FAA’s $875M AI Bet: Why a Startup Won Air Traffic Control

The Deal That Shouldn’t Have Happened

The FAA just handed Air Space Intelligence, a Boston startup backed by Andreessen Horowitz, an $875 million, 12-year contract to build the SMART system (Strategic Management of Airspace, Routes, and Trajectories) — essentially replacing the air traffic control backbone that routes 45,000 daily flights across U.S. airspace. Initial deployment targets fall 2026, with full rollout in 12-24 months.

This is the kind of procurement that makes defense contractors file protests and senators ask pointed questions. A startup — not Raytheon, not Lockheed, not even Palantir (who was in the running) — won the contract to modernize arguably the most safety-critical infrastructure in civilian aviation. The decision is either visionary or reckless, and which one depends entirely on whether Air Space Intelligence can scale their airline dispatch tool to handle the entire national airspace without anyone dying.

What SMART Actually Does

The current air traffic control system works on a 15-minute conflict prediction window. Controllers see problems when planes are already converging. SMART extends that window to two hours by ingesting flight schedules, weather data, and real-time airspace state, running 4D trajectory modeling, and suggesting route adjustments before conflicts materialize. The system is advisory — controllers retain final authority — but the goal is workload reduction, not replacement.

Air Space Intelligence’s Flyways AI already runs at Alaska Airlines, where it’s been optimizing dispatch decisions since 2021. Alaska claims savings of “tens of thousands of flight hours” and roughly a million gallons of fuel annually. That’s a single airline with predictable routes, manageable complexity, and humans in the loop who can override bad recommendations without catastrophic consequences.

Scaling that to the FAA’s use case means coordinating every airline, every charter, every cargo hauler, and every private jet across 29 million square miles of controlled airspace, in real time, with weather that changes faster than models can update, and failure modes that ground the national economy. The technical gulf between “help Alaska’s dispatchers pick better routes” and “be the autopilot for the entire U.S. air traffic system” is the difference between a microservice and a mission-critical distributed system with no acceptable downtime.

Why This Probably Won’t Work (And Why It Might)

The FAA’s own Government Accountability Office audit from 2025 found that 37% of the agency’s 138 air traffic control systems are “unsustainable” and another 39% are “potentially unsustainable.” The modernization push isn’t optional — legacy systems are literally failing. The question isn’t whether to upgrade, but whether to bet on unproven AI or extend decades-old vendor lock-in with incremental improvements from the usual suspects.

Air Space Intelligence won because they demonstrated actual results in production. Alaska’s Flyways system isn’t vaporware; it’s deployed, it works, and it saves money. That track record beats PowerPoint decks from incumbents who’ve spent 20 years and billions delivering systems that are late, over-budget, and obsolete on arrival. The FAA is gambling that a startup with a working product can scale faster than a defense contractor can innovate.

The failure modes are obvious. The system could generate plausible-but-wrong recommendations that controllers trust because the AI has been right 99% of the time. It could optimize for fuel efficiency at the expense of safety margins. It could fail gracefully in testing and catastrophically under real-world edge cases — thunderstorms that spawn faster than the model’s refresh rate, military airspace that closes without warning, airport equipment failures that cascade into system-wide delays. And if any of that happens at scale, the political blowback will end both the SMART program and Air Space Intelligence’s runway.

But if it works? If SMART actually prevents the congestion bottlenecks that currently cost airlines billions in delays, if it reduces controller workload enough to address the staffing crisis, if it makes U.S. airspace measurably safer and more efficient — then this contract will look like the FAA finally figured out how to procure modern software instead of paying Raytheon to maintain COBOL.

The Real Test Isn’t Technical

The technical challenges are solvable. Predictive trajectory optimization is hard, but it’s not unsolved research — it’s an engineering problem with known approaches and measurable outcomes. The real test is whether the FAA can deploy this system without either rushing it into production before it’s ready (to meet political deadlines) or delaying it into irrelevance (to satisfy risk-averse bureaucrats).

Industry stakeholders are already saying “you’re going too fast” about the September deployment target. Controllers’ unions will push back on anything that looks like automation replacing human judgment. Airlines will demand proof that the system won’t introduce new delay vectors. And Congress will scrutinize every dollar if the project hits snags.

The FAA is betting that moving fast and trusting a startup is less risky than moving slow and trusting the incumbents who got us into this mess. That’s a bet I’d take — but only if “fast” doesn’t mean “skipping the boring parts like testing edge cases and building proper fallbacks.”

FAQ

Q: Will SMART replace air traffic controllers?
No. The system is advisory — it suggests route adjustments to prevent downstream conflicts, but controllers retain final decision authority. The goal is workload reduction (extending conflict prediction from 15 minutes to 2 hours), not automation. That said, if it works well enough, future versions will inevitably push toward more autonomy.

Q: Why did the FAA pick a startup over established defense contractors?
Air Space Intelligence had a working product (Flyways AI) already deployed at Alaska Airlines with measurable results (saving tens of thousands of flight hours and ~1M gallons of fuel annually). Incumbents like Raytheon have decades of delays and cost overruns on FAA modernization projects. The FAA gambled that proven execution at small scale beats unproven promises at large scale.

Q: What happens if SMART fails in production?
If the system generates bad recommendations that controllers trust, or if it fails under real-world edge cases (e.g., rapidly changing weather, equipment failures), the consequences range from system-wide delays to potential safety incidents. Political blowback would likely kill the program and damage the FAA’s credibility on AI procurement. The deployment timeline (fall 2026 initial, 12-24 months full rollout) will be the critical test of whether they’re moving fast or recklessly.