I feel like the Microsoft interest in DeepSeek is such a signal. Many valley tech companies are already building their AI on open source, if Microsoft jumps in and everyone figures out that running V4 natively is significantly cheaper, and the actual performance difference is negligible (I’m sure whilst Fable 5 was available, it could do awesome thing things, however 99.9% of my use case was covered with mostly Sonnet and a bit of Opus), why WOULDNT you do it.
Made even worse by one of the biggest data centre pushers themselves realising maybe it’s not all necessary, could finally change things.
Google has 13 products with over a billion users. Many produce cash, cash and more cash. OpenAI has an albotross around its neck. Anthropic found THE niche, the question is will it matter
Especially in about 12 months when ever model basically is sonnet 4.6 or better in quality. People will choose the cheapest.
Great roundup, Gary! I’ll add the (unsurprising to those here) 404 Media coverage of a paper showing that bad actors can poison deep learning results with just 13 words in a Reddit post. AEO is the new SEO and it’s even more sus, but no one can ignore the chaos.
This is why I've never once intentionally used an LLM chatbot or any of the major products in the AI space. I don't want to have to think about who has been either subtly or unsubtly try to manipulate me using the model as an intermediary.
I don't think either OpenAI or Anthropic have a future - Anthropic almost certainly has almost as bad numbers. I strongly suspect neither will actually IPO and I think your estimate of 30c on the dollar of acquisition cost is high. It'll be more of a bankruptcy where Amazon, Microsoft and Oracle split the carcass for the nominal price of the compute commitments both have made to them
I have to say, Microsoft considering DeepSeek (I assume they'll end up being pressured by the Trump regime to not do that, but use another US model) sends a message to all the Anthropic fanbois who continue to say "I have to spend gobs of money because Opus is SO GOOD!" - without ever actually having tried the top of the line open weight models like the new Kimi K2 Coder.
As an aside, I posted this comment in a Note earlier:
QUOTE:
I’m still not sure why “reasoning over large codebases” is actually all that. Humans don’t do it well, but then they tend not to do it at all if it can be avoided. Instead, humans zeros in on the section of a large codex that is already understood to be important, based on a CONCEPTUAL understanding - not a TEXT understanding - of the entire codex.
Conceptual understanding is exactly what LLMS don’t do at all. let alone well. So instead, they call it “context” - and that’s why we have LLMs with a million token “contexts.”
END QUOTE
As I keep saying, "good enough" for cheap is good enough.
Also, note that Anthropic now being sued for false advertising on its token limits for its subscription plans. Which is directly related to the issue of cost:
Anthropic Sued Over Alleged False Advertising on Claude Max Subscription Usage Limits
The lawsuit claims that the AI company is misleading customers about its $200-per-month Claude plan.
Look how well open source worked out for IBM owning Red Hat (Linux). They will stick with OpenAI, perhaps as a discount. Most likely, they will end up owning OpenAI.
<… except to say that another dark horse bidder is Elon. Wouldn’t that be an interesting twist?>
It would be "revenge best served cold" for Elon after he lost his case against Altman. Of course he would be putting all that paper money in an even bigger furnace, but Elon probably doesn't care. If he can spin the SpaceX story as a stock worth 100x revenue, I'm sure he could find a way to make it seem that buying OpenAI at cents on the dollar was a smart business move.
"I'm sure he could find a way to make it seem that buying OpenAI at cents on the dollar was a smart business move." Unfortunately, I think you're right about that.
Perhaps we are all missing the bigger signal. Local models are just as good as frontier models. There is zero difference between frontier models. All models are equally good at giving confident and authoritative answers even though the answers are incorrect. Which is often. Microsoft doesn’t want to pay for the tokens. Big surprise. If your tokens are throw away I don’t want to pay for them either. Nadella is telling the market that explicitly.
Update: The above called out enthusiast has conceded that the latest Qwen 3 seems to be quite a step up towards parity: "Lokale KI ist jetzt so gut wie ChatGPT? | c't 4004 #29" (https://www.youtube.com/watch?v=A5HJ1YnbTIk)
Again, I wouldn't know myself, being a "KI Hater".
A signal will be clear that the longer-term structuring of the LLM frontier players with other parts of the AI industrial complex is starting is "if" the OpenAI IPO is delayed. It is becoming clear that for some period of time as token-based pricing enters the market (and raises the price of AI services) there will be an over-capacity (or insufficient demand to keep prices up) in the market. The question will quickly become; how do we (players in the industry) address this? If countries in the EU, Britain and other countries proceed to build their own native LLM systems, then the growth anticipated from those places will hit the U.S. players expansion plans hard. Growth and M&A strategies will need to be revised.
"Growth and M&A strategies will need to be revised" That's putting it mildly, especially with the growing aversion of us European against anything US (with the possible exception of Harley Davidson motor bikes and US muscle cars in general, which have a cult following).
And I notice the absence of the very USAien term "profit" in your and other statements ;-)
I always felt that Google was in the best position to get the general public on AI. They simply have the best ecosystem by a longshot. Gmail is ubiquitous. Their office suite is fantastic. I'm not sure how their search abilities help their LLMs, but I assume they must, right???
All things being equal, if Google AI platforms are anywhere close to everyone else, then for the general public it is a no brainer to stay with Google.
I hope google will focus on offering specific targeted AI tasks to users. This is what the general population will actually use. We do not need platforms that can do everything, but rather those than can help us do the things we actually do.
"Gmail is ubiquitous." I am not using it, I just got the account when I bought my Samsung S2 a long time ago.
"Their office suite is fantastic." I am not using it. LibreOffice is more than excellent for my use cases.
"search abilities help their LLMs, but I assume they must, right?" Please read up on the german court in Munich making Google responsible for the results of their own fucked up LLM.
Your's maybe an US centric view, but from the current state, being from Europe, I would put my short term money on Microsoft, which has itself pretty well entrenched in european government organisation. My company switched to O365, now we have to mark every document with "sensitive" to avoid it being loaded up onto some MS platform. Of course, MS pinky-swears all our secrets are save with them.... yeah, I heard that one before.
I do not take issue with anything you wrote. I will revise my comment to say that Google and Microsoft are both in very strong positions. Apple will find how to incorporate AI into their ecosystem as well, but I don't imagine them being the same kind of player at all. That being said, when the OpenAI fire sale happens, Apple may very well be a major buyer of their assets.
True, as they really have the money, and don't need to borrow (as Google has recently done). If they get either OpenAI or Anthropic for 1...10c to dollar, this might even be good business.
We should all stop using these Silicon Valley apps. Try Swiss made Lumo. Works well and respects your privacy. And oh, they don’t send money to Trump either, which is a huge plus.
I am wondering if there would be a good moment to short OpenAI and/or companies that depend on revenue from them… if (rather when) it goes down, the companies it has commitments to would be in trouble. Oracle might be the most vulnerable here
I feel like the Microsoft interest in DeepSeek is such a signal. Many valley tech companies are already building their AI on open source, if Microsoft jumps in and everyone figures out that running V4 natively is significantly cheaper, and the actual performance difference is negligible (I’m sure whilst Fable 5 was available, it could do awesome thing things, however 99.9% of my use case was covered with mostly Sonnet and a bit of Opus), why WOULDNT you do it.
Made even worse by one of the biggest data centre pushers themselves realising maybe it’s not all necessary, could finally change things.
Google has 13 products with over a billion users. Many produce cash, cash and more cash. OpenAI has an albotross around its neck. Anthropic found THE niche, the question is will it matter
Especially in about 12 months when ever model basically is sonnet 4.6 or better in quality. People will choose the cheapest.
Don't you mean an AIbotross?
(Oh, if only Substack let us have serif fonts, then AIbotross and Albotross Wouldn't look the same.)
Ai-bot rusts?
ᴀɪʙᴏᴛʀᴏss
ᴀʟʙᴏᴛʀᴏss
⠀
When Substack doesn't let you — hack at it until u make it.
⠀
\ /
oVo
\___XXX___/
__XXXXX__
/__XXXXX__\
/ XXX \
V
We need Pink AI-lefonts for all things AI
Great roundup, Gary! I’ll add the (unsurprising to those here) 404 Media coverage of a paper showing that bad actors can poison deep learning results with just 13 words in a Reddit post. AEO is the new SEO and it’s even more sus, but no one can ignore the chaos.
This is why I've never once intentionally used an LLM chatbot or any of the major products in the AI space. I don't want to have to think about who has been either subtly or unsubtly try to manipulate me using the model as an intermediary.
I don't think either OpenAI or Anthropic have a future - Anthropic almost certainly has almost as bad numbers. I strongly suspect neither will actually IPO and I think your estimate of 30c on the dollar of acquisition cost is high. It'll be more of a bankruptcy where Amazon, Microsoft and Oracle split the carcass for the nominal price of the compute commitments both have made to them
I wrote about the contrast between SpaceX and the other two here - https://ombreolivier.substack.com/p/one-ipo-achieved-two-flame-outs-likely?r=7yrqz - and go into more detail as to why I don't think those IPOs will happen, but no moat and no road to cashflow positive is the TL;DR
Falling to 50% market share of a rapidly growing market is still possibly a good strategic place to be if they can adapt. "if".
OpenAI's problem is that its early success went to its head. It tried to be everything and hasn't found anything with a real payoff.
I have to say, Microsoft considering DeepSeek (I assume they'll end up being pressured by the Trump regime to not do that, but use another US model) sends a message to all the Anthropic fanbois who continue to say "I have to spend gobs of money because Opus is SO GOOD!" - without ever actually having tried the top of the line open weight models like the new Kimi K2 Coder.
As an aside, I posted this comment in a Note earlier:
QUOTE:
I’m still not sure why “reasoning over large codebases” is actually all that. Humans don’t do it well, but then they tend not to do it at all if it can be avoided. Instead, humans zeros in on the section of a large codex that is already understood to be important, based on a CONCEPTUAL understanding - not a TEXT understanding - of the entire codex.
Conceptual understanding is exactly what LLMS don’t do at all. let alone well. So instead, they call it “context” - and that’s why we have LLMs with a million token “contexts.”
END QUOTE
As I keep saying, "good enough" for cheap is good enough.
Also, note that Anthropic now being sued for false advertising on its token limits for its subscription plans. Which is directly related to the issue of cost:
Anthropic Sued Over Alleged False Advertising on Claude Max Subscription Usage Limits
The lawsuit claims that the AI company is misleading customers about its $200-per-month Claude plan.
https://www.cnet.com/tech/services-and-software/anthropic-sued-alleged-false-advertising-claude-max-subscription-usage-limits/
Look how well open source worked out for IBM owning Red Hat (Linux). They will stick with OpenAI, perhaps as a discount. Most likely, they will end up owning OpenAI.
<… except to say that another dark horse bidder is Elon. Wouldn’t that be an interesting twist?>
It would be "revenge best served cold" for Elon after he lost his case against Altman. Of course he would be putting all that paper money in an even bigger furnace, but Elon probably doesn't care. If he can spin the SpaceX story as a stock worth 100x revenue, I'm sure he could find a way to make it seem that buying OpenAI at cents on the dollar was a smart business move.
"I'm sure he could find a way to make it seem that buying OpenAI at cents on the dollar was a smart business move." Unfortunately, I think you're right about that.
Perhaps we are all missing the bigger signal. Local models are just as good as frontier models. There is zero difference between frontier models. All models are equally good at giving confident and authoritative answers even though the answers are incorrect. Which is often. Microsoft doesn’t want to pay for the tokens. Big surprise. If your tokens are throw away I don’t want to pay for them either. Nadella is telling the market that explicitly.
"Local models are just as good as frontier models" According to the AI enthusiasts of "ct3003" this is not the case. (https://youtube.com/@ct3003?si=l8J_wWMXFLiIPH45)
I wouldn't know myself, but I watch that channel to see what new wonders AI brings to us ;-)
Update: The above called out enthusiast has conceded that the latest Qwen 3 seems to be quite a step up towards parity: "Lokale KI ist jetzt so gut wie ChatGPT? | c't 4004 #29" (https://www.youtube.com/watch?v=A5HJ1YnbTIk)
Again, I wouldn't know myself, being a "KI Hater".
Bought by Elon wouldn't just be a neat twist but an almost Shakespearean denouement
A signal will be clear that the longer-term structuring of the LLM frontier players with other parts of the AI industrial complex is starting is "if" the OpenAI IPO is delayed. It is becoming clear that for some period of time as token-based pricing enters the market (and raises the price of AI services) there will be an over-capacity (or insufficient demand to keep prices up) in the market. The question will quickly become; how do we (players in the industry) address this? If countries in the EU, Britain and other countries proceed to build their own native LLM systems, then the growth anticipated from those places will hit the U.S. players expansion plans hard. Growth and M&A strategies will need to be revised.
"Growth and M&A strategies will need to be revised" That's putting it mildly, especially with the growing aversion of us European against anything US (with the possible exception of Harley Davidson motor bikes and US muscle cars in general, which have a cult following).
And I notice the absence of the very USAien term "profit" in your and other statements ;-)
I always felt that Google was in the best position to get the general public on AI. They simply have the best ecosystem by a longshot. Gmail is ubiquitous. Their office suite is fantastic. I'm not sure how their search abilities help their LLMs, but I assume they must, right???
All things being equal, if Google AI platforms are anywhere close to everyone else, then for the general public it is a no brainer to stay with Google.
I hope google will focus on offering specific targeted AI tasks to users. This is what the general population will actually use. We do not need platforms that can do everything, but rather those than can help us do the things we actually do.
Yeah, about that:
"Gmail is ubiquitous." I am not using it, I just got the account when I bought my Samsung S2 a long time ago.
"Their office suite is fantastic." I am not using it. LibreOffice is more than excellent for my use cases.
"search abilities help their LLMs, but I assume they must, right?" Please read up on the german court in Munich making Google responsible for the results of their own fucked up LLM.
Your's maybe an US centric view, but from the current state, being from Europe, I would put my short term money on Microsoft, which has itself pretty well entrenched in european government organisation. My company switched to O365, now we have to mark every document with "sensitive" to avoid it being loaded up onto some MS platform. Of course, MS pinky-swears all our secrets are save with them.... yeah, I heard that one before.
I do not take issue with anything you wrote. I will revise my comment to say that Google and Microsoft are both in very strong positions. Apple will find how to incorporate AI into their ecosystem as well, but I don't imagine them being the same kind of player at all. That being said, when the OpenAI fire sale happens, Apple may very well be a major buyer of their assets.
True, as they really have the money, and don't need to borrow (as Google has recently done). If they get either OpenAI or Anthropic for 1...10c to dollar, this might even be good business.
We should all stop using these Silicon Valley apps. Try Swiss made Lumo. Works well and respects your privacy. And oh, they don’t send money to Trump either, which is a huge plus.
I am wondering if there would be a good moment to short OpenAI and/or companies that depend on revenue from them… if (rather when) it goes down, the companies it has commitments to would be in trouble. Oracle might be the most vulnerable here
Boooooo! To the statistical maniac who made that graph and decided that Gemini should be medium blue, and DeepSeek should be medium-dark blue.
Not to mention release of GLM 5.2 which is within spitting distance of OAI and Anthropics best models while being open and costing far less!
When the comments are just as enlightening as the text itself. Double win.