The rising cost of AI sovereignty
With datacenters eating up all the fabrication capacity AI sovereignty is becoming out of reach
September 21, 2026
I've added a second DGX Spark to my little basement lab. Together the two of them can run the official full Deepseek v4 Flash weights, nvidia's own NVFP4 Qwen 3.8 Flash Next weights, GLM, etc. Models that are as good as giant frontier models were at the beginning of the year and essentially "good enough" for the vast majority of use cases. These class of models are only going to keep getting better too. Good enough for 6 agents, all in around 300 watts of juice when working full tilt.
This is of course exciting. No matter what happens to AI token costs, network outages, cloud provider failures, geopolitical strife, etc, you can truly own your own AI agents. Companies can truly own the models that power their agentic products, fine tune against them, and choose to retire / upgrade those models when they want, not when a provider requires it. Given how critical AI is becoming to so many businesses products and individual's workflows, this idea of AI sovereignty is critical.
However it is becoming increasingly unaffordable. DGX Sparks will soon be essentially 50% higher than their launch price less than a year ago. 5090s, RTX 6000 Pros, Mac Studios, have all seen similar increases. Datacenter GPUs are even harder to obtain at reasonable prices. The ability to have AI sovereignty is rapidly becoming out of reach for small businesses and individual users.
Now you might be saying that buying hardware never made financial sense, so it getting worse doesn't matter. Truthfully yes, you can buy tokens from providers outside of the big three at prices that are impossible to beat with owned hardware. Even Deepseek directly would take years to break even on the hardware cost of pair of Sparks or a Mac Studio - and third-parties hosting their models are even cheaper. That hardware will be retired before then.
What you lose is privacy* and AI sovereignty. What if the provider you're working with deprecates the model you've built a product against? I've seen it first hand. When you train a workforce to work with an agentic solution they love, even a minor model version bump will throw people. They will notice the differences. Suddenly the way they talk to it doesn't quite work. Sure you can prompt engineer, build a better harness, etc, but you're always chasing.
What if that provider cranks the costs because AI usage has increased globally, because it is only getting started. What if they block your country? What if several block your country? Suddenly AI dependence becomes incredibly risky for a business, and problematic for individuals who have adopted agentic workflows for their daily tasks. This isn't doomerism, the geopolitics of AI are already starting to bubble and you had better believe nation states are acquiring hardware to hedge against an AI blockade.
A lot of folks will say this is a case to not adopt AI, but be realistic: pandora's box is open. Industries, the job market, and daily life are already changing. The solution instead is AI sovereignty. Own the hardware your AI runs on and resist a fully consolidated world where this immense power sits in the hands of very few.
* Outside of OpenAI and Anthropic who sign zero retention deals for HIPAA customers you are essentially giving all of your prompts to a third party who is absolutely going to review them. The days of data being lost in a sea of other data is not really a thing in the era of AI. Reputable token providers will use this data for safety, quality, and model improvements. Maybe a fine trade-off depending on your industry and/or risk profile. Less reputable token providers? Good luck.