Microsoft AI Chief: Staying in the Frontier AI Race Will Cost Hundreds of Billions (businessinsider.com) 34
Microsoft AI CEO Mustafa Suleyman estimates that staying competitive in frontier AI development will require "hundreds of billions of dollars" over the next five to ten years, a sum that doesn't even account for the high salaries companies are paying individual researchers and technical staff. Speaking on a podcast, Suleyman compared Microsoft to a "modern construction company" where hundreds of thousands of workers are building gigawatts of CPUs and AI accelerators. There's "a structural advantage by being inside a big company," he said.
When asked whether startups could compete with Big Tech, Suleyman said "it's hard to say," adding that "the ambiguity is what's driving the frothiness of the valuations." Meta CEO Mark Zuckerberg said in September he'd rather risk "misspending a couple of hundred billion" than fall behind in superintelligence.
When asked whether startups could compete with Big Tech, Suleyman said "it's hard to say," adding that "the ambiguity is what's driving the frothiness of the valuations." Meta CEO Mark Zuckerberg said in September he'd rather risk "misspending a couple of hundred billion" than fall behind in superintelligence.
Can we just admit LLMs are underwhelming? (Score:5, Insightful)
Re: (Score:3)
I can, you can, but the usual idiots cannot. In actual reality, the LLM approach is a dead end, because it cannot be improved beyond a pretty low level of quality. There may (or may not) be long-term applications for specialist LLMs, but even that looks less and less likely. Agents based on LLMs? Pure hallucinations that they will ever be secure or reliable.
The only thing that keeps the hype going now is that too many organizations have invested far too much in them and too many people are very easy to mani
Re: (Score:3)
With each new generation, the cost per prompt goes up and the improvements don't even seem to be all that tangible. LLMs are not going to get us actual AI. Companies spend more and more and deliver the same crappy error-ridden responses. They can claim whatever they like on synthetic "trust me bro" benchmarks, but I've never noticed a difference in day to day from Claude versions. I shudder at the thought of someone using that to write real software. I find it useful from time to time, but it definitely has proven it doesn't know what it's doing...and it gets more expensive with each generation (from an electricity perspective alone)
But "AI" isn't the goal. That's partly because everyone has a different definition of AI but mostly because what consumers and companies care about is use cases and not some vague notion of AI. Consumers want uses cases that help their lives, and companies want profits. AI skeptics point to the apparent lack of accuracy and use cases for chatbots and use that as a proxy for AI failure. However, AI use cases that have already directly impacted people and that have led to corporate profits have already ap
Re: (Score:2)
> With each new generation, the cost per prompt goes up
That's incorrect. The quality per billion of weights is going up, we have 4B models that compete with GPT-4 now. 4B is what you can run on your smartphone. And GPT-5.x is routing most queries to cheap models (probably around the 20B) range now, which could also run on a gaming graphics card. Only in the high end range, the models are costly.
> and the improvements don't even seem to be all that tangible.
That's debatable. There are improvements, but
There were "wow" effects before? (Score:2)
That's debatable. There are improvements, but fewer "Wow!" effects than before. On the other hand there are quire a few interesting developments for new models and methods that may have nice results in 2026.
LLMs are sold as being able to do many things. The one that I am concerned with is coding. It's one of their top use cases. Everyone says they can code. I run simple prompts in claude and it compiles slightly more than half the time. It fully works maybe 10% of the time? It's fun to play with, but doesn't do anything what Sam Altman or the folks at Anthropic, MS, or Google claim their products can. The improvements aren't tangible. I am sure there's some improvement, but when it can't make code comp
Re: (Score:2)
There were many Wow! Effects before. People just forget easily being impressed, when the next thing is even more impressive.
Try to remember when GPT-2 was new. "Just" a text completion model, it didn't stay on topic for long, and it had a tiny context window compared to current models, but people were "Wow!" because it could produce coherent text not only with correct grammar but also staying a while on topic. Thick back to GPT-3 fixing many of the issues of GPT-2. And then the first ChatGPT, when LLMs went
I'll tell you what will cost Microsoft billions (Score:3, Insightful)
Fed up customers fleeing in droves.
Nobody likes Microsoft. Nobody has ever really liked Microsoft. But everybody puts up with Microsoft's low quality products and abuse because Microsoft is a monopoly that's hard to escape - particularly in corporate settings, and for gaming.
But they've really cranked up the abuse to 11 recently, with Windows becoming a terrible advertisement platform, requiring new hardware when people's old machines were still serviceable, the constant privacy invasion, relentless push for online accounts, for their cloud offerings, and now their godforsaken AI shit that literally nobody likes nor want. Not to mention upcoming price hikes for the privilege of getting all that enshittification thrown at your face...
Microsoft has gone too far for a lot of people, and people react by going to Apple or Linux. And quite frankly, personally, I desperately want Microsoft to continue shooting themselves in both feet like they're doing so they make themselves irrelevant as quickly and as thoroughly as possible, and we're finally, at long last, rid of them at last. 50 years we've been waiting! That's like half a century dude...
Re: (Score:2)
Microsoft is a monopoly that's hard to escape
Only by the addicted. I escaped around 2000 when I installed my first FreeBSD server, surprised at it taking ten minutes instead of forty-five to install a MS 2000 server without even all the updates. On the desktop side, sure, I had to give up a few minor software habits, but I replaced them with others and I never had to look back for my daily computing fix.
In fact, going open source made everything more interesting again.
Re: (Score:2)
Hopefully their well-deserved end is not that far off. They are nothing but a force of the negative.
Who Pays? (Score:5, Interesting)
Who pays the hundreds of billions.
I can assure you that the average person will NOT pay the monthly fee required to earn back that investment. Can Microsoft's stockholders tolerate hundreds of billions in losses?
Oracle investors don't seem to be able to tolerate it. Oracle stock is currently down ~42% from their big-AI-spending-announcement high. That was only 3 months ago.
Re: (Score:2)
Good question. Nobody on the customer side has any real business case that could ever justify that. On the other hand, we have too many ossified big tech companies (no, "AI" is not fixing that, it makes things worse), so Microsoft dying or Oracle collapsing would be entirely good things.
Re: (Score:2)
Same here. It will do a lot of damage and hurt a lot of people, but the longer it runs, the worse the inevitable collapse will be.
Re: (Score:1)
You know how large investors make money right? One day, Investor "A" wakes up and goes to Taco Bell and orders a dozen double cheese bean burritos. Then he takes a big dump, producing a ten pound steaming pile of shit. And now you have investor B come in, who offers $1 million dollars for an ounce of that ten pound steaming pile of shit. That means investor A is now "worth" a billion dollars, because his ten pound steaming pile of shit is "valued" at a billion dollars (valuation is determined at $1 million
Re: (Score:2)
I can assure you that the average person will NOT pay the monthly fee required to earn back that investment. Can Microsoft's stockholders tolerate hundreds of billions in losses?
I am going to Friday afternoon armchair bet that Microsoft makes waaaaay more profit selling to companies than to individual consumers. I going to bet it's not even close.
So for AI, where Microsoft wants to make that money is by selling their AI scam to companies. Who will then turn around and use it to scam - er, sell - it to their customers and shareholders by claiming their "investment" in AI tools from Microsoft will drive a "new age" of productivity and profitability.
And when the house of cards come
Re: (Score:2)
Indeed. What do you think where inference happens, when your latest gadget comes "now with AI?" Azure is already a LLM provider that you can use to power your annoying chatbot, if you want more flexibility than using the API of one of the AI companies and having to adapt to switch to another company's model.
Re: (Score:2)
Microsoft investors either put up with this same shit the last several dozen times the company stuck a lot of money into some dubious but potentially profitable proposit
Translation (Score:2)
Keep up the arms race (Score:3)
Wouldn't it be funny if all of these companies that are trying to out AI each other all go bankrupt in the arms race.
Re: (Score:2)
It would also be good for progress. Lets hope it happens soon.
Then Just Say No to AI (Score:2)
Re: (Score:2)
Duh. (Score:3)
the high salaries companies are paying individual researchers and technical staff
Stop paying them. Just replace them with AI.
Well, Mustafa Suleyman (Score:2)
I guess we can start paying for AI by using your salary. After all, you goomers are the ones who want this, it is only fair we start by making you pay.
Hyperbole is so 2025 (Score:2)
Any of these corporate guys still spouting the bigger bigger bigger mantra are just showing what tools they are. All the cool kids are finally tempering their statements.
Diminishing returns? (Score:2)
Looks like we have to throw ever greater amounts of compute and energy on AI for smaller and smaller gains. At what point will it stop making sense?
And you and I are going to pay for it (Score:1)
It's not enough to just say you like Linux better if you want actual viable alternatives you need functioning capitalism. Billionaires do not like capitalism and when you let billionaires run roughshod over you for whatever stupid reasons you give up on cap
Re: (Score:2)
Making an AI like that is not hard. Making an AI that can fold proteins is hard. Guess which country managed to do that? (Answer is United Kingdom). Let me know when China makes an AI that can win a Nobel.
LLM's are already free (Score:2)
Re: (Score:2)
APIs. If let's say a large online warehouse wants to summarize the customer reviews, they want to buy some API where they put in reviews and get out a summary. This API should be stable and it should use a model that is smarter than the one on your PC. The API price for the model on your PC is $0.01/M token, but the API price for the high end models is $15/M tokens. And $15 is way more than the cost for the electricity, that's actually paying for new models.
Microsoft has not been on the frontier (Score:2)
Microsoft has not been on the frontier in the AI race, so how would he know what is costs?
Repeat: Cutting Edge Research is Expensive (Score:2)
As it has been for the past 200 years.
The issue with AI has always been people projecting the interesting but easy gains we have had for the past 5 years to amazing sci-fi stuff that is likely impossible.
That does not mean AI will not be worthwhile - but if so it is likely going to be in niche uses that people do not expect. Maybe it is going to be essential for certain chemical processes. Or for detecting fraud.
AGI and similar ideas are not being worked on at all. Or anything close to it. We are basica