I think todays latest models would satisfy most users, so it makes me wonder how soon until we can run it on our machines without having to subscribe and with China handing us freebies? I think they have dug well past the point of profitable returns before that happens, so I can see why big tech companies are scrambling to get into bed with US government more than ever as a backup plan when/if their big gamble fails. And now our corrupt government and its 3 letter agencies, they have bargaining power to use this tech and relax privacy efforts and add more backdoors for "safety and security" whereas in the past it was easier for these companies to say no. A match made in hell that we are all roped into now.
> I think todays latest models would satisfy most users, so it makes me wonder how soon until we can run it on our machines without having to subscribe and with China handing us freebies?
Because of the DRAM shortage, not anytime soon. But if you’re willing to pay for inference somewhere else, Deepseek Flash 4.1 is good enough [1]. It might be going to be extra cheaper soon, too [2].
Well once their current majority buyers go bust (potentially without paying for their outlandish orders) what's left of the memory cartel + the Chinese will have to sell memory to real people again.
There are probably cases where it's true though. 4K screens have been declared good enough, computers have come with 16GB of ram for forever now, etc.
It could be less that the current models are good enough forever but that if the prices for proprietary models get too high, people will switch to open or self hosted options. We have not even begun to exhaust the possibilities of the current models.
That quote apocryphal btw. No evidence Gates ever said it. And even if he did, does the fact that it was wrong imply the converse is true? Will everyone always need a better model at ANY cost?
If model development stopped today it would still take decades for the productivity benefits to diffuse through society.
He might not have said it (in public at least), but he became a billionaire by building a whole business model on it, by exploiiting the ecosystem with that limitation. So it doesn’t really matter that he now denies to have said it, actions speak louder than words.
I wouldn't buy my own at that price, but I imagine that that would make it available at cloud providers at prices/speeds that would still be world changing
The vast majority of LLM usage probably would pass evals equally well on a smaller model
I see tons of devs fretting about how only Opus is "smart" enough for the React CRUD web app they are building even though Sonnet and Chinese models are demonstrably equivalent for most real world developer use cases
I think this is why Anthropic is pushing so hard towards the platform side of things. They are trying to build their own source control, document storage, artifacts etc... That's why people will buy it because it becomes an operating system for businesses enabled by AI.
Still some way to go until we run "it" on our own machines. Even with current "mid to high end" hardware you barely reach the same quality and speed with a free LLM compared to what Opus or Astra delivers. In numbers: A 3k Euro Macbook with 32 GB RAM allows you to run a (hard guess only) up to 16 GB Model. Alone! Not speaking about sub agents here. Because you still need memory for your everday work and other programs, cache, context etc.
Well, if there is demand, it will be possible over time. 3D graphics also went from each frame being generated for hours on a compute farm to real time 3D games running on a phone.
Apple is very good at what they do. They will move mountains and invent new technologies to achieve their goals. But the goal is not the technology itself -- it's the experience.
People complained Macs had underpowered hardware. That's because Apple didn't need overpowered hardware to achieve the experience they wanted. Same for iPhone, and how it still has less memory than Android flagships -- by as much as half in some cases, or even more. But they simply don't need more for the experience. (Though recently they have felt the RAMpocalypse)
I believe the same will happen here, and everywhere. Apple will not release a frontier model just because they can. They won't pull ahead just for the sake of it. But there is a high likelihood that they will create much a better experience, that doesn't need a model to be at the frontier.
Of course, they also make considerably powerful machines now. Pushing your hardware to the limit is a significant experience these days, and Apple sees the market for it. Even if they do not create the next frontier model themselves, we may eventually find one running on Apple hardware anyway.
I also think that it is not just user experience, but developer experience that stands to benefit. Apple's developer experience has been rotting since Jobs died, but I still hold out hope that they will pull through. Call me foolish for that, though, if you must.
The real headlines here:
- $450B of non-cancellable cloud commitments over the next 10 years >> the revenue gravy train cannot stop or Anthro is toast.
- 20% of revenue is from just 2 customers >> If their highly concentrated customers get smart and start looking at low cost (meta, grok, google) or open weight (nvidia, inkling, reflection) models it's all bad news for Anthro. This isn't happening now but it's the sort of thing that barely moves until it moves all at once.
As others have mentioned $72b of $80b headline is a non cash (but still relevant) accounting line.
I thought I'd give his podcast a listen a while back because it's good to keep an open mind.
Halfway through one episode, he mentioned his contact in the industry who was telling him about how the agents are actually useless and people have to do all this cleanup work to use the output they generate, and of course Ed promised his audience that's the entire industry and that the collapse was imminent.
That was right at the moment that the models were getting so good that we could "one shot" an issue.
I knew then and there this guy was hot air and I never took him seriously since.
> I knew then and there this guy was hot air and I never took him seriously since.
He's definitely not a software guy, and has been dead wrong on the software use-cases for LLMs. They are really, really bad in most other areas though (because we haven't built decent harnesses for broader "knowledge work").
We'll get there, but it'll take a decade or two. My plan is to ride this wave into retirement/death/immanence of the machine god, whichever comes first.
Grok is 100% not comparable. Like, we have a cursor subscription at work, and when my boss starts hassling me about token spend I honestly prefer to use composer rather than Grok.
It gaslights me on a regular basis and just straight up tells me it's done things, and then when I ask why this thing we fixed isn't working it tells me it hasn't fixed it.
It's actually incredible how bad it is. All the world's compute, and they still can't compete with even the Chinese vendors, never mind the labs and Meta/Google.
Hasn't Apple themselves partnered with Google? Why not switch to one of the Gemini coding offerings? (Yes, they may be inferior for now, but that's not guaranteed to persist.)
> "...included a $72 billion accounting charge reflecting an increase in the estimated value of bonds that could eventually convert into Anthropic shares,"
I think all this means is that Anthropic's valuation went up. Anthropic owes holders of convertible debt[0] a fixed number of their post-IPO shares. Since Anthropic's valuation rose, the book value of the shares they owe to these note-holders went up.
On paper, this is a "loss", because the valuation of their debt increased when they recomputed it. In practice, these notes function like equity.
This is a positive signal that looks bad because the accounting is counterintuitive.
Reuters reported first in September that Anthropic posted a net loss of $42 billion in 2025 and plans to spend $518 billion on cloud, computing and infrastructure obligations in coming years.
$16B revenue and hoping to spend $500B in the coming years.
Its revenue grew 12-fold year to year, and 50% just last quarter. Last quarter it reached profitability if you exclude convertibles etc. And if you won’t it is still well on track to be profitable this/next year.
It is actually faring much better than other tech startups ipo’ing over the last 20 years.
Very absurd notion given that most economists recognize personal computers and the internet have barely impacted labor productivity and global GDP outside of consumption spending
Embedded and industrial computing however have had a profound impact
Could real GDP. Maybe...? I mean people need to dive into the details. How many products/services can be sold...? its not infinite dont be stupid.
Its easy talking about this stuff at a high-level, when you start going into the details you begin to get very uncomfortable. The math right now, simply does not math!
Firms will not wait forever to see what the value of frontier AI is - these scare tactics are wearing off.
> given that most economists recognize personal computers and the internet have barely impacted labor productivity and global GDP outside of consumption
Only that three quarters of the world economy [1]? (The only other component on a global scale is investment.)
Also, the claim is bullshit. Solow said his famous bit about seeing computers everywhere but in national productivity statistics in 1987. Yet the 1990s saw that growth [2]. We saw the paradox again in the 2010s [3]. (Total factor productivity grew 0.8% annually 2000 to 2007, 0.1% '11 to '19 and then roared back to 0.7% and 0.6% in 2024 and 2025, respectively [4].)
There are good arguments against AI increasing productivity to the point that it becomes revenue generating to AI in the near term. Mis-stating the Solow paradox is not one of them.
Robert Gordon pointed out that the 90s productivity gains attributable to IT where overwhelmingly attributable to the manufacturing of computer hardware itself and secondarily the computerization of retail logistics
Daron Acemoglu's prediction is that Total Factor Productivity gains from AI will be a <0.55% increase over the next 10 years
Krugman’s only long-term treatment of the topic in that article, looking “at 10-year rates of growth in labor productivity” concludes “that information technology did indeed produce a bump in economic growth between the mid-1990s and the mid-2000s, but one that was relatively modest and short-lived.”
When he eyeballs short-term (i.e. one-year) TFP growth he fails to see a pattern. But look above. This isn’t a short-term signal. It’s a short-lived bump in a long-term trend. There is proper statistical treatment of this compounding effect that reaches a clear and obvious answer.
I’ll also note the asymmetry in his treatment of the washing machine, which he judges on penetration and supposition (“circa 1920, only about one in five U.S. households had washing machines; by 1970, almost everyone had one, or access to one. circa 1920, only about one in five U.S. households had washing machines; by 1970, almost everyone had one, or access to one. Don’t you think that made a big difference?”) while with the internet we’re meant to visually discern a signal in a noisy and itself-derived source. Use internet and mobile-phone penetration as the washing-machine sign and ask the same rhetorical question.
I’m curious for your Gordon link. You seem to have posted Krugman twice.
> Acemoglu's prediction is that Total Factor Productivity gains from AI will be a <0.55% increase over the next 10 years
That’s fucking huge! Look at annual TFP growth above. 0.1% for a decade. In ‘24 and ‘25 rising to 0.7 and 0.6 percent.
An economy with TFP growing at 0.6% growing 0.55 percentage points faster winds up 12% larger in ten years; the one that starts at 0.1% winds up 7% larger. You’re pooh pooing a number that represents $7 to 13 trillion in surplus annual production after a decade. Integrating under the curve we find both worlds produced about $25 trillion more in that decade, though again, the compounding effect of that higher TFP level will continue compounding for decades.
55 basis points of new TFP is an incredibly bullish cash for AI, and single-handed sustains multi-trillion dollar valuations with a sub-10% take rate of caused growth.
There is no direction this can head in other than inevitable enshittification and price increases that invariably push the majority of users towards open weight models
“The vast majority of this year's net loss, $72 billion, resulted from accounting charges tied to changes in the value of convertible notes and other instruments the company issued to raise funding, the Journal report said.”
Interesting, would love to know the details of this.
Basically the $72B are updating the value of what their convertible notes would be worth if they IPOed today at their current valuation. It's a non cash expense but it does matter because it dilutes all current shareholders and thus lowers earnings per share.
What we need to see is the statement of cash flows.
Whats happening to that cash balance?
People are missing the point IMO - the reinvestment needs is what matters. A) how large are they? B) How will they change over-time? C) How does a slow down in capabilities and closing of the gap by China + price-competition affect their cash flows from operations net of the reinvestment (e.g. will it draw down their cash balance).
You can only hold the perception of increasing capabilities for so long - humans are very adaptive as we are seeing and scare-tactics do not work forever.
Total loss is about 96 billions, they deduced the top (~16 billions) from the bottom to get the net loss (~80 billions). What is "Other Expenses" here? The article says nothing about that. I am not sure to get this right.
Keep throwing money into the wishing well. This has been, since day one, either a gamble or a complete con, depending on how in touch with reality you are.
just use xhigh setting to support them through this AI winter brrrrr.
they are obviously a con ( opinion). look at last slop report. opus 5.5 gets _worse_ when u put higher attention settings. u get banned for using better harnesses etc. etc.
they could do really well. a lot of avid supporters. But i think if they do not change their tactics then openAI, GLM for those who can run it, or others, might start taking even those (pardon my french) religious fans.
Wasn't their previous revenue increase more than twice as big?
That would very much not be great for Anthropic, they need to be accelerating growth to keep valuation, doubly so as interest rates on corporate bonds continue to rise.
Accounting thing in the sense that those are expenses that have not been paid in cash, but with options for the future Anthropic shares.
So they might not be very relevant for the solvency of Anthropic, but that means that if their IPO flops they will also pull down others with them.
Regardless how they have paid for them, the expenses of Anthropic remain many times greater than their revenues. By accepting to not be paid now, their creditors are also gambling on their future success.
> The loss, up from $13 billion a year earlier, was driven by $72 billion in accounting charges tied to changes in the value of financial instruments Anthropic issued to raise money, the report said, citing a draft prospectus for its IPO.
Does that mean they lost $8bn in actual cash? I'm trying to understand how that matches up to their earlier reported claims that Q2 and Q3 were "profitable".
(Oh wait, this $80bn loss is reported for "the first half" of 2026, so it's not including the supposedly profitable Q3.)
The overwhelming majority of the net loss in the first half of this year, $72 billion, resulted from accounting charges linked to changes in the value of convertible notes and other instruments Anthropic has issued, according to the report.
So this is really just capped notes converting. Total nothing burger.
Trying to be sure I get this straight. This would be convertible notes, which are kind of like bonds that were issued earlier on which get returned as some amount of stock/equity, but that amount can change because the value of the company early on can be pretty uncertain. Right?
So the "loss" happens, because they were issued assuming one company valuation before, but the company is worth a lot more now, so the equity exchange on the conversion is now more "expensive", and thus a loss? But ultimately it just means the earlier investors are getting their equity cheaper than what an equivalent would be worth today, right?
Sounds like not too bad of a problem to have if I'm understanding things right. It implies Anthropic is worth more than expected, so they should have potentially received more money back in the day, but the whole point of those convertible notes is the future value is highly speculative. In the end if that's your problem, its because you're potentially crushing it harder than expected at least from a valuation viewpoint.
Anyone please correct me if I'm wrong. I am not in finance, and am definitely looking to learn and understand more.
Lots of people really want to convince the public there is a bubble.
Hmmm...
My advice: if you actually want to know the truth of what is going on, get a general understanding of finance and read the actual financial statements/filings and their disclaimers. It's very easy to write articles like this that cherry pick certain numbers out of context to support a specific narrative. Reuters is in the process of liquidating their reputation in the same way that forbes and business insider and many others publications have.
I think todays latest models would satisfy most users, so it makes me wonder how soon until we can run it on our machines without having to subscribe and with China handing us freebies? I think they have dug well past the point of profitable returns before that happens, so I can see why big tech companies are scrambling to get into bed with US government more than ever as a backup plan when/if their big gamble fails. And now our corrupt government and its 3 letter agencies, they have bargaining power to use this tech and relax privacy efforts and add more backdoors for "safety and security" whereas in the past it was easier for these companies to say no. A match made in hell that we are all roped into now.
> I think todays latest models would satisfy most users, so it makes me wonder how soon until we can run it on our machines without having to subscribe and with China handing us freebies?
Because of the DRAM shortage, not anytime soon. But if you’re willing to pay for inference somewhere else, Deepseek Flash 4.1 is good enough [1]. It might be going to be extra cheaper soon, too [2].
[1]: https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/
[2]: https://news.ycombinator.com/item?id=50163277
Well once their current majority buyers go bust (potentially without paying for their outlandish orders) what's left of the memory cartel + the Chinese will have to sell memory to real people again.
> I think todays latest models would satisfy most users
This reminds me the famous (although unproven) "640K ought to be enough for anybody."
There are probably cases where it's true though. 4K screens have been declared good enough, computers have come with 16GB of ram for forever now, etc.
It could be less that the current models are good enough forever but that if the prices for proprietary models get too high, people will switch to open or self hosted options. We have not even begun to exhaust the possibilities of the current models.
That 16GB forever is 2 years and was walked back in march of this year.
That quote apocryphal btw. No evidence Gates ever said it. And even if he did, does the fact that it was wrong imply the converse is true? Will everyone always need a better model at ANY cost?
If model development stopped today it would still take decades for the productivity benefits to diffuse through society.
He might not have said it (in public at least), but he became a billionaire by building a whole business model on it, by exploiiting the ecosystem with that limitation. So it doesn’t really matter that he now denies to have said it, actions speak louder than words.
I've been thinking since Opus 4.5 came out, that if I had near unlimited usage of that I could do some immensely interesting things.
If Cerebras burned Opus 4.5 to a chip that I/anyone could buy for $5k, that would be worldchanging.
it's probably closer to 50k than 5k
I wouldn't buy my own at that price, but I imagine that that would make it available at cloud providers at prices/speeds that would still be world changing
His counterfactual?
The vast majority of LLM usage probably would pass evals equally well on a smaller model
I see tons of devs fretting about how only Opus is "smart" enough for the React CRUD web app they are building even though Sonnet and Chinese models are demonstrably equivalent for most real world developer use cases
seriously. today's models are good enough because they're... today's models.
Today’s models are not good enough, they need constant supervision to come close to humans.
I think this is why Anthropic is pushing so hard towards the platform side of things. They are trying to build their own source control, document storage, artifacts etc... That's why people will buy it because it becomes an operating system for businesses enabled by AI.
Doing some googling against DeepSeek for which more stats are openly available: We need
~508GB+ Uncompressed memory
~256Gb with 3-bit compression (about 3 bits per weight)
~128Gb with 1-bit compression (1.58 to 2.3 bits per weight)
So uncompressed is still about an order of magnitude out of most home systems' reach, but with compression it's not far off.
When you spend 1k/month of LLM api calls. That doesn’t sound all too unreasonable to have locally in the near future.
Still some way to go until we run "it" on our own machines. Even with current "mid to high end" hardware you barely reach the same quality and speed with a free LLM compared to what Opus or Astra delivers. In numbers: A 3k Euro Macbook with 32 GB RAM allows you to run a (hard guess only) up to 16 GB Model. Alone! Not speaking about sub agents here. Because you still need memory for your everday work and other programs, cache, context etc.
Well, if there is demand, it will be possible over time. 3D graphics also went from each frame being generated for hours on a compute farm to real time 3D games running on a phone.
Regulations can't save them. Everyone will just use open weights.
Apple will do it eventually
Apple isn't doing this. Apple is creating experiences, not frontier models.
Infrastructure.
Apple is a brand.
To have a half eaten apple logo on your computer is the most important thing for many people.
They tried to put that logo on openai’s services but got scammed.
Im sure there was a time nobody thought Apple would design their own CPU's.
Foolish poster.
You misunderstand my comment.
Apple is very good at what they do. They will move mountains and invent new technologies to achieve their goals. But the goal is not the technology itself -- it's the experience.
People complained Macs had underpowered hardware. That's because Apple didn't need overpowered hardware to achieve the experience they wanted. Same for iPhone, and how it still has less memory than Android flagships -- by as much as half in some cases, or even more. But they simply don't need more for the experience. (Though recently they have felt the RAMpocalypse)
I believe the same will happen here, and everywhere. Apple will not release a frontier model just because they can. They won't pull ahead just for the sake of it. But there is a high likelihood that they will create much a better experience, that doesn't need a model to be at the frontier.
Of course, they also make considerably powerful machines now. Pushing your hardware to the limit is a significant experience these days, and Apple sees the market for it. Even if they do not create the next frontier model themselves, we may eventually find one running on Apple hardware anyway.
I also think that it is not just user experience, but developer experience that stands to benefit. Apple's developer experience has been rotting since Jobs died, but I still hold out hope that they will pull through. Call me foolish for that, though, if you must.
The real headlines here: - $450B of non-cancellable cloud commitments over the next 10 years >> the revenue gravy train cannot stop or Anthro is toast.
- 20% of revenue is from just 2 customers >> If their highly concentrated customers get smart and start looking at low cost (meta, grok, google) or open weight (nvidia, inkling, reflection) models it's all bad news for Anthro. This isn't happening now but it's the sort of thing that barely moves until it moves all at once.
As others have mentioned $72b of $80b headline is a non cash (but still relevant) accounting line.
> $450B of non-cancellable cloud commitments over the next 10 years
Where have I heard about this before? Oh right, months ago by Ed Zitron.
"Where have I heard about this before? Oh right, months ago by Ed Zitron."
Please dont say that name here, the witch hunters will out with their forks!!
I thought I'd give his podcast a listen a while back because it's good to keep an open mind.
Halfway through one episode, he mentioned his contact in the industry who was telling him about how the agents are actually useless and people have to do all this cleanup work to use the output they generate, and of course Ed promised his audience that's the entire industry and that the collapse was imminent.
That was right at the moment that the models were getting so good that we could "one shot" an issue.
I knew then and there this guy was hot air and I never took him seriously since.
Call this pitchforks.
> I knew then and there this guy was hot air and I never took him seriously since.
He's definitely not a software guy, and has been dead wrong on the software use-cases for LLMs. They are really, really bad in most other areas though (because we haven't built decent harnesses for broader "knowledge work").
We'll get there, but it'll take a decade or two. My plan is to ride this wave into retirement/death/immanence of the machine god, whichever comes first.
Just remember to come back to this comment in a year or two.
> (meta, grok, google)
Grok is 100% not comparable. Like, we have a cursor subscription at work, and when my boss starts hassling me about token spend I honestly prefer to use composer rather than Grok.
It gaslights me on a regular basis and just straight up tells me it's done things, and then when I ask why this thing we fixed isn't working it tells me it hasn't fixed it.
It's actually incredible how bad it is. All the world's compute, and they still can't compete with even the Chinese vendors, never mind the labs and Meta/Google.
This is a non issue as nvidia would circular finance this thing.
> If their highly concentrated customers get smart and start looking at low cost
The two customers are Apple and FB. The former will never switch and the latter already has.
Hasn't Apple themselves partnered with Google? Why not switch to one of the Gemini coding offerings? (Yes, they may be inferior for now, but that's not guaranteed to persist.)
Also Apple had thrown long term partners and suppliers under the bus many times in the past once there was a more convenient alternative.
> "...included a $72 billion accounting charge reflecting an increase in the estimated value of bonds that could eventually convert into Anthropic shares,"
I think all this means is that Anthropic's valuation went up. Anthropic owes holders of convertible debt[0] a fixed number of their post-IPO shares. Since Anthropic's valuation rose, the book value of the shares they owe to these note-holders went up.
On paper, this is a "loss", because the valuation of their debt increased when they recomputed it. In practice, these notes function like equity.
This is a positive signal that looks bad because the accounting is counterintuitive.
[0] https://en.wikipedia.org/wiki/Convertible_bond
> This is a positive signal that looks bad because the accounting is counterintuitive.
It's not entirely positive, it means that they have a bunch of debt.
But yeah, that number is not reflective of their core business.
If they can pay for training/inference out of revenue then they'll most likely survive. If they can't, then most likely they won't.
Why leave out the time scales? 16.8B revenue for Q3 of 2026 and planning to spend 518B over 10 years (which would average to 13B quarterly)
Its revenue grew 12-fold year to year, and 50% just last quarter. Last quarter it reached profitability if you exclude convertibles etc. And if you won’t it is still well on track to be profitable this/next year.
It is actually faring much better than other tech startups ipo’ing over the last 20 years.
If revenue continues to increase like this we can expect Anthropic to make up 110% of the world's GDP very very soon (Rubs hand together).
Perhaps in US they already signed up everyone, but in Poland I see new people/orgs signing up every day.
Or world GDP to accelerate.
Very absurd notion given that most economists recognize personal computers and the internet have barely impacted labor productivity and global GDP outside of consumption spending
Embedded and industrial computing however have had a profound impact
Not disagreeing with you just curious if you can cite sources for my own research.
Correct nominal GDP will barely move.
Could real GDP. Maybe...? I mean people need to dive into the details. How many products/services can be sold...? its not infinite dont be stupid.
Its easy talking about this stuff at a high-level, when you start going into the details you begin to get very uncomfortable. The math right now, simply does not math!
Firms will not wait forever to see what the value of frontier AI is - these scare tactics are wearing off.
> given that most economists recognize personal computers and the internet have barely impacted labor productivity and global GDP outside of consumption
Only that three quarters of the world economy [1]? (The only other component on a global scale is investment.)
Also, the claim is bullshit. Solow said his famous bit about seeing computers everywhere but in national productivity statistics in 1987. Yet the 1990s saw that growth [2]. We saw the paradox again in the 2010s [3]. (Total factor productivity grew 0.8% annually 2000 to 2007, 0.1% '11 to '19 and then roared back to 0.7% and 0.6% in 2024 and 2025, respectively [4].)
There are good arguments against AI increasing productivity to the point that it becomes revenue generating to AI in the near term. Mis-stating the Solow paradox is not one of them.
[1] https://data.worldbank.org/indicator/NE.CON.TOTL.ZS?end=2025...
[2] https://direct.mit.edu/rest/article-abstract/85/4/793/57428/...
[3] https://en.wikipedia.org/wiki/Productivity_paradox
[4] https://www.conference-board.org/pdf_free/TED-SummaryTables-...
Krugman talks a bit about the subject here
Robert Gordon pointed out that the 90s productivity gains attributable to IT where overwhelmingly attributable to the manufacturing of computer hardware itself and secondarily the computerization of retail logistics
Daron Acemoglu's prediction is that Total Factor Productivity gains from AI will be a <0.55% increase over the next 10 years
https://web.archive.org/web/20230405154054/https://www.nytim... https://web.archive.org/web/20230405154056/https://www.nytim...
> Krugman talks a bit about the subject here
Krugman’s only long-term treatment of the topic in that article, looking “at 10-year rates of growth in labor productivity” concludes “that information technology did indeed produce a bump in economic growth between the mid-1990s and the mid-2000s, but one that was relatively modest and short-lived.”
When he eyeballs short-term (i.e. one-year) TFP growth he fails to see a pattern. But look above. This isn’t a short-term signal. It’s a short-lived bump in a long-term trend. There is proper statistical treatment of this compounding effect that reaches a clear and obvious answer.
I’ll also note the asymmetry in his treatment of the washing machine, which he judges on penetration and supposition (“circa 1920, only about one in five U.S. households had washing machines; by 1970, almost everyone had one, or access to one. circa 1920, only about one in five U.S. households had washing machines; by 1970, almost everyone had one, or access to one. Don’t you think that made a big difference?”) while with the internet we’re meant to visually discern a signal in a noisy and itself-derived source. Use internet and mobile-phone penetration as the washing-machine sign and ask the same rhetorical question.
I’m curious for your Gordon link. You seem to have posted Krugman twice.
> Acemoglu's prediction is that Total Factor Productivity gains from AI will be a <0.55% increase over the next 10 years
That’s fucking huge! Look at annual TFP growth above. 0.1% for a decade. In ‘24 and ‘25 rising to 0.7 and 0.6 percent.
An economy with TFP growing at 0.6% growing 0.55 percentage points faster winds up 12% larger in ten years; the one that starts at 0.1% winds up 7% larger. You’re pooh pooing a number that represents $7 to 13 trillion in surplus annual production after a decade. Integrating under the curve we find both worlds produced about $25 trillion more in that decade, though again, the compounding effect of that higher TFP level will continue compounding for decades.
55 basis points of new TFP is an incredibly bullish cash for AI, and single-handed sustains multi-trillion dollar valuations with a sub-10% take rate of caused growth.
It reached profitability if you exclude all costs (so it isn’t profitable lmao)
There is no direction this can head in other than inevitable enshittification and price increases that invariably push the majority of users towards open weight models
I thought all of this machine learning business was about minimizing the loss.
The loss without accounting adjustments was $8.7 billion for H1.
The other $72 billion was non-cash adjustment.
Revenue was $16.3 billion for H1 2026, up from $1.1 billion for H1 2025.
Not actually that bad.
“The vast majority of this year's net loss, $72 billion, resulted from accounting charges tied to changes in the value of convertible notes and other instruments the company issued to raise funding, the Journal report said.”
Interesting, would love to know the details of this.
Basically the $72B are updating the value of what their convertible notes would be worth if they IPOed today at their current valuation. It's a non cash expense but it does matter because it dilutes all current shareholders and thus lowers earnings per share.
yes it affects book equity
What we need to see is the statement of cash flows.
Whats happening to that cash balance?
People are missing the point IMO - the reinvestment needs is what matters. A) how large are they? B) How will they change over-time? C) How does a slow down in capabilities and closing of the gap by China + price-competition affect their cash flows from operations net of the reinvestment (e.g. will it draw down their cash balance).
You can only hold the perception of increasing capabilities for so long - humans are very adaptive as we are seeing and scare-tactics do not work forever.
This is from the article by the wall street journal:
https://imgur.com/a/XRPOXLa
Total loss is about 96 billions, they deduced the top (~16 billions) from the bottom to get the net loss (~80 billions). What is "Other Expenses" here? The article says nothing about that. I am not sure to get this right.
Keep throwing money into the wishing well. This has been, since day one, either a gamble or a complete con, depending on how in touch with reality you are.
It can be two things!
if the gamble is also con, it collapses to being just a con
just use xhigh setting to support them through this AI winter brrrrr.
they are obviously a con ( opinion). look at last slop report. opus 5.5 gets _worse_ when u put higher attention settings. u get banned for using better harnesses etc. etc.
they could do really well. a lot of avid supporters. But i think if they do not change their tactics then openAI, GLM for those who can run it, or others, might start taking even those (pardon my french) religious fans.
"Anthropic's third-quarter revenue rose 45% to $16.8 billion"
This is great for Anthropic, but who is increasing their AI spending in this climate?
Wasn't their previous revenue increase more than twice as big?
That would very much not be great for Anthropic, they need to be accelerating growth to keep valuation, doubly so as interest rates on corporate bonds continue to rise.
It’s actually not great, that’s a slowdown
The biggest issue is concentration risk.
The corporate partners, who are also often investors, and have an interest in ensuring the world sees genAI as utterly required.
Microsoft, SK Group, Meta, Google, Oracle, etc.
I dont know, with the exception of SK, all these companies have announced reduced AI budgets and/or move to own inference.
Maybe they did it like SpaceX and sold their compute to a competitor and are booking that as sales.
R.I.P. Expected mind blowing move, I am all eyes, all ears. Gotta learn from this one
And 5% of that loss is dedicated for the 'public good'
1792524879 | Anthropic S-1 Shows Red-Hot Growth, Risks and Computing Costs | https://www.wsj.com/tech/ai/anthropic-revenue-growth-ipo-acc... | https://news.ycombinator.com/item?id=50171533 | 1 comment
1792535113 | Anthropic S1 Shows $80B Loss This Year, WSJ Reports | https://www.youtube.com/watch?v=6OnuO6Y5OQY | https://news.ycombinator.com/item?id=50173522 | 1 comment
1792535227 | Anthropic recorded a staggering $80B net loss in the first half of 2026 | https://twitter.com/Polymarket/status/2112634886070919426 | https://news.ycombinator.com/item?id=50173557 | 20 comments
1792537741 | Anthropic Does Not Understand Why Its Draft Bill Was Unacceptable and Troubling | https://thezvi.substack.com/p/anthropic-does-not-understand-... | https://news.ycombinator.com/item?id=50173938 | 1 comment
1792538303 | Anthropic Loses $8.7B, and Reports $80.7B | https://ailately.com/articles/anthropic-80-billion-paper-los... | https://news.ycombinator.com/item?id=50174015 | 0 comments
1792572675 | Anthropic disclosed an $80.7B net loss for the first half of 2026 | https://www.wsj.com/tech/ai/anthropic-revenue-growth-ipo-acc... | https://news.ycombinator.com/item?id=50177795 | 9 comments
1792584875 | Anthropic loses $80B in six months | https://www.telegraph.co.uk/business/2026/10/20/anthropic-lo... | https://news.ycombinator.com/item?id=50179617 | 2 comments
Probably stocks based compensation or something like that.
Important to note that $72bn of that loss is some accounting thing and not really relevant wrt. the solvency of Anthropic
Accounting thing in the sense that those are expenses that have not been paid in cash, but with options for the future Anthropic shares.
So they might not be very relevant for the solvency of Anthropic, but that means that if their IPO flops they will also pull down others with them.
Regardless how they have paid for them, the expenses of Anthropic remain many times greater than their revenues. By accepting to not be paid now, their creditors are also gambling on their future success.
> The loss, up from $13 billion a year earlier, was driven by $72 billion in accounting charges tied to changes in the value of financial instruments Anthropic issued to raise money, the report said, citing a draft prospectus for its IPO.
Does that mean they lost $8bn in actual cash? I'm trying to understand how that matches up to their earlier reported claims that Q2 and Q3 were "profitable".
(Oh wait, this $80bn loss is reported for "the first half" of 2026, so it's not including the supposedly profitable Q3.)
Most probably went to Musk’s GPUs.
Fable 5.5 better be incredible.
From Bloomberg:
The overwhelming majority of the net loss in the first half of this year, $72 billion, resulted from accounting charges linked to changes in the value of convertible notes and other instruments Anthropic has issued, according to the report.
So this is really just capped notes converting. Total nothing burger.
Trying to be sure I get this straight. This would be convertible notes, which are kind of like bonds that were issued earlier on which get returned as some amount of stock/equity, but that amount can change because the value of the company early on can be pretty uncertain. Right?
So the "loss" happens, because they were issued assuming one company valuation before, but the company is worth a lot more now, so the equity exchange on the conversion is now more "expensive", and thus a loss? But ultimately it just means the earlier investors are getting their equity cheaper than what an equivalent would be worth today, right?
Sounds like not too bad of a problem to have if I'm understanding things right. It implies Anthropic is worth more than expected, so they should have potentially received more money back in the day, but the whole point of those convertible notes is the future value is highly speculative. In the end if that's your problem, its because you're potentially crushing it harder than expected at least from a valuation viewpoint.
Anyone please correct me if I'm wrong. I am not in finance, and am definitely looking to learn and understand more.
They said they have $72 BILLION dollars in note depreciation and you believed that?
lol
Anyone can lose a $1, it takes something special to drop $80B.
Lots of people really want to convince the public there is a bubble.
Hmmm...
My advice: if you actually want to know the truth of what is going on, get a general understanding of finance and read the actual financial statements/filings and their disclaimers. It's very easy to write articles like this that cherry pick certain numbers out of context to support a specific narrative. Reuters is in the process of liquidating their reputation in the same way that forbes and business insider and many others publications have.
Do you have evidence that Reuters is misleading? Please prove it, otherwise this is just hearsay
How are you going to find trustworthy financial statements of a company that's private and can cherry pick what they release?
its not a public company, there's no s1 or 10q to read
They cherry pick financial data to make themselves look healthier than they are, which is all you need to know.
Terrible post.