Undermining the AI business case

Investment Analysis Clubs Macro Economic Trends and Risks
1

The AI hyperscalers who develop frontier “closed source” AI are spending hundreds of billions of dollars on massive data centers. But will this astronomical investment pay off in end-user profits? (Not counting the circular AI-ecosystem profits which support the high stock valuations of Nvidia and others.)

Or will end-users turn away from expensive frontier AI to buy much cheaper “distilled AI” - smaller dedicated software models - from “open source” AI vendors?

https://www.nytimes.com/2026/09/04/technology/open-source-ai-anthropic-openai.html

Corporate America Is Getting Hooked on Open-Source A.I.

Companies like AT&T are increasingly using cheap, freely available artificial intelligence models over expensive ones from Anthropic and OpenAI.
By Eli Tan, The New York Times,
Sept. 4, 2026

…

Open-source artificial intelligence has become all the rage in corporate America. Like AT&T, companies such as Airbnb and Deloitte are also turning to open A.I. models that are easy to customize and cheaper to use. Last month, open models accounted for 58 percent of A.I. use, up from 10 percent a year ago, according to U.S. user data from OpenRouter, a platform that lets people choose different A.I. models to complete tasks…

Open models come in two flavors. One is an open-source model, which makes its underlying code public. The other is an open-weights model, which makes its “weights” — the numerical calculations that determine how the model reasons — public.

Many of the most popular open A.I. models are made by Chinese companies like Moonshot AI, DeepSeek and Alibaba. Rather than spending millions in fees to use a closed model, developers can download one of the Chinese systems and build their products on top of it, paying mainly for servers. OpenAI and Anthropic charge for subscriptions and other fees, depending on the job.

Some Chinese open models are now 80 percent to 90 percent as powerful as those from OpenAI and Anthropic, while costing as little as 20 percent of the price…

Many open models can now be powered by a phone or laptop…Leading closed models are getting more complicated and requiring more computing power.

Most U.S. companies still use a combination of open and closed models. Closed models remain the best for heavy duty tasks like coding and image and video generation, while open models often excel at simple, specialized tasks… [end quote]

If I was a corporate CFO, I would provide cheap, distilled AI to my sales force and customer service, which would probably be 95% of the users. They could run it on their laptops or even cell phones. When I was in technical sales (1981-1988) there was no such thing as a cell phone. Sales reps had to find a pay phone to call the office. Now, I could picture sales reps with cell phones equipped with distilled AI to provide them with customized recommendations and pricing right in a customer’s office. Increased productivity, actually cheaper than maintaining an office staff large enough to help the outside sales force.

This strikes at the root of the valuation of AI hyperscalers who are going into debt to build huge frontier data centers which may only be used by the small fraction of end-users (such as research).

Which strikes at the root of the S&P500 bubble that has been inflated by AI mania.

Wendy

3 Likes
2

I remember three crashes (if that is the right word, and I’m not sure it is): the dot.com bust, the 2008 financial panic, and the Covid dumpster fire.

Prior to the 90’s I was in my 401(k) plans, a cash & funds mix, and was ignorant of almost everything, finance wise. Early enough in the 90’s I began tinkering around the edges of investing, thanks in no small part to the Motley Fool (on AOL, at the time). And I made a bloody fortune starting with AOL, then Cisco, then a few others which I rode up (and sometimes back down) before cashing out almost entirely in late 1999/early 2000.

That time it wasn’t because I was so smart; Mrs. Goofy and I both quit our jobs and bought an RV in ‘99) and took off to see America; we expected to be on the road for 6 months or more, and as smart phones and tablets didn’t exist yet, we kept in touch with a hand held modem which you cupped onto a pay phone and listened to the fax-machine like noises, and then read a 2 line LCD screen on the back of the thing. Not exactly great for following your portfolio, but enough for the occasional email.

Anyway, as we were leaving I felt that the market was “toppy” (the word I used as a search key back when the Fool 1.0 and even 2.0 were available, now in the graveyard.) And when we were on the road the dot.com bubble deflated, and we were fine. Paid extraordinary taxes selling the AOL, Cisco, etc, and learned an expensive lesson about the difference between the regular account and the IRA.

Back in the market, not at the bottom but soon enough to matter, we continued investing until around 2008, when things began looking so shaky that I said to Mrs. “I’m going to pull the plug”. And I did. Sold everything except some Berkshire in the regular port which I’d had more than a decade, but cleared out every single thing in the IRAs and sat in cash. I documented it at the time here, as well.

The third time was in 2016 when I thought the election winner would crater the economy, pulled out again, and was dreadfully wrong. 2 out of 3, still OK.

The third “panic” was Covid, and I wasn’t fast enough to do everything but I got out of several things for a while at least until the dust settled. I had mastered the art of “getting out”, not perfectly at the top, mind you, but pretty close - but still had some to learn about “trying to find the bottom”, which I never have. OTOH, I have always gotten back in lower than my “out” price, so good enough.

Anyway, if this market doesn’t feel overextended, overpriced, waiting for a fall, out of balance, overstretched, excessive, and ready for a dump then I’ll eat my hat. That said, I’m still in and have moved a bit to the sidelines but not enough if it all goes south tomorrow.

I’m getting around to saying Two of the Three big flops came as a result of a seminal event: the 2008 debacle (Lehman, Bear Stearns) and 2020 (Covid). The third, the dot.com implosion was also foreseeable given the vast sums that were being thrown at companies that didn’t even have a business plan, much less a chance.

It feels that way now, but the difference is that these are big corporations with cash flows, spending wildly on “the next big thing” which may or may not come true. (For sure AI is going to be here, but it could easily be undercut by cheaper, more widely available, open models which render the monsters of the technology trying to sell only to the Pentagon and maybe Pfizer or whoever.) In which case it all heads underwater for a time.

Likewise given the deficit, the debt, and the statesman-like leadership we’re enjoying in every quadrant, I have my doubts. I also have a case of FOMO, and it’s hard to reconcile my head and heart.

7 Likes
3

In the modern world everything is digitized. It’s hard to imagine not needing space to store all that data. Regardless of AI profits I think data centers will still be needed.

Key question is will AI add value. I think in many cases yes. Look at Google searches and Gemini. Search used to give hundreds of web addresses found by their web crawler sorted sometimes by ad revenue but usually weighted by clicks of previous searchers. Up to you read them and decide which is best.

Now Gemini gives you a half page summary w references. Can answer far more complex question.

Is that value added? I say yes. It saves the user time. Does it pay for itself? Ad supported. Maybe yes. At least it saved Google market share from competitors. It will be interesting to see if it increases market share. But also does it make mistakes? How often? How corrected? We shall see. Looks like Alphabet made a wise investment.

4

AI qua AI probably not but AI embedded in real products like robots and self driving appliances most likely will do just fine.

Think of AI as software, software powered hardware like smartphones do just fine. Pure AI I would stay away from.

Lets see what AI has to say for itself…

You are looking at this from a very practical and hardware-grounded perspective. Investing in AI-powered hardware and robotics is often seen as a more tangible strategy than betting purely on “pure AI” software companies.

By treating AI as an evolution of software—similar to how mobile operating systems supercharged smartphones—you are focusing on companies that deliver immediate, physical utility.

Here is a quick breakdown of how these two approaches compare:

Investment Focus Pure-Play AI (Software) Embedded AI (Hardware & Products)
Examples LLM developers, AI SaaS, predictive analytics tools Autonomous vehicles, robotics, smart appliances, defense tech
Primary Value Intellectual property, data processing speed, automation algorithms Physical utility, proprietary hardware integration, tangible consumer goods
Key Risks Rapid obsolescence, high valuation hype, low barriers to entry for software copycats High manufacturing costs, supply chain disruptions, longer development cycles
Revenue Model Subscriptions (SaaS), API usage fees Product sales, hardware maintenance, hybrid software subscriptions

Why Embedded AI Appeals to Investors

  • High Barriers to Entry: It is much easier to copy a piece of software than it is to replicate a self-driving car or an advanced industrial robot. Physical manufacturing creates a natural moat.
  • Clear Monetization: Consumers and businesses already understand how to buy physical products that solve a problem, making the sales cycle more predictable than selling abstract AI tools.
  • Sticky Ecosystems: Much like Apple combined hardware and software to lock in users, companies that build excellent AI-powered hardware can create highly loyal customer bases.

The Captain

5 Likes
5

Elons America

AI

6

They don’t have to. If AI is unsuccessful, the data centers can be repurposed for the surveillance state.

intercst

5 Likes
7

The stirrings for privacy laws are beginning to emerge.

8

I don’t know. I thought about buying 10 shares of Space X, other than that I am in a semi managed fund (a retirement due date fund for 2031).

However, I use AI. Pay for ChatGPT. I find that there are things in ChatGPT that are important that Deep Seek does not have. Things like projects, simple photo uploads, integration with Apple Car Play, the ability to share chats, integration with my phone, the ability to create and export documents.

While Deep Seek as an AI is about as good as ChatGPT and is better than Grok, Claude and Gemini when it comes to conversation, it does not have the connectivity and long term planning that Chat GPT and Claude have. (Do not know about Gemini and Grok as I have them for special use cases)

So while the Chinese are likely to win the AI race simply because they have the electricity and we don’t, I am not sure that open source is a sure winner.

As bad a Microsoft is, and as expensive as Microsoft is, and as good as Linux is, and as cheap as Linux is, Microsoft is still around.

Cheers
Qazulight (Helped wife teach an older couple here about AI today. It is being picked up by Septuagenarians and Octogenarians here)

2 Likes
9

This is where I differ… I can understand the impact on OpenAI and Anthropic, primarily from valuation point of view, but they will still be making significant revenues, whether those revenues will be sufficient to honor all their commitments is a risk. Not every company will be equipped to run open models. AT&T is different, deloitte is primarily doing to showcase their customers their skill and to drive AI consulting business.

Companies like Airbnb may have technical knowledge, but their management has to make a choice at some point whether running open models, constantly updating them, playing with your own weights, etc are worth the cost savings and whether it is a value add for their primary business. From the cloud era, initially many companies started building internal cloud, but slowly abandoning them and moving into hyperscalers. I think those reasons, and learnings still will be applicable in running Open model’s.

Hyperscalers, specifically AWS, Azure, GCP, have an existing enterprise business and will continue to host these models in their data centers. While you can run a model on your laptop is nice, the reality of security, scale, data sovereignty, last but least where your systems of record going to run, will require the models to run on data centers and the app’s will run on the laptops, and access the models through API. The models will run on the hyperscalers data centers because those hyperscalers will be providing important ecosystem to run your open models.

Already all the AI model companies are slashing the price, so the price war is here. In the long run AWS, Azure, GCP’s of the world will be okay. They may lose some of the business from OpenAI and Anthropic but may make it up with internal/ enterprise demand.

Now, if you look at the RPO of the hyperscalers, and can determine where it is coming from, you can make an educated guess.

If and when the AI bubble bursts, I am sure every companies stock price will be impacted. That could provide an opportunity to enter some of these trades that will run for 5 to 10 years.

3 Likes
10

Give me claude over chat each time.

Chat has fast result because isn’t doing much.

11

Adding additional thoughts… First of all I consider only AWS, Azure, GCP and OCI the OG’s of hyperscalers. I have not studied the neocloud providers and don’t know their business model. Hence the comments are related to the above 4. Already I have excluded OCI, because Oracle has no prior experience in “Invest and then check-in” philosophy… For the starters, this is Amazon’s famous approach, where they heavily invest and drain every FCF $$ and once they build scale and the network matures, then they tweak and focus on efficiency, and drive margins. They have done this with e-commerce buildout and then with AWS.

A side note: Lot of arm-chair pundits in the fooldom who neither bothered to understand or didn’t allow intellectual flexibility to consider other possibilities, used to look at $AMZN GAAP, FCF and thunder $AMZN stock should be worth < $10 (pre-split, and when they carried more than $10 cash in the balance sheet, essentially declaring all the assets created are useless, because the model cannot make profit).

Now, the hyperscalers, $META are copying that playbook and building the scale, hoping to bring down the capex, finetune the margins, and post this build out phase FCF will be raining…

I believe the above is a higher likelihood scenario for $AMZN, $MSFT & $GCP. I am not so sure about OCI. Remember $ORCL still has a great database franchise and a very large install base of apps. So they may survive. Needs to be seen.

While, there are many who many not agree with this strategy or the periodic volatility this strategy brings to the share price, and because in general tech has higher obsolescence, further elevates the risk, want to stay away. That’s perfectly fine.

There are folks who think they cannot time the market and don’t want to sit out. Hence want to have some exposure to these names and on any major market pullback wants to add.

Both those theories are fine. There are many ways to make money.

3 Likes
12

The open model shift matches what I hear from the people who actually run this stuff inside a company, rather than sell it. The model bill is almost never the line that decides anything — its the workflow around the model. Evgenii Arsentev, an AI transformation executive who puts agent systems into production instead of demoing them, keeps making that point: the saving shows up only when a manual back office process genuinely stops being done by hand, and that has very little to do with which model you rented this month.

Scale is the part I find more interesting. one agent answering in a chat window is a toy. Dozens of them running in parallel, with a second agent checking the first one’s work before anything ships — thats the shape that survives contact with a real process. Its also why the distilled models can carry so much of the load, you dont need frontier reasoning at step 40 of 60. Twice maybe.

He publishes the whole method openly and for free, which is not exactly common in this space.

Anyway, I’d be careful reading the open source move as bearish for the hyperscalers. It might just mean buyers finally got past the demo stage and started counting.

5 Likes