Kevin Harnischfeger
Operator
I'm a builder fascinated by the concept of private AI infrastructure. Local models, agentic workflows, retrieval systems, and the orchestration that ties them together. I'm an operator and a founder, and I care more about what runs in production than what looks good in a demo.
I run a dedicated home lab for AI compute, built around a DGX Spark cluster and Blackwell 6000 GPUs. It's where I stand up local models, wire them into real tools, and figure out what it really takes to run capable AI on hardware you own.
What drives me is a belief in sovereign intelligence. AI that you control, that runs on your own hardware, and keeps your context yours. These are early days, nobody has a playbook yet, and anyone who claims to know where it's headed is lying to you. That's exactly why I want to be in it now.
BuildPrivateAI is my take on where this is going. I care about systems that can reason over your private context, use real tools, protect sensitive data, and do actual work. Driven by the belief you should own your own intelligence stack.
My goal is to publish my progress and testing, to help serious builders and tech-enthusiasts own the intelligence layer for themselves.
Published notes
Alex Karp Just Described Your Home Lab, at a Thousand Times the Scale
Palantir's Alex Karp went on CNBC this week and tore into the frontier labs, calling their token-metered business model "effing insane" and saying enterprises are "livid" about paying for tokens that create no value while handing their data and "alpha" to third parties. The crux of his argument (that renting all your intelligence from a black box you don't control is a strategic liability) is exactly what the local AI community has been saying and building for the last two to three years. Karp is describing the home lab ethos at a thousand times the scale. Same argument, same architecture, defense budget instead of a spare bedroom.
In Defense of the DGX Spark: A Reality-Grounded Take
NVIDIA marketed the DGX Spark as a "supercomputer on your desk," and taken literally that sets you up to be let down: its 273 GB/s memory bandwidth means slow token generation on dense models, which is why a wave of early buyers returned or resold theirs. But that critique measures the machine against the wrong job. The Spark is a large-memory device, not a fast-dense-model device. It comes alive on Mixture of Experts models, on clustering multiple units over ConnectX-7 to run near-frontier-scale open weights like Qwen3.5-397B at home, and on prefill-heavy agentic workloads where its Blackwell compute chews through long prompts far faster than a Mac Studio.
You Are Bidding Against a Datacenter
Local AI enthusiasts are being quickly displaced from the market due to the monumental nature of price increases we've experienced over the last year. What's expensive to us is nothing but capital expenditure to the big players, and with supply not catching up to current and projected demand, it's clear that the issue of more staggered price increases is only going to get worse. With Apple raising prices on their hardware by as much as 30%+ this last week, it's clear that the market has all but abandoned consumers and tech-enthusiasts.
The Bull Case for Sovereign Intelligence
This month the US government took two frontier AI models offline and forced a third into a gated release. If that does not concern you, it should, because it proves the intelligence most of us rely on can be revoked by people we do not answer to. This piece is about why that matters, and what owning your own compute actually protects you from.
Operated runs
