The authors document a systematic reconfiguration of defense procurement, citing Anduril's takeover of the IVAS contract, Palantir's Maven Smart System deal, Scale AI's Thunderforge, and OpenAI's policy reversal on military use. They argue that while legacy primes like Lockheed and RTX still hold platform contracts, the decisive software layer — which determines what those platforms do — has been ceded to Palo Alto, El Segundo, and Denver.
By submitting the Watson Institute paper to Hacker News and driving it to 163 points, this submitter amplifies the paper's central thesis that Silicon Valley is quietly displacing traditional defense contractors in the most consequential parts of the procurement stack.
The authors highlight how Founders Fund, a16z's American Dynamism practice, and Sequoia have funneled billions into defense-adjacent startups while former Pentagon officials rotate into these companies and executives rotate onto DoD, DIU, and In-Q-Tel advisory boards. They frame this personnel and capital circulation as the mechanism enabling more than 300 venture-backed defense-tech firms to emerge in just a few years.
The authors are careful to note that Lockheed, RTX, and Northrop retain the major platform contracts, so this is not a wholesale displacement of the traditional complex. Instead, the shift is that the software layer — increasingly the decisive part of any weapons system — has migrated to Silicon Valley, meaning developer labor now sits at the heart of modern warfighting.
The Watson Institute at Brown University published a Costs of War paper — authored by researchers Roberto González and Heidi Peltier — arguing that the American military-industrial complex is undergoing its most significant restructuring since the Cold War, and that the new center of gravity is not Bethesda or Arlington but Palo Alto, El Segundo, and Denver. The paper walks through the money trail: Anduril's $22B valuation and its takeover of the Army's IVAS mixed-reality contract from Microsoft; Palantir's $480M Maven Smart System deal and its expanded Army enterprise agreement; Scale AI's Thunderforge program with the Pentagon's Chief Digital and AI Office; OpenAI's newly minted defense partnerships after quietly deleting the "military and warfare" clause from its usage policy in January 2024; and the $9B Joint Warfighting Cloud Capability contract split between AWS, Microsoft, Google, and Oracle.
The authors also document the lobbying and personnel side. Founders Fund, Andreessen Horowitz's American Dynamism practice, and Sequoia have poured billions into a cluster of "defense-adjacent" startups — Shield AI, Epirus, Saronic, Vannevar Labs, Hadrian, Saildrone. Former Pentagon officials rotate into these companies; company executives rotate into advisory boards at DoD, the Defense Innovation Unit, and In-Q-Tel. The paper counts more than 300 venture-backed defense-tech firms active in 2024–2025, up from a handful at the start of the decade.
The framing is deliberately provocative: the authors call this a "reconfiguration, not a replacement" of the traditional complex. Lockheed, RTX, and Northrop still hold the platform contracts. But the software layer — the part that increasingly decides what the platforms do, when, and against whom — has been ceded to Silicon Valley. And the software layer is where developer labor lives.
Defense procurement used to be slow, cost-plus, and structurally hostile to the kind of iterative development that dominates modern software. The old primes were organized around it: giant program offices, decade-long timelines, hardware bill-of-materials thinking. What changed is not that DoD suddenly loves agile — it's that DoD stopped being able to buy the AI, autonomy, and cloud capabilities it needs from anyone else. The paper cites the 2023 National Defense Industrial Strategy explicitly acknowledging this dependency. Once you accept that the Pentagon cannot build a modern targeting stack, an autonomous drone swarm controller, or a multi-domain fusion system without commercial AI labs, the leverage inverts. The commercial side sets the price and the pace.
The paper's sharpest observation is that this creates a novel kind of dual-use problem. The traditional complex separated defense work from civilian work by company, by campus, by clearance level; the new complex runs the same models, on the same clouds, with the same engineers, and just changes the tenant. GPT-4 class models trained on public web data are the same models being fine-tuned for military intelligence summarization. The Kubernetes cluster running your consumer product is architecturally identical to the one running JWCC. Anthropic's Claude, per the company's own announcements, is now available to US intelligence and defense customers through AWS GovCloud. This isn't Lockheed building a bomb factory in the desert — it's an API key with a different billing address.
On the labor side, the paper documents that the Google Maven walkout of 2018 has largely lost. Microsoft employees protested the HoloLens IVAS contract; it went forward and grew. OpenAI's policy change happened with essentially no internal reporting. Palantir, Anduril, and Scale openly recruit engineers on the pitch that defense work is patriotic and technically hard. The authors argue — and the numbers back them — that the marginal software engineer in 2025 is much more willing to work on military contracts than the marginal software engineer in 2018 was, and that the compensation premium is now meaningful enough to reshape hiring at big-tech competitors.
The procurement rule fight is the least visible but possibly most important piece. The paper documents active lobbying by the American Dynamism cohort to expand Other Transaction Authority, raise the Small Business Innovation Research ceilings, and carve out software-specific acquisition pathways that bypass the Federal Acquisition Regulation. If those changes pass, the DoD's software spend can move at commercial speed — which is exactly what allows a Series C company to displace a $100B prime.
Three practical implications for people who ship code.
First, your dependency graph has quietly become a defense supply chain. If your production stack touches OpenAI, Anthropic, AWS, Azure, Google Cloud, Palantir Foundry, or Databricks, you are running on infrastructure whose largest growth customer segment is now the US government's national-security apparatus. That is not a moral claim — it is a reliability and roadmap claim. Feature priorities, latency SLAs, region availability, and export-control postures are increasingly being set with a federal customer in mind. Expect more FedRAMP-shaped constraints leaking into commercial tiers, more "sovereign" region requirements, and more contractual language about end-use.
Second, the ethics conversation has moved from the CEO to the individual contributor. The old model — a company either takes defense work or it doesn't, and employees choose accordingly — is breaking. In the new model, the same platform team supports both a consumer chatbot and an intelligence-community deployment, and the person writing the eval harness may not know which tenant a given fine-tune belongs to. If you care about where your work ends up, you now have to ask the question at code-review time, not at offer-letter time, because the disclosure signal has shifted from company-level to contract-level and most contracts are not public.
Third, the career-optionality math has flipped. Defense-tech was, until recently, a specialty that closed doors on your resume. Now Anduril, Palantir, and Scale are among the fastest-growing engineering employers in the country, with compensation competitive with FAANG, and their alumni are being aggressively recruited by every AI infrastructure company. If you were writing off the sector, the paper is a useful nudge to at least understand the landscape, because the recruiters certainly do.
The Brown paper is not a call to arms in either direction — it is an argument that the transformation has already happened and the public conversation is running years behind. The likely 2026 story is not whether Silicon Valley enters the defense market; it is which of the new entrants survives the first real budget contraction, which acquisition-reform bills pass, and whether any of the big model labs draw a public line about specific military use cases the way DeepMind briefly tried to. For working engineers, the honest read is that neutrality is getting harder to maintain by accident. The infrastructure you rely on has a customer list now, and that customer list includes the Pentagon. Decide what you think about that on purpose, or the default will decide for you.
Serious question: should companies not participate in contracts from their country's defense department? Or is the argument that it's only US companies which shouldn't, because the US is uniquely evil?
Transforming? You guys do know where google got it's start right? SV has been DOD funded from the start.
It's up to us, as tech workers to push back against this. I quit my job at Microsoft due to their complicity with Israeli war crimes and oppression (and certainly the US government was involved too). It's tough to leave a high paying job, but I feel good about it every single day. We don&#
some interesting tidbits from the longer report:>Keyhole, a small San Francisco-based company that developed software for creating three-dimensional models of the earth’s surface [was] provided seed funding in 2003 [by a CIA-backed VC firm, In-Q-Tel], and within two weeks, military and intelligen
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I do think that for those of us who got into tech in the last couple of decades, the history can be rather obscured, but from the Fairchild semiconductor days, the military was a main customer, right?I think there's a real question: would the world be a better place had Fairchild and peers e.g.