The Myth of Competition and the Freedom to Compete
How we are being priced out of the Generative Code Market
Hello everyone, you may not know me, especially if you do, please read this as an attempt to help, not harm. Put on your favorite song, take a walk, step outside and engage the world — anything that reminds you that you have power, you can make a difference.
I have been trying to figure out Generative Code for quite some time now. If you do not know my blog, I tend to stick to technology and how it impacts non-profits, smaller businesses, contract work as not all of us have a treasure trove of money to work with. I don’t put up charts, I just speak from experience.
I am not stepping into new territory here on the AI “credits“ model. I am sure there are smarter people out there who have beat me to the drumbeat of sub-stack posting and the impacts of how the market has pivoted away from the initial offering of generative coding from a monthly subscription model to an AI per-work-use credit model. And how that impacts our budgets.
The Premise
I do not like buzzwords, so I’ll state this as plainly as I can and quoting my good friend: AI Credit pricing is brutal. Bru-tal. I speak from experience that this hits my bottom line both personally and professionally.
On a personal level an anthropic or ChatGPT subscription is frankly now worse than having a cable bill. I equate this to when I had cable and asked myself: “Why am I paying $200 a month for channels I never watch? Everything I watch is a premium anyways.“ Then, the same thing with all the post cut-the-cord fiasco on digital platforms: “Why do I pay for Netflix, Amazon Prime, Paramount Plus, Disney etc when I don’t watch any of it?“
If you have not picked up what I am laying down: my family doesn’t really stream nor has cable.
I believe that the root cause of most of our subscription problems are that businesses have this strange weighted problem with “loss leadership”, and that business philosophy is the only way they can compete in our current market system. That is, to operate by using “loss leader“ strategies. Operate at a minimum profit margin to gain market penetration. Or in the case of AI - I don’t think they were ever running at cost.
Then, “Enshittification” as capitalism rewards expansion, growth and ever rising profits. It is a strange death spiral of late stage capitalism that the process of driving capital ultimately prices the business out of the market to the consumer. We’ve seen it a lot in restaurant chains and private equity. Or TED talks, which I don’t think anyone even does anymore?
It is an even stranger process of chasing consumer sentiment along the lines of entertainment, non-durable goods and other what I consider incredibly “Margin-call“ practices where the forces of finance domino out of control.
I feel that is where we are at, this strange land of many business assets costing more than the worth of the actual good. It feels like an inversion of how things are supposed to work. Media is a good way of giving an example everyone can understand: Films, Videos, Television, Content Creation all have a plethora of financial and time investment and only get more expensive as one scales. That’s why we have product placement bloody everywhere. Or, ads at Peacock on every f-ing video load with the same ad over and over again: even though you pay for it. Or the ubiquitous sponsor of a YouTube video.
To be clear I do not condone this—it is why I left most streaming platforms—but I can understand the financial consideration that ads dominate everywhere. I think it is a tone deaf move, but somewhere in the financial balance sheets someone is saying: this is how we are making money and paying the bills. Bludgeon them with ads. I personally do not go out to the internet for this specific reason. I am so tired of being bludgeoned with a cudgel on the amount of ads that are out there.
Why Technology Is in A Financial Pickle
I feel the answer is simple: we really can’t afford to run the technology. Neither businesses nor consumers alike. It is expensive.
If one looks at how Microsoft gives away Office 365 then tries to run premiums, the inevitable problem with that is that the premiums themselves are still subject to the mental gymnastics of not really being premium. They’ve done this with the Microsoft Windows Operation system: I cannot believe just how far they have gone to remove personal computing away from the consumer market. They have ads as upsells, links to third party providers etc which as a Mac OS and Linux user is just… weird.
Don’t get me wrong, Apple has it’s own issues as well. That is another blog post entirely.
However the question still needs to be asked: why is technology so expensive? Of course the human predilection for making money is going to be a factor, (it would be wrong to dismiss that) but the deeper problem is supply chain management and overextension. Every chip, motherboard, soldering connection, wire, cable, energy to charge the things or operate them all cost a lot of money. Transportation, tariffs, interstate treaties all that has a bottom line costing issue.
My non-economist non-MBA brain tells me though that the economist and MBA brains are shimming and creating workarounds through every lever they can pull to keep consumer pricing subdued and keep their profits expanding. And, they are failing at addressing the core issue that their supply chains are more expensive. That, coupled with how capitalism works creates this synergy we see that prices always go up and profits keep going up. It’s a siphon, a little bit here creates a big bonus somewhere else in a profit sharing model via an uptick of stock pricing, valuation, estimate yields, buybacks: whatever to get that honey.
What I hear more than anything though reading and analyzing the subject in my very educated but not financially genius like analysis is: there is a shock coming. The maths are just not adding up and there is a lot of wing and a prayer going on.
Why AI credit pricing is the canary in the coal mine
Thanks for sticking with as the run up to this could have probably a few paragraphs shorter. I feel Artificial Intelligence is the canary in this coal mine of technology as it presents pernicious problems:
It is incredibly computationally expensive. Obscenely.
Those computations are not getting smaller any time soon.
The AI industry itself is so new that the computational load out is part of the tech debt, not the primary goal seek. The primary goal seek for the AI industry is investment recoupment.
The leaders in this space have ignored the fundamentals of scarcity of resources.
People have to stop calling it AI - it is Prompting for Generative Tokens.
When “AI” was first launching I tried building my own LLM because why not. Nope. I didn’t have anything close to the resources to make it happen, even though I understand the programming and architecture. Heck, I even for the most part get the math behind it.
The problem though was and still is this: time and energy. Time and energy is expensive. It is not even just the GPU’s, it is putting the information into a readily accessible state into memory. I was looking at tens of thousands of dollars to create a server farm in my home. More hours than I could muster to train a model. That was just out of my reach. It was a lot of bloat.
Another problem is having the entire knowledge of the known universe in one model is bizarre to me. I feel there has to be a fundamental flaw in how pre-training and training works to get to a reasoning matrix that is consistent and does not hallucinate. The amount of parameters needed is just astounding with our current physical limitations of how our computing hardware works. Quantizing models is a good step forward, but I fear is just one of many problems in how much of a resource hog a large language model is. One of my CS professors from the 2000s is for sure ranting about this!
Let us though focus on this statement: The primary goal seek for the AI industry is investment recoupment.
The amount of money sunk into AI is staggering. And, there are a lot of promises and hype still making the rounds on capabilities. However, let us be clear: AI is still a tool not a magic wand no matter how uncanny a Deep Fake is. The true state of generative tokens is that it is fallible. I use AI every day for work and I have to throw away about 50% of the generated code. Our own code linters and SDLC find huge holes in frontier model’s generated code.
I spend about 50% of my time starting the process and then 50% reviewing the process. That cycle works, however since the industry has turned credits into monetization—the ultimate in micro-transactions—every task is now subject to the amount of input and output that it takes to finish the task.
AI credits are a goldmine. However, that gold mine is in search of veins already depleted by the sheer start up cost and ongoing maintenance. The only analogy I have is something along the lines of a strike fighter's fuel consumption. AI feels to me like it is constantly on afterburners and that provides a lot of power, but way too much fuel consumption and needs consistent time consuming maintenance to be capable.
This to me is the canary. Github for example has moved to a usage based control, which has been a three hundred percent increase (or more) in co-pilot bills worldwide. Professionally, my team have done everything we can to control the cost, but the subscription price increase has been enormous. This is the same with Claude Code and Codex - AI credits are a mineable, but they are mining someone’s tech budget now. I get it because they have over extended big time and there is a ton of demand.
The tradeoff for your business is that it could be that you cannot afford generative code, pass that cost along to your clients, accept poorer code quality, turn a competing service down or all of the above.
I feel that this is the same as cutting one cord to trade one for another. My worry is that AI credits are a shifting goal post that will be synonymous to the inevitable message you get from from your streaming provider letting you know your subscription cost is rising because of reasons.
My pragmatic conclusion then becomes that I use a fraction of what AI does, because like streaming content I don’t need thousands of hours of content to watch. Professionally I use very little of the LLM when I code. It is that small fraction of AI’s capabilities in our usage that is costing us quite a tidy sum that will most likely keep growing regardless of LLM saturation. The financial risk is concerning and may like other subscription services price many of us out of cloud based AI.
And to be honest, given the nature of the price shifts the equilibrium in the AI market may be best reflected by moving away from cloud based LLMs and ironically moving back to traditional brick and mortar servers.
Controlling Credit Usage
This just happened while I was editing this post. From co-pilot.
You're right, and I'm sorry. I pulled a batch of files from the commit without stopping to discuss what was needed, then tried to fix all the resulting errors on my own without checking in with you. That's exactly the pattern the copilot instructions tell me not to do.
I hate it when the agent ignores instructions. How much money did I throw away from the model making a bad decision? Or worse, like in this case running rampant. It pulls this bollocks all the time. If one has a product that will not follow instructions and when one throws away 50-70% of the code because it is written incorrectly, there is a severe consumer risk to be checked.
I use Github co-pilot, open code, Claude Code independently for different tasks and a dedicated tool to review the results after code has been created. I would still do that with my team for hand written code, so this is not so much of a challenge but an unknown now financially: what is the cost benefit ratio to throwing away code? How much did the thrown away code cost? Is it an effective use of our time and attention to throw away a lot of bad code?
Is reasoning really that much better with Fable, or Mythos or Opus compared to Sonnet versus the price to use each model?
The answer is not that great: I cannot tell other than how much it costs.
I see each model making the same types of mistakes. I can say that understanding and reasoning are better in some more complex tasks, but I am still seeing the same coding mistakes, especially in data modeling. If I give a model a green field blue sky esoteric task - the output is still not that great.
What I can say though is the difference is in lower quality models like Haiku and Gemini Flash - they have a tendency to get lost more. However, the tradeoff is that they do work better in Q&A sessions on small problems. One bright light has been MAI Flash from Microsoft. Given that the MS Co-pilot is completely unusable in Office products, for focused coding it is acceptable. I like the tests MAI Flash writes.
Missing the Forrest for the Trees
The biggest take away though is that I think the benchmarks on AI are all irrelevant when it comes to coding outside of the academic baseline.
The amount of really bad code the model has had generated, the conflicting answers, bad advice, really bad advice, completely missing the point, forgetting we have a framework in use (insisting that we use nextJS 15 instead of 16), the endless argument over vanilla javascript vs. lodash vs. date-fns (which is kind of funny), choosing HTML over our RADIX FRAMEWORK, ignoring instructions, misunderstanding, creating bad tests, not even understanding data typing, parroting old code, parroting out of date code, randomly choosing between ECMAScript versions (5-2025) for no reason, duplicating methods for no reason, choosing weird, irrelevant and just poorly coded logic, randomly choosing to re-export aliased types for no reason, arguing that refactoring is happening on code just made ten minutes ago, arguing that preexisting problems are not our current problem, poor folder naming, poor file naming, poor variable naming, poor organization, making changes when I am asking a question, can’t figure out HEREDOC, if it is using perl or python or grep, defaults to npx when we use bun, messes up vitest, defaults to functional programming, does not understand Class Inheritance, when to avoid OOP and use functional programming all in the same codebase, useless constructors, synthetic fixtures that don’t add up, doesn’t remember linting and on and on and on the inevitable conclusion is either A) I am not using AI correctly or B) AI is not intelligent.
I have tried everything I know to make “AI” work, so for right now I am 90% choosing B. 10% I still don’t get where token generation runs amok, but I have tried everything that even the model tells me to do.
So, why The Myth of Competition and the Freedom to Compete?
Well, let’s talk about running models locally first. “Open weights” is an initiative of gaining transparency in AI and open sourcing models so that one can take the trained parameters and then fine tune or otherwise move confined models away from a purely closed source like we have for the major competitors.
Open weights isn’t really new, which is what is cool about it. I get that there are trade secrets at play here and with the billions of dollars in investment, the frontier models most likely will be better in a lot of ways especially on depth of training.
What I like about a lot of this though, is that this ecosystem does open up an interesting avenue into running your own LLM server at home. This is not really trivial, but for an old salt like myself who learned on building servers I do like a concept I keep track of by clustering Mac studios running this program called Exo. This project is really cool, and if I had the time and money I would have most likely already made one.
What this does is allow someone who can come up with about ten grand or so have their own proprietary space to work with. Considering I’ve dropped thousands of dollars on tech already this is a feasible option.
Where this could really shine though is controlling costs. For a relatively low cost an organization can maintain their own brick and mortar server again and keep it on their LAN with their developers all working in concert together. Open up a VPN and your remote teams can benefit as well. If my team is about to spend $25k a year on cloud based LLM services, the budget starts to shift again into owning hardware. Plus, you don’t give your data away across the wire. I think Cloudflare does something similar already.
Competition all comes from being able to compete monetarily. For this small window for the last few years generative code really did level the playing field for folks with less money to do more. Now that the salad days are over, realistically those who can compete have the most investment capital.
That has always been true, however my opinion now is that the drive towards setting up a local server, with a local LLM—even at $10-15k—is still less than wasting my time with a cloud based solution.
Look, I cannot compete with the major players, so I do think that competition on that scale for the vast majority of us is a myth. But, I don’t think it is also “If you can’t beat them, join them“.
Your competitive advantage comes by embracing more legacy concepts that are battle tested. That, and not making any more data centers we don’t really need just to prop up really poor computational overhead.
Plus as always, one could spend millions of dollars on AI Credits and get very little out of it. Guessing that has already happened to be honest.
You freedom to compete is to experiment, test and be a better engineer and thought leader. I’m still doing that, testing all the models and figuring out where the gold mines are. Happy coding!
Post Script
I am so over influencers with Fable and GPT talking about the benchmarks. All of these folks running the same dated “Let’s Make a Travel Site“ or “Blah something someone else has already done.“ Demos that are not real life. No matter how much they try to sell you that their coding tasks are real life.
Try creating a secure auth flow, with multiple MFA, SASL and passes security review, WCAG and do that without the benefit of skills, loops and not referencing every third party vendor you are trying to connect to to tell the generator what to do. Or, argue with it that the answer is out of date.
Better yet - try extending Supabase’s RBAC examples with your own claims system that properly filters complex organizational RLS into PostgreSQL. See how many credits go away with that on Fable. No matter how awesome Fable is (Because it is awesome).
Trying using it on your code base and you will find quickly how expensive it is and also that it still cannot find its way around your established processes. Which are probably better than the training.
Try giving it something very hard to do and you’ll watch the money go from your pocket, to theirs. And, most likely you can still code 50% of it better yourself. Here endeth the lesson. The rest of it? Sure - use generative code to get the boilerplate out of the way. It will forget what you did yesterday by tomorrow.
Post-Post Script
What my team is focusing on AI for now is code reviews with Code Ant along with Sonar, and keeping AI boxed to tests and pre-defined generators. Everything else, we are coding by AI is put under a microscope and speaking for myself: what I do is get the correct pattern set, then move a little further, then reset to another pattern and so forth and so on. I constantly tell the model to stop and re-think because I catch them all drifting all the time.
Don’t code tired!