AI and Machine Learning: Why I Spent 80 Hours a Week Studying It So You Don't Have To
Frontier AI is priced as a loss-leader today and a metered bill tomorrow. For small defense firms, owning the model is the escape — cheaper, and CUI-safe.
People assume the hours I put into machine learning are a hobby. They aren't. They're a hedge — mine, and the hedge I'm building for every small defense business I advise. I do the study so you don't have to, and so you don't get caught on the wrong side of the most expensive bait-and-switch in technology right now.
AI is a wave small business cannot afford to miss. But "not missing it" is not the same as "renting it from someone else on their terms." The way most firms are adopting AI today quietly hands a competitor — the model vendor — a seat inside your cost structure and your data. There's a better way to catch the wave, and it starts with understanding how the pricing you see today is engineered.
The Bait Is Cheap — the Hook Is Coming
Look at what the frontier labs are actually doing. OpenAI, Anthropic, and the rest are burning enormous amounts of cash to put their models in your hands cheaply. A flat monthly subscription is a loss-leader — it is priced to build a habit, not to recover the cost of serving you. That is a deliberate strategy, and it is not charity: the goal is dependence. Get the workflow wired into their model, then move the price toward what the compute actually costs.
The trap isn't the subscription. It's what comes after it. Real production AI — the agents, the retrieval systems, the document pipelines that actually save a small team time — runs on tokens, and tokens are metered. The more useful the tool becomes, the more of them you burn, and the bill scales with your success, not your budget. A subscription is a fixed line item you can plan around. Usage-based token pricing is a variable one that grows exactly when you lean on it hardest.
Now put that against the reality of a small defense firm's monthly burn. You are already carrying the weight — SAM.gov registration, proposal costs, white papers, teaming, and CMMC Level 2 compliance stacked on top of payroll. It is information overload with a price tag attached. You cannot bolt an open-ended, usage-metered AI bill onto that burn rate and call it a strategy. The moment the subsidy ends, the firms who built their edge on someone else's meter are the ones who feel it first.
Own the Server, Own the Model, Own the Tokens
Here is what most people don't realize: you don't have to rent. The open-weight models — the ones you can download and run yourself — are now good enough for the work most small businesses actually need done. Drafting, summarizing, searching your own documents, monitoring a market. And they run on hardware you control.
Once you own the server, you own the model. Once you own the model, you own the tokens — there is no per-token meter, no surprise invoice, no vendor deciding next quarter that your workflow is now worth triple. The economics invert. Instead of a variable cost that grows with your usage, you have a fixed, capital investment that gets cheaper per unit of work the more you use it. That is the difference between renting a cliff edge and owning the ground you stand on.
This is why I spend the 80 hours. Not to chase the newest model announcement — to understand which open-weight models are good enough, what hardware actually runs them, and how to stand the whole thing up so a small business owns its AI instead of leasing it. It is unglamorous, technical work. It is also the difference between an AI capability you control and a monthly bill you don't.
The Compliance Dividend Nobody Mentions
For a defense firm, owning the model isn't only about cost. It's about where your data lives.
The moment you paste a capability statement, a proposal draft, or anything touching Controlled Unclassified Information into a commercial AI service, you've shipped that data off-premises — to a third party, under their terms, to be processed on infrastructure you can't see or attest to. Under CMMC Level 2, that is not a convenience. It's a control you now have to explain.
An on-premises model closes that gap by construction. The data never leaves your environment. There's no third-party processor to assess, no data-handling clause to reconcile, no question about where your CUI went. You get the productivity of AI and a cleaner compliance story — instead of trading one for the other. For a small defense business, that isn't a nice-to-have. It's often the only way to use AI on the work that actually matters without creating a new finding.
The Deeper Skill: Reading the Signal Early
Cheaper, owned, compliant AI is the tool. Here's what you point it at.
In the defense market, the public event is almost always late. The clean announcement — the posted solicitation, the press release, the award — is the end of a process, not the start. The useful work happens earlier, in the messy signal: committee language, budget marks, acquisition guidance, data-rights clauses, vehicle timing, and the quiet reshuffling of agency priorities. Small businesses lose when they wait for the tidy announcement. They win when they learn to read the mess before it resolves.
That reading is exactly the kind of continuous, document-heavy monitoring that AI is built for — and exactly the kind of workload that would bankrupt you on a metered token plan and expose your CUI on a commercial one. Own the model, and you can run that monitoring around the clock, across your own sensitive material, at a fixed cost. The tool and the strategy are the same story: the firms that win are the ones who read the signal early and own the machine that reads it.
Why You Can't Build This on a Home Computer
Here's the honest part, and the part I want to be careful about. Downloading an open-weight model is step one, not the finish line. A blank model is a brilliant intern with no memory of your world — fast, capable, and useless on a defense pursuit until you give it something to reason about.
What turns a raw model into analysis is what you feed it and how long you shape it. The system I run has ingested more than 50 terabytes of data and over 100 million documents — budget justifications, contract histories, policy, regulation, filings, market records — and I've put more than 100,000 hours into shaping the models to reason the way I reason about an opportunity. There is no magic button. There is no prompt you can paste into a frontier chatbot that conjures analysis grounded in 50 terabytes and 100 million documents you don't have. Prompt engineering rearranges words; it cannot manufacture a corpus or the judgment that reads it.
And the judgment is the other half. These models are trained to weigh a signal the way I learned to — from policy work inside the Pentagon, from time on Capitol Hill watching language quietly become money, from years flying C-17s and living the logistics and supply-chain reality most briefings only describe, and from a father-in-law who built more than one small business from nothing and taught me what a burn rate feels like from the inside.
I'll say the immodest part plainly, then take the air back out of it: I don't think there's another model out there that reasons about a small defense pursuit the way this one does. Not because I'm smarter than the frontier labs — they have more compute and more brilliant people than I ever will. It's that I've spent years pointing a system at the one problem they will never prioritize: yours. That isn't something a subscription rents you, and it isn't something a competitor prompts their way into overnight.
The Risk
The risk is waiting until the market has already priced the signal in. By the time the clean announcement arrives, your competitors have already adjusted their teaming conversations, their compliance posture, their pipeline assumptions, and their proof points. And the second risk compounds the first: building your new AI edge on rented infrastructure, so that the day the subsidy ends or the token price moves, your advantage — and your data — belong to someone else.
The Monday Move
Do two things this week.
Pick one pursuit and map the decision chain now. One customer, one vehicle, one timing window. Who shapes the requirement, who influences the funding, who owns the contract path, and what proof they'd need before the solicitation ever appears. That turns a saved link into capture work.
Then price your AI honestly. Whatever AI you're using or planning to use, ask the one question the vendors would rather you skip: what does this cost me at 10x the usage, and where does my data go when I run it? If the honest answer is "an open-ended bill and off-premises CUI," you've found your next move.
The Signal
AI is not the thing you subscribe to. It's the thing you own. The small defense firms that come out ahead won't be the ones with the flashiest model — they'll be the ones who read the market signal early and own the machine that reads it, at a fixed cost, on their own compliant ground. Rent the wave and you're exposed on price and on data. Own it, and it compounds in your favor.
That's the work I do the 80 hours for. If you'd rather not spend them yourself, that's exactly where DoD Industry Advisor comes in.
DoD Industry Advisor
DoD Industry Advisor helps small defense teams turn policy, budget, compliance, and market signals into earlier capture decisions — and helps them build AI capability they own instead of renting. Mike Komorous brings defense policy, acquisition, contracting, aviation, and hands-on AI/on-premises machine-learning experience to that work.