is going to be a deal breaker.

No I mean inference. The idea is that inference demand will be massive and a race to the bottom with razor thin margins.

Training costs can be amortized over the entire lifetime of the model, but if you lose money on inference or can't offer competitive usage limits for subscribers, there's no amortizing that.

While we have you here, could you please clarify a point in your privacy policy?

> For data collected from the UI or other usage: We retain the personal information described in this privacy notice for as long as you use our Services

I have two quick questions:

1. Why are UI prompts and responses kept for the entire life of the account?

2. When an account is closed, is the data actually deleted or just de-identified?

5x improvement of energy efficiency in just GPUs translates to more like 50% reduction of power usage, with is significant but doesn't warrant a 80% reduction in pricing. Especially since Nvidia will charge more for the same card - they have been pricing things pretty aggressively.

And on the DC side they will be building to a power and heat budget. If Vera Rubin changes the power density per rack equation that may have some impact. But thinking rationally if the flops per kw-sq ft are lower than Blackwell, no problem. If they are a lot better then even if the kw per sq ft is higher you can just space the racks out a little

> Or maybe construction of the physical DCs is behind schedule, so today's Blackwells are sitting around unused, waiting for power and networking tomorrow. Then they're in a bit of trouble.

Other reporting says this is very much the case. Stargate barely has some of the land cleared, but the buildings were supposed to be finished and have GPUs installed over the course of 2026.

There's also the indicator of Nvidia giving out billion-dollar deals to other companies such that they could commit to buying even more Blackwells to keep production going. The chips from those new deals don't have anywhere to go, everyone already spent their cash on getting shipped chips that they're still installing today (apparently some are even in warehouses)

There are two generations and 4.5 years between A100 and B200.

A100 has 312 TFLOPS of FP16 for 250W, i.e., 1.25 TFLOPS/W.

B200 has 2250 TFLOPS of FP16 compute for 1000W, i.e., 2.25 TFLOPS/W.

This is ~34% growth per generation and ~14% per year. It's hard to believe it will be 400% per generation this time

It might be 400% in the one thing everyone is interested in.

you think in FP16. nobody uses FP16 for inference anymore. 400% probably for FP4/INT4 computation.

Tensor core performance is inversely proportional to precision across all generations (i.e., reducing precision by a factor of 2 increases OPS by a factor of 2). 8-bit precision will give you the same improvement ratio. A100/H100 didn't support 4-bit if I remember correctly.

So FP4/INT4 will likely improve the same 30% OPS/W. You could get a separate improvement by reducing precision, but going 1-bit for 4x improvement feels unlikely for now.

The issue is really about if the DC is water-cooling capable

One could argue that the headline is correct then. Today is tomorrow’s yesterday.

Hey Reiss, I just checked Synthetic. So nice to see indie providers for smaller LLMs. I am personally building products to run only with small (actually < 20b) models. My aim is for laptop usage. Would love to know what plans you have for models smaller than you have currently. Industrial use is all about smaller models IMHO

> The problem appears to be that Oracle is building today's DCs... Tomorrow.

By the time Vera Rubins will be available on scale, will they immediately be put into DCs, or will tomorrows chips be running.. the day after tomorrow?

Thank you for Synthetic.new

I moved over from OpenRouter and it's been breezy. I hope you are sustainable at $30/month and are successful!

Interested to know more about your inference start up? How you guys operating, do you own hardware or use the cloud?

> Oracle should probably just buy the rack-scale Vera Rubins when they come out instead of Blackwells and roll them into their new DCs. Tada! Tomorrow's DCs, tomorrow.

Or we‘ll get a supply problem and they get nothing or not enough. Tomorrow’s DC, never. Tada!

I love how you explained this

Supply chain at volume makes it hard to to little else.

Isn't this a problem for everyone buying Nvidia GPUs at scale?

I think the difference is that the other hyperscalars are doing this out of the enormous cash rivers produced by their other profitable businesses, at a rate less than that at which profits are flowing in, whereas Oracle is funding it out of debt with AI capex in 2026 projected to reach levels nearly as high as their expected revenue (not profits) in the same period.

If the hardware refresh rate makes a substantial share of data center cost function more like opex than capex, the companies funding it out of operations (especially from operations of what are essentially monopoly businesses, in the sense pricing power), even if it isn’t the operations it power specifically, are fine in the near-to-intemediate term (barring exogenous shocks to those other businesses), whereas Oracle, funding it by a debt bonanza, is in a different position.

Google, Amazon, Meta, etc don't have to wait 12 or 24 months for their big data center to open. They already have lots of DCs to cram all the NVidia cards into, right now.

Definitely not true. Meta is building tents for GPU's

> Meta is building tents for GPU's

And Starlink / xAI is going to shoot them into space. We are simultaneously living in the future and the past.

> And Starlink / xAI is going to shoot them into space.

I highly doubt that. They claim they want to shoot them into space, but I don’t believe a word of it until I see it happen (and see it work). It’s no more real than hyperloop.

DCs in space is hype but actually makes no rational sense when you figure the size of radiators you'll need, and while solar cells are more efficient in space, they aren't that much better.

The Google paper (https://arxiv.org/pdf/2511.19468) didn’t seem too concerned with radiator mass/size when I skimmed it, but maybe I just missed it. My understanding is that if you run the chips relatively hot (and maybe boost with heat pumps? But then you’re not quite as solid state, and maintenance is tough up there), the radiation ability increases enough such that you can make the radiators slightly smaller than the solar panels, and they’d sit on the dark side of the panels. Many people like to point to the ISS system and scale that up, but there’s a big difference between a system assembled in space and meant to keep humans at human temps vs mass manufactured on the ground and keeping things around 100C.

Well, the sun is always up in space, so yeah they're at least 3x better from that alone.

Tents? Like.. where sleep in the woods in?..

I think it's less a matter of space and more a matter of power availability

Only the ones that require profit from the GPUs.

Next servers might need more power or different cooling. Then your DCs are just big concrete rooms.

All DCs are big concrete rooms that can supply so much power per sq area and remove so much heat per sq area (the two related of course since the heat comes from dissipating the power). Variation is just in density of whatever sort of fancy resistor you plan to put in the concrete room.

Thanks! "thinking resistors" will be the name of my future AI-punk band.

I hope the lawnmower goes bankrupt with this and the hostile WB take over.

Don't anthropomorphize the lawnmower.

From the consumer perspective the last thing I want is a Netfli-xation of WB…

I also don't want the next Bari Weiss in charge of WB

It was good, actually, that she suppressed accurate news unfavorable to the current President, tanking ratings and the network's credibility. I want more news outlets to decline. Except of course, my favorite, which doesn't count. It says it's fair and balanced; they wouldn't lie, would they?

I hope this is satire

On the contrary, Netflix would have been decent because WB is bigger than them (in terms of IP, existing content, brands, etc) and would have probably (at least that's what they said publicly) left it mostly alone. And it's weird how people assume that just because Netflix produce a ton of content, all of it is low quality. A lot of it is for people half paying attention, but there is plenty of actually good stuff. Them having WB would improve them on both fronts.

Nepo baby is coming with a political angle and wants control of the news media part of WB. The American media landscape is already without much competition nor diversity in political views, now there would practically be none.

> I hope the lawnmower goes bankrupt with this and the hostile WB take over.

Unfortunately there is no chance of that happening.

At his level of personal wealth there is no realistic scenario that leads to personal bankruptcy. In our current capitalist society once you're into the billions you're "too big to fail" and you have unlocked the infinite money glitch.

The only consolation is the lawnmower is 81 and thus is going to be dead soon (even the mega-wealthy can't plastic surgery themselves out of this outcome, at least not yet) and he can't take any of it with him. But all indications point to his progeny having aspirations to be even more damaging to society than he has been.

> The only consolation is the lawnmower is 81 and thus is going to be dead soon

Reminder to lay up your treasures in heaven.

Thomas Piketty would love to have a word.

Piketty’s central argument is that when the rate of return on capital (r) exceeds the rate of economic growth (g), wealth concentrates over time into fewer and fewer hands. This is his now-famous r > g inequality.

The implication is that capitalism, left to its own devices, doesn’t naturally spread wealth around. It does the opposite. The relatively egalitarian period of the mid-20th century (roughly 1930s-1970s) was the historical exception, driven by two world wars, the Great Depression, and deliberate policy choices like progressive taxation. The longer historical pattern, which Piketty traces with extensive data going back to the 18th century, is one of increasing concentration.

His practical prescription is a global progressive tax on wealth (not just income) to counteract this tendency. He acknowledges this is politically difficult but argues it’s the most straightforward mechanism to prevent a return to the kind of patrimonial capitalism that defined the Gilded Age and the Belle Époque, where inherited wealth dominated and social mobility was minimal.

The book’s real contribution was less the theoretical claim (which economists had gestured at before) and more the empirical work. Piketty and his collaborators assembled an unprecedented dataset on wealth and income distribution across multiple countries and centuries, which gave the argument a weight that prior discussions lacked.

None of that applies to individually wealthy people, who have a long history of going bankrupt.

Which ones? The Sacklers are a prime example of how impossible it is to actually go bankrupt; considering they harmed millions of people, had the government step in and still remain one of the wealthiest families in the US.

x=1 when n=1, therefore all x=n

Who was the last billionaire that went bankrupt, involuntarily?

"Filed for bankruptcy" != "out of money" in the ordinary plebeian sense.

These are the kind of criminals where the judges will let them stay under home arrest in their twenty-bedroom mansion, have their chauffeur drive them around in a car worth more than my entire life savings, etc... because it would be "unconscionable" for them to lose the life that they're accustomed to. I.e.: Affluenza.

Just look at Prince Andrew or whatever he's called now. He raped children and his rightful punishment would be to sit in a jail cell with no access to anything even resembling his lavish digs, instead he's luxuriating in a lifestyle you and I would envy.

I can list far, far more examples of billionaires or mere hundred-millionaires living luxuriously after committing capital crimes or "going bankrupt" than not.

Find me an ex-billionaire living out of a motor home, then I'll cede your point.

> These are the kind of criminals where the judges will let them stay under home arrest in their twenty-bedroom mansion…

Another egregious example of this sort of thing:

> Robert H. Richards IV was convicted of rape, the wealthy heir to the Du Pont family fortune […] received an eight-year prison sentence in 2009 for raping his toddler daughter, but the sentencing order signed by a Delaware judge said “defendant will not fare well” in prison and the eight years were suspended.

— https://www.cnn.com/2014/04/02/justice/delaware-du-pont-rape...

Talk about a two-tier justice system.

I am aware; "barely" is exclusively used to describe items that surpass the threshold.

I don't even know thr point you're arguing as the intro to the listicle concurs with my main argument:

>> It is very rare for a person to achieve the status of billionaire and then lose it.

>The book’s real contribution was less the theoretical claim (which economists had gestured at before) and more the empirical work.

Empirical work... like conveniently ignoring the fact that there's far less old money billionaires than we'd expect?

>For these lucky people, the experience of the Vanderbilts and their contemporaries offers a cautionary tale. At the turn of the 20th century, America’s census recorded about 4,000 millionaires, note Victor Haghani and James White, two wealth managers, in their book, “The Missing Billionaires”. Suppose a quarter of them had at least $5m (the richest had hundreds) and had invested it in America’s stockmarket. Had they then procreated at the average rate, paid their taxes and spent 2% of their capital each year, their descendants today would include nearly 16,000 old-money billionaires. In reality, it is a struggle to find a single one who traces their fortune back to the first Gilded Age.

https://www.economist.com/finance-and-economics/2025/06/12/h...

> In reality, it is a struggle to find a single one who traces their fortune back to the first Gilded Age.

This is a good point because there are no oil billionaires and things like trusts, family offices, offshoring etc. actually pose no challenge to accurately numerating and identifying people that ‘have’ or effectively control over a billion dollars at their discretion because they all just sign up for the list.

Of course there’s the Panama Papers and the Paradise Papers but that doesn

There's 100+ people the FBI had tabs on for sex trafficking related to Epstein.

So far the only individual that has been meaningfully punished has been Ghislaine Maxwell.

This seems like a prime example of being too big to fail. The FBI puts on kid gloves whenever a rich person is accused of wrong doing.

>There's 100+ people the FBI had tabs on for sex trafficking related to Epstein.

>So far the only individual that has been meaningfully punished has been Ghislaine Maxwell.

That factoid is meaningless without the rate of prosecutions/convictions for people that FBI "had tabs on".

What rate are you looking for?

With J6, in the matter of 2 or so years the FBI has secured over 1000 convictions.

When it wants to, the FBI can move very quickly.

They also had all those guys on camera doing crimes in Washington DC and bragging about it on Twitter.

FBI was literally sitting on Epstein files for years. They have chosen not to prosecute. When the state tried to investigate Epstein, FBI came in, took control over with claim they will share investigation results ... and then did nothing.

I seem to remember him dying in a jail cell?

Yes, which happened only because a journalist broke a story about how FBI was not investigating Epstein for years and years. It was media who forced that to happen, after decades of abuse FBI was aware of.

Speaking of which, the previous conviction, the super sweet deal Acosta gave to Epstein before is also an example of elite unaccountability.

FBI and friends protected Epstein until it became impossible.

Ok but can't you say that about most non-street-level convictions? That they required some amount of complaining for a detective to do their job?

J6 is not a strong counterexample, IMHO. Part of the problem with Epstein is "proof beyond a reasonable doubt," for which evidence is needed--and, it appears, hard to come by. Whereas with J6, there were thousands of hours of footage showing the crimes being committed (and in many cases bragged about), which made prosecutions much easier.

>With J6, in the matter of 2 or so years the FBI has secured over 1000 convictions.

Again, large numbers, but no context. How many people did you think were at the riots? 10k? 50k?

Moreover, Jan 6th was an event that definitely happened. The same can't be said for whatever happened at Epstein's island. The island exists, Epstein's a convicted sex offender, and people flew there, but associating with sex offenders isn't a crime, no matter how despicable it might seem.

It's sexism in action; the woman gets punished while "boys will be boys." Prove me wrong. Epstein himself is probably still alive in Tel Aviv anyway.

> It's sexism in action; the woman gets punished while "boys will be boys." Prove me wrong.

Epstein died in his cell. If Maxwell preferred death to punishment she could've also killed herself. Also it's well documented that women receive less harsh punishment in court vs men for the same crimes, so yeah, it's sexism but not in the way you insinuate.

> Epstein himself is probably still alive in Tel Aviv anyway.

Yes, and it's Maxwell's lookalike that's serving the sentence, while she's enjoying herself in Argentina. See how quickly you can derails discussion with such absurd claims without any substance?

The "all in podcast"

Your original comment involved claims of "infinite money glitch". Now you're walking it back to "rich people are rich"?

In the ballpark ($2-4 billion in the 1970s): https://priceonomics.com/how-the-hunt-brothers-cornered-the-...

Ten Billionaires Who Went Broke

https://finance.yahoo.com/news/10-billionaires-went-broke-15...

> At his level of personal wealth there is no realistic scenario that leads to personal bankruptcy. In our current capitalist society once you're into the billions you're "too big to fail" and you have unlocked the infinite money glitch.

This is plainly false. There are plenty of example, even recently, of billionaires losing their fortunes or going bankrupt. Often they come with criminal prosecution because they get desperate and try illegal ways to hang on to their wealth. Sam Bankman-Fried, Elizabeth Holmes, and several other examples come to mind.

There are a lot of stories of billionaires getting too risky with their investments or too concentrated in businesses and losing the majority of their wealth. The Barclay story, Jim Justice, the old Peloton CEO.

It’s not a common outcome because you have to try hard to screw up that badly when you have over a billion dollars in wealth. Parking it anywhere in common investments would leave you and your ancestors set forever.

> Sam Bankman-Fried, Elizabeth Holmes, and several other examples come to mind.

Billionaires that were dumb enough to attempt to screw even bigger billionaires. Sure you can find exceptions to the rules, but Ellison isn't going to be one of those.

Not only will he die, he is obsessed with his mortality and thinks about death every day. That's no way to spend your life. He's a massive fool.

> Haven't been following Bryan Johnson, eh?

he's gonna die just like the rest of us, just with a slightly odder uncanny valley look for himself when he goes.

Willing to bet good money he won’t outlive someone’s grandpa who smokes two times a day

I don't believe that Stargate is "yesterday's data center". It's being built in multiple phases and Oracle has access to Nvidia's roadmap. They know 200 kW/rack is coming. The newer phases could easily be built out to support Rubin and Feynman.

200 kW/rack is absolutely insane to me. The power consumption of these facilities is just...ridiculous.

With respect to consumption, it’s pretty efficient vs older traditional servers, though I know workloads like that aren’t completely fungible. Nonetheless it bears keeping in mind that a single GB200 NVL72 rack provides 1.4 ExaFLOPS of AI compute (at FP4 precision, ideal circumstances, but this is envelope math all around). So it’s power efficient, for what it is.

Oh, I have no doubt it is functionally efficient. I'm just amazed given the system deployments I've been party to, and the tiny amount of per rack energy usage comparatively speaking given the functionality of those systems.

Like, what in the good god damn are we using all this energy for?

> Like, what in the good god damn are we using all this energy for?

Bad AI porn, terrible AI music, AI scams and completely devastating the labor market.

And based on the recent Anthropic/Pentagon rift... I guess also creating autonomous kill-bots and doing mass surveillance.

Just a bunch of super cool stuff.

You left out overthrowing governments with customized targeted propaganda, jamming citizen discussion with noise, artificially creating and nourishing contrarian cells in democratic societies. The machines will now be programming people.

>and completely devastating the labor market.

? lol

Since you and me and everyone else will foot the elecricy bill. Energy consumption or efficiency is not a concern.

It's the water use that concerns me

In theory the water stays clean and can be reused. But I assume these cheapskates will go for evaporative cooling everytime? Then yeah, we need laws against that.

I guess when you're dissipating upwards of a gigawatt of power at a single site boiling water starts to look attractive. It's a pretty impressive curveball; I definitely would never have predicted "an evil corporation boils off all the local drinking water" to be a legitimate concern. I'm pretty sure that's too absurd a plot point for even a children's movie.

> an evil corporation boils off all the local drinking water

Nestle jumps into my mind whenever I want to think of an evil corporation and water together.

Yeah because it's cheaper they go evaporative. That's an easy fix by just making it more expensive.

People talk about the water usage like it's an intrinsic feature of datacenters; it's not. You just make it more expensive so they are forced to conserve. But you wait till you have to so you don't push them to build elsewhere.

I keep hearing people claiming that water is just as much as issue as energy for operating these DCs, but that just doesn't make any sense to me. However, I haven't had to step inside a DC for almost two decades.

Continuously dissipating 1 gigawatt of energy by boiling room temperature water would require approximately 1.38 million liters of water per hour.

Seems like the environmentally responsible thing to do be to build the datacenter near the coast and use the waste heat to desalinate water. Or at least dissipate the heat into the ocean rather than boiling off an inland freshwater supply.

And kill the local aquatic life as you raise the temp beyond their happy place?

Setting aside a small patch of ocean for the task seems like a much better plan than the current practice. Provided you dump it in a place with a decent current any adversely affected area should be exceedingly small.

Keep in mind that the sun is constantly dumping energy on us. Absorption averaged across the entire earth is ~200 W/m^2. Assuming I didn't misplace some zeros somewhere then a gigawatt corresponds to ~5 km^2 of ocean surface. That's the daily flux. Penetration falls off exponentially so 75% of that only ever makes it ~10 m down.

I think the takeaway here is the utterly incomprehensible scale of the ocean.

run it through a turbine and generate electricity to power the datacenter - infinite energy and infinite ai unlocked.

This idea is probably more worth it in middle eastern countries given that 90% of their water comes from Desalination Plants. But given the recent war within region, I don't really expect Datacenters to be built within the region for quite a long time.

The compute power is also ridiculous.

Some of the reason for the high density is that you need devices physically close to each other to share such bandwidth. It’s not because we’re limited by the physical building space, because we can construct buildings all day long. Sending bits around at ultra high speed is hard and you need to keep all of the devices physically close to avoid having your interconnect costs explode.

Interestingly the realm in which I have domain experience has similar constraints, but based primarily on physical transport latency and less on bandwidth. There has been a move in some spaces towards hyper-dense deployments, but it’s a very small amount of the total compute capacity due to other limitations.

Still, the world I’m used to operating in is typically 5-10 kVA/rack.

There are box trucks with less power consumption.

So what's the theory that goes with this about why cnbc are reporting that openai are walking because they want newer nvidia hardware? CNBC are clueleess? People at openai are lying to cnbc? cnbc are fabricating stories while drunk?

There has to be some theory to explain the story to be consistent with this comment.

Something is probably happening but I don't know what it is. Maybe this is really a negotiation over price.

I agree with you more than I agree with the parent comment.

To use the hit HBO TV show silicon valley analogy, it is far more likely that "the bear is sticky with honey" will happen at Oracle than at Open AI. Some kind of game of telephone gone wrong at some point and now the people responsible at Oracle must double down in order to kick the can to the next quarter and not appear clueless.

Statutory disclaimer: I am not affiliated with either Open AI or Oracle and have no insider information. All of this is mere conjecture and has no basis in reality.

Yes and CNBC is comically rife with payola content. I just want to know who’s buying.

Diedra is a solid reporter with pretty good access and understanding.

Diedra Bosa is a good journalist.

What happens to older datacenter GPUs? Do they have a second life somewhere outside of datacenters?

I could see Nvidia adding terms of sale requiring disposal rather than resale.

Plenty of enterprise server hardware (racks, servers, RAM, disks) does have an active secondhand market after 3-5 years of use, but I think GPUs are too specialized for it to be viable. I doubt anyone has the setup to run a H200 in their home rig.

I also don't think companies are going to have mandatory replacement cycles for GPU hardware the same way they do for everything else, because:

1. It is an order or magnitude (or more) more expensive.

2. It isn't clear whether Moore's law will apply to the AI GPU space the same way it has for everything else.

Unless Nvidia can launch a new chip every 2-3 years with massively improved performance-per-watt at a lower price no one is going to rush to recycle the old one.

> Unless Nvidia can launch a new chip every 2-3 years with massively improved performance-per-watt at a lower price no one is going to rush to recycle the old one.

That's exactly the point.

Performance/watt is increasing so much gen-to-gen that it makes no longer sense to run older hardware.

Not my words, Jensen's.

Are you saying that the person selling shovels thinks you should buy a new shovel? I guess they must be the expert.

Are you saying that there hasn't been massive improvement in performance per watt metrics for datacenter GPUs, which directly affects performance and profitability of said DCs?

So you are saying that the ceo of the company that builds the chips is saying that it makes sense to change them each generation?

When the warbirds are on the wing, sell anti-aircraft systems to both sides.

you can absolutely run e.g. datacenter-level A100 at home, there are adapters from the SXM to the PCIe socket. Haven't seen people running SXM versions of H100s this way but this could be due to the price factor only

Well technically true, I would wager that the home lab is going to require increasingly distinct and unusual adaptations to retrofit the hardware to home use.

New stuff is all liquid cooled by default and that's a paradigm shift for your average home lab.

I'm less aware of exactly what's happening on the power side of things but I think some of the architectures are now moving to relatively high voltage DC throughout and then down converting it to low voltage right before it's used. So not exactly just plug-and-play with your average nema15 outlet.

Well by the time the become obsolete you can run that computing on a Mac with no special cooling so I really doubt they will be of any use. Maybe in some parts of the world where electricity is cheap. If someone wants to really find out perhaps watching the crypto ASICs stories could help.

> I doubt anyone has the setup to run a H200 in their home rig.

There are PCIe versions of these right? And another comment is saying there are PCI adapters too. It "only" requires 600 to 700W. It's not out of reach for everybody.

If the used regular server market is any indication, you can find, after a few years, a lot of enterprise gear at totally discounted prices. CPU costing $4K brand new for $100 after a few years: stuff like that.

A friend has got a 42U rack and so do some homelab'ers. People have been running GPU farms mining cryptocurrencies or doing "transcoding" (for money).

It's not just CPUs at 1/40th of their brand new price: network gear too. And ECC RAM (before the recent RAM craze).

I'm pretty sure that if H200 begin to flood the used market, people shall quickly adapt.

> Unless Nvidia can launch a new chip every 2-3 years with massively improved performance-per-watt at a lower price no one is going to rush to recycle the old one.

I agree with that. But if they resell old H200s, people are resourceful and shall find a way to run these.

Would it even require a particularly high level of resourcefulness? Purchase the GPU along with the mobo that slots it. It's not as though companies typically swap out CPU and GPU while keeping the rest of the box.

The SXM mobo is huge because it takes 8 GPUs. It requires 10 kW of power etc.

You'd be better off with the SXM-PCIe adapters.

They should max out a bit below 6 kW? The H100 SXM5 is 700 w which would place the system at 5.6 kW plus change. Too much for a standard circuit but well within the bounds of a residential appliance.

It's a monolithic 8U rackmount appliance so perhaps a dishwasher would make for a decent size comparison?

Definitely no good if you rent but homeowners should have little to no difficulty. The sort of people interested in such gear usually have multi kW racks already.

Where do you find such deals

Start on eBay and learn the off-lease companies, and start watching them directly.

K80's are up on eBay.

Last I checked AWS is still offering g4dn instances that run on NVIDIA T4 GPUs, which were first released in 2018. I think most people underestimate how long superscalers can keep these things running profitably after they depreciate, and you probably don’t want anything they throw away.

My last employer is still running a bunch of otherwise discontinued g3 instances with 2015 era GPUs.

It's likely the GPU boards are designed for water cooled data center racks and might not fit in a regular PC case. It's also possible the PCB the GPU's are mounted to might not be standard PCIe cards that fit into an ATX case.

I bought a used NEC SX Aurora TSUBASA (PCIe x16 board that looks like a GPU board) and realized it has no fans. The server case it is designed to fit into is pressurized by fans forcing air through eight cards on a special 4 + 4 slot motherboard. I have to stack and mount three 40mm fans on the back.

How is software compatibility with the tsubasa BTW? Is it worthwhile to play with?

Honestly, I bought it more for collection than actually using. If I ever do anything with it, it will be from Plan 9. However, searching around there is plenty of documentation on the architecture in addition to a github repo: https://github.com/veos-sxarr-NEC

They are build to physically last 5-7 years in 24/7 datacenter use, but they have effective lifetime just 3-4 years, then their value has deprecated and electricity and infrastructure cost dominates. Meta did a benchmark where 9% of the chips failed every year, 'infant mortality' is much higher in the first 3 months of use.

9% is an absurd failure rate for solid state electronics. Particularly considering the profit margins. I assume it's related to the power densities involved. Would you happen to recall the source?

It's pretty bad: https://www.datacenterdynamics.com/en/news/meta-report-detai...

Jensen said they added a lot of RAS in Blackwell which kind of admits Hopper wasn't reliable enough.

Depending on the elemental composition, it could definitely be worthwhile to recycle wherever scale is practical. For giant datacenters and companies using hundreds of thousands or millions of gpus, that adds up to a lot of gold and other valuable elements.

In order to take advantage of that, someone needs to be positioned to process all that material economically, and to make the logistics achievable by the big players. If it costs Facebook $10million to store and transport phased out gpus vs just sending them to a landfil, they're not going to do it. If they get $100k for recycling - probably not going to do it. If they pocket $5 million, they will definitely contract that out, especially if it costs $50 million to build out the infrastructure to handle it.

Probably a good company idea - transport, disposal, refurbishment of out of cycle GPUs and datacenter assets, creating a massive recycling pipeline for recapturing all the valuable elements is a pretty good niche.

We already have that. It's called ebay.

I've written about this elsewhere but I predict there will be a significant secondary market for repurposing parts of datacenter GPUs (for example, RAM chips) by desoldering them and soldering them onto new PCBs that fit PC/consumer use cases.

That might work to some extent with the DDR5 DIMMs connected to the CPUs, but is thoroughly impossible for the HBM DRAM stacks packaged with the GPUs.

I wish there to be an active market like what Gamer's Nexus covered in China:

https://www.youtube.com/watch?v=1H3xQaf7BFI&t=1577s

in the States.

It's all about the cost of labor. In the US you could not find someone capable of soldering BGA chips for a price that makes sense doing.

They are thrown away. The average lifespan of a data center GPU is 3 years. Meaning they fail.

You can already buy A100/H100s on eBay. While it might not ever be economical to run these at home (cost of electricity), but it's plenty fun.

It seems like GPUs with a high utilization rate (60%+) degrade after 1-3 years: https://www.tomshardware.com/pc-components/gpus/datacenter-g...

Would be interested to know if others have takes on this.

I previously ran 150,000 AMD gpus in all conditions at 100% utilization for years. I currently have a multi-million $ cluster of enterprise AMD GPUs.

A couple real world points:

1. They generally don't just fail. More likely a repairable component on a board fails and you can send it out to be repaired.

2. For my current stuff, I have a 3 year pro support contract that can be extended. Anything happens, Dell goes and fixes it. We also haven't had someone in our cage at the DC in over 6 months now.

I have to maintain our GPU's. Generally the worst parts are the watercooling pressure, the HVAC, and the power. I can run it stable only at 300W per CPU, the normal max is 310W. Now with throttling to 300 it's stable for a year, before it burned two mainboards already, with lots of downtimes.

My experience is that power problems stem from not having good power and/or poor airflow.

I'm convinced that this is why we haven't had any issues in our current location. Zero outside air, zero dust, insanely well built zero expense spared airflow and power supply / management.

No affiliation (I wish), but: https://gptshop.ai

This site apparently sources ex-enterprise(-only) systems and puts them into desktop style enclosures.

You send them back to Nvidia or a third party e-waste recycler at end of life. Sometimes they're resold and reused, but my understanding is that most are eventually processed for materials.

We literally do not know because it hasn't been 5 years yet...

V100 and A100 are over 5 years old and you can find used ones on eBay.

The H100 and H200 are not.

The data center owners aren't the ones selling new GPUs.

They are selling access to GPU computation. Seeking their used GPUs would flood the market with cheap competitors using their old GPUs.

If the GPUs were competitive with their own, they wouldn't be selling them off.

This article misses the point that people are still actively running older compute.

This feels a bit overdone. OpenAI has had problems with every compute partner they've ever had. It's just not a solvable problem, who would they go to to allegedly get next-gen chips quicker?

I think the more interesting question is how much longer does oracle have and at what point does a hostile takeover make sense.

Their databases are heavily used in government, banking and other large industries which have been slower to adapt to change and strugglyto migrate away. At what point does purchasing oracle to gain customer share, existing data centres and the opportunity to migrate to your cloud platform make more sense than competing?

They still have a high market value. However, the debt they will need to service will result in ongoing price increases which will encourage people to migrate away. Over time they will struggle to service the debt and a buyout will be the best of the bad options.

I mean, it's still pretty unclear to me how IBM continues to operate. They run like Oracle but with 20 year old tech.

Last I looked (about a year ago,) IBM is 3 things: Software licensing, services and Big Iron. The companies' revenue is split roughly equally among these three areas.

An interesting perspective of IBM is its relative position. It's leveled off at about $60b/y, after a lengthy decline. It is far overmatched by many big tech companies today in terms of revenue.

It's a niche business, serving niche demands. I think IBM's moat is that most of its business is highly uninteresting: industrialized box ticking work, deeply entangled by contracts and a strong need for continuity by its customers.

I actually had a recent encounter with one of IBM's products. A commercial B2B REST API I created was analyzed by an IBM vulnerability scanning platform on behalf of a major US municipality. It didn't find anything actually critical, but there were some worthwhile points in the report, and working around a false positive was a frustration. The product, in this case, is diffusion of responsibility.

On the Big Iron end, IBM isn't really selling hardware. They selling an ecosystem: services, software, support, continuity (over decades,) etc. It pleases me that they chose to stick to Power: it's nice to know Itanium didn't kill off every enterprise RISC platform.

Maybe, one day, some major quantum computing breakthrough happens at IBM. As far as I can see, that's the only play they have that could change their trajectory. In the meantime, they have a large software portfolio and plenty of institutions that will keep signing contracts long after I'm gone.

iirc they still do a lot of chip research and fabrication equipment. They were able to keep that side of the business going at least.

They're one ~ 3 main companies in the US that will sell you quantum computers, and the only one offering a quantum PaaS.

They do a lot of stuff. Also own Hashicorp now, so they have things like: Ansible / RedHat Linux (already owned), Terraform, Consul, Nomad, Packer, etc. A lot of "let's build modern infra" tooling.

Mainframes aren't going anywhere.

Well they're really heavy.

They are, but slowly

Why would you buy the old gen GPUs instead to commit to buy the newest and best GPU available in 2 years? anyone knows that electronics depreciate fast. Unless they get them at discount this is really stupid. It’s like buying the best TV or iPhone at full price and keep it in storage for 2+ years.

Is it possible that the supply of used GPUs available to home builders will somehow increase as the result of this?

Data centres are actually prohibited from using consumer level GPUs via license restrictions. The GPUs they use are largely SXM (server connector) and if you did somehow get one of the PCIe variants (with enormous power and cooling needs) most don't even support gaming APIs.

Yeah, it used to be true that server GPUs at least somewhat resembled their gaming counterparts (i.e. Nvidia Tesla server components from 12+ years ago); they were still PCIe cards, just with server-optimized coolers, and fundamentally shared the same dies that the gaming and professional cards used.

That stopped being true many years ago though, and the divergence has only accelerated with the advent of AI datacenter usage. The form factor is now fundamentally different (SXM instead of PCIe); you can adapt an SXM card to PCIe with some effort [1], but that may not even be worthwhile because 1. the power and cooling requirements for the SXM cards are radically different than a desktop part and more importantly 2. the dies are no longer even close to being the same. IIRC, Blackwell AI chips straight up don't have rasterization hardware onboard at all; internally they look like a moderate number of general SMs attached to a huge number of tensor core. Modern AI GPUs are fundamentally optimized for, well, mat-mults, which is not at all what you want for gaming or really any non-AI application.

[1] https://l4rz.net/running-nvidia-sxm-gpus-in-consumer-pcs/

The only thing that matters is stonk++

too bad stonk is down 23% this year. i think they are doing it wrong

I never thought I would see the day, but my stodgy, lumbering company just banned new Oracle databases. Everyone hates Oracle, and only does business out of necessity. I think more and more companies are trying to extricate themselves from Oracle legal, so Oracle needs a new way to leech onto corporations for the coming decades. AI is the best play in sight.

Did you guys go out to celebrate? It's not too late for Ding Dong, the Bitch is Dead.

If you're Oracle it's not necessary a bad thing if you build an antiquated data center. Isn't much of their customer base legacy customers they are rent-seeking from in perpetuity? Those people are never going to be doing cutting edge AI. They will do what they have always done: adopt new technologies right at the nadir of the Trough of Disillusionment.

"The North Korea of the computer industry" indeed

It’s a huge gamble but they have no choice but to take it. Most their software will be rendered obsolete by AI (I’ve vibecoded replacements saving millions already, companies everywhere are doing this right now).

So they have to hope they’re a part of the future in the AI capacity because their SaaS business is going to take a big hit.

YTD performance didn’t fully bake this reality in. It was seen as them having 2 huge revenue streams, the market is realizing that AI is a threat to SaaS and baking that into stonks

The actions of oracle lately seem extremely misaligned to maximize stonks - it's extremely political, more than is necessary to merely keep in the good graces of the current administration.

the winners of the railroad boom were the buyers of distressed assets, not the dreamers that built them

Not directly related to the article itself, but aren't all debt by definition tomorrow's debt? i.e. debt is money to be paid off in the future.

Can a datacenter practicably upgrade to the next generation of GPU every year?

If it wants to stay competitive, yes.

I don’t understand the content of most of the comments here - am I being thick?

>> Oracle is the only one using debt to build the data center

Stargate is backed by the US gvt hence why they're comfortable to put that under debt financing

Learn from the best!

> The missing part is that current gpus are already money making machine in 2026

Are they? Unless you are Nvidia that is very far from the case.

OpenAI's current revenue is $25 billion a year. They are expected to spend $600 billion on infrastructure in the next 4 years to sustain and grow that revenue.

Amazon, Google, Microsoft and Meta are spending a combined $650 billion on infrastructure in 2026 alone.

The story is the same across the rest of the industry.

None of these investments are immediately profitable. And it remains to be seen whether they eventually will be or not.

Anthropic in 2026 only added several billions of revenue. This is insanely fast. In my company llm cost are already eating hiring budgets to a certain extent. We don’t buy gpu. We are paying to those who will.

> None of these investments are immediately profitable.

If you're OpenAI spending $100M on a training run they're not.

But if you're Oracle renting out GPUs to little guys doing inference, they are.

25 bln from just one company . There will be 6-7 companies like this . And they just scratched the surface . The penetration in many areas is almost 0. Yet.

This is general compute hardware as I understand it. It will not go unused no matter what happens. If new algorithms appear that reduce the number of calculations needed per token for an llm they are probably still good. It's not like silicon advances are accelerating.

If it's built in stages each state will have never variants of hardware I imagine.

General purpose compute with an odious partner. I would pay a premium to not fear Oracle lawyers auditing my usage.

I'm definitively not interested in buying oracle servers but I imagine AWS and Microsoft are also building out aggressively.

What the article did not mention is that oracle founder, executive chairman and biggest stockholder larry ellison is currently bankrolling his kid David's bid to monopolize the entire US news industry so that they are more friendly to Trump, Netanyahu and various other right wing ideologists.

David Ellison is fueling his buying spree with debt guaranteed by his dad's oracle shares. The various assets David has bought are already suffering losses of viewership because viewers are turned off by their new ideological slant.

Usually debt investors are not worried if the stock price is high. Debt has precedence over equity, so if the stock price is riding high, the CEO can always be convinced to print more shares to service the debt. The Oracle stock price has not been doing that hot lately, however. As the article said, it is 50% down. Still ORCL has 430 Billion market cap in comparison with 130 Billion of debt. It seems manageable. But stock prices can move very fast. Ironically, the war in Iran, which David's new news sources keep supporting is causing ORCL stock to go down which can bring down David's new media empire.

David just purchased Warner Bros for about 110. A lot of that (40 billion) is also guaranteed by daddy's ORCL shares. Warner Bros owns Comedy Central, which sadly has been one of Americas most dependable news sources.

The house of cards is still standing but its getting awfully wobbly.

Perhaps oracle going bust can be the silver lining to an AI bubble bursting

to me, seems the page is gone. This could be a related item:

https://www.msn.com/en-us/money/general/as-oracle-plans-thou...

TFA says nothing about taxpayer money - this is about Oracle taking on debt...

what taxpayer money?

The inevitable bailout.

Finally someone said it

While Trump is in power, the bail out is a sure thing.

Wow, I just saw in the article that NVIDIA called a new chip Vera Rubin (https://en.wikipedia.org/wiki/Vera_Rubin, also https://en.wikipedia.org/wiki/Vera_C._Rubin_Observatory). How is it allowed for a commercial company to assign a name of a known person to a product ?

People have sued over this sort of thing. Apple's Power Macintosh 7100 was originally codenamed "Carl Sagan":

https://en.wikipedia.org/wiki/Power_Macintosh_7100

Sagan sued. Engineers at Apple changed the name to BHA: "Butt-Head Astronomer".

He sued again. The final codename was "LAW: Lawyers are Wimps".

It's a codename. The product will be called "R100" or "R200" etc. (And "RTX6090" etc for the consumer versions).

different vera rubin, common mistake

I mean Nvidia has been naming thier chips after scientists for a while now Hopper, Blackwell etc. Names are not copyrightable, you can literally create a toaster and call it Einstein. It doesn't mean you're doing anything illegal. There are some exceptions like if the name is actually used by a brand (like Tesla now) or if the person is alive/recently dead, or if you claim they are in someway endorsing your project. Like claiming "Einstein always toasted bread with the Einstein toaster!" is not okay.

The way Nvidia does it is actually super respectful and it's honestly better to use names like these instead of ULTRA PRO MAX 5x etc.

Would you want the commercial company use your name, or the name of your relative?

Sure, why not? Especially if it’s honoring my contribution to science.

The observatory is named in honour of Vera Rubin. That makes sense. The commercial company deciding to name their new generation of chips does not (at least to me).

Tesla???