U.S. government locked AI budgets two years ahead, yet racing compute costs that arrive monthly
Congress budgets on two-year cycles, but LLM subscriptions and compute bills change monthly. The real solution isn't cutting costs, but making budgets fungible across models and use cases.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Software engineering costs have actually collapsed, not just gotten cheaper
Koven describes a State Department project extracting insights from roughly 3,000 daily diplomatic cables using NLP—originally requiring a six-person team over six months. Today, he says, two people could do the same work—or better—in weeks. This isn't hyperbole: he has run the same work with both old and new tools and compares them directly. Software engineering costs have truly collapsed; you see it in how government now processes information.
— Barnett KovenValidate diplomatic AI by running it against expert analysts
Koven proposes a concrete evaluation: feed the same prompt to both the model and the actual Russia analyst, then compare conclusions. If they align, both got it right—or both fell into the same mirror imaging (unconsciously assuming the adversary reasons like you do). If they diverge, the divergence itself is the real discovery: it exposes where the model breaks down, or where the human missed something.
— Barnett KovenEntity resolution can expose shell companies trying to evade export controls
Export control enforcement isn't about whether to regulate, but how to spot the shell companies actually evading controls among many legitimate subsidiary structures. Koven describes AFS's approach: entity resolution merges Bloomberg, S&P, and government data into one master corporate relationship graph. From that graph, you can identify a previously unknown shell company, trace it through seven layers of intermediaries, and land on a People's Liberation Army procurement office.
— Barnett KovenGenerative AI wastes money when classical methods would work better
Meta fired people for not burning enough tokens; SemiAnalysis reports an analyst burning $50k/quarter. Koven says enterprises should do the opposite: let proven classical machine learning handle 80–90% (it's certain, it works, it's cheap), and reserve generative and agentic AI only for where classical methods fail. This keeps total cost down, locks determinism where you need it, and—he calls it the secret—gets better results.
— Barnett KovenBudgets that lock two years early can't survive monthly compute bills
U.S. federal budgets, in practice, lock in roughly two years before the money spends—but LLM subscriptions, token consumption, and cloud bills move monthly. The only realistic escape, Koven says, is building fungibility inside the budget: tokens from model A swap with tokens from model B; one use case borrows from another. Roll up budget granularity from individual offices to whole departments. Add FinOps to stay accountable. Stop locking down line items in advance.
— Barnett KovenClassified AI networks favor attackers because defense costs rise exponentially
Once you're in IL6 or higher, the latest model doesn't drop the same day. Real concern isn't compute—it's whether data leaks or whether the adversary has already poisoned the model. Koven is blunt: if he were on the other side, he could predict what U.S. analysts want the model to do, then plant a trap in advance. He frames this as an asymmetry: attack cost rises linearly over time, defense cost rises exponentially. In classified space, the attacker wins by default.
— Barnett KovenPublication bias caused the U.S. government to waste billions on failed strategies
Koven's work in counterinsurgency research: early studies said development aid reduces violence. When he reanalyzed using spatial diffusion models from cancer research, he found the opposite—aid mostly displaced or amplified violence to nearby villages. But he and peers had tried dozens of variants that missed statistical significance and never got published. Other teams duplicated the failed work. The U.S. government kept funding aid blindly. The cost wasn't just dollars; it was military casualties.
— Barnett KovenAbandoned research data holds solutions to today's critical supply chain crises
DOE's Genesis mission tackles a blind spot: research failure data is never kept or shared. Example: beryllium can be extracted from many ores; a century ago, certain sources were abandoned because costs were high. Now, with rivals dominating supply chains, those old experiments might be critical. Tailings once dumped as waste could merit re-extraction. RBAC/ABAC is the technical unlock: role- and attribute-based controls let you open data row by row, column by column, between national labs, universities, and partners without losing control.
— Barnett KovenIn their own words · checked verbatim
the Department of State sends back on order of 3,000 cables between sort of post and sort of headquarters in D.C. every day
Barnett Koven3:05
We pay people in the U.S. government all the time to go really deep on Russia, really deep on China, whoever the adversary de jure is.
Barnett Koven9:33
80 percent of the work that a high end sort of data science team is actually sort of relatively run of the mill data engineering
Barnett Koven16:48
how can we sort of bring in sort of the traditional non generative AI ML that has been tried and true for decades, have that be 80 or 90% of the solution because it works and works well
Barnett Koven22:02
This sort of her sort of most powerful insight was inaction is in and of itself risky.
Barnett Koven27:15
this is a bit of a gun armor race where in the cost of the sort of offense increases linearly over time, where the cost of sort of the defense increases exponentially.
Barnett Koven28:17
And what I ended up finding was that on average, you actually only serve to displace and possibly exacerbate the violence.
Barnett Koven32:28
beryllium is a great example, can be actually made from a number of different types of ore, some of which have not been sort of commercially viable in the past
Barnett Koven37:43
Figures
| State Department daily diplomatic cables | ~3,000/day | 3:05 |
| AFS full-time federal employees served | ~15,000 | 18:54 |
| Accenture global workforce | ~800,000 | 18:54 |
| Genesis mission company partners | 26 | 29:17 |
Glossary
- Mirror imaging
- Unconsciously projecting your own decision-making logic onto an adversary, leading to misreading their true intentions.
- Entity resolution
- Linking and merging records across disparate databases that refer to the same real-world entity.
- RBAC/ABAC
- Role- or attribute-based access control: automatically deciding who sees which data, down to row and column level.
- IL6
- U.S. government intelligence information security level 6, a classification tier above standard government cloud environments.
- FinOps
- Cloud financial operations: tracking and accountability for compute and token spending.
- Genesis mission
- Department of Energy initiative partnering 26 technology companies to advance federal AI capabilities.
How to listen
Anyone tracking how the U.S. government procures and deploys AI, concerned about export control enforcement, or interested in research data sharing and open science: practitioners and policy researchers.
The first 2 minutes—personal background from counterinsurgency PhD research into consulting—is low-signal and skippable.