It's the 3rd of June and at work emails are already flying around about the github copilot billing change, which has already stopped some colleagues AI-assisted work - it's affecting internal policy and requiring budgets to be set (= money suddenly needs to be found).
And to be honest our interns are still better than the AI... _and_ they have a conscience.
Probably not far behind you. The June 15 change only meters programmatic use, SDK and CI and third-party agents. Interactive Claude Code is still flat, which is where I live, so my bill stays boring a bit longer.
But it's the third tightening this year, all in the same direction. The flat plan is what's getting carved up. When, not if.
The intern line holds though. Conscience and an off switch, neither ships in the SDK.
I think you touch on really good points here. The issue is AI has to replace humans to match the lofty valuations, investments and ultimately revenue that have been made and have been promised. The current economic equation around AI is based on this. It feels like we are at a turning point when the ‘value’ if AI is increasingly questioned against a tighter lens…feels like we are headed for a correction
That's the real fault line. The money was raised on replacement, cut the headcount, keep the output, and the product does the opposite: it amplifies people, it doesn't remove them. So when the value gets questioned against a tighter lens, what breaks isn't AI, it's the part of the valuation that assumed humans were leaving.
A correction here doesn't mean the tech was fake. It means it was priced as a labor replacement and it's actually a labor multiplier, and those are very different numbers. The multiplier is real. The replacement was the story that raised the round.
Absolutely. I can see the technology persisting but probably not done of the companies who we associate with it today. With a correction I also see future investment in AI being cut after people face been burnt and are more risk adverse. A correction will also impact the whole ecosystem from data centres, semis to private credit and equity. For some reason this inter connectedness reminds me more of 2007/8 than 2000
I'm fine with paying by token usage personally. So is $dayjob as a whole (hough as with you we currently have a monthly plan because Anthropic offer it). But $dayjob doesn't waste tokens gratuitously. We use claude to write code or documents to our specs and that is far faster than us writing them ourselves but no one is being pushed to use AI for everything. As you say treat it like a keen but naive intern and it does fine
The question is whether Anthropic, OpenAI etc. can make money providing services above cost to all of us. I'm not sure whether they can
I suspect the subscriptions are subsidized just as hard, we're being seated at the utility, not sold it. Cheap now, indispensable soon, then the price moves and it's pay or walk, except walking means relearning how to work without the thing you built your week around.
Which answers your question sideways: the labs don't have to be profitable above cost yet. They have to be unleavable first. Profitability comes after the lock, not before it. The flat plan that feels like a deal today is the on-ramp.
Yeah, but you say you are sticking with older models that consume fewer tokens and do just fine. You have also previously posted about running your own private AI instead of using claude or ChatGPT.
If an older model on local hardware turns out to be good enough for a large proportion of tasks then the price Anthropic and ChatGPT can charge is going to be capped at a slight premium to the cost of running your own. That may not be a price that is profitable
Right on the price ceiling, but local has rough edges today. Qwen3.6-27B at Q4 on my M3 Pro 36GB started choking around 128-256K context, it's dense so the KV cache eats memory fast, and Mac runs about a fifth of server throughput. And the models are empty out of the box, you have to set up the prompts and the scaffolding and burn time tuning before it's useful. Wiring it into ForgeCode or Pi is its own job on top.
Solvable though. A year of upgraded subscription roughly buys a 64GB Mac Mini, set it once and you're done. So yes, it caps the price they can charge, just not to zero and not without the setup time.
Separately, Codex at $20 has been good for reviewing my plans, catches stuff I used to flag to Claude by hand. The cheap tier pays off when it's pointed at a real task.
Not that I intend to run a model on GPU-less server, but because it suggests that there's a good number of things that can be usefully tweaked on a local deployment that will significantly improve performance
I’m not a complete devotee of Ed Zitron, but if the numbers he is reporting are even halfway accurate (and I think they are probably more like within 10% or so) then profitability is a pipe-dream.
The Capex costs are simply enormous and require AI as a whole to generate on the order of $100B a year in gross profit to get a positive ROI within the next decade
It's the 3rd of June and at work emails are already flying around about the github copilot billing change, which has already stopped some colleagues AI-assisted work - it's affecting internal policy and requiring budgets to be set (= money suddenly needs to be found).
And to be honest our interns are still better than the AI... _and_ they have a conscience.
Probably not far behind you. The June 15 change only meters programmatic use, SDK and CI and third-party agents. Interactive Claude Code is still flat, which is where I live, so my bill stays boring a bit longer.
But it's the third tightening this year, all in the same direction. The flat plan is what's getting carved up. When, not if.
The intern line holds though. Conscience and an off switch, neither ships in the SDK.
I think you touch on really good points here. The issue is AI has to replace humans to match the lofty valuations, investments and ultimately revenue that have been made and have been promised. The current economic equation around AI is based on this. It feels like we are at a turning point when the ‘value’ if AI is increasingly questioned against a tighter lens…feels like we are headed for a correction
That's the real fault line. The money was raised on replacement, cut the headcount, keep the output, and the product does the opposite: it amplifies people, it doesn't remove them. So when the value gets questioned against a tighter lens, what breaks isn't AI, it's the part of the valuation that assumed humans were leaving.
A correction here doesn't mean the tech was fake. It means it was priced as a labor replacement and it's actually a labor multiplier, and those are very different numbers. The multiplier is real. The replacement was the story that raised the round.
Absolutely. I can see the technology persisting but probably not done of the companies who we associate with it today. With a correction I also see future investment in AI being cut after people face been burnt and are more risk adverse. A correction will also impact the whole ecosystem from data centres, semis to private credit and equity. For some reason this inter connectedness reminds me more of 2007/8 than 2000
I also saw a "usage limits" email today.
I'm fine with paying by token usage personally. So is $dayjob as a whole (hough as with you we currently have a monthly plan because Anthropic offer it). But $dayjob doesn't waste tokens gratuitously. We use claude to write code or documents to our specs and that is far faster than us writing them ourselves but no one is being pushed to use AI for everything. As you say treat it like a keen but naive intern and it does fine
The question is whether Anthropic, OpenAI etc. can make money providing services above cost to all of us. I'm not sure whether they can
I suspect the subscriptions are subsidized just as hard, we're being seated at the utility, not sold it. Cheap now, indispensable soon, then the price moves and it's pay or walk, except walking means relearning how to work without the thing you built your week around.
Which answers your question sideways: the labs don't have to be profitable above cost yet. They have to be unleavable first. Profitability comes after the lock, not before it. The flat plan that feels like a deal today is the on-ramp.
Yeah, but you say you are sticking with older models that consume fewer tokens and do just fine. You have also previously posted about running your own private AI instead of using claude or ChatGPT.
If an older model on local hardware turns out to be good enough for a large proportion of tasks then the price Anthropic and ChatGPT can charge is going to be capped at a slight premium to the cost of running your own. That may not be a price that is profitable
Right on the price ceiling, but local has rough edges today. Qwen3.6-27B at Q4 on my M3 Pro 36GB started choking around 128-256K context, it's dense so the KV cache eats memory fast, and Mac runs about a fifth of server throughput. And the models are empty out of the box, you have to set up the prompts and the scaffolding and burn time tuning before it's useful. Wiring it into ForgeCode or Pi is its own job on top.
Solvable though. A year of upgraded subscription roughly buys a 64GB Mac Mini, set it once and you're done. So yes, it caps the price they can charge, just not to zero and not without the setup time.
Separately, Codex at $20 has been good for reviewing my plans, catches stuff I used to flag to Claude by hand. The cheap tier pays off when it's pointed at a real task.
I haven't had the bandwidth to play with it, or anything similar, but I read this post and the previous two in the series with interest - https://point.free/blog/gemma-4-on-a-2016-xeon/
Not that I intend to run a model on GPU-less server, but because it suggests that there's a good number of things that can be usefully tweaked on a local deployment that will significantly improve performance
I think with the amount of investment the window to profitability is significantly tighter vs other tech initiatives, like AWS
See also https://garymarcus.substack.com/p/ais-black-friday?r=7yrqz&utm_campaign=post-expanded-share&utm_medium=post%20viewer
I’m not a complete devotee of Ed Zitron, but if the numbers he is reporting are even halfway accurate (and I think they are probably more like within 10% or so) then profitability is a pipe-dream.
The Capex costs are simply enormous and require AI as a whole to generate on the order of $100B a year in gross profit to get a positive ROI within the next decade