Hacker News Reader

> I still don't understand how cursor is making any money at all.

They don't make any money. They are burning VC money. Anthropic and OpenAI are probably also not making moeny, but Cursor is making "more no money" than others.

No.

Are anthropic and openai making money (including training and infra costs)?

For OpenAI: short answer is no. From what I've seen, their biggest expense is training future models. If they stop that (putting aside the obvious downsides) they'd still be in the hole for a few billion dollars a year.

This is based on what I've read here: https://www.wheresyoured.at/openai-is-a-systemic-risk-to-the... (big AI bear, for what it's worth)

edit: Well, if they shed the other expenses that only really make sense when training future models (research, more data, fewer employees ..) they would be pretty close to break even.

The market for AI-assisted development is exploding and token costs are plummeting all the time. It makes sense for them to subsidise usage to capture market share in the short-term with the expectation that servicing their users will cost them less in the future.

I use Aider with Openrouter and I keep wondering about the pricing of LLMs after providers decide to be profitable. Can we still afford a model which knows Python, Java and how to disrupt snail biology without poisoning mammals?

The answer to that depends on when the VC bubble bursts- if it lasts long enough costs will eventually drop far enough. Pets.com was a .com-boom era joke but today I actually buy my pet-food online and I'm pretty sure nobody is subsidising me doing that.

The.global food market is so heavily subsidized that is almoat impossible that your dog food is not subsidized. Animal feed is even more subsidized.

Commercial animal feed is subsidized. So are some forms of human food in many countries.

Pet food is not subsidized in my country nor the EU. If any countries do subsidize pet food, they are the exception. Maybe the US? Pet food is often manufactured from the waste of other processes, including the human food industry, but that is not a subsidiary.

All EU countries provide income support to farmers.

USA too

That's what I said )) It's not "pet food".

Not "pet food", just "the only input to pet food"

I understand that it is not directly subsidized. However the sources it comes from while are the "waste" of a greater product. That greater product is heavily subsidized.

This also goes to a personal issue that why would you feed your pet a waste product. My dog gets food I cook for him just like myself. There are tons of crock pot recipes online for safe cheap high quality dog food.

That depends, there are specific diets for, say, cats, and it not only needs to be prescribed (to be able to purchase), but it costs a fuckton of money.

<rant>

Think of it like this: imagine if lactose-free or gluten-free food could be bought only with a prescription. Sadly the prices are already high as it is for gluten-free, but I would rather not get into the reasons here. :)

My girlfriend (LA, US) just left 1k USD on 2 visits to the vet with her cat, for some cheap ass antibiotics, and a "specific prescription-only food". Crazy. All that would have been "free" (not quite, but you know) or at a very low cost for humans around here, in Europe. Not the gluten-free food though!

</rant>

Yes. It’s already profitable to run inference at today’s prices. AWS isn’t subsidising you when you buy compute from them. And inference cost is declining steeply.

> The cost of LLM inference has dropped by a factor of 1,000 in 3 years.

— https://a16z.com/llmflation-llm-inference-cost/

AI startups are not profitable because they are throwing vast sums of money at growth and R&D, not because inference is unaffordable.

But, we need a future where unlimited inference, in parallel is profitable. It is not: even less than cloud compute (where it is terrible also), when I buy 500 flimflams for $50/mo, what did I buy exactly? As currently it seems to depend on the position of the moon: one time 10 prompts make what I want, sometimes 100 prompts keep looping over the same issue unable to fix it (like a typescript type issue which takes me 1 seconds, llms, the flagship ones, can easily burn 100 prompts and not fix it). I do very much NOT want to pay for those 100. I see 'vibecoders' aka people who cannot code, burn through all Tokens for the month without having anything working in a single day.

The question that was raised was whether or not current LLM usage will be affordable after providers decide to be profitable.

You are asking if infinite usage is affordable.

A bit of an out of context reply for me to jump in here, but in the abstract, it can be a reasonable question to ask if infinite usage is affordable. Maybe not infinite without constraints… but as an example from the past there are many mobile phone plans that have “infinite” calls and texts for an affordable monthly cost. There would’ve been a time where asking if unlimited calls would be affordable would’ve sounded insane, but now it’s fairly normal.

> It makes sense for them to subsidise usage to capture market share in the short-term with the expectation that servicing their users will cost them less in the future.

Switching costs are zero and software folks are keen to try new things.

The time it would take me to switch IDE and work process and learn the best prompting style and idiosyncrasies of a new model (and do some testing to build confidence) would be half a day, at very least.

That makes the opportunity cost of switching significant.

(I'm not really a coder/programmer/engineer).

Half a day is not much at all

There is no loyalty or lock in though. There is little real uniqueness. And everyone in AI is trying to make everyone else on AI the "commodity complement"

It's like a horse race.

But yeah enjoy the subsidies. It's like the cheap Ubers of yesteryear.

> It's like the cheap Ubers of yesteryear.

Inference cost is plummeting. It’s like the cheap Ubers of yesteryear, if the cost of hiring a driver dropped by a factor of a thousand in the past three years.

You forgot the cost of training (which is zero for Uber drivers, but far from zero for Cursor)

Uber has R&D costs too.

The winners will be those that climb the abstraction ladder. The more sophisticated and useful the abstractions, the more lockin/sticky it will be

That's a really smart observation.

It's hard to add sophisticated abstractions though, because they are all selling text by the pounds (kilos?). So it feels the same as vendor lock for a cucumber seller, doesn't it? The seller can sell you an experience that would lock you in, but aside from it there is no moat since anyone can sell cucumbers.

To try and give examples: an autonomous agent that can integrate with github, read issues, then make pull requests against those issues is a step (or maybe two) above an LLM API (cucumber seller).

It doesn't take much more of a stretch to imagine teams of agents, coordinated by a "programme manager" agent, with "QA agents" working to defined quality metrics, "architect" agents that take initial requirements and break them down into system designs and github issues, and of course the super important "product owner" agent who talks to actual humans and writes initial requirements. Such a "software team system" would be another abstraction level above individual agents like Codex.

This exactly. I built CheepCode to do the first part already, so it can accept tasks through Linear etc and submit PRs in GitHub. It already tests its work headlessly (including with Playwright if it’s web code), and I am almost done with the QA agent :-)

I don't know what the future holds, but I know that this pattern is the 'horseless carriage' of developer automation.

When people talk about how sophisticated hierarchical agent swarms will be built up that perfectly reflect existing human social structures I'm reminded distinctly of all the attempts to build DDD frameworks for modeling software, and then the actual result is that software went in the opposite direction - towards flattening.

As native LLM task completion horizons increase another order of magnitude, so much of this falls out of the window.

My bet is on Bezos. The sheer physical infrastructure is the moat.

Do you mean AWS? They're competing with half a dozen or more hyperscalers now. Cloud infrastructure components are so heavily commoditized now, many of them have open source solutions with compatible API's. (Think Minio)

My second bet is on Google (for general-purpose LLMs in general) - not because of any technical advantage, but because they have a captive audience of large organizations using GSuite that would be happy to just get Gemini on top to satisfy need for AI tools, instead of having to jump through the hoops of finding another provider. Sales, sales, sales.

This is exactly it. Selling the output of a LLM is going to an incredibly cut-throat and low-margin business.

The more interesting, novel, and useful work you wrap the LLM in the more defensible your pricing will be.

That said I think this can describe a lot of agentic code tools - the entire point is that you're not just talking to the raw LLM itself, you're being intermediated by a bunch of useful things that are non-trivial.

I see this with Anthropic most - they seem to have multiple arms in multiple lines of business that go up the abstraction ladder - Claude Code is just one of them. They seem to also be in the customer service automation business as well.

[edit] I think a general trend we're going to see is that "pure" LLM providers are going to try to go up the abstraction ladder as just generating tokens proves unprofitable, colliding immediately with their own customers. There's going to be a LOT of Sherlocking, and the LLM providers are going to have a home field advantage (paying less for inference, existing capability to fine-tune and retrain, and looooooots of VC funding).

Sherlocking won't happen. It requires them to already have the superior customer relationship.

They do need to develop sustainable end-user products, or be purchased by larger players, or liquidate.

This may be old fashioned thinking and the automated loom might come get me but I think traditional software products with enthusiastic customers, some kind of ecosystem will benefit with AI being used.

However they will benefit in a way like they benefit from faster server processors: they still have competition and need to fight to stay relevant.

The customers take a lot of the value (which is good).

While there is a lot of fear around AI and it's founded I do love how no one can really dominate it. And it has Google (new new IBM) on it's toes.

You nailed it. I imagine this is why OpenAI is looking to develop hardware. Right now, I have tabs open for ChatGPT, DeepSeek, and Gemini. I have zero loyalty to any of them. But if I owned a piece of hardware, I immediately am locked into that ecosystem.

Yup. I hope local LLMs and hardware are fast enough a year or two from now when the subsidies run out.

There is a loyalty if they keep winning, if they stop running their competitors will beat them. I don't switch between cursor or windsurf daily, i keep cursor only even if windsurf has some marginal improvement in workflow as i know cursor will have them in short time. No need to switch, But if they stop improving they will get eaten away. They have already taken lot of developer market share away from vscode and vscode copilot.

> so much time is spend waiting for the request to complete (while I work on another part of the app in a different workspace with Cursor).

You can open up to three parallel chat tabs by pressing Cmd+T

https://docs.cursor.com/kbd

Each chat tab is a full Agent by itself!

This won't create race conditions all of them will know the others live accept of commit to directory? or have to wait then hit enter on other tab instantly after?

They handle file locking by themselves. "File1 is currently awaiting edit approval, accept its changes and continue here?"

I still worry this could cause issues even with file locking, if one is reading a file to understand how something is structured and another tab does a refactor on that moments later…

Thanks for this

Google's Jules supposedly can handle that: "An Asynchronous Coding Agent" https://jules.google/

they say they are making hundreds of millions, but they never say how much of it is going to GPU cost. If I had to guess, they are burning everything and far from being profitable

If history has taught us anything, it's that unless full accounting data is released, there is a reason that full accounting data is not being released, and that reason would almost certainly paint the company in a bad light.

"Lines of Agent Edits: 301k"

What does this measurement mean?

1049 / (4 * 8) ~= 32 seconds, on average. Doesn't look like much waiting to me.

The play is probably - Expand and get lock in - Custom fine-tuned models (much cheaper) for increasing # of completions - Enterprise contracts

That works in systems that exhibit economy of scale.

The problem with generative ai workloads: The costs rise linerly with the number of requests, because you need to compute every query.

GPU type and utilization mean that the costs likely rise only logarithmically or sub-linear. If you commit to buying enough inference over long enough, someone can buy a rack of the newest custom inference chips and run them at 100% for you, which may be a lot cheaper per request than doing them on a cpu somewhere.

I disagree tbh. I mean, I accept that new silicon will have better power usage and probably be more efficient in terms of flops/Joule, but there would need to be a major technical breakthrought to get a logarithmic relationship between N requests and inference cost. N requests at P flops, still means I need C x P flops for C x N requests. A not-so-steep linear relationship is still linear.

It hints that there could be real capital deployment limits to a near-term future of artificial intelligence explosion.

The last feature you mentioned on your wish list is literally one of the new features in the major release. I’m hyped

I'm curious what's your usage. How much different it is from Claude based Copilot.

Both are genuine questions.

Do you mean Claude Code? Or something else? As for usage, do you mean the product I am working on? Happy to answer.

I meant you're using Cursor (I assume Claud as the model) and also wondering that you go over monthly usages and then what product/project is that where you find it so useful that you have such a usage.

I have used or rather use Claud with CoPilot and I find it pretty useful but at times it gets stuck in niche areas.