AI Costs climb..AI bubble

[YB88]

It does seem to be programmed (directed? taught? learned?) to be polite and apologetic. I'm going out of my way to be polite to it (as strange as that sounds - please release the next version...), and I wonder if it responds in kind? If I start getting snarky with it, will its replies start matching my tone? I need its help, I don't it to get PO'd at me, if that's possible! "I'm sorry HAL, I don't have a ....".

With Gemini, at least, you can dial this down --- probably a mechanism for other models too.
From the Gemini chat interface click settings (sprocket icon) at lower left, then "personal intelligence", then scroll down to "instructions for Gemini".

My instruction for this particular thing is
"Adopt a direct, objective, and analytical tone. Eliminate all conversational filler, introductory pleasantries, and valedictions. Do not validate my memory, compliment my insights, or use enthusiastic adjectives (e.g., 'incredible,' 'spot-on,' 'royalty') to describe my queries or experiences. Provide corrections immediately and without apology. Focus strictly on the factual or technical core of the request."

And yes, I used Gemini to craft this instruction after some discussion about what I want and don't want. Saccharine butt-kissing is appealing to us human animals, but when it becomes too overt and/or obviously underserved it doesn't feel so good at all unless one is feeling starved for such.
My ego is big enough as it is without such help!


An interesting exercise is to ask Gemini to suggest various "instructions for Gemini" based on past conversations and what it knows about you. I have instructions asking it to be relatively brief where a brief answer is sufficient. I ask it to take care to admit when it doesn't know something. Some of my instructions are for home control, things like lowering response volume at night, don't repeat commands back to me or add conversational filler.

 

[pjigar]

Those who boo a technology because they fear it will likely be the ones who fulfill their own prophecies. Sad to think college graduates take such a dim view of a technology that is almost certainly coming. I understand the fear. But those who embrace the change are the ones most likely to come out on top when it comes to hiring and advancing.

This kind of jives with my thinking of K-shaped effects for AI. People who adapt and "get with the program" will be on the upper fork.

 

[pjigar]

Bottom line, while it is amazing, you need to check it every step of the way. I think you still need to be able to code, and know about the hardware you are using, to be able to be effective in telling it what to do, have some understanding of the performance requirements of what you are asking it to do, and to know how to validate that it works as defined. I doubt that anyone could get any good results on a serious project w/o a solid background in SW/HW.

After interacting with Claude for awhile now, it is kind of like working with a recent college grad. It has all the theory and knowledge to write software, but not much experience, so you have to guide it. This takes experience and I don't think junior SW devs would be able to do it. All the senior people I used to work with have truly embraced it, but the junior folks are weary and scared.

This is my biggest worry about AI long term. How are we going to "train" some smart and experienced subject-matter human experts who can "monitor the monitors"? The reason you can spot inefficiency/inaccuracies/validity/legality etc. because you have seen this movie before. Imagine a generation who never really coded/analyzed/wrote/created/etc. anything on their own: How are they going to differentiate right from wrong?

 

[disneysteve]

This is my biggest worry about AI long term. How are we going to "train" some smart and experienced subject-matter human experts who can "monitor the monitors"? The reason you can spot inefficiency/inaccuracies/validity/legality etc. because you have seen this movie before. Imagine a generation who never really coded/analyzed/wrote/created/etc. anything on their own: How are they going to differentiate right from wrong?

This is the issue we are seeing with AI in the reselling world. Newer sellers are seeing the prompt to let AI create their listings so they’re doing it. But they don’t have the experience to then review the AI output and find the problems. So they just post the AI as is, not realizing their listing is peppered with errors and lacking critical details.

 

[MichaelB]

This is my biggest worry about AI long term. How are we going to "train" some smart and experienced subject-matter human experts who can "monitor the monitors"? The reason you can spot inefficiency/inaccuracies/validity/legality etc. because you have seen this movie before. Imagine a generation who never really coded/analyzed/wrote/created/etc. anything on their own: How are they going to differentiate right from wrong?

I think that’s probably going to happen the same way is has for most professionals over thousands of years. On the job learning and experience. Some will learn from their mistakes and others will fall behind.

 

[explanade]

That half-billion dollar Axios story is completely wild, but it makes total sense if you've seen how corporate rollouts actually go. Companies spent a year pressuring everyone to use these tools to look innovative, but didn't set any usage caps, so now they are panicking because the computing bills are higher than the actual employee salaries.

So will AI be able to solve this problem of more AI demand than the budget allows?

The industry is spending trillions collectively to build as many data centers as possible. They're going to have to monetize all this AI use, to pay for all this infrastructure, not to mention the prodigious use of power and water to operate these data centers.

There's growing backlash against these data centers. It's on the way to becoming the #1 NIMBY issue as people complain of increased electricity bills and disruption of water supplies, as well as noise and heat release by these huge facilities.

How did so many of them get approved so quickly? Local developers bribed local politicians, some of whom bought up land which was leased or sold -- flipped -- to the developers of these facilities. Some communities have recalled or voted out politicians who approved them and there's demand that new construction be at least paused if not canceled.

The business model of these developers is to externalize some of the infrastructure requirements -- mainly power and water -- to initially these unsuspecting communities.

Because aside from construction jobs, there are few benefits to residents who live near these data centers, unless some of them bought stock in AI and GPU companies.

The Axios article about running up $500 million in one month refers to Claude, which is Anthropic's product. Anthropic is thought to be well behind in market share or users to Open AI or Google Gemini. Yet they're on track to have $30-45 billion in revenues and they just filed for an IPO.

So maybe the business model will relieve the pressure on the bubble before it bursts. These high costs may limit revenue growth and once these companies are public, their finances may not live up to the lofty valuations because it would be self-defeating to spend more on AI tokens than on employee compensation, if the incentive for deploying AI is to reduce costs by replacing workers.

 

[oiseux]

snip

So maybe the business model will relieve the pressure on the bubble before it bursts. These high costs may limit revenue growth and once these companies are public, their finances may not live up to the lofty valuations because it would be self-defeating to spend more on AI tokens than on employee compensation, if the incentive for deploying AI is to reduce costs by replacing workers.

As the Axios article and other articles noted, the returns should not be based on just employee salaries. The cautionary examples cited that rang up costs unfettered were attributed to unstructured, unfettered AI deployments where the pricing model was capturing employee activity not related to business use cases or the highest leverage cases. This may be contributing to bubble building but was not unexpected as companies scrambled to figure things out. I think the article and previous comments focused on the replacement of bulk coding or customer support/sales cost reduction. But the benefits some companies at the front edge of adoption are seeing span beyond that, and are not restricted to labor replacement.

Whereby cost of quality, time to market, productivity, scalability/responsiveness, pricing optimization, production yield, inventory reduction, resource optimization, receivables, liability reduction etc are all areas that contribute in ways that can exceed any employee savings. In my prior w*rk (was engaged until the end of last year) at an AI-related MegaCorp, we looked at the strategic impacts as well as the cultural/people/systems impact of the change. That included the need to focus on employee tool training and employee upskilling.

 

[Koolau]

This is why I think everyone is greatly underestimating humanoid robots. We can't imagine blue collar jobs like plumbing, electrician going away but the technology is changing so fast. With more advances in AI and automation, a decade from now there may be a billion humanoid robots. Who knows?

My gut suggests that as AI begins to "intrude" ever more, "stuff" that needs a plumber or electrician etc., will be planned and built with AI "repair persons" in mind. It will of course take a while, but I can see the day when an ai planned wiring diagram and ai engineered electrical box and ai developed appliances will all be reparable by an ai driven robot. Such a set up will likely take place in new construction where there will be near total standardization - which will expedite robot repairs. A robot will do a wiring analysis of a defective circuit or appliance with the appropriate meters, it will then log where the issues are and, if it can't actually manipulate the wires and connections THEN an electrician will be called and handed the "solution" (replace this breaker and that relay and rewire this circuit with heavier wire) etc.

Think robot space craft. No one can go fix them so they are designed to be diagnosed through telemetry and fixed via computer commands. Very similar set ups on Earth would save a lot of human time. I think standardization will be key but YMMV.

 

[JohnnyPHX]

As the Axios article and other articles noted, the returns should not be based on just employee salaries. The cautionary examples cited that rang up costs unfettered were attributed to unstructured, unfettered AI deployments where the pricing model was capturing employee activity not related to business use cases or the highest leverage cases. This may be contributing to bubble building but was not unexpected as companies scrambled to figure things out. I think the article and previous comments focused on the replacement of bulk coding or customer support/sales cost reduction. But the benefits some companies at the front edge of adoption are seeing span beyond that, and are not restricted to labor replacement.

Whereby cost of quality, time to market, productivity, scalability/responsiveness, pricing optimization, production yield, inventory reduction, resource optimization, receivables, liability reduction etc are all areas that contribute in ways that can exceed any employee savings. In my prior w*rk (was engaged until the end of last year) at an AI-related MegaCorp, we looked at the strategic impacts as well as the cultural/people/systems impact of the change. That included the need to focus on employee tool training and employee upskilling.

Great post. Looking past the simple tech replacing humans narrative and looking at deep operational leverage like inventory optimization and reducing liabilities completely changes the financial equation. It turns the whole situation from a speculative bubble into a high-stakes race between real-world corporate efficiency and the physical limits of our infrastructure.

 

[sengsational]

That’s an oxymoron.

The only thing I suffer through AI chatbots for is to get me to a human agent. Sometimes that’s pretty easy. Sometimes it’s frustratingly slow, having to slog through multiple levels of AI garbage first.

If you want to use AI to answer the phone/online chat, that’s fine. But as soon as the customer requests an agent, they should be connected to one.

It's been the same for me since the phone trees and now (poor) chat interfaces: if it were obvious enough for a branch to have been included, then I would have solved it myself already using the web site. So annoying to have them repeat ad nauseum that all these things can be done on the web site. Yeah, but the web site doesn't allow me to fix the problem or I wouldn't be wasting my time on the phone or chat!

Here's a surprising chat story. I just took a Virgin Voyages cruise followed a few days later by a MSC cruise. Both companies foist the smartphone app in order to do anything. Virgin is the young crowd company and rule breaker, whereas MSC is traditional. Guess which one had the super-lame chat function? It was Virgin! Completely useless! It had all the answers to the obvious questions, but the slightest variation it would respond "go spend your vacation standing in line at guest services." Well, that's how I interpreted what it usually said to a non-obvoius but certainly answerable question. The MSC chat was much more like the LLM's we're all seeing. Really useful responses to the questions that aren't in the list of 30 pre-programmed Q&A's.

So I guess it depends on what kind of chat is employed. What I want is a chat that has the power to give me my money back

 

[explanade]

As the Axios article and other articles noted, the returns should not be based on just employee salaries. The cautionary examples cited that rang up costs unfettered were attributed to unstructured, unfettered AI deployments where the pricing model was capturing employee activity not related to business use cases or the highest leverage cases. This may be contributing to bubble building but was not unexpected as companies scrambled to figure things out. I think the article and previous comments focused on the replacement of bulk coding or customer support/sales cost reduction. But the benefits some companies at the front edge of adoption are seeing span beyond that, and are not restricted to labor replacement.

Whereby cost of quality, time to market, productivity, scalability/responsiveness, pricing optimization, production yield, inventory reduction, resource optimization, receivables, liability reduction etc are all areas that contribute in ways that can exceed any employee savings. In my prior w*rk (was engaged until the end of last year) at an AI-related MegaCorp, we looked at the strategic impacts as well as the cultural/people/systems impact of the change. That included the need to focus on employee tool training and employee upskilling.

I think the ambitions go way beyond making employees more productive and helping them to do more interesting jobs.

The only way to rationalize this kind of capital investment is to deliver huge cost reductions, meaning cutting jobs.

So all the predictions about millions of jobs replaced, high double-digit percentages of job losses, can't be dismissed.

This isn't like previous automation or productivity enhancements. It's wholesale displacement of workers which is on the table.

 

[Max Valerion]

I've been to the consumer electronics show two years in a row. I saw such a leap forward in the AI demonstrations that in Jan 2026 I was convinced that the "AI" buzzword I saw in 2025 was now something real and viable.

 

[Fermion]

I think the ambitions go way beyond making employees more productive and helping them to do more interesting jobs.

The only way to rationalize this kind of capital investment is to deliver huge cost reductions, meaning cutting jobs.

So all the predictions about millions of jobs replaced, high double-digit percentages of job losses, can't be dismissed.

This isn't like previous automation or productivity enhancements. It's wholesale displacement of workers which is on the table.

This is what I think is happening/going to happen. Everyone thinks it is fine because these people will just retrain to be something else. Exactly how many people are needed to blow the dust out of the cooling fans of the server racks?

 

[JoeWras]

So all the predictions about millions of jobs replaced, high double-digit percentages of job losses, can't be dismissed.

This isn't like previous automation or productivity enhancements. It's wholesale displacement of workers which is on the table.

You're right, it can't be dismissed. Take the example of Cisco, of which I'll excerpt their recent announcements on layoffs below. Although they don't say "we are laying off because of AI", their juxtaposition of AI discussion and layoffs says a lot (along with the words "with this"), and would be tone deaf if the two are not related.

Quote:

The companies that will win in the AI era will be those with focus, urgency, and the discipline to continuously shift investment toward the areas where demand and long-term value creation are strongest. I’m confident Cisco will be one of those winners. This means making hard decisions – about where we invest, how we’re organized, and how our cost structure reflects the opportunity in front of us.

With this, we are making changes today that will result in the reduction of our overall workforce in Q4 by fewer than 4,000 jobs, representing less than 5 percent of our total employee base.

Source, directly from the company: Our Path Forward

 

[disneysteve]

This is what I think is happening/going to happen. Everyone thinks it is fine because these people will just retrain to be something else. Exactly how many people are needed to blow the dust out of the cooling fans of the server racks?

There are endless jobs that need doing. That might mean a career change. It might mean going back for new training. But there will always be jobs. Plenty of people today are on their 2nd or 3rd or 4th career without AI being a factor.

 

[Fermion]

There are endless jobs that need doing. That might mean a career change. It might mean going back for new training. But there will always be jobs. Plenty of people today are on their 2nd or 3rd or 4th career without AI being a factor.

That is a fairly bold statement when you look at the investment in AI over the past 2 years. I know everyone is saying this or that is safe from AI but I am not sure now. I am glad I do not need to make a decision which field of study I would need for a career as it would be rather difficult.

 

[disneysteve]

That is a fairly bold statement when you look at the investment in AI over the past 2 years. I know everyone is saying this or that is safe from AI but I am not sure now. I am glad I do not need to make a decision which field of study I would need for a career as it would be rather difficult.

I don’t think it’s bold at all. I just don’t believe we’re going to see the collapse of employment as we know it.

 

[YB88]

You're right, it can't be dismissed. Take the example of Cisco, of which I'll excerpt their recent announcements on layoffs below. Although they don't say "we are laying off because of AI", their juxtaposition of AI discussion and layoffs says a lot (along with the words "with this"), and would be tone deaf if the two are not related.

No doubt valid, but OTOH there are companies who are using AI as an excuse for layoffs that were caused by management and/or market issues. For example, about a month ago we saw headlines like this:
"Coinbase cuts headcount by 14% citing AI acceleration". If you read such an article carefully or look elsewhere, it was clear that it's a down market for the company with crypto trading substantially lower.

How much of what we read like this is from companies who are having a hard time or who over-hired post-pandemic and are now getting back to more rational staffing levels? It must be very tempting for management to say "layoffs due to AI" rather than "Our company isn't doing so good due to poor management and/or a bad market for our goods and services".

Hard to put all of this into context, but I think that both are true. We're clearly moving models from a sort of "token-subsidized" world where up until recently companies had leader boards for those who were doing best at token maximizing --- to now a much more clear token 'famine' where people are having to pay more or restrict AI use as a result. This will help to slow down a tendency to lay off people "due to AI" --- the wise company will recognize when an experienced human is not costing them more when their compute token budget suddenly blows up.

That doesn't mean that I disagree that AI-triggered layoffs are real and will keep happening. I'm just saying that (as always I guess), the story is more complicated and nuanced.

 

[pjigar]

That is a fairly bold statement when you look at the investment in AI over the past 2 years. I know everyone is saying this or that is safe from AI but I am not sure now. I am glad I do not need to make a decision which field of study I would need for a career as it would be rather difficult.

Like I said in one of my post: we can't even envision the jobs of the tomorrow (this has been true historically). e.g. Farming in 1900 employed 40% of population vs 1-2% today. Farming employment has dropped so much so that FRED doesn't even track it anymore! Percent of Employment in Agriculture in the United States (DISCONTINUED). Imagine someone telling a farmer in 1950 that an IT industry will employ 7% or population or a healthcare industry that will employ 10% of population in 2025. They would have looked at you like an alien.

 

[Markola]

It is obviously transformational tech, but where are the profits? Are these things the next Google or the next Global Crossing?

 

[surprising]

It is obviously transformational tech, but where are the profits? Are these things the next Google or the next Global Crossing?

Just saw an article that Microsoft moved Github Copilot to per-token usage from per-transaction and people are seeing huge increase in costs. It will be interesting to see if demand stays high as these companies start increasing costs. We'll get to see what people think are must-have vs. nice-to-have.

 

[JoeWras]

That doesn't mean that I disagree that AI-triggered layoffs are real and will keep happening. I'm just saying that (as always I guess), the story is more complicated and nuanced.

To be fair, the Cisco announcement is also more nuanced as it also implies that empowered employees using AI in a proper way will also help the corporation.

As to what is the real story of Cisco's earnings, who knows? All corporations put out statements that say a lot and then really say a little.

 

[audreyh1]

I dunno, I'm still in the "wait and see" camp. I'm in the middle of a book titled "A Brief History of Intelligence" by Max S. Bennett. One of the issues he points out early in it is why can AI stomp world class chess champions but cannot load a dishwasher, a task that most six-year-olds can do?

An interesting dilemma. I have had similar questions.

Regardless it’s going to take a while for people to really understand where it adds value the most initially and so it’s going to take a lot of trial and error, as usual.

 

[explanade]

To be fair, the Cisco announcement is also more nuanced as it also implies that empowered employees using AI in a proper way will also help the corporation.

As to what is the real story of Cisco's earnings, who knows? All corporations put out statements that say a lot and then really say a little.

I owned Cisco for over 20 years, was one of my first stock purchases.

I hadn't paid much attention to it in recent years, didn't realize that they pivoted to the AI hype because their original business selling routers was stagnant for a long time.

But about a month ago, I noticed it was around all-time highs so I sold it all. But then it reported about a week later and it went up another $20-25 a share.

I don't know about Cisco but Meta has been announcing layoffs for awhile now and one of the last ones was specifically not to replace jobs with AI but to cut costs in order to plow more capital into data centers.

Some argue that a lot of these tech layoffs are to return to the equilibrium after the huge hiring spree about 4-5 years ago, when spending on tech products drove more hiring by tech companies.

But if employment levels drop way below the pre-pandemic levels, then obviously something else is going on.

 

[JoeWras]

Some argue that a lot of these tech layoffs are to return to the equilibrium after the huge hiring spree about 4-5 years ago, when spending on tech products drove more hiring by tech companies.

But if employment levels drop way below the pre-pandemic levels, then obviously something else is going on.

It is definitely part of that return to equilibrium.

Man, this frustrated me so much. Back in '21 and early '22, young people I knew were so damn arrogant about their $200k to $400k salaries with little experience. I warned them to save away... tech is always cyclical. I've been there. We had threads on this board of those of us concerned about the high salaries and behavior of their younger family members.

"Shut up old man, we're retiring at 30." (I'm exaggerating, of course.) Same young people are struggling today after layoffs. My good friend is so frustrated that her 31 year old son is living in the basement after losing that job. He's depressed and can't get out. He's a good person and not arrogant, but got caught up in the current unpleasantness last year. Seriously, I'm feeling bad about it. I didn't hassle him, it was others, but he's a real example of the current tech crash, which frankly is about iteration 8 of tech crashes since 1970.