Clock Speed: The Pentagon Learned to Buy AI Fast. Buying It Right Is the Hard Part.

01Correcting the Meme

First, the ledger, because the viral framing gets both numbers wrong. The FY2026 national-defense topline is not $725 billion — the FY26 NDAA alone is a $925 billion policy bill, and the full national-defense budget runs near the trillion mark. (Our "$725 Billion Question" was about something else entirely: the private-sector data-center buildout that Chinese open-weight models threaten to strand.) And "zero working systems" hasn't been true for years: Project Maven's targeting-support line has been operational since the late 2010s and, per reporting, frontier commercial models now run inside classified environments through fielded systems, generating real operational products. The honest indictment is subtler and worse: the Department has islands of working AI surrounded by an ocean of process — hundreds of thousands of annual contract actions executed by the same overloaded workforce, under the same FAR and DFARS, at the same speed as before. The marquee AI deals got senior-leader attention and white-glove contracting. The question this article cares about is what happens to everything that doesn't.

02The Machine McNamara Built

To see why AI breaks the system, look at when the system was built and for what. The Planning, Programming, Budgeting (and later Execution) system — PPBE — was installed by Robert McNamara in 1961 to bring rational, multi-year discipline to buying aircraft carriers and missile wings: capital assets with decade-long lives whose requirements could be specified in advance. It worked, for that. The modern consequence is a money clock that runs like this: a program office identifies a need, fights it into a service's Program Objective Memorandum roughly two years before the fiscal year in question, survives budget review, waits for Congressional authorization and appropriation (frequently late, under continuing resolutions that freeze new starts), and finally obligates funds — two to three years from idea to first dollar, on a good day. Layer on the requirements bureaucracy (JCIDS), the regulations (FAR plus DFARS), and the security accreditation gauntlet (the Authorization to Operate), and the median major defense program historically runs the better part of a decade from need to fielding.

Now set that against the technology. Frontier AI models turn over roughly every six months; capabilities that define the state of the art at POM-build are two to five generations obsolete by the time the appropriated dollar arrives. A budget justification written around a specific model, benchmark, or vendor is describing a museum piece by execution year. This is not a criticism of anyone in the building — it is a clock-speed mismatch between an annual-appropriations constitutional machine and a technology on a Moore's-law-adjacent curve. Every reform described below is, at bottom, an attempt to gear these two clocks together.

03The Speed Revolution, 2025–26

Give the building its due: the last eighteen months produced the most aggressive assault on that machinery in modern memory, executed almost entirely through the escape hatches Congress had already cut — Other Transaction agreements, commercial pathways, and executive direction.

04What Speed Doesn't Fix

Now the load-bearing section. A contract ceiling is a hunting license, not a fielded capability, and between the two stands everything the 2025–26 reforms only partially touch.

The money clock still runs PPBE time. OTAs spend money faster once it exists; they don't make it exist. The congressionally chartered Commission on PPBE Reform delivered its final recommendations in 2024 — consolidated budget lines, faster reprogramming, more flexible appropriations — and implementation remains piecemeal. Until it lands, every AI portfolio still submits its homework two years early, and every continuing resolution still freezes new starts while model generations turn over.

Color of money strangles software. Appropriations law sorts dollars into research (RDT&E), procurement, and operations & maintenance — categories drawn for hardware. A deployed model that needs continuous retraining, evaluation, and fine-tuning is simultaneously all three, which means program managers burn months lawyering which pot may legally pay for an update the vendor could ship tonight. The single-appropriation software pilots Congress authorized point the right way; they remain pilots.

The ATO wall is cultural, not technical. The January memo directs ATO reciprocity — one organization's security authorization honored by others — which is also what predecessor memos directed. Authorizing officials carry personal accountability for breaches and near-zero reward for speed; until incentives change, "aggressively eliminate blockers" meets the immovable object of a GS-15 who signs the risk. The Barrier Removal Board is the most interesting new tool precisely because it targets this layer monthly, with waiver authority. Watch whether its minutes fill with waivers granted or exceptions defended.

Nobody knows how to test this stuff. The test-and-evaluation enterprise certifies systems by verifying specified behavior against requirements. Generative models are nondeterministic by construction — same input, different outputs — and their failure modes (hallucination, prompt injection, drift) have no MIL-STD. Fielding at three-million desktop scale ahead of a settled T&E doctrine is a bet that operational feedback beats laboratory rigor. It might be the right bet. It is a bet.

Sustainment is the void. Ships get depot availabilities; aircraft get programmed depot maintenance; models get… nothing, yet, in the budget structure. AI degrades in place — data drifts, adversaries adapt, the world changes — and continuous retraining is a sustainment bill no service has historically programmed. The first program to die of unfunded model-maintenance will make this famous. Better to fund it before then.

05The New Failure Mode: Procurement as Policy Battlefield

Move fast enough and you discover a problem the old slow system never surfaced: when the government buys frontier AI as a commercial service, the contract terms become de facto national policy — and the vendors have policies of their own. The collision is no longer hypothetical. Per public reporting and legal commentary: the Department's January strategy asserted an "any lawful use" posture for acquired AI; at least one frontier lab's agreements had, by its account, always excluded certain use cases (mass domestic surveillance and fully autonomous weapons among them); when the lab declined to drop those restrictions, the Department in February 2026 designated it a supply-chain risk — the first frontier AI company so designated — and the company is now suing in the Northern District of California, even as its models reportedly continue operating inside fielded classified systems. However that case resolves, the structural lesson stands: the Department is now dependent on a handful of commercial labs whose usage terms embed value judgments, and it possesses exactly two levers — negotiate or coerce — each with costs. Planes never came with acceptable-use policies. Frontier models do, and a procurement system that treats that as a compliance nuisance rather than a governance question will keep landing in court. (Full disclosure of the obvious: this article was drafted with commercial AI assistance; readers can weigh that as they see fit. The reporting cited is public.)

06The Mirror: How the Other Side Buys

The competitive backdrop — the through-line of this whole series — is that China's acquisition problem is differently shaped. Military-civil fusion collapses the vendor boundary: the PLA doesn't "procure" Kimi or GLM-class capability through arm's-length contracts so much as absorb it through mandated cooperation, state-linked labs, and an open-weight ecosystem that publishes frontier-adjacent models for anyone — including every defense institute — to download and fine-tune. That is precisely the commoditization dynamic from our $725 Billion Question, operating in the military domain: when the model is free and open, the acquisition problem reduces to compute, data, and integration talent. None of this makes the PLA an efficient buyer — its procurement has its own pathologies of corruption, duplication, and inspection-driven metrics, and "absorbed" is not "fielded" there either. But the asymmetry is real: the US pays a market premium and a governance premium for the world's best closed models; China pays an integration cost for good-and-improving open ones. The US system's advantage — genuinely superior frontier capability and a legally accountable industrial base — only cashes out if the machinery in Section 04 can field it faster than the gap closes.

07What "Buying It Right" Would Look Like

The fixes are unglamorous and mostly known; the test is execution. Finish PPBE reform — consolidated AI/software appropriation lines and rapid reprogramming authority, so portfolios can chase the technology inside a budget year. Fund models like fleets: a sustainment account for retraining, evaluation, and drift monitoring, programmed from day one. Measure the right clock: publish time-to-ATO and time-to-field for every AI effort the way the Department publishes depot throughput — what leadership measures monthly is what the Barrier Removal Board will actually remove. Build the T&E doctrine now, with red-teaming and operational evaluation standards for nondeterministic systems, before the first fielded failure writes it in blood and inspector-general reports. Buy outcomes and evaluations, not model names — contracts specified against benchmarked task performance survive model turnover; contracts specified against "Model X v3" fund antiques. And settle the governance question deliberately: decide, as policy, which use-case restrictions the Department will accept from commercial labs and which it won't, instead of litigating it vendor by vendor. Speed was the easy revolution. This is the one that decides whether the trillion dollars buys a fighting advantage or the world's most expensive pilot program portfolio.

The deepest irony of the whole file: the Pentagon spent forty years being told it buys technology too slowly, finally built the muscles to buy at commercial speed — and immediately discovered that speed was never the binding constraint. The binding constraint is that a machine designed to purchase objects is now purchasing something closer to a workforce: capabilities that learn, drift, expire, argue back through their terms of service, and turn over five times per budget cycle. McNamara's clock can be geared to that, but only by rebuilding the gearbox — the money, the accreditation, the testing, the sustainment, the governance — and not just the storefront. The contracts are signed. The desktops are counted. Now comes the part that was always going to be hard.