Funny how they include nvidia, micron, and AMD revenue as “AI revenue” and that it represents that majority of industry revenue but presumably a big chunk of everyone else’s spend. Almost might as well include electric utility revenue as AI Revenue by that metric
The accounting is pretty iffy there though. Like the top line implies Amazon lost $334bn on AI but they aren't really an AI company and made a $16bn profit last quarter at AWS up 64% on the previous year. Also a larger profit as a company as a whole.
AI is more like an ongoing research project than it is a product. GPUs are currently the product that is profitable and Nvidia will keep funding labs to keep buying GPUs.
Yes, they said that about a lot of companies that dumped product until the competition was eliminated, then abused their monopoly position and resultant political connections to do whatever they wanted. Their sites are worse than ever. Literally worse than the day after launch.
This is exactly how Cloud Computing looked in 2012-2018. Dumping huge $ into computing buildout that wasn't profitable yet. All those co's: Amazon, GCP, Azure paid off immensely and are ridiculously profitable.
No, it wasn't. Cloud computing was almost immediately profitable.
And Amazon was famously "unprofitable" for their first 9 years because they were investing all their very real profits into a form of capital that the US tax code didn't recognize.
You are being downvoted but only a few years ago Google contemplated whether to kill GCP altogether because of lack of traction in the market or revenue goals.
IIRC Amazon chose to reinvest early revenue to grow AWS intentionally, and the revenue curve eventually evened out and obviously surpassed expenses. The problem with the AI buildout is that there's not a ton of evidence that these companies are approaching profitability, we can't even know because they're private.
Amazon's incubation of AWS was methodical and transparent. OpenAI is saying that they don't expect to be profitable until at least 2030, with over a trillion dollars in committed spend before that point. It's the largest "trust me bro" play in human history.
My main issue with this timeline is that AI still has trouble transitioning to the real world. It predicts for 2029:
> There are swarms of insect-sized drones that can poison human infantry before they are even noticed; flocks of bird-sized drones to hunt the insects; new ICBM interceptors, and new, harder-to-intercept ICBMs. The rest of the world watches the buildup in horror, but it seems to have a momentum of its own.
Does anyone really predict insect drones _in production_ 3 years from now, to the degree that we need bird drones to hunt the insect drones? How the hell are these things powered?
Lean/math/millenienium prizes are "grindable" [0]. Wake me up when AI is making order-of-magnitude improvements in ungrindable real world tasks like batteries, hypersonic engine manufacturing, and stealth/silent motors that you can't hear.
> Does anyone really predict insect drones _in production_ 3 years from now
The industrial expansion timelinen as described in AI 2027 is way too compressed; I don't think any AI doomer believes that. The dynamics are plausible though, even without China stealing the weights.
What theft? Modern humans learned from the works of past humans, LLMs learned from the works of modern humans, other LLMs learned from those LLMs. Given that the methods by which Anthropic and OpenAI obtained training data remains a legal gray area, it's not so cut and dry to me that distillation is stealing. It's just a continuation of the long tradition of building knowledge on the knowledge of others.
I take it you didn't read AI 2027; it doesn't philosophize about cultural property - it simply narrates that a turning point is China acquiring the Agent's weights through hacking. "Stealing" in this context is simply the strictly legal term.
Please elaborate: what dynamics? This is rather vague. My point is that AI can't grind real world physics/chemistry/engineering. What dynamics are in play here?
This doesn't solve grinding? If it did, we'd see results by now (Published: 27 January 2026), and there's months of delay between submission and publication.
I'm willing to be wrong, but I'm just not seeing anything worth doomering over. There are multiple companies throwing AI at materials discovery; a research paper about a "data-driven framework" is about as unthreatening to my thesis as it gets. I'm willing to cede the point if say, Radical releases ~3 new materials that have commercial applicability and ~3x some useful metric e.g. tensile strength, but until then, to my amateur eye, it looks like AI+Real World is missing its ChatGPT moment.
Gonna note that biology is even less grindable than ordinary chemistry/physics/engineering. We'll have advancements in other real-world fields before AI-powered bioengineering becomes A Thing.
I think the idea is that AI itself is going to massively accelerate its own development, and AI with real-world competence is coming very soon. At the rate things are going, it wouldn't suprise me at all if we had mass production of AI-designed systems in the next six months, actually.
> AI with real-world competence is coming very soon
Disagree, Moravec's Paradox remains undefeated. How long has Elon promised self-driving cars? Or how long have we been seeing humanoid-robot-walking demos? Laundry folding demos?
Let's assume that a magical AI powered robot hand lands tomorrow. How long do you think it'll take to ramp up assembly/production/distribution/sourcing/materials/QA for, say, a million of them? Never mind the legal/integration/maintenance time.
And even once those have landed, and assuming they are ALL put to work on iterating on research testing to build out super-high-capacity-drone-batteries, how long do you think it would take for that to evolve into killer-insect-drones? BTW you should know that even though high capacity silicon carbon batteries exist, they haven't supplanted other Li-ion batteries for a host of reasons. A million things stand between a technology working in a lab and surviving the real world.
And during this entire process, the (geo)political/social/legal process will be churning away, changing the societal landscape in which the killer-robots land. Not to mention that mechanistic interpretability is (likely) somewhat grindable. A lab just needs to dump a billion dollars of compute into it after it declares AGI.
I'm unconvinced we get killer insect drones before we crack mech-interp.
And even once we get killer insect drones, they need to somehow have no kill-switch and then literally exterminate everyone on the globe? Can you see why I have a problem with doomers predicting human extinction within the decade?
I'm putting words in your mouth and not replying to precisely what you said, just the general vibe. LMK if I overstepped.
>Agent-2, more so than previous models, is effectively “online learning,” in that it’s built to never really finish training. Every day, the weights get updated to the latest version, trained on more data generated by the previous version the previous day.
That seems to me to be a natural progression, from discrete models to models that are just continuously improved. Maybe we'll end up with different models with different rates of improvement rather than static differences in performance, and methodologies for that improvement will be the thing we care about. Maybe over time, even benchmark tests will be primarily concerned with that kind of efficiency.
I think a huge debate right now is the relative value of the "frontier" models from Western companies at the cutting edge, vs distilled versions of those models that are good enough and exponentially cheaper coming from China. But a paradigm of 'always training' means an always active, always advancing frontier, which is a stronger moat than a one-off model that's more advanced for a few months.
One of the most biased claims IMO in AI 2027 is that a huge portion of the geopolitical and existential risk argument is hinged on the notion that China just steals the US frontier weights.
Look what's coming out of China, they are catching up on performance and surpassing the US in efficiency. They're on a different level when it comes to open releases of weights.
I don't, to me the entire premise is a bit flawed at it's core (ASI), and I read it like bad science fiction with China playing the bad guy just a narrative crux so we get to the acceleration timeline and warring nation states.
The US/China divide is one thing, but if I'm asked whether I trust OpenAI or Deepseek, as companies, more, I'm not sure how I'd answer. I suppose my default is to distrust whoever is in the lead, since the lead is power, and power corrupts.
Interesting but it degenerates into sci-fi tropes if you look at the extrapolations. Reminds me of 90s writing about what the Internet was going to do.
The Internet ended up being both more incredible and more mundane than predicted.
One of the first things I learned in ML is that the model can only learn from the information in X. If X doesn't contain enough information to determine y, no amount of compute can fully recover it.
That's why I'm skeptical of grand claims about AI. Scaling can make models much better, but it can't create information that isn't there. An AI system can be extremely useful without becoming superhuman.
And what do you think about recent developments, for example Navier-Stokes? I always thought the same but now I have doubts, but maybe it is just psyops from openai.
It is speculated that there was context leak from a NYU Professor's chat I think? I contribute to SciPy heavily and I've seen AI shit the bed a few times now. So 13 million lines of lean? Hell no dude!
Sorry, but the notion that creative writing and robotaxi's exist as proof of anything is like saying my child can drive and write; and while true, the measure of that ability is not at the level of the best humans. It's average at best.
All of the early data shows that self-driving cars from Waymo are safer than the average attentive driver. This is not the average driver, since of the 36,000 deaths a year from accidents in the US, about 12,000 of them are due to alcohol or impaired driving.
As for writing: AI is not nearly as good as the average professional writer, but they are definitely better than the average citizen of the United States, considering that 21% of adults are not functionally literate in the US. AI has no problem writing at the undergrad level.
Depends what your definitions of "requires" and "full" are. On mine there's some nagging on the scale of once every tens to hundreds of seconds if it thinks I'm not paying attention. You're willingly living a more stressful and unsafe life if you're still manually driving your car in 2026.
Fun to read this again and see actual parallels. The 2030 Takeover section is such a ludicrous leap, however. None of the supply chain infrastructure, energy, or Moravec's Paradox realities are ever addressed. Turn the page and suddenly humanity is largely annihilated with a Corgi-esque human breed kept as pets. How did these robots emerge from utter rhetorical nothingness? Robocalypse impossible? Perhaps not. By 2030? an intellectually embarrassing farce worthy of a facepalm.
I wonder when the people will get that intelligence is not only directed at the external, but only really starts when you look at the internal (joy, pleasure, traumas, taboos, awkwardness, abuse etc.). Look up the word "interoception".
https://isaiprofitable.com/
I like this one better
Funny how they include nvidia, micron, and AMD revenue as “AI revenue” and that it represents that majority of industry revenue but presumably a big chunk of everyone else’s spend. Almost might as well include electric utility revenue as AI Revenue by that metric
The accounting is pretty iffy there though. Like the top line implies Amazon lost $334bn on AI but they aren't really an AI company and made a $16bn profit last quarter at AWS up 64% on the previous year. Also a larger profit as a company as a whole.
Oh look, the answer is the same when I looked a couple months ago. Interesting
AI is more like an ongoing research project than it is a product. GPUs are currently the product that is profitable and Nvidia will keep funding labs to keep buying GPUs.
[dead]
I don’t understand how they track these figures.
For example: META made a bit more than $5 bn revenue since 2022?
Hey now, some of them reach -50% profit margins!
You almost have hope for them from the total number until you realize the vast majority of that green is made up of NVIDIA
On the other hand, people said this about Amazon, Google, Meta/Facebook, and loads of other huge tech things that are now wildly profitable.
It hasn't been a reliable predictive metric so far. All runaway growth tech sectors and businesses tend to look economically insane until they're not.
> and loads of other huge tech things that are now wildly profitable.
They also said it about loads of other huge tech things that died. Pointing at the ones that made it is literally survivor bias.
Yes, they said that about a lot of companies that dumped product until the competition was eliminated, then abused their monopoly position and resultant political connections to do whatever they wanted. Their sites are worse than ever. Literally worse than the day after launch.
This is exactly how Cloud Computing looked in 2012-2018. Dumping huge $ into computing buildout that wasn't profitable yet. All those co's: Amazon, GCP, Azure paid off immensely and are ridiculously profitable.
No, it wasn't. Cloud computing was almost immediately profitable.
And Amazon was famously "unprofitable" for their first 9 years because they were investing all their very real profits into a form of capital that the US tax code didn't recognize.
GCP reported its first quarterly operating profit in Q1 2023, roughly 15 years after Google began offering cloud computing.
Frontier labs are also reinvesting their tens of billions of revenue back into infra scaleout, sounds like Amazon.
If 2 of 3 were like this, and Azure numbers were never split out, want to take a guess what the economics of the third was like too?
Yes, there exist cloud computing companies that took a long time to be profitable. IDK if Oracle ever was.
Cloud computing was still profitable the entire time. Amazon alone could cover all of GCP expenses.
You are being downvoted but only a few years ago Google contemplated whether to kill GCP altogether because of lack of traction in the market or revenue goals.
Apples to oranges
the same type of people will eventually hate it when AI actually starts profiting at which point they will ask for redistribution.
damned if you do.
damned if you don't.
if all of humanity's knowledge was redistributed to AI labs to sell back to us, it's only fair that some of the profits get redistributed back.
IIRC Amazon chose to reinvest early revenue to grow AWS intentionally, and the revenue curve eventually evened out and obviously surpassed expenses. The problem with the AI buildout is that there's not a ton of evidence that these companies are approaching profitability, we can't even know because they're private.
Amazon's incubation of AWS was methodical and transparent. OpenAI is saying that they don't expect to be profitable until at least 2030, with over a trillion dollars in committed spend before that point. It's the largest "trust me bro" play in human history.
My main issue with this timeline is that AI still has trouble transitioning to the real world. It predicts for 2029:
> There are swarms of insect-sized drones that can poison human infantry before they are even noticed; flocks of bird-sized drones to hunt the insects; new ICBM interceptors, and new, harder-to-intercept ICBMs. The rest of the world watches the buildup in horror, but it seems to have a momentum of its own.
Does anyone really predict insect drones _in production_ 3 years from now, to the degree that we need bird drones to hunt the insect drones? How the hell are these things powered?
Lean/math/millenienium prizes are "grindable" [0]. Wake me up when AI is making order-of-magnitude improvements in ungrindable real world tasks like batteries, hypersonic engine manufacturing, and stealth/silent motors that you can't hear.
[0]: https://www.dwarkesh.com/p/the-next-paradigm
> Does anyone really predict insect drones _in production_ 3 years from now
The industrial expansion timelinen as described in AI 2027 is way too compressed; I don't think any AI doomer believes that. The dynamics are plausible though, even without China stealing the weights.
What theft? Modern humans learned from the works of past humans, LLMs learned from the works of modern humans, other LLMs learned from those LLMs. Given that the methods by which Anthropic and OpenAI obtained training data remains a legal gray area, it's not so cut and dry to me that distillation is stealing. It's just a continuation of the long tradition of building knowledge on the knowledge of others.
I take it you didn't read AI 2027; it doesn't philosophize about cultural property - it simply narrates that a turning point is China acquiring the Agent's weights through hacking. "Stealing" in this context is simply the strictly legal term.
It stuns the mind that anyone can compose such a sophism.
Literally 2 days ago, from a former Anthropic researcher:
> The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.
https://x.com/hilbertspaess/status/2097476203863224394
> The dynamics are plausible
Please elaborate: what dynamics? This is rather vague. My point is that AI can't grind real world physics/chemistry/engineering. What dynamics are in play here?
Solve the grinding then: https://www.nature.com/articles/s41524-026-01964-8
This doesn't solve grinding? If it did, we'd see results by now (Published: 27 January 2026), and there's months of delay between submission and publication.
I'm willing to be wrong, but I'm just not seeing anything worth doomering over. There are multiple companies throwing AI at materials discovery; a research paper about a "data-driven framework" is about as unthreatening to my thesis as it gets. I'm willing to cede the point if say, Radical releases ~3 new materials that have commercial applicability and ~3x some useful metric e.g. tensile strength, but until then, to my amateur eye, it looks like AI+Real World is missing its ChatGPT moment.
I think if organized crime couldn’t conquer the world, ai doesn’t have a chance. I mean it would be defeating by the enemy yelling “mitochondria !!!!”
I have much more faith that AI would come up with a way to bioengineer ordinary mosquitoes than have any chance at a tiny insect sized robot
Agree.
Gonna note that biology is even less grindable than ordinary chemistry/physics/engineering. We'll have advancements in other real-world fields before AI-powered bioengineering becomes A Thing.
Room temperature superconductivity, developed by AI
Poulsen treatment for the rich (https://hyperioncantos.fandom.com/wiki/Poulsen_treatments)
The race is for things that does not exist today.
I think the idea is that AI itself is going to massively accelerate its own development, and AI with real-world competence is coming very soon. At the rate things are going, it wouldn't suprise me at all if we had mass production of AI-designed systems in the next six months, actually.
Sup fellow Alex.
> AI with real-world competence is coming very soon
Disagree, Moravec's Paradox remains undefeated. How long has Elon promised self-driving cars? Or how long have we been seeing humanoid-robot-walking demos? Laundry folding demos?
Let's assume that a magical AI powered robot hand lands tomorrow. How long do you think it'll take to ramp up assembly/production/distribution/sourcing/materials/QA for, say, a million of them? Never mind the legal/integration/maintenance time.
And even once those have landed, and assuming they are ALL put to work on iterating on research testing to build out super-high-capacity-drone-batteries, how long do you think it would take for that to evolve into killer-insect-drones? BTW you should know that even though high capacity silicon carbon batteries exist, they haven't supplanted other Li-ion batteries for a host of reasons. A million things stand between a technology working in a lab and surviving the real world.
And during this entire process, the (geo)political/social/legal process will be churning away, changing the societal landscape in which the killer-robots land. Not to mention that mechanistic interpretability is (likely) somewhat grindable. A lab just needs to dump a billion dollars of compute into it after it declares AGI.
I'm unconvinced we get killer insect drones before we crack mech-interp.
And even once we get killer insect drones, they need to somehow have no kill-switch and then literally exterminate everyone on the globe? Can you see why I have a problem with doomers predicting human extinction within the decade?
I'm putting words in your mouth and not replying to precisely what you said, just the general vibe. LMK if I overstepped.
P.S. we do have mass production of AI designed systems today e.g. https://en.wikipedia.org/wiki/Evolved_antenna There's nuance to this whole thing.
Mom said it was my turn to post this
>Agent-2, more so than previous models, is effectively “online learning,” in that it’s built to never really finish training. Every day, the weights get updated to the latest version, trained on more data generated by the previous version the previous day.
That seems to me to be a natural progression, from discrete models to models that are just continuously improved. Maybe we'll end up with different models with different rates of improvement rather than static differences in performance, and methodologies for that improvement will be the thing we care about. Maybe over time, even benchmark tests will be primarily concerned with that kind of efficiency.
I think a huge debate right now is the relative value of the "frontier" models from Western companies at the cutting edge, vs distilled versions of those models that are good enough and exponentially cheaper coming from China. But a paradigm of 'always training' means an always active, always advancing frontier, which is a stronger moat than a one-off model that's more advanced for a few months.
Kimi K3 isn’t cheaper.
Then don't take my comment as directed toward that example. Charitable interpretation for the win.
One of the most biased claims IMO in AI 2027 is that a huge portion of the geopolitical and existential risk argument is hinged on the notion that China just steals the US frontier weights.
Anthropic documented even more proof of that yesterday:
https://www.anthropic.com/threat-intelligence-report-septemb...
Distillation is not infiltrating systems and exfiltrating weights.
Biased how? China has a long history of corporate espionage.
Everyone knows the Chinese are capable of whatever they put their minds to. But stealing IP to skip some steps is part of the system.
Look what's coming out of China, they are catching up on performance and surpassing the US in efficiency. They're on a different level when it comes to open releases of weights.
I don't, to me the entire premise is a bit flawed at it's core (ASI), and I read it like bad science fiction with China playing the bad guy just a narrative crux so we get to the acceleration timeline and warring nation states.
good thing US is completely clean
Of course it's a spectrum that all advanced countries exist on, but if you think China and the US are on the same end of the spectrum, that's strange.
The US/China divide is one thing, but if I'm asked whether I trust OpenAI or Deepseek, as companies, more, I'm not sure how I'd answer. I suppose my default is to distrust whoever is in the lead, since the lead is power, and power corrupts.
Yeah Sam Altman doesn't inspire trust. Dario and Musk are a bit better in this regard. I think they both say what they really think. Altman... no way.
What country is supposed to be counterexample here, with zero history of corporate espionnage?
That's what a lot of American companies say.
Wait until you hear about Amazon's practices.
I'm probably going to be downvoted for this, but this makes this whole industry look like a bunch of snake oil salesmen and charlatans.
Interesting but it degenerates into sci-fi tropes if you look at the extrapolations. Reminds me of 90s writing about what the Internet was going to do.
The Internet ended up being both more incredible and more mundane than predicted.
I don't like grandiose claims about AI.
How do you know if you are sticking your head in the sand versus being reasonable?
One of the first things I learned in ML is that the model can only learn from the information in X. If X doesn't contain enough information to determine y, no amount of compute can fully recover it.
That's why I'm skeptical of grand claims about AI. Scaling can make models much better, but it can't create information that isn't there. An AI system can be extremely useful without becoming superhuman.
And what do you think about recent developments, for example Navier-Stokes? I always thought the same but now I have doubts, but maybe it is just psyops from openai.
It is speculated that there was context leak from a NYU Professor's chat I think? I contribute to SciPy heavily and I've seen AI shit the bed a few times now. So 13 million lines of lean? Hell no dude!
So often people shy away from making predictions, which is a shame. I always love when people make the attempt and use their imagination.
I am an AI booster and have visibility into a number of these models, and this is ridiculous.
They underestimate AI's existing impact on some job markers and overestimate it's impact in such short a timeframe.
Dario likes this timeline. Good for the IPO valuation.
Sorry, but the notion that creative writing and robotaxi's exist as proof of anything is like saying my child can drive and write; and while true, the measure of that ability is not at the level of the best humans. It's average at best.
All of the early data shows that self-driving cars from Waymo are safer than the average attentive driver. This is not the average driver, since of the 36,000 deaths a year from accidents in the US, about 12,000 of them are due to alcohol or impaired driving.
https://publichealth.jhu.edu/2026/the-safety-data-on-autonom...
As for writing: AI is not nearly as good as the average professional writer, but they are definitely better than the average citizen of the United States, considering that 21% of adults are not functionally literate in the US. AI has no problem writing at the undergrad level.
Safer going under 35 mph
So you've never tried FSD.
The one that requires full-attention from the driver? No.
Depends what your definitions of "requires" and "full" are. On mine there's some nagging on the scale of once every tens to hundreds of seconds if it thinks I'm not paying attention. You're willingly living a more stressful and unsafe life if you're still manually driving your car in 2026.
Sounds less stressful.
sounds about right
Fun to read this again and see actual parallels. The 2030 Takeover section is such a ludicrous leap, however. None of the supply chain infrastructure, energy, or Moravec's Paradox realities are ever addressed. Turn the page and suddenly humanity is largely annihilated with a Corgi-esque human breed kept as pets. How did these robots emerge from utter rhetorical nothingness? Robocalypse impossible? Perhaps not. By 2030? an intellectually embarrassing farce worthy of a facepalm.
Now that's some load-bearing seam if I ever seen one
> It’s informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes.
In other words: bias. Tons and tons of self-congratulatory, glue sniffing bias.
> Hacker News Guidelines
> Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something.
That wasn't a shallow dismissal. It's literally the sentence that renders the rest of the article not credible.
[dead]
At which point do we get to fight in the anti clanker uprising?
I wonder when the people will get that intelligence is not only directed at the external, but only really starts when you look at the internal (joy, pleasure, traumas, taboos, awkwardness, abuse etc.). Look up the word "interoception".