Are AI Subscription Models Sustainable?

Originally published at: Are AI Subscription Models Sustainable? - TidBITS

When it comes to AI, what should be available for free? We’re accustomed to getting many core Internet services, such as search and email, for free, but running anything at scale requires massive investment and operating costs. To give you a sense of the scale, OpenAI recently reported 700 million active weekly users of ChatGPT, four times the number accessing the service at the end of last year.

You can use ChatGPT, Claude, Gemini, and others to an extent without paying. But all of them come with a $20-per-month paid mode that offers better models, higher or no limits, and additional features. (For power users who need even more capabilities, there are premium tiers ranging from $100 to $250 per month that provide enhanced features, priority access, and significantly higher usage limits.) AI companies are willing to provide sufficient capabilities to showcase what their chatbots can do, but they must attract enough users to cover their infrastructure costs and move toward a sustainable business model.

The Evolution of Internet Business Models

This freemium model, although relatively common now, is historically quite unusual in its direct connection between revenue and perceived value to users. In the early days of the Internet, websites like Yahoo and Google instead followed in the footsteps of more magazine-like sites in supporting themselves with advertising.

(Here’s where I point out that TidBITS pioneered advertising on the Internet in 1992, before websites were even a thing (see “TidBITS Sponsorship Program,” 20 July 1992). The idea was too obvious for me to claim significant influence on the evolution of the Internet, but I’m still sorry that Internet advertising has become so prevalent, trashy, and prone to abuse.)

I’d argue that this early reliance on advertising stemmed from the value proposition not being high enough to support monthly subscription fees, the desire of companies to focus on attracting users over initial profit, and the lack of a reliable payment system infrastructure at the necessary scale. Whatever the reasons, the advertising model has dominated mainstream Internet services for decades.

But it’s much harder to display sensible ads alongside chatbot conversations, and the value proposition is much higher. Many people consider $20 per month a reasonable fee to access the benefits of a powerful chatbot, particularly now that chatbots can search the Web for current information that goes beyond whatever ended up in their training data.

App-Based AI Subscription Challenges

Many people are also happy to pay for AI-powered features in other apps. For instance, the popular Mac launcher Raycast can supercharge its capabilities with AI for $10 or $20 per month (the higher tier offers advanced models), the Lex.page online word processor integrates an AI editorial assistant for $18 per month, and Notion AI adds an assistant that helps with writing, analysis, and more in Notion’s $20 per month Business plan. These fees are necessary because the developers of Raycast, Lex, and Notion all have to pay the model providers—such as OpenAI, Anthropic, and Google—for API access to their models. In essence, they’re repackaging the models’ capabilities in a new form and passing on the costs.

Although it may seem that charging for AI functionality is an indication of a functional business model, much of what these apps provide is convenience and context. They’re offering custom interfaces, fine-tuning prompts, and leveraging local context, but at the base level, the models are doing the heavy lifting. That leads to a few challenges that may impact users:

  • Subscription stacking: In the short term, users may accumulate multiple $20 monthly subscriptions, resulting in significant ongoing costs. That’s already an issue with app subscriptions, but most of those are well under $20 per month.
  • Constant feature justification: These apps must continuously persuade users that their additional functionality warrants another subscription when pasting into a chatbot would produce similar results. Plus, many of these AI-enhanced apps overlap in functionality, particularly in terms of writing features, making it more challenging to justify multiple subscriptions.
  • Competition from model providers: The model providers have a significant cost advantage if they want to compete directly. My experience with Gemini in Google Docs is that it isn’t nearly as useful as Lex, but Google has the resources to change that if it wants to.
  • Pricing vulnerability: Third-party apps are at the mercy of model providers’ pricing decisions. If OpenAI or Anthropic were to raise their API fees, apps would need to either absorb the costs (threatening their viability) or pass them on to users (potentially losing customers).

There are two other approaches supported by some apps, such as Raycast and DEVONthink:

  • API access: You can pay for direct API access to a particular chatbot, which you connect to a third-party app using an API key. Instead of a monthly subscription fee, you pay for each prompt. It might take a month or two of usage to discover whether API access is more or less expensive than the subscription.
  • Local models: If you have a sufficiently beefy Mac with Apple silicon, you could use Ollama to install and run local models. Local models require significant disk space and memory, and likely won’t be as powerful as those that run in the cloud. Again, testing would be required to determine their utility, but there would be no cost to using them.

How Apple Could Fit into the Picture

Interestingly, these challenges point to a huge opportunity for Apple. Imagine if Apple Intelligence were built on an LLM good enough to compete with ChatGPT, Claude, and Gemini, and that could perform Web searches like Perplexity. Enabling Apple developers to leverage such capabilities could significantly enhance apps for the iPhone, iPad, Mac, Apple Watch, Vision Pro, and even Apple TV. If Apple Intelligence were sufficiently compelling and broadly adopted by developers, Apple could even use it to boost Services revenue by offering it as a separate subscription that would unlock AI features across many apps throughout the ecosystem.

Apple isn’t the only company with this opportunity. Google and Microsoft have already built impressive AI capabilities and are working to extend them more broadly across their own ecosystems. Meanwhile, companies like OpenAI (with a rumored browser), Anthropic (which just started talking about a Claude browser extension), and Perplexity (with its beta Comet browser) are expanding beyond being mere model providers by building agentic browsers that turn the Web itself into their platform.

Given this intense competition in AI, it’s concerning that Apple seems to be rearranging deck chairs with Liquid Glass instead of prioritizing the development or acquisition of a top-notch LLM for its next-generation operating systems under Apple Intelligence. It’s not that Liquid Glass precludes work on a competitive version of Apple Intelligence, but the former is happening, and the latter isn’t shipping.

1 Like

Another option, for those who want free AI (and more important, privacy) is that you can build and run your own copy locally. Most of the popular LLMs are available as open source or have open source versions.

The Ollama site provides lots of different models that you can download and run. Instructions and links are on their GitHub page.

Of course, you need suitable hardware. The model-names all have a suffix that identifies the size of the model. According to the GitHub instructions, a “7B” model (7 billion nodes) requires at least 8GB of free RAM to run. A larger “13B” model requires 16GB of free RAM. A “33B” model requires 32GB of free RAM.

But there are also several smaller models that you could run on your Mac. For example:

  • Google’s Gemma3 model (I think this is the open source model that Gemini is based on) has a few small sizes:
    • 270M: The model is one of the smallest: 292 MB, so it can probably run on a system with “only” 512 MB free RAM.
    • 1B: This model is 815 MB, and should probably run on a system with 1 GB of free RAM.
    • 4B: This model is 3.3 GB, and should probably run if you have 4 GB of free RAM.
  • Meta’s llama 3.2 is also relatively small:
    • 1B: The model is 1.3 GB and can probably run with 1.5-2 GB free.
    • 3B: The model is 2.0 GB and can probably run with 2.5 GB free
  • Compared with the much larger llama 3.1:
    • 1.8B: The model is 4.9 GB and will probably require 5-6 GB free RAM
    • 70B: The model is 43 GB, and will require a computer with 64 GB
    • 405B: This model is 244 GB. You’re going to need a computer with 256 GB RAM. This means a pretty high-end Mac (maxed-out M3 Ultra Mac Studio).

Those small models are small enough you might even be able to run them on a modern Raspberry Pi.

I haven’t actually tried any of these on my computer, but if you want to play with LLMs and don’t trust someone else’s cloud-based service, this might be a viable option. Or at minimum, something fun to play with.

5 Likes

I was annoyed enough at how Microsoft and Google increased pricing on Office and Workspace respectively, justufying it using AI features that people apparently weren’t willing to pay for voluntarily in sufficient numbers.

As an aside, it’s interesting to think about how much we spend on subscriptions for things like broadband networking, mobile phones, and streaming services that weren’t significant expenditures or even in existence in relatively recent living memory.

The one thing that bugs me in your article here, Adam, is your implication that Apple is working on Liquid Glass to the exclusion or detriment of AI. The only way this would be the case would be if Apple is dumping funding into Liquid Glass that could be going to other things, such as Apple Intelligence, and there’s little indication that they are.

Liquid Glass is almost completely handled under the auspices of the interface design team. This team would be touching AI only at the top-most level, and that’s not where the issues are right now. Apple Intelligence is another team altogether, and acquiring or leasing a third party’s LLM wouldn’t be handled by developers or designers at all until the deal is on paper. The work going into Liquid Glass isn’t what’s holding up their AI efforts.

Now, Apple is definitely focusing attention on Liquid Glass instead of Apple Intelligence, but at the current juncture, that’s not necessarily a bad thing. Unlike most of the tech world, Apple likes to keep its developments under wraps until they believe they’re ready. So focusing attention (and criticism) on this new interface design paradigm keeps outsiders from digging too deeply into whatever the current progress on developing or acquiring better AI is. Yes, it’s a distraction, but the distraction is for us, not the people in the company.

1 Like

Now that I think about it, while there are many science fiction stories and films that feature AI assistants, I can’t think of a single one there’s mention of a subscription for the AI service. It it just assumed that everyone or the main character has an AI assistant.

…with one exception: Her showed Theodore (Joaquin Phoenix) buying an operating system upgrade, which came with the new AI.

1 Like

Yeah, I had that thought at one point while writing and forgot about while massaging the text. It’s certainly not the case that Liquid Glass work is preventing Apple Intelligence work, but it is the case that we’re getting new operating systems with almost no new Apple Intelligence features worth mentioning (a bit of Visual Intelligence). Let me see if I can work that caveat back in…

1 Like

Yes, but in the past, we had subscriptions for other things.

Growing up, we subscribed to a daily newspaper ($0.25 per daily issue, $1.25 for Sunday). This came to $2.75 per week or $143 per year. Those 1980 dollars are equivalent to $560 in today’s dollars.

We also subscribed to magazines. I had 3-4 magazine subscriptions that cost about $10-20 per year. Other family members had their own subscriptions. Overall, we probably paid around $15 for each of 8 subscriptions, or $120 per year. Those 1980 dollars are $470 in today’s dollars.

In other words, the expense hasn’t changed (and in my case, has gone down) over the past 45 years. What has changed is what we’re spending the money on.

I’ve always assumed that the software came with the robot. But you’re right that in reality (as we understand it today), the owner would be paying a subscription fee for frequent software updates and those unable/unwilling to pay would end up with an inferior robot or a bricked device.

4 Likes

Wondering when Apple will play its cards with AI, with whatever third party provider, until its own is up to speed. Thats kinda been the assumed plan but not seeing it so far.

I’m also wondering if folks will have any loyalty to an AI platform, what the churn is like. Are folks trying out different ones? I took a years subscription to Perplexity Pro. I wonder what I’ll do when renewing comes around.

This is a very charitable interpretation of Apple’s actions. I admire your optimism.

I even mostly agree. I just worry about how many things they are doing lately that seem to be complete distractions from what is supposed to be their core competencies.

And I keep in mind that the only reason anyone even talks about Apple and AI is that they have embarrassed themselves with Siri for a decade, and even moreso recently.

It’s hard for me to believe they have a handle on even the strategy they want to take, given their public statements and missteps over the past year plus.

3 posts were split to a new topic: What are Apple’s core competencies and should the company focus on them?

It’s a good question, but I suspect people will be pretty loyal because of the “vibes.” People were unhappy with GPT-5 because it felt different than GPT-4o; imagine how unhappy they’ll be with a completely different chatbot purely because it will feel different.

1 Like

Looking at the Big Picture, this is an interesting tack in the business of Internet. I don’t use any of the ‘AI’ products but it find it interesting that they are offering paid service so early in their development.

(Do the ‘AI’ services have advertising? are they offering paid variations on top of same rubbish Google et al do?)

I wonder how things would have been different if Google and other search engines had had paid tiers instead of developing this ‘advertising/surveillance’ model.

Recently I tried the trial period of Kagi search service but didn’t find their results much better than DuckDuckGo, which has been my default for a while now. I suppose I should look at it again and maybe try paying for it for a while and see…

Now THIS is a tremendous idea for a Take Control book: Take Control of Your Local AI Model

What do you say, Joe Kissell?

1 Like

11 posts were split to a new topic: Down on AI at the moment

I have an M1 Macbook Air with 8 GB of ram and 256 GB of SSD running the developer beta (3 or 4) of Tahoe and I used the suggestion from Gary, who hosts MacMost.com, to create a chatbot using Apple Intelligence on device . I was fairly impressed with the results.

I played around with it giving it questions in calculus , algebra, cultural questions (who was Seinfeld’s neighbor - answer: Kramer ) and physics. Only errors were names of actors in Severance, a tricky integral and unit conversion (kw to hp - it got watt to hp right). It also confused the names of several equations used in relativistic quantum mechanics. On the math problems, it gave the steps that it used to solve it. This is a very impressive start for an on-device AI!

I am somewhat cool on AI using remote servers, mainly for privacy reasons, but am very interested in local, on-device approaches which should get better as Apple’s already impressive chips improve even more.

There is no support (yet) for Apple models on-device, but Beyond Better runs locally and supports local (on-device) models via Ollama. With the ‘local-only’ mode combined with Ollama models, Beyond Better can operate completely isolated and free of the cloud.

I believe that Tahoe allows on-device Apple models, though accessing them needs a work around. I haven’t looked at other on-device models but was fairly impressed by Apple’s on my 8 GB machine. So, this will be officially available very soon.

It’s a buyer’s market for Apple. This time Apple has got it right and the pundits are living in la la land. OpenAI is a very risky bet as AI is its only play whereas most of the other biggies have many other sources of revenue.

But Apple is getting the hardware all wrong. It could compete very well for self-hosting AI except it charges way too much for RAM and SSD. It should dramatically lower the price of RAM and make the SSD user installable. An AMD-based MiniPC with Strix Halo Ryzen AI Max+ 395 and Radeon 8060S with 128GB built in RAM and user installable SSD up to 8TB costs half the price of a similarly equipped MacStudio. And USB 4 v2 (Thunderbolt 5) is coming any day. In addition, AMD is getting ready to max out the NPU for AI and perhaps even produce NPU boards. This has the potential to dramatically lower the energy requirements for certain AI use cases.

Apple could dramatically increase its AI hardware sales if it – you know – began to “think differently” again.