I spent the last two weeks struggling with AI costs going up like crazy. I realized I had been treating AI tokens as if they were free. Like many others, I fell into the trap of using premium, Ferrari-level models to perform supermarket-run tasks. I convinced myself it was necessary because the API costs were negligible.
But the subsidy era is over.
The 2026 price hikes, driven by rising HBM memory costs, new energy taxes, and heavier compliance mandates, are not just inflationary pressures. They are the market correcting years of massive economic malinvestment. For years, VC subsidies allowed us to build systems without genuine demand, treating compute as infinite. From a Misesian perspective, building where there is no market demand for that resource consumption doesn't make us more productive; it makes us poorer by wasting scarce resources.
I started a business with my own savings, not venture capital. I thought this made me immune to "grow at all costs" mentalities using VC money in non-essentials that won't make the product profitable, but I was wrong. I had adopted the habit of using the most powerful, expensive models for every single task because it was easy, convincing myself that was essential to guarantee performance (it was bullshit). It wasn't until the costs started biting into our margins that I realized I wasn't just being wasteful; I was being lazy.
I started experimenting with "Model Tiering". I shifted routine processing to smaller, faster, and cheaper models and reserved flagship models only for complex reasoning where the value-add actually justifies the cost. The result was surprising: the performance difference was negligible for 80% of our tasks, but also the latency dropped and now I get things done even faster.
I’ve come to realize that real wealth is generated when you align resource usage with actual demand. Every time we create demand artificially, fueling growth with VC money, we create malinvestments that the market eventually corrects (and that's never nice).
To stay lean in this environment, I've adopted three constraints:
1. Local Compute Sovereignty: If a task can run on local hardware, it should. Subscribing to an API for something that could be a local asset is now a liability.
2. Prompt Minimalism: I am aggressively stripping filler and redundant context. In an era of high variable COGS and especially with the energy scarcity still not being resolved, every unnecessary token is a direct drain on capital.
3. Prioritize Human Action: AI is a tool, not a strategy. The most valuable builds are those where human insight directs the machine, not where we abdicate strategic judgment to a prompt.
I’m focusing on building a profitable, sustainable business model, not manufacturing fake growth. I’ve learned that the skill of making a product profitable is far more valuable than the skill of raising money to scale a loss-making one. I am finally being forced to account for the actual scarcity of intelligence, and frankly, whether we like it or not, it is making us better humans.
But the subsidy era is over.
The 2026 price hikes, driven by rising HBM memory costs, new energy taxes, and heavier compliance mandates, are not just inflationary pressures. They are the market correcting years of massive economic malinvestment. For years, VC subsidies allowed us to build systems without genuine demand, treating compute as infinite. From a Misesian perspective, building where there is no market demand for that resource consumption doesn't make us more productive; it makes us poorer by wasting scarce resources.
I started a business with my own savings, not venture capital. I thought this made me immune to "grow at all costs" mentalities using VC money in non-essentials that won't make the product profitable, but I was wrong. I had adopted the habit of using the most powerful, expensive models for every single task because it was easy, convincing myself that was essential to guarantee performance (it was bullshit). It wasn't until the costs started biting into our margins that I realized I wasn't just being wasteful; I was being lazy.
I started experimenting with "Model Tiering". I shifted routine processing to smaller, faster, and cheaper models and reserved flagship models only for complex reasoning where the value-add actually justifies the cost. The result was surprising: the performance difference was negligible for 80% of our tasks, but also the latency dropped and now I get things done even faster.
I’ve come to realize that real wealth is generated when you align resource usage with actual demand. Every time we create demand artificially, fueling growth with VC money, we create malinvestments that the market eventually corrects (and that's never nice).
To stay lean in this environment, I've adopted three constraints: 1. Local Compute Sovereignty: If a task can run on local hardware, it should. Subscribing to an API for something that could be a local asset is now a liability. 2. Prompt Minimalism: I am aggressively stripping filler and redundant context. In an era of high variable COGS and especially with the energy scarcity still not being resolved, every unnecessary token is a direct drain on capital. 3. Prioritize Human Action: AI is a tool, not a strategy. The most valuable builds are those where human insight directs the machine, not where we abdicate strategic judgment to a prompt.
I’m focusing on building a profitable, sustainable business model, not manufacturing fake growth. I’ve learned that the skill of making a product profitable is far more valuable than the skill of raising money to scale a loss-making one. I am finally being forced to account for the actual scarcity of intelligence, and frankly, whether we like it or not, it is making us better humans.