NextFin News - Lockheed Martin is bringing OpenAI's frontier models onto the F-35 program, enlisting the ChatGPT maker to work through complex math and physics problems tied to the stealth fighter's advanced sensor capabilities. The disclosure, made by Sarah Hiza, Lockheed's senior vice president of technology and strategic innovation, is one of the clearest signals yet that the Pentagon's largest weapons supplier is willing to let a commercial artificial-intelligence laboratory touch the engineering core of its flagship platform.
The admission arrived in an interview aired Oct. 2, 2026, and it lands at a moment when the F-35 program is under cost and schedule pressure that in-house engineering alone has failed to relieve. The deeper story is not a single contract. It is a change in where the defense-industrial moat sits — from owning the entire stack inside one security perimeter to owning the integration, testing, and certification layer that makes commercial models usable in a weapons system.
The Disclosure: OpenAI Inside the Sensor Stack
Hiza's disclosure was specific: OpenAI is working alongside the F-35 team on the mathematical and physical modeling behind advanced sensors. That is the same sensor-fusion stack that is the aircraft's defining advantage — the system that takes raw radar, infrared, and electronic-warfare inputs and decides what the pilot is actually looking at. This is not an administrative chatbot rollout or a public-relations pilot. It is work on the equations that let the jet classify a threat.
The framing matters because Lockheed has spent years building its own AI stack behind its own wall. It created Astris AI, a subsidiary formed in December 2024 to help defense companies adopt artificial intelligence inside secure, hardened infrastructure. It opened the AI Fight Club, a large-scale synthetic testing environment announced in June 2025, where AI systems are stress-tested across air, land, sea, and space domains before they are trusted with real missions; the inaugural event ran May 6, 2026, in Bethesda, Maryland. In February 2026, the company flight-tested Project Overwatch at Nellis Air Force Base in Nevada, an AI-enhanced combat-identification model that independently flagged targets on the F-35 pilot's display — the first time, the company said, that a tactical AI model generated an independent combat ID in flight.
So why reach outside the fortress for the hardest problem on the jet? Hiza also revealed the scale of the answer: Lockheed is running 55 different large language models across its business, taking a deliberately model-agnostic approach rather than betting on a single vendor. A company that intended to own the model layer would standardize on one or two. A company that intends to own the integration layer shops aggressively, validates ruthlessly, and hardens whatever it finds.
The technology is being applied to everything from internal operations to autonomous weapons systems, with a testing regime designed to validate models before deployment. That testing layer — not the model itself — is where Lockheed now believes its durable advantage lives.
The Pressure on the F-35 Program
The timing is not accidental. The F-35 is the most-produced stealth fighter in history, with more than 1,300 aircraft in service across 12 nations — including Germany, whose first F-35A completed final assembly in August 2026. But the program's modernization pipeline is straining under its own weight.
The Government Accountability Office reported this year that Block 4, the jet's major modernization effort involving more than 75 enhancements to weapons, sensors, electronic warfare, and networking, now runs more than $6 billion over original estimates and five years behind schedule. The Pentagon is redefining and scaling back parts of the Block 4 effort rather than delivering everything as first planned. The lifetime math is starker still: sustaining the fleet across its 77-year life cycle is projected to cost at least $1.58 trillion, pushing the program's total lifetime cost above $2 trillion. The Pentagon's own 2025 modernized selected acquisition report put the cost to acquire all planned aircraft at $536.3 billion, up from $485.2 billion in the prior year's estimate.
Every month that sensor software slips is a month the world's most-produced stealth fighter is flying with a fusion system that is harder to update than the threat environment demands. Adversaries are not waiting on a procurement calendar. Commercial frontier models offer something an in-house team cannot easily replicate: a capability curve that compounds at the pace of the global AI race, not the pace of a defense acquisition cycle.
There is also a budgetary reality behind the decision. The Pentagon's FY2026 request included a record $13.4 billion for AI and autonomy — the first dedicated line item for these capabilities — but that sum must cover $9.4 billion for aerial drones, $1.7 billion for maritime autonomous platforms, $734 million for underwater systems, $210 million for autonomous ground vehicles, and $1.2 billion for software and cross-domain integration. Even a record defense AI budget has to be split across every domain; a frontier-model laboratory concentrates comparable capital on a single stack. No single defense prime's internal research budget can match the aggregate R&D of the commercial AI sector, and the gap is structural, not cyclical.
The Testing Regime: Why Lockheed Tests Before It Deploys
The part of this story that does not fit the defense-prime-outsourcing-its-brain narrative is the infrastructure Lockheed has built to make outsourcing safe. Hiza discussed how Lockheed tests AI before deploying it, and the AI Fight Club is the physical expression of that discipline: a synthetic environment large enough to simulate global-scale operations, where an AI agent can be flown through thousands of scenarios in air, land, sea, and space before it ever touches a real aircraft.
Project Overwatch shows what that pipeline looks like in practice. During the February test, an AI and machine-learning model resolved identification ambiguities among emitters, improving situational awareness and reducing the pilot's decision latency. Engineers then used an automated tool to label new emitters, retrain the model on the new emitter class within minutes, and reload the updated model for the next flight — all inside the same mission-planning cycle. The point was not just that the model worked; it was that the model could be changed faster than the threat could adapt.
"This is a demonstration of 6th Gen technology brought to a 5th Gen platform," said Jake Wertz, vice president of F-35 combat systems at Lockheed Martin Aeronautics. "Equally important is our ability to re-program the AI model on the ground and have those updates available for the next sortie — an essential step toward maintaining a tactical edge in a rapidly evolving threat environment."
That is the old model at its best: build it, flight-test it, patch it between sorties. The OpenAI arrangement points to a new one: source the best available intelligence from the commercial frontier, validate it in your own synthetic proving ground, and fuse it into a system the customer already trusts. The moat is no longer the model. It is the proving ground.
The Second-Order Play: From Sensor Math to Drone Swarms
The first-order read of this news is narrow: OpenAI helps Lockheed model sensors better, and the F-35 sees more clearly. The second-order implication is broader, and it is the one the market has not fully priced. Hiza also discussed the growing role of autonomy and crewed-uncrewed teaming — the doctrine in which a manned fighter commands a swarm of uncrewed collaborative combat aircraft.
That doctrine lives or dies on the same math OpenAI is now helping to solve. A drone wingman is only as good as the manned jet's ability to fuse what the wingman sees, classify it, and hand off a target faster than an adversary can jam the link. Better sensor modeling on the F-35 does not just improve the F-35. It improves the entire manned-unmanned team that orbits the F-35's decisions. The transmission channel runs from a commercial model, through a prime's sensor stack, into a fleet architecture that multiplies one pilot's reach across a swarm.
This is where the structural-shift call becomes concrete. As Boeing builds the Air Force's next-generation F-47 fighter under the NGAD program and the Navy advances its F/A-XX, the value in sixth-generation air dominance is shifting from airframe performance — which has largely converged across competitors — to the software and sensor-fusion layer that turns a formation of crewed and uncrewed aircraft into a single decision network. Lockheed is not just buying a better model for one jet. It is positioning the F-35, a fifth-generation airframe with more than 1,300 units already flying, to serve as the brain of a sixth-generation fleet.
That is the real asymmetry: the F-47 will take years to reach the fleet in meaningful numbers. The F-35 is already there, in twelve countries, today. If commercial AI can keep its fusion stack ahead of the threat, the existing fleet becomes a platform for capabilities that were supposed to belong to the next generation.
The Counter-Thesis: Security Walls Still Have Teeth
The strongest argument against reading this as a revolution is the simplest one. The F-35 is among the most classified programs on the planet, and OpenAI is a commercial company with public safety commitments that explicitly restrict some military uses of its technology. OpenAI's published national-security principles bar the use of its technology for mass domestic surveillance, for directing autonomous weapons systems, and for high-stakes automated decisions. A sensor-modeling engagement that stays on the unclassified side of the math — improving how raw emissions are translated into tracks and correlations — can fit inside those guardrails. Work that touches fire-control logic or classified source data may not.
There is also Lockheed's own hedging to weigh. Running 55 models is not just flexibility; it is insurance against vendor lock-in and against the day a frontier laboratory refuses a defense application on policy grounds. The company has already watched a smaller rival move first: Anduril Industries announced an OpenAI partnership in December 2024 focused on counter-uncrewed-aircraft systems, training models on Anduril's threat and operations library. Lockheed is not following Anduril into counter-drone work. It is going after something harder, and something closer to the jet's core value proposition — but it is doing so with one hand tied behind its back by classification and policy.
The counter-thesis has teeth, but it does not overturn the direction of travel. Even a bounded, unclassified engagement on sensor math establishes a precedent: the prime contractor's security perimeter can include frontier commercial models. Precedents in defense procurement tend to widen, not narrow. And the economic logic is one-way: once a program learns it can buy a capability in weeks that would have taken years to build, it does not volunteer to go back to building.
The Growth Contrast: Primes Versus the Software Layer
The financial evidence for the structural shift sits in the growth rates. Lockheed raised its 2026 sales guidance to a range of $79.75 billion to $81.75 billion in its second-quarter 2026 earnings, expecting high single-digit to low double-digit growth in the second half of the year, with free cash flow of $7 billion to $7.2 billion. Its backlog grew $64 billion year over year on a book-to-bill ratio of 3.2 to 1 — a transition from demand capture to multi-year production scaling.
Compare that with Palantir Technologies, the pure-play defense software name that has become the market's proxy for AI adoption inside government. Palantir reported first-quarter 2026 revenue of $1.633 billion, up 85 percent year over year, with U.S. revenue of $1.282 billion, up 104 percent. It raised full-year 2026 revenue guidance to about $7.65 billion, implying roughly 71 percent growth. One company is growing at defense-prime speed, on a base ten times larger. The other is growing at software speed, on a base that is still a fraction of a prime's.
The spread is the investment signal. If the defense industrial base is truly shifting its AI stack toward commercial and frontier models, the primes capture the integration margin but cede the model-margin growth to the software layer. Lockheed's bet — that it can own the integration, testing, and certification layer — is a bet that the customer will keep paying the prime for trust, even as the intelligence inside the system comes from outside the perimeter.
What to Watch: Scenarios and Signals
The falsifying signal is concrete and near-term. A formal Lockheed-OpenAI contract announcement specifically covering F-35 sensor work would confirm the structural shift; a Department of Defense restriction blocking commercial frontier models from classified flight programs would disprove it. Watch the next selected-acquisition report on the F-35 for whether software-modernization timelines improve after this engagement, and watch whether Lockheed's AI Fight Club begins advertising frontier-model-derived systems in its test catalog.
Three scenarios frame the path:
- Base case: The OpenAI engagement stays bounded to unclassified sensor math, delivers incremental improvements to emitter resolution and retraining speed, and becomes one line in Lockheed's model-agnostic portfolio. The F-35 fusion stack improves at the margin; the stock continues to track defense budgets and delivery schedules. Trigger: a routine contract notice or a technical paper, with no policy controversy.
- Upside case: The sensor modeling accelerates Block 4 software timelines, the capability extends into crewed-uncrewed teaming, and Lockheed publicly demonstrates a frontier-model-derived fusion upgrade across the 1,300-aircraft fleet. The integration moat re-rates the stock as a software-enabled platform rather than a metal-bending prime. Trigger: a program-milestone announcement tying AI-derived modeling to a Block 4 software drop, or a live demonstration of manned-unmanned teaming at a major exercise.
- Downside case: Classification or OpenAI's own usage policies block the work from touching the classified flight program, the engagement stays confined to internal operations, and the F-35's sensor modernization remains on its delayed Block 4 track. The headline proves to be more about AI adoption optics than about the jet's capability. Trigger: a DoD policy memo restricting commercial frontier models in classified programs, or a Block 4 selected-acquisition report showing no schedule improvement.
For the broader defense complex, the asymmetry is clear. The beneficiaries of this shift are the primes with the certification infrastructure and the classified data pipelines to absorb commercial models; the exposed are the smaller AI vendors that can build a model but cannot clear a security perimeter. In the short term, Lockheed's shares will track budgets, deliveries, and geopolitical risk sentiment. In the long term, the valuation question is whether Lockheed's moat has really moved from the airframe to the fusion layer — and whether OpenAI's math can help the F-35 see farther before the next adversary learns to hide.
The real story is not that Lockheed asked OpenAI for help. It is that the most secretive weapons program in the world now treats a commercial AI laboratory as part of its engineering base — and that boundary, once crossed, is unlikely to be redrawn.
Explore more exclusive insights at nextfin.ai.
