A couple of related developments over the last few months changed the AI landscape (seascape?) significantly: industry-wide AI subscription price increases, and the (AI-induced) consumer RAM shortage. Both of these phenomena impacted my wallet in the same negative way.
AI Subscription Prices
In Switching AI Tools, I wrote about how I walked away from my Warp AI subscription after they changed their pricing model in a way that made continuing with them impractical for me. I moved on to GitHub Copilot, which Microsoft owns. GitHub Copilot was working out very well for me: it was actually cheaper, and it provided better functionality than Warp. The base $10/month GitHub Copilot subscription provided me with as much AI usage as I needed. And, since it is owned by Microsoft, GitHub Copilot had very nice integration with the Microsoft developer tools that I use, VS Code and Visual Studio.
As mentioned in Back to Work, after I switched to GitHub Copilot, I started a new part-time job. I had to get up-to-speed on a large codebase, and I leaned on GitHub Copilot pretty hard to help me with that. I thought it did a remarkably good job with summarizing and explaining code to me, even with some low-level, close-to-the-hardware nuts-and-bolts code that I didn’t think it would be very familiar with. I also used it to suggest improvements, review changes, and as a smart search-and-replace tool. But I never ’turned it loose’ on the code or vibe-coded new features with it. As explained in Vibe Coding vs AI Coding, I only used it to do AI coding, and I always inspected the minor changes it actually made for me very carefully. Although I was using AI more heavily in my new job than I had previously in my personal projects, I never exhausted my monthly quota of requests, even when I was on their cheapest subscription.
That happy time did not last very long.
After a month or so, GitHub Copilot changed its pricing model too, and much more drastically than Warp did. They changed from charging per request to charging per token – your subscription bought an allotment of tokens instead of an allotment of requests. And since complicated requests could involve many thousands of tokens (input tokens that you type, context tokens from your code, and output tokens all count), it made some AI usage patterns hugely more expensive. Just a few requests could drain the base $10/month token allotment. The internet erupted in fury over this pricing change, and I have to admit I was a bit upset also. It felt quite unreasonable. I get that Microsoft may have been losing money with the old scheme, and was probably subsidizing it to build up a user base, so I would understand a price increase of say, 50% or even 100%. But this felt more like a price increase of 10x or more.
A few (AI-gathered) links on these price increases:
The New Normal
My first reaction was of course to just change AI providers again. But it quickly became clear that the industry was going through a sea change to higher prices. Other AI providers had already followed the same pattern – they were losing money because they had been subsidizing the AI compute costs, and they all decided to stop doing that, pretty much in unison. At the time that GitHub Copilot changed, there were maybe one or two providers that still had bargain prices, but the writing was on the wall that they would soon raise their prices too. If not because they had to, just because they could.
So I did not switch, at least not yet. I grudgingly stayed put and paid Microsoft more money for the $40/month plan. In the Warp case, I could walk away because I was annoyed and the tool was no longer pulling its weight for me. This time I could not as easily move on because I had the new part-time job and I needed AI to help me with that. I didn’t really have the time or ambition to go shop for a new AI provider and integrate it into my work. And I had become too accustomed to GitHub Copilot’s nice integration with the Microsoft developer tools that I used.
Even with the more expensive subscription, I found I had to limit my AI usage a bit, even though I was hardly an AI ‘power user’ to begin with.
While I am swallowing the higher subscription cost, I am still hoping that local AI on my new MacBook Pro will eventually be good enough that I can drop the subscription. As explained in Upgrade Time, the possibility of running local AI was one of the motivations behind that purchase – more SSD, more RAM, and enough of a machine to at least make local AI a possibility.
Still, I am not getting my hopes up because:
-
Local AI models are improving, but they are still not as capable or as convenient as the best hosted models.
-
Even if the models are good enough, the local tooling ecosystem still has to be smooth enough to use every day. I have not yet found the time to connect the Microsoft tools to local AI, so I don’t know if it will end up being as convenient as GitHub Copilot is now.
-
The most constraining thing for running local AI is still the amount of RAM - even with the much larger RAM in the MacBook Pro that I bought, I can’t run the largest available models locally.
Which leads me to the next unfortunate recent development, The Great RAM Shortage of 2026.
My Precious RAM
So not only am I paying more money to use AI datacenters in the cloud, but those same AI datacenters are forcing consumer RAM and SSD prices to soar.
The short version is this: AI datacenter demand pulled the memory industry toward manufacturing high-margin HBM High Bandwidth Memory instead of consumer RAM chips. Less consumer RAM and SSD chips available meant much higher prices for consumers. Note the graphs in that reference may not look impressive because of the vertical scale used, but most of the consumer RAM prices have gone up by at least 300%.
A few (AI gathered) references for the RAM pricing pressures and AI-memory demand trend:
-
TrendForce: AI Server Demand Continues to Support Memory Prices in 3Q26
-
Tom’s Hardware: Samsung and SK hynix warn AI-driven memory shortages could last until 2027+
I have been considering also upgrading my main Windows PC, but I did not manage to get this done before the price increases, so I’m going to hold off for a while. There are at least some hints that consumer RAM and SSD prices have stabilized, so maybe they will start going back down to more reasonable levels soon.
My Current Stance
My current AI and hardware situation is not ideal, but it is functional.
For my day-to-day AI usage, I got annoyed by Warp, but it turned out that their pricing changes were just on the front edge of an industry-wide trend. I moved to GitHub Copilot, but GitHub Copilot soon became more expensive too. I stayed on GitHub Copilot anyway because I needed it. Now I am trying to make local AI on the MacBook Pro work well enough to become my escape hatch.
If my local-AI experiment works out, I will be delighted. If it does not, I will probably keep paying Microsoft and muttering about it the whole time.
Meanwhile, I did manage to upgrade my MacBook before the RAM price increases really hit, so that was fortunate. But the RAM shortage has caused me to postpone my Windows PC upgrade indefinitely.
