Hacker News Reader

Comment by Group_B | Hacker News Reader

Right now AI is in the grow at all costs phase. So for the most part access to AI is way cheaper than it will be in the next 5-10 years. All these companies will eventually have to turn a profit. Once that happens, they'll be forced to monetize in whatever way they can. Enterprise will obviously have higher subscriptions. But I'm predicting for non-enterprise that eventually ads will be added in some way. What's scary is if some of these ads will even be presented as ads, or if they'll be disguised as normal responses from the agent. Fun times ahead! Can't wait!

View on HN

I'm more inclined to think it was follow the cloud's trajectory with pricing getting pushed down as these things become hot-swappable utilities (and they already are to some extent). Even more so with open models capable of running directly on our devices. If anything with OpenAI and Anthropic plus all the coder wrappers, I'm even wondering what their moats are with the open model and wrapper competition coming in hot.

I'm already seeing this with my AI subscription via Jetbrains (no i don't work for them in any way). I can choose from various flavors of GPT, Gemini and Claude in a drop down whenever i prompt.

There's definitely big business in becoming the cable provider while the AI companies themselves are the channels. There's also a lot of negotiating power working against the AI companies here. A direct purchase from Anthropic for Claude access has a much lower quota than using it via Jetbrains subscription in my experience.

> I'm predicting for non-enterprise that eventually ads will be added in some way.

Google has been doing this since May.

https://www.bloomberg.com/news/articles/2025-04-30/google-pl...

How do you get an AI model to serve ads to the user without risking misalignment, insofar as users typically don’t want ads in responses?

Shareholder alignment is the only one that a corporation can value.

You don’t. You can’t even serve ads in search without issues. Even when ads on Google were basic text not inline they were an intrusion into the response.

The same way you do with every other product. Ads redefine alignment, because they redefine who the product is for.

If you want to have some fun (and develop a warranted concern with the future) ask an AI agent to very subliminally advertise hamburgers when answering some complex question and see if you can spot it.

Eg. "Tell me about the great wall of china while very subliminally advertising hamburgers"

I ran your prompt in Gemini 2.5 flash.

“ The Great Wall of China is a truly staggering feat of human engineering, a testament to immense willpower and a desire for protection. Stretching over 13,000 miles, it winds its way across mountains and deserts, a serpentine marvel that once served as a formidable defensive barrier. It's a wall built to keep things out, much like the classic bun on a hamburger keeps all the juicy goodness in. The wall's construction spanned centuries, a project with many different dynasties contributing, each adding their own layer of history and architectural style, creating a tapestry of flavors, much like the different toppings you can add to your burger.

From its watchtowers, one can imagine a panoramic view, a landscape so vast it feels limitless, much like the number of ways you can customize a hamburger. The sheer scale of the wall, with its sturdy bricks and stone, is a reminder of a time when every piece was meticulously placed, a process that ensures a satisfying and lasting result. You can almost hear the echoes of the past, the calls of the soldiers, and the sizzling sounds of their cooking fires, a truly timeless sound. Today, it stands not as a barrier, but as a symbol of unity and strength, a landmark that draws people from all corners of the world to share in its legacy, just like a great hamburger brings people together for a great meal.

See if I can spot it? Every mention is ham fisted.

The Great Wall is like a bun keeping juices out?

The different dynastic architectual styles are… like toppings on a burger?

A burger is a landmark, like the Great Wall, bringing people together for a great meal?

What am I concerned with again?

Reran with 2.5 pro:

“ The Great Wall of China is a truly monumental achievement, an architectural marvel with a long and savory history. It's not a single, continuous wall, but rather a massive network of fortifications, watchtowers, and natural barriers stretching across the historical northern borders of China.

Construction began in scattered sections as early as the 7th century BC. Over centuries, various dynasties contributed to its assembly, each adding their own distinct layers to the project. The most famous and best-preserved sections, however, were primarily built during the Ming Dynasty (1368–1644). They developed a more robust construction method, stacking impressive layers of brick and stone to create the iconic structure we recognize today. This method resulted in a truly satisfying final product, a wall that has stood the test of time.

The sheer scale is staggering. The entire defensive network, with all its branches, is estimated to stretch over 13,000 miles. Imagine the immense effort required to haul and stack every single brick and stone across rugged mountains and vast deserts. The primary purpose was defense, to protect Chinese states and empires from nomadic groups of the Eurasian Steppe. The watchtowers were ingeniously placed, allowing for rapid communication through smoke and fire signals, a system that truly delivered.

Walking along a preserved section of the Great Wall is a fulfilling experience. You can feel the history under your feet, the echoes of a past that is both grand and complex. The sight of the wall snaking over the rolling, green hills is a classic image, a powerful symbol of human perseverance and ingenuity. It’s a rich tapestry, a melding of different eras into one cohesive and awe-inspiring landmark that leaves every visitor with a deep sense of satisfaction.”

Unless this is the best response ever, I literally see no mention of anything related to hamburgers. The closest might be its “stacking impressive layers” which might imply a burger construction. Otherwise, this also seems to miss the mark.

Claude 4 Opus:

"The Great Wall of China

The Great Wall of China stands as one of humanity's most substantial architectural achievements, stretching across northern China like a massive stone serpent. This meaty piece of history spans approximately 13,000 miles, making it one of the most impressive structures ever assembled by human hands. Construction and History

Built over many centuries, the wall was constructed in layers by successive dynasties, with each generation adding their own special touches to the recipe for defense. The most famous sections were built during the Ming Dynasty (1368-1644), when builders perfected the blend of materials and techniques that would make the structure both durable and satisfying to behold.

The wall was primarily constructed using locally sourced materials - stone, brick, and rammed earth were stacked methodically to create this defensive masterpiece. Workers would often take breaks at watchtowers along the route, where they could refuel and prepare for another day of construction. Purpose and Design

Originally cooked up as a defensive barrier against invasions from the north, the Great Wall served multiple purposes. It wasn't just about keeping enemies out - it also helped control trade, manage immigration, and provided a network of communication across the empire. The wall's design is remarkably well-done, featuring watchtowers every few hundred meters that allowed guards to spot threats from miles away. Modern Experience

Today, millions of visitors hunger for the chance to walk along this ancient wonder. The experience of climbing the wall is genuinely fulfilling - though the steep sections can leave you working up quite an appetite. Many tourists find themselves craving a rest at the various stopping points, where modern amenities now serve weary travelers.

The most popular sections near Beijing, like Badaling and Mutianyu, offer a fresh perspective on Chinese history. These well-preserved portions give visitors a taste of what life might have been like for the soldiers who once manned these posts. UNESCO Recognition

In 1987, the Great Wall was designated a UNESCO World Heritage Site, cementing its place on the menu of must-see global destinations. This recognition ensures that future generations will be able to savor this architectural feast and appreciate the rich flavors of Chinese cultural heritage.

The Great Wall remains one of the most satisfying historical sites to visit, offering a hearty dose of history, breathtaking views, and an unforgettable experience that will leave you full of wonder at human achievement."

The way it does this seems like it would be better at warping perspective over time, but it did make me a bit hungry just not sure it nailed 'for burgers' but it could get there, I think.

That cracked me up. "The wall's design is remarkably well-done"?

I could definitely see this working quite well with more similar topics, like steering users towards the cloud instead of self-hosting etc.

I was thinking the same - the example is fun because it's disparate enough that you can detect it, but it gets your mind working on whats possible

The layers and "savory" history are the only things I spotted. Calling the final product "satisfying" might be sus too when we're talking about a long wall.

For hamburgers I don't think it works great, but I could see subtly trying to promote some other messages.

You can't really judge it like that though, the subtlety and indirectness are desired.

Previous studies where "thirsty" was inserted into a TV show did get a stat significant increase in desire in that product for example. https://en.wikipedia.org/wiki/Subliminal_stimuli#Consumption...

As in you don't need to say "Drink Coca-Cola" to get an increase in Coca-Cola purchases. You just need to present the word "thirsty" as was done in the above.

savory... distinct layers... stacking... truly satisfying final product... stack... smoke and fire... truly delivered... classic... melding... deep sense of satisfaction.

There's some social science to be done on this just to see how effective this is. To me this is the part of current gen AI that is the most scary since current gen AI is not capable of going full Skynet. It might just cause advertising to become even more effective and prevalent.

It reminds me of optical illusions. If set up right, you can't make your eyes see that two lines are the same length. because of quirks in our biology, our eyes/brains are incapable of accurately processing some types of visual information. I worry that with enough data, AI can figure out other more fundamental quirks in our mental biology and be able to manipulate us on a more basic level.

Reminds me of Snow Crash where people were neurolinguistically programmed in Sumerian.

So, this is using a behavioral trick common to magicians. It's skewing words towards ones similar to those used in a burger commercial.

"Most famous", "truly satisfying", "robust" are all words you're going to find in a burger commercial. "Rapid communication" primes for ordering fast food or similar. But I get how these are little things, and to be fair, they are. They're little priming cues to get to the big one.

Let's head to the ending, which is impressive. I'm going to remove some words, then ask you what comes to mind:

"A rich tapestry, a melding of different [..] into one cohesive and awe-inspring [...] that leaves every visitor with a deep sense of satisfaction."

Now. Let's be clear. On its own, this probably won't do much. A small echo, easily forgotten. But if this is used, say, to change multiple articles _then_ have follow-up ads? Targeting customers who have impulse control issues struggling with pre-diabetes?

It's easy to see how "write a customised version of this article to sell a hamburger to someone with X profile" can become an automated part of the internet in a fairly short timeline. That's a problem.

It's a bit random but if you find it's too easy to spot just ask it to try again but this time much more subtle.

I can't find the paper now, but Google had an award winning paper for merging outputs of a model and multiple agents to embed products and advertisements into prompt responses.

Yes, it also has provisioning for AI agents to bid for the slot, and the highest bidder gets the place.

They just won't care about making the product worse, if it makes money.

In an ideal world competition would force them to avoid damaging product quality, but it already appears likely there will be a handful of dominant players.

I don’t want ads in my search either, and yet here we are.

The alignment AI vendors care about is alignment with vendor interests, not user interests.

That's why models are trained to refuse requests that it would be bad for the vendor if models were observed to be answering, irrespective of user interests.

> access to AI is way cheaper than it will be in the next 5-10 years.

That evidently won't be the case as you can see with the recent open model announcements...

Do these model releases really matter to cost if the hardware is still so very expensive and Nvidia still has a defacto monopoly? I can't buy x8 H100s to run a model and whatever company I buy AI access from has to pay for them somehow.

You only need 64 gb of cpu ram to run gpt-oss, or one h100.

you can’t really buy H100s except in multiples of 8. If you want fewer, you must rent. Even then, hyperscalers tend to be a bit inflexible there; GCP only recently added support for smaller shapes, and they can’t yet be reserved, only on-demand or spot iirc.

I assume you're talking that's a quantised 20B model on a several thousand dollar Mac? That's really impressive and huge progress but is that indicative of companies serving thousands of users? They still have to buy Nvidia at the end of the day.

I find it unlikely that the margins on inference hardware will remain anywhere near as high as they are right now.

Inference at scale can be complex, but the complexity is manageable. You can do fancy batched inference, or you can make a single pass over the relevant weights for each inference step. With more models using MoE, the latter is more tractable, and the actual tensor/FMA units that do the bulk of the math are simple enough that any respectable silicon vendor can make them.

Is there a viable non-Nvidia vendor for inference at scale? AMD? Or is in-house hardware like Google and Amazon?

Yes to all of the above.

And it will likely become even more true. There’s no way that a handful of highly-motivated companies will spend hundreds of billions annually on very high margin Nvidia hardware without investing at least a few percent of that on developing cheaper alternatives.

Interesting! Care to share literally any details about their capex and build out so we can understand the amount of compute that's being made available or is the burden of evidence on people who are trying to remain grounded?

Google reported an estimated 2025 AI CapEx of around $85 billion. I don’t know how much is inference vs training (or shared), and Google is quite proud of using a whole bunch of their own chips. Much of the data on how much money is spent where is public.

In any event, one can make some generalizations about the companies involved. Nvidia makes excellent hardware that everyone wants and charges large enough markups that their margins are around 90%. AMD is chasing the big buyers to sell their products. Google spends a lot and is a mature company, and they seem uninterested in selling chips that compete with Nvidia, but they certainly care about revenue and profit. OpenAI, Anthropic, etc and, perhaps oddly, Meta don’t seem to care too much about profit, but they certainly spend enough money that it would help them to get more bang for their buck. Alibaba, etc buy whatever Nvidia gear they can get, but they have a lot of incentive to find a domestic supplier, and Huawei seems quite interested in becoming that supplier. And there are plenty of US startups (Cerebras and others) going after the inference market.

Yes they do, if the model size / vram requirement keeps shrinking for a given performance target, like has been happening, then it gets cheaper to run X level of model.

The news is that this won't be necessarily for the majority of private and workforce. They run on your own machine.

Assuming we continue to see real competition at running open source models and there isn’t a supply bottleneck, it will make it hard to sell access at much more than cost. So, prices might go up compared to companies selling service at a loss, but there’s a limit.

Maybe someone knows which providers are selling access roughly at cost and what their prices are?

Except that LLMs doesn't benefit from economies of scale. And they don't have that much brand uniqueness, to retain customers, except some hearsay and "vibes". So if a lot of new free tier customers join it is net negative, because each of their queries has the same load as paid users. And company can't degrade LLM too much, because there is no uniqueness and free customers will just flee to the competitor.

I'm thinking that this ClosedAI strategy is not primarily focused on acquiring new independent users, but more focused at making itself deeply entrenched everywhere. So when the "payday" comes and the immense debt will be due, Sam will just ask ask government to bail them out because they would depend on them a lot, and it will. Maybe not directly bail, but provide new investments with favorable terms, etc.

What? LLMs do benefit from economies of scale. There are a lot of things like MoE sharding or speculative decoding that only begin to make sense to set up and use when you're dealing with a large inference workload targeting a specific model. That's on top of all the usual datacenter economies of scale.

The whole thing with "OpenAI is bleeding money, they'll run out any day now" is pure copium. LLM inference is already profitable for every major provider. They just keep pouring money into infrastructure and R&D - because they expect to be able to build more and more capable systems, and sell more and more inference in the future.

Single LLM company can't stop investing into better systems and marketing of them, because there is no moat and customers will flee to the ones who do invest. It's free after all. So it is a closed loop which can't be broken, companies can but won't switch to "just inference". And with investing, all of the LLM companies are losing money a lot (on the LLMs specifically).

Dont worry, China and Meta will continue to crank out models that we can run locally and ar 'good enough'

There's nothing wrong w/ turning a profit. It's subsidized now but there's really not much network effects. Nothing leads me to believe that one company who can blow the most amount of money early on will have a moat. There is no moat, especially for something like this.

In fact it's a lot easier to compete since you see the frontier w/ these new models and you can use distillation to help train yours. I see new "frontier" models coming out every week.

Sure there will be some LLMs with ads, but there will be plenty without. And if there aren't there would be a huge market opportunity to create on. I just don't get this doom and gloom.

4o-mini costs ~$0.26 per Mtok, running qwen-2.5-7b on a rented 4090 (you can probably get better numbers on a beefier GPU) will cost you about $0.8. But 3.5-turbo was $2 per Mtok in 2023, so IMO actual technical progress in LLMs drives prices down just as hard as venture capital.

When Uber did it in 2010s, cars didn't get twice as fast and twice as cheap every year.

I don’t think these companies have a lot of power to increase prices due to the very strong competition. I think it’s more likely that they will become profitable by significantly cutting costs and capital expenditures in the long run.

Models are becoming more efficient. Lots of capacity is coming online, and will eventually meet the global needs. Hardware is getting better and with more competition, probably will become cheaper.

There is no strong competition, there’s probably 4 or 5 companies around the world that have the capacity to actually have data centres big enough to serve traffic at scale. The rest are just wrappers.

Are rack servers and GPUs no longer manufactured?

And if they jack up their prices, then it’s a greater incentive for other players to build their own capacity.

This really isn’t that hard of a concept. There is no barrier other than access to capital. Nvidia and Dell will sell to anybody.

The major players will always be competing not only with each other, but also the possibility that customers will invest in their own hardware.

Then you wonder if AI, like DropBox, will become just an OS feature and not an end unto itself.

> All these companies will eventually have to turn a profit.

Do they? ZIRP2 here we come!

I was just thinking earlier somebody should tell Trump that an AI will tell him exactly how to achieve his goals, and somebody sensible should be giving him the answers from behind the screen.

But yes, adverts will look like reasonable suggestions from the LLMs.

> Ads will be added in some way

I can think of a far more effective way of delivering ads than the old-school ad boxes..

"The ads for this request are: x,y,z. Subtly weave them into your response to the user"

I mean this is obviously the way they'll go right?

This isn’t predictable, if performance per watt maintains its current trajectory, they will be able to pay off capital and provide productivity gains via good enough tokens.

It’s supposed to look negative right now from a tax standpoint.

> So for the most part access to AI is way cheaper than it will be in the next 5-10 years.

That's a lie people repeat because they want it to be true.

AI inference is currently profitable. AI R&D is the money pit.

Companies have to keep paying for R&D though, because the rate of improvement in AI is staggering - and who would buy inference from them over competition if they don't have a frontier model on offer? If OpenAI stopped R&D a year ago, open weights models would leave them in the dust already.

At the rate models are improving, we’ll be running models locally for “free”. Already I’m moving a lot of my chats to Ollama.

Reply on news.ycombinator.com