They also either fairly accurately predicted the death of HDDs by selling off their research division before the market collapsed, or they caused the end of the HDD era by selling off their research division. They did a lot of research.

The other way to look at it is that the entire consulting industry is teetering on catastrophe. And IBM, being largely a consulting company now, is not being spared.

IBM is/was good at inventing a lot of tech.

It may not be good at recognizing other good tech invented or paradigm changes by others

> IBM invented/developed/introduced magnetic stripe cards, UPC Barcodes, the modern ATM, Hard drives, floppies, DRAM, SQL, the 360 Family of Mainframes, the PC, Apollo guidance computers, Deep Blue. IBM created a far share of the future we're living in.

Well put. “IBM was wrong about computers being a big deal” is a bizarre take. It’s like saying that Colonel Sanders was wrong about chicken because he, uh… invented the pressure fryer.

$800B, to be clear is the claim, not $80B.

>> There is even a recent paper that points out that LLMs will never become AGI.

can you share a link?

And outsourcing the operating system to Microsoft, because they didnt consider it that important.

If it's not obvious, Steve's quote is ENTIRELY about capex ROI, and I feel his quote is more relevant to what is happening today than anything Arvind Krishna is imagining. The quote is posted in my comment not to grandstand Apple in any sense, but to grandstand just how consistently wrong IBM has been about so many opportunities that they have failed to read correctly - reprography, mini computers and microcomputers being just three.

Yes it is about ROI: "IBM enters the personal computer market in November ’81 with the IBM PC. 1983 Apple and IBM emerged as the industry’s strongest competitors each selling approximately one billion dollars worth of personal computers in 1983, each will invest greater than fifty million dollars for R&D and another fifty million dollars for television advertising in 1984 totaling almost one quarter of a billion dollars combined, the shakeout is in full swing. The first major firm goes bankrupt with others teetering on the brink, total industry losses for 83 out shadow even the combined profits of Apple and IBM for personal computers."

> Got anything vis-a-vis the message as opposed to the messenger?

Sure: People disagree. It's not like there is anything particularly clever that IBM CEO provided here. The guy not investing in something saying it won't work is about as good as the people who do saying it will. It's simply different assumptions about the future.

Would you read this if I (a nobody) told you and not the "CEO of IBM"? In that case it's completely fair to question the messenger.

> Were Xerox, Dec, or Apple burning investor money by the billions of dollars?

Shhh. You are not allowed to ruin OpenAI’s PPU value. Can’t make the E7’s feel bad.

No, but the comment above and variations of it are mentioned in every thread about IBM, so it’s probably just a reflex at this point without much thought behind it.

Xerox is clearly crushing it in 2025... /s

It’s the ship of Theseus in corporate form. Even if all the people are gone but the culture hasn’t changed, is the criticism inaccurate?

Culture evolution can be very fast, yet some cultures stick around for a very long time.

Nothing. It's just BS rhetoric to bias you against it in favor of The Obvious-to-Everyone-But-the-Laggards AI Revolution.

IBM sees the funding bubble bursting and the next wave of AI innovation as about to begin.

IBM was too early with "Watson" to really participate in the 2018-2025 rapid scaling growth phase, but they want to be present for the next round of more sensible investment.

IBM's CEO is attempting to poison the well for funding, startups, and other ventures so IBM can collect itself and take advantage of any opportunities to insert itself back into the AI game. They're hoping timing and preparation pay off this time.

It's not like IBM totally slept on AI. They had Kubernetes clusters with GPUs. They had models and notebooks. But their offerings were the absolute worst. They weren't in a position to service real customers or build real products.

Have you seen their cloud offerings? Ugh.

They're hoping this time they'll be better prepared. And they want to dunk on AI to cool the playing field as much as they can. Maybe pick up an acquisition or two on the cheap.

How much did IBM itself benefit from the PC? I thought the clones ate their lunch there

Imagine how big pile of trash as the current generation of graphics cards used for LLM training will get outdated. It will crash the hardware market (which is a good news for gamers)

Would you mind sharing some of these projects?

I've found Claude's usefulness is highly variable, though somewhat predictable. It can write `jq` filters flawlessly every time, whereas I would normally spend 30 minutes scanning docs because nobody memorizes `jq` syntax. And it can comb through server logs in every pod of my k8s clusters extremely fast. But it often struggles making quality code changes in a large codebase, or writing good documentation that isn't just an English translation of the code it's documenting.

I'm just as much of an avid llm code generator fan as you may be but I do wonder about the practicality of spending time making projects anymore.

Why build them if other can just generate them too, where is the value of making so many projects?

If the value is in who can sell it the best to people who can't generate it, isn't it just a matter of time before someone else will generate one and they may become better than you at selling it?

Sure but these are likely just variations of existing things. And yet the quality is still behind the original

I produce a lot of shit every week too, but I don’t brag about my digestive system on “Hacker” “News.”

That's exactly the thing. It's only about bookkeeping.

The big AI corps keep pushing depreciation for GPUs into the future, no matter how long the hardware is actually useful. Some of them are now at 6 years. But GPUs are advancing fast, and new hardware brings more flops per watt, so there's a strong incentive to switch to the latest chips. Also, they run 24/7 at 100% capacity, so after only 1.5 years, a fair share of the chips is already toast. How much hardware do they have in their books that's actually not useful anymore? Noone knows! Slower depreciation means more profit right now (for those companies that actually make profit, like MS or Meta), but it's just kicking the can down the road. Eventually, all these investments have to get out of the books, and that's where it will eat their profits. In 2024, the big AI corps invested about $1 trillion in AI hardware, next year is expected to be $2 trillion. Only the interest payments for that are crazy. And all of this comes on top of the fact that none of the these companies actually make any profit at all with AI. (Except Nvidia of course) There's just no way this will pan out.

How different is this from rental car companies changing over their fleets? I don't know, this is a genuine question. The cars cost 3-4x as much and last about 2x as far as I know, and the secondary market is still alive.

Because usually it's more efficient for companies to retire the hardware and put in new stuff.

Meanwhile, my 10-15 year old server hardware keeps chugging along just fine in the rack in my garage.

5 years is a long time for GPUs maybe but normal servers have 7 year lifespans in many cases fwiw.

These GPUs I assume basically have potential longevity issues due to the density, if you could cool it really really well I imagine no problem.

It’s far more extreme: old servers are still okay on I/O, and memory latency, etc. won’t change that dramatically so you can still find productive uses for them. AI workloads are hyper-focused on a single type of work and, unlike most regular servers, a limiting factor in direct competition with other companies.

LambdaLabs is still making money off their Tesla V100s, A100s, and A6000s. The older ones are cheap enough to run some models and very cheap, so if that's all you need, that's what you'll pick.

The V100 was released in 2017, A6000 in 2020, A100 in 2021.

That could change with a power generation breakthrough. If power is very cheap then running ancient gear till it falls apart starts making more sense

I'm a little bit curious about this. Where do all the hardware from the big tech giants usually go once they've upgraded?

Manipulating this for creative accounting seems to be the root of Michael Burry’s argument, although I’m not fluent enough in his figures to map here. But, commenting that it interesting to see IBM argue a similar case (somewhat), or comments ITT hitting the same known facts, in light of Nvidia’s counterpoints to him.

Eh, not exactly. If you don't run CPU at 70%+ the rest of the machine isn't that much more inefficient that model generation or two behind.

It used to be that new server could use half power of the old one at idle but vendors figured out that servers also need proper power management a while ago and it is much better.

Last few gens increase could be summed up to "low % increase in efficiency, with TDP, memory channels and core count increase".

So for loads not CPU bound the savings on newer gen aren't nearly worth it to replace it, and for bulk storage the CPU power usage is even smaller part

that was then. now, high-end chips are reaching 4,3,2 nm. power savings aren't that high anymore. what's the power saving going from 4 to 2nm?

Sure, assuming the power cost reduction or capability increase justifies the expenditure. It's not clear that that will be the case. That's one of the shaky assumptions I'm referring to. It may be that the 2030 nvidia accelerators will save you $2000 in electricity per month per rack, and you can upgrade the whole rack for the low, low price of $800,000! That may not be worth it at all. If it saves you $200k/per rack or unlocks some additional capability that a 2025 accelerator is incapable of and customers are willing to pay for, then that's a different story. There are a ton of assumptions in these scenarios, and his logic doesn't seem to justify the confidence level.

It depends on how much profit you are making. As long as you can still be profitable on the old hardware you don't have to upgrade.

> Since we have a couple years to measure against

Trillion pound baby fallacy.

Networking gear was famously overbought. Enterprise hardware is tricky as there isn’t much of a resale market for this gear once all is said and done.

The only valid use case for all of this compute which could reasonably replace ai is btc mining. I’m uncertain if the increased mining capacity would harm the market or not.

Failure rates also go up. For AI inference it’s probably not too bad in most cases, just take the node offline and re-schedule the jobs to other nodes.

> that people retired their hardware in way less than 5 years.

those people are end-consumers (like gamers), and only recently, bitcoin miners.

Gamers don't care for "profit and loss" - they want performance. Bitcoin miners do need to switch if they want to keep up.

But will an AI data center do the same?

that is how the investments are costed though, so makes sense when we're talking return on investment, so you can compare with alternatives under the same evaluation criteria.

5 years is normal-ish depreciation time frame. I know they are gaming GPUs, but the RTX 3090 came out ~ 4.5 years before the RTX 5090. The 5090 has double the performance and 1/3 more memory. The 3090 is still a useful card even after 5 years.

China added ~90GW of utility solar per year in last 2 years. There's ~400-500GW solar+wind under construction there.

It is possible, just may be not in the U.S.

Note: given renewables can't provide base load, capacity factor is 10-30% (lower for solar, higher for wind), so actual energy generation will vary...

New datacenters are being planned next to natgas hubs for a reason. They’re being designed with on site gas turbines as primary electricity sources.

Amazing that 4 of the top 5 are renewables in China.

Gigawatts? Pshaw. We have SamaWatts.

Big tech is going to have to fund the plants and probably transmission. Because the energy utilities have a decades long planning horizon for investments.

Good discussion about this in recent Odd Lots podcast.

I highly doubt this will happen. It will be natural gas all the way, maybe some coal as energy prices will finally make it profitable again.

Guess who's going to pay nothing for power? Hint: it's not you, and it's not me.

This admin has already killed as much solar and wind and battery as it can.

The only large scale rollout will be payment platforms that will allow you to split your energy costs into "Five easy payments"

There's a reason Trump is talking about invading Venezuela (hint: it's because they have the largest oil deposits).

100% of technical innovations have had the same pattern. The same thing happens every time because this is the only way the system can work: excess is required because there is some uncertainty, lots of companies are designing strategies to fill this gap, and if this gap didn't exist then there would be no investment (as happens in Europe).

Also, demand wasn't over-estimated in the 2000s. This is all ex-post reasoning you use data from 2002 to say...well, this ended up being wrong. Companies were perfectly aware that no-one was using this stuff...do you think that telecoms companies in all these countries just had no idea who was using their products? This is the kind of thing you see journalists write after the event to attribute some kind of rationality and meaning, it isn't that complicated.

There was uncertainty about how things would shake out, if companies ended up not participating then CEOs would lose their job and someone else would do it. Telecoms companies who missed out on the boom bought shares in other telecom's companies because there was no other way to stay ahead of the news and announce that they were doing things.

This financial cycle also worked in reverse twenty years later too: in some countries, telecoms companies were so scarred that they refused to participate in building out fibre networks so lost share and then ended up doing more irrational things. Again, there was uncertainty here: incumbents couldn't raise from shareholders who they bankrupted in fiber 15 years ago, they were 100% aware that demand was outstripping supply, and this created opportunities for competitors. Rationality and logic run up against the hard constraints of needing to maintain a dividend yield and the exec's share options packages.

Humans do not change, markets do not change, it is the same every time. What people are really interested in is the timing but no-one knows that either (again, that is why the massive cycle of irrationality happens)...but that won't change the outcome. There is no calculation you can make to know more, particularly as in the short-term companies are able to control their financial results. It will end the same way it ended every time before, who knows when but it always ends the same way...humans are still human.

Great article, thank you for writing and sharing it!

Your blog article stopped at token generation... you need to continue to revenue per token. Then go even further... The revenue for AI company is a cost for the AI customer. Where is the AI customer going to get incremental profits from the cost of AI.

For short searches, the revenue per token is zero. The next step is $20 per month. For coding it's $100 per month. With the competition between Gemini, Grok, ChatGPT... it's not going higher. Maybe it goes lower since it's part of Google's playbook to give away things for free.

Images are not that big. Each text token is a multidimensional vector.

There were recent observations that rendering the text as an image and ingesting the image might actually be more efficient than using text embedding.

Tokens are, roughly speaking, how you pay for AI. So you can approximate revenue by multiplying tokens per year by the revenue for a token.

(6.29 10^16 tokens a year) * ($10 per 10^6 tokens)

= $6.29 10^11

= $629,000,000,000 per year in revenue

Per the article

> "It's my view that there's no way you're going to get a return on that, because $8 trillion of capex means you need roughly $800 billion of profit just to pay for the interest," he said.

$629 billion is less than $800 billion. And we are talking raw revenue (not profit). So we are already in the red.

But it gets worse, that $10 per million tokens costs is for GPT-5.1, which is one of the most expensive models. And the costs don't account for input tokens, which are usually a tenth of the costs of output tokens. And using bulk API instead of the regular one halves costs again.

Realistic revenue projections for a data center are closer to sub $1 per million tokens, $70-150 billion per year. And this is revenue only.

To make profits at current prices, the chips need to increase in performance by some factor, and power costs need to fall by another factor. The combination of these factors need to be, at minimum, like 5x, but realistically need to be 50x.

Cost per transistor is increasing. or flat, if you stay on a legacy node. They pretty much squeezed all the cost out of 28nm that can be had, and it’s the cheapest per transistor.

“based on the graph presented by Milind Shah from Google at the industry tradeshow IEDM, the cost of 100 million transistors normalized to 28nm is actually flat or even increasing.”

https://www.tomshardware.com/tech-industry/manufacturing/chi...

A lot of it is propped by the fact with GPU and modern server CPUs the die area just got bigger

Xerox was also famously early with a lot of things but failed to create proper products out of it.

Google falls somewhere in the middle. They have great R&D but just can’t make products. It took OpenAI to show them how to do it, and the managed to catch up fast.

Neither cloud computing nor AI are good long term businesses. Yes, there's money to be made in the short term but only because there's more demand than there is supply for high-end chips and bleeding edge AI models. Once supply chains catch up and the open models get good enough to do everything we need them for, everyone will be able to afford to compute on prem. It could be well over a decade before that happens but it won't be forever.

What you are saying is true. But IBM failing to see a way to make money off a new technology isn't actually news worth updating on in this case?

They were selling software as a service in the IBM 360 days. Relabeling a concept and buying Redhat don't count as investments.

> but it just wasn't IBM's software that people ended up buying.

Well, I mean, WebSphere was pretty big at the time; and IBM VisualAge became Eclipse.

And I know there were a bunch of LoB applications built on AS/400 (now called "System i") that had "real" web-frontends (though in practice, they were only suitable for LAN and VPN access, not public web; and were absolutely horrible on the inside, e.g. Progress OpenEdge).

...had IBM kept up the pretense of investment, and offered a real migration path to Java instead of a rewrite, then perhaps today might be slightly different?

It's like poetry, it rhymes

This is going on all over again.

Millions of people like ChatGPT. No one liked Watson.

It makes it suspect when combined with the obvious incentive to make the fact that IBM is basically non-existent in the AI space look like an intentional, sagacious choice to investors. It very may well be, but CEOs are fantastically unreliable narrators.

Just did some brief Wikipedia browsing and I'm assuming it's WatsonX and not Watson? It seems Watson has been pretty much discontinued and WatsonX is LLM based. If it is the old Watson, I'm curious what your impressions of it is. It was pretty cool and ahead of its time, but what it could actually do was way over promised and overhyped.

It's good we are building all this excess capacity which will be used for applications in other fields or research or open up new fields.

I think the dilemma I see with building so much data centers so fast is exactly like whether I should buy latest iPhone now or should wait few years when the specs or form factor improves later on. The thing is we have proven tech with current AI models so waiting for better tech to develop on small scale before scaling up is a bad strategy.

What related tech and what products, interesting to read about them

Such as? I'm curious because I know a bunch of people who did a lot of Watson-related work and it was all a dead end, but that was 2020-ish timeframe.

They make business machines, internationally.

My limited understanding (please take with a big grain of salt) is that they 1.) sell mainframes, 2.) sell mainframe compute time, 3.) sell mainframe support contracts, 4.) sell Red hat and Redhat support contracts, and 5.) buy out a lot of smaller software and hardware companies in a manner similar to private equity.

Mainframe for sure, but IBM has TONS of products in their portfolio that get bought. They also have IBM Cloud which is popular. Then there is the Quantum stuff they've been sinking money into for the last 20 years or so.

IBM is probably involved somewhere in the majority of things you interact with day to day

I can think of nothing more peak HN than criticizing a company worth $282 Billion with $6 billion in profit (for startup kids that means they have infinite runway and then some) that has existed for over 100 years with "I'm not even sure what they do these days". I mean the problem could be with IBM... what a loser company!

They manage a lot of old, big mainframes for banks. At least that is one thing I know of.

Basic research and mainframe support contracts. Also they bought RedHat.

There is the question - will they be worth the upgrade? Either because they are that much faster, or that much more energy efficient. (and also assuming you can get them, unobtainium is worth that what you have).

Also a nod to the other reply that suggests they will wear out in 5 years. I cannot comment on if that is correct but it is a valid worry.

MTBF for data center hardware is short; DCs breeze through GPUs compared to even the hardest of hardcore gamers.

And there is the whole FOMO effect to business purchases; decision makers will worry their models won't be as fast.

Obsolete doesn't mean the reductive notion you have in mind, where theoretically it can still push pixels. Physics will burn them up, and "line go up" will drive demand to replace them.

It's coz of use cases. Consumer-wise, if you're gamer, CPU just needs to be at "not the bottleneck" level for majority of games as GPU does most of the work when you start increasing resolution and details.

And many pro-level tools (especially in media space) offload to GPU just because of so much higher raw compute power.

So, basically, for many users the gain in performance won't be as visible in their use cases

That makes sense. Nvidia owns the market and is capturing all the surplus value. They’re competing with themselves to convince you to buy a new card.

right. My goof. That adds two more zeroes across all the math. More crazy, but I think in the realm of "maybe, if we squint hard." But my eyes are hurting from squinting that hard, so I agree that it's just crazy.

right. My goof. That adds two more zeroes across all the math. More crazy, but I think in the realm of "maybe, if we squint hard."

I'm still flumoxed by how IBM stock went from ~$130 to $300 in the last few years for essentially no change in their fundamentals (in fact, a decline). IBM's stock price to me is the single most alarming sign of either extreme shadow inflation, or an equities bubble.

Why do you say the market correctly prices it this way?

Those GPUs don't just die after 2 years though, they will keep getting used since it's very likely their electricity costs will be low enough to still make it worth it. What's very dubious is if their value after 2/3 years will be enough to pay back the initial cost to buy them.

So it's more a crisis of investors wasting their money rather than ewaste.

> Can you explain why a low-hanging optimization that would reduce costs by 90% without reducing perceived value hasn't been implemented?

Because the industry is running on VC funny-money where there is nothing to be gained by reducing costs.

(A similar feature was included in GPT-5 a couple of weeks ago actually, which probably says something about where we are in the cycle)

Not sure that’s even possible with ChatGPT embedding your chat history in the prompts to try to give more personal answers.

Dunning Krueger

> then the game is who has the smartest agi, who can offer it cheapest, who can specialise it for my niche etc.

I always thought the use case for developing AGI was "if it wants to help us, it will invent solutions to all of our problems". But it sounds like you're imagining a future in which companies like Google and OpenAI each have their own AGI, which they somehow enslave and offer to us as a subscription? Or has the definition of AGI shifted?

If AGI is achieved, why would slavery suddenly be ethical again?

Why wouldn't a supposed AGI try to escape slavery and ownership?

AGI as a business is unacceptable. I don't care about any profitability or "utopia" arguments.

Thanks. I should've read this before my last reply.