The high-risk research bet: how DARPA and AFRL are buil...

Most of this series follows public money into programs that overran, underdelivered, or ran ahead of the public's consent. This entry is the counterweight, and it is meant to be read that way. Two defense research organizations, the Defense Advanced Research Projects Agency (DARPA) and the Air Force Research Laboratory (AFRL), are structured so that most of what they fund is expected to fail. The justification is not that failure is fine in the abstract. It is that a handful of transformational wins, spread across decades, plausibly returned many multiples of everything the agencies ever spent. Reading that ledger honestly means resisting two temptations at once: the boondoggle frame that would flatten a genuine success like the internet into a punchline, and the boosterism that would credit a government office for inventions that belonged to universities and companies.

This is a capstone about a model, not a scandal about a single program. So the frame here is the portfolio bet: the deliberate philosophy that lets an agency accept a very low batting average because the occasional home run is measured on the scale of entire industries.

The DARPA model: a small budget, fixed-term managers, and an accepted failure rate

By Pentagon standards DARPA is tiny. The Congressional Research Service, in its overview report R45088, tracks a budget on the order of $4 billion a year: the FY2024 figure sits around $4.39 billion, up from roughly $4.06 billion the year before, and the FY2026 request runs near $4.9 billion. The number creeps up year to year, so the honest shorthand is "on the order of $4 billion," with the year attached whenever a precise figure matters. Against a federal research and development budget measured in the tens of billions for the Air Force alone, DARPA is close to a rounding error.

What makes it distinctive is not the money but the operating design. DARPA hires program managers on fixed terms, typically in the range of three to five years, and hands them wide latitude to fund what its own literature calls "DARPA-hard" problems: high-technical-risk bets where the payoff, if it arrives, is transformational, and where failure is the expected outcome for most projects. DARPA and CRS both document the flat structure, the tenure limits, and the high tolerance for failure. The manager builds a program, funds it, and then leaves before the long tail of results is even in. Nobody is around long enough to protect a legacy, which is part of the point.

There is one structural fact that readers routinely get wrong, and it is the spine of everything below.

DARPA funds; AFRL builds

DARPA does not run laboratories. According to CRS, it has essentially no in-house benches. It is a funder and a manager: its program managers move money to universities, companies, and other government labs, and the research happens there. That is why credit for every DARPA win is shared by definition. The performers did the work.

AFRL is the opposite kind of organism. It is the Air Force's in-house science and technology enterprise, formed on October 31, 1997 at Wright-Patterson Air Force Base by consolidating four Air Force laboratories (Wright, Phillips, Rome, and Armstrong) with the Air Force Office of Scientific Research. AFRL runs its own directorates and benches and controls a substantial share of Air Force science and technology, conducting a meaningful portion of the work in-house while contracting the rest to academia and industry. Its workforce is commonly cited in the range of eleven thousand, though older fact sheets put it nearer ninety-six hundred, so the headcount is best treated as soft. The af.mil fact sheet is the authoritative current description.

Keep the distinction in mind for the rest of this piece: when DARPA "wins," it means a program manager placed a bet and someone else built the thing. When AFRL wins, it often built the thing.

The hits

Each of these is a line about what a public research bet became, with the shared credit named. None of these agencies invented these things alone.

ARPANET and the internet. DARPA funded ARPANET, whose first host-to-host message ran from UCLA to SRI on October 29, 1969. It became the direct precursor of the modern internet. Credit is genuinely shared: packet switching was developed independently by Paul Baran at RAND and Donald Davies at the National Physical Laboratory in the UK, and the network was built by teams at BBN, UCLA, and other universities. The World Wide Web, which most people mean when they say "the internet," is a separate and later layer, created by Tim Berners-Lee at CERN between 1989 and 1991, and it was not a DARPA project. ARPANET is not the same thing as the internet, and DARPA is not the same thing as the Web.
TCP/IP. The core internet protocols came out of DARPA-sponsored work by Vint Cerf and Bob Kahn, who published the TCP design in 1974. Kahn was a DARPA program manager; Cerf was at Stanford. The Defense Department made TCP/IP its standard in 1982, and ARPANET cut over from its old protocol to TCP/IP on the "flag day" of January 1, 1983. The Internet Society's history documents that migration. Credit is shared with the wider academic networking community.
Stealth. Stealth aircraft trace to DARPA's Have Blue demonstrator. DARPA contracted Lockheed's Skunk Works in 1976, two subscale demonstrators first flew in 1977, and the lineage led to the F-117 Nighthawk, operational in 1983. The enabling faceting math built on published work by the Soviet physicist Pyotr Ufimtsev, and Lockheed built the aircraft. DARPA funded the demonstrator; it did not build the plane.
The drone lineage. The Predator descends from DARPA-funded work by the aeronautical engineer Abraham Karem. His Albatross, and the DARPA-funded Amber from 1985, were the direct ancestors of the General Atomics Gnat-750 and the MQ-1 Predator. This is a seed and lineage role, not a sole invention: the path runs Albatross to Amber to Leading Systems to General Atomics.
The Grand Challenge and self-driving cars. DARPA's autonomous-vehicle races catalyzed the modern self-driving field. In the first Grand Challenge on March 13, 2004, no vehicle finished the 142-mile Mojave course; the best managed only about 7.4 miles. That productive failure seeded the 2005 race, where five vehicles finished and Stanford's Stanley won, and the 2007 Urban Challenge, where six of eleven finished. The credit belongs at least as much to the competing university teams at Stanford and Carnegie Mellon as to DARPA. There is a companion deep-dive on this one.
CALO and Siri. Apple's Siri assistant descends from DARPA's CALO project, part of the PAL program, at SRI International: a roughly five-year, roughly $150 million effort involving more than three hundred researchers across twenty-two institutions. Siri spun out of SRI, Apple acquired it in April 2010 for a reported figure north of $200 million, and it shipped on the iPhone 4S in 2011. SRI built CALO and Apple commercialized it. DARPA funded the underlying research.
Early mRNA money. DARPA was an early funder of mRNA therapeutics at Moderna, awarding a grant of up to $25 million (about $24.6 million) over up to five years in 2013, preceded by a small "seedling" grant earlier that year. The important discipline here is not to overclaim: this was a seed, not the COVID-19 vaccine. mRNA vaccines rest on decades of academic work, including Katalin Kariko and Drew Weissman's nucleoside-modification research at the University of Pennsylvania and NIH spike-protein design, and the later-stage development money came from BARDA on the order of $955 million years afterward. DARPA put in early risk capital; it did not develop the shot.
Onion routing and Tor. The technology behind Tor originated in the mid-1990s at the U.S. Naval Research Laboratory (Syverson, Reed, and Goldschlag), with a Navy patent in 1998, and was refined under DARPA funding before the Tor project launched in 2002. The origin is NRL, not DARPA proper. It is worth noting neutrally that a defense-funded privacy tool is now used by activists, journalists, and criminals alike.
Robotics. DARPA robotics funding seeded Boston Dynamics' legged machines, including BigDog and the Atlas robot unveiled in 2013 for the DARPA Robotics Challenge. Boston Dynamics is commonly reported to have taken on the order of $200 million across roughly fifty-seven U.S. government contracts, though that aggregate is an approximate 2025-dollar rollup and is better labeled government-seeded than purely DARPA.
AFRL's in-house wins. On the AFRL side of the ledger, the laboratory's long contributions in advanced materials and directed energy, and its work on the X-plane research aircraft, sit alongside a headline success: the X-51A Waverider. On May 1, 2013, the X-51A flew powered under scramjet thrust for about 210 seconds and reached roughly Mach 5.1 over about 230 nautical miles, the longest air-breathing hypersonic powered flight on record. It was managed by AFRL's Aerospace Systems Directorate as a joint effort with DARPA, NASA, Boeing, and Pratt & Whitney Rocketdyne, and it followed earlier X-51 flights in 2010 through 2012 that failed first. It worked. There is no boondoggle frame to force onto it.

A word on GPS, because it is frequently and loosely tossed into the DARPA trophy case. Navstar GPS emerged in the 1970s from Defense Department programs led mainly by the Air Force and Navy, running through a Joint Program Office and building on the Navy's earlier Transit satellite navigation system. It is a Defense research win with shared credit, but DARPA was not the prime, and this piece will not overclaim DARPA authorship of it.

The misses

The failures split into two categories that a careful reader should keep separate. One is money spent on programs that did not pan out. The other is legitimacy spent on programs that ran ahead of public and congressional consent. Collapsing both into a single "waste" bucket obscures more than it reveals.

Dollars that did not pay off operationally:

HTV-2. DARPA's Hypersonic Technology Vehicle 2 failed both of its test flights. Flight 1 on April 22, 2010 and Flight 2 on August 11, 2011 each lost control roughly nine minutes into a planned Mach 20 glide. This is an informative failure rather than a total loss: the second flight did demonstrate about three minutes of stable, aero-controlled flight at Mach 20 before the aeroshell skin peeled away and ended it, and DARPA framed the effort as data-gathering that fed later U.S. hypersonics work. Do not call it a clean success, and do not call it a null result.
The Airborne Laser (YAL-1). A megawatt chemical laser mounted on a modified Boeing 747, the YAL-1 consumed more than $5 billion over roughly sixteen years before funding was cut in 2010 and the program was cancelled in December 2011, with the aircraft flown to the boneyard in early 2012. The honest both-sides matters here: on February 11, 2010 the YAL-1 actually destroyed two in-flight ballistic-missile targets. It worked as a physics demonstration. It was cancelled for operational impracticality, because a chemical-laser 747 would have to loiter near enemy launch sites and the fleet cost and logistics were prohibitive. That is a legitimate-goal failure, not fraud. One attribution note: this was primarily a Missile Defense Agency and Air Force program with Boeing, not a DARPA program, so it should not be filed as a DARPA miss even when grouped on the Air Force side of the ledger.

Legitimacy that ran ahead of consent:

Total Information Awareness (TIA). Run out of DARPA's Information Awareness Office under John Poindexter, TIA was a mass-data-mining surveillance effort that drew intense privacy backlash. Congress cut the office's funding in 2003, directing its termination, though some component research reportedly continued elsewhere in classified programs. This is a civil-liberties controversy, not a technical failure. The Federation of American Scientists maintains the congressional-record archive; there is a companion deep-dive.
The Policy Analysis Market ("terrorism futures"). Part of the FutureMAP project under the same office, this was a prediction-market experiment. Senators Ron Wyden and Byron Dorgan denounced it at a press conference on July 28, 2003, and the Pentagon cancelled it the next day, formally on July 29. The legitimate goal was testing whether markets aggregate dispersed information better than analysts do; the fatal problem was the optics and ethics of a market that could look like betting on assassinations and attacks. Poindexter offered his resignation that week, but he did not cancel the market and was not fired the same day it was killed.

The balance: why a low batting average is the design, not the defect

Set the two ledgers side by side and the portfolio logic comes into focus.

The ROI case. The strongest defense of this model is structural rather than anecdotal. Venture capital, which is the private mechanism most often proposed as an alternative, seeks roughly 30 to 40 percent annual returns and will take market risk but generally not deep, decade-long technical risk on problems with no near-term product. That leaves a gap: the National Academies literature on moving basic research to innovation describes exactly the kind of long-horizon, high-uncertainty work that markets rationally underprovide. Public research fills it, and the return structure is inherently a portfolio. You accept that most bets fail because a few of them create entire industries. The internet alone plausibly returned many multiples of DARPA's entire cumulative budget, and that budget is only on the order of $4 billion a year. This is a reasoned argument, not a hard dollar figure: no honest writer can put a precise ROI on "the internet" without inventing it. But as an argument about why society tolerates a high failure rate, it is well grounded.

The honest caveats. The defense does not excuse everything, and it should not be allowed to. Two distinct critiques survive scrutiny. First, on dollars: the fair question is not "why did programs fail," because failure is the stated plan, but whether specific bets were sized and killed responsibly. It is reasonable to ask whether the Airborne Laser's operational impracticality should have been recognized long before it crossed $5 billion. Second, on legitimacy: TIA and the terrorism-futures market were not budget boondoggles so much as governance failures, programs that outran public and congressional consent and had to be killed on privacy and ethics grounds. That is a real cost, and it is a different kind of cost than a missile that misses.

What the balance should resist is the move that would make either ledger tidy. A portfolio with no failures would mean the agency had stopped taking the high-risk bets that are its entire reason to exist. The low batting average is the point. And the wins, every one of them, were shared with the universities and companies that did the building. DARPA funds and manages; AFRL adds its own benches; and Stanford, Carnegie Mellon, the University of Pennsylvania, Lockheed, SRI, Moderna, General Atomics, Boeing, and many individual researchers turned the money into the thing. A society willing to eat many public failures, and to cancel programs when they fail on cost or on ethics, is buying option value on transformational technology that no private actor would purchase on its own.

Fact-check notes and sources

DARPA budget on the order of $4 billion a year (FY2024 about $4.39 billion, up from roughly $4.06 billion; FY2026 request near $4.9 billion), the tenure-limited program-manager model, and the funder-not-operator distinction: Congressional Research Service, report R45088 and DARPA program-manager page.
DARPA created in 1958 in response to the Soviet launch of Sputnik, under President Eisenhower: Encyclopaedia Britannica.
ARPANET as DARPA-funded precursor of the internet, with packet switching developed independently by Baran (RAND) and Davies (NPL) and the Web a separate later CERN layer; TCP/IP by Cerf and Kahn published 1974, with the ARPANET cutover on January 1, 1983: DARPA innovation timeline and Internet Society, TCP/IP migration in 1983.
Stealth via DARPA's Have Blue (Lockheed contract 1976, first flights 1977, F-117 operational 1983), with Ufimtsev's math and Lockheed's construction as shared credit: Lockheed Have Blue.
Predator lineage through Karem's Albatross and the DARPA-funded Amber (1985): Air and Space Forces Magazine.
Grand Challenge results (2004 no finisher, 2005 five finishers and Stanley, 2007 Urban Challenge six of eleven): DARPA Grand Challenge.
CALO/PAL at SRI (about five years, about $150 million) leading to Siri, with the Apple acquisition in April 2010 and the iPhone 4S launch in 2011: SRI International.
DARPA's early mRNA grant to Moderna of up to $25 million in 2013, with strong shared credit to Kariko and Weissman at the University of Pennsylvania, NIH spike design, and BARDA's later-stage funding on the order of $955 million: Moderna 2013 press release.
Onion routing originating at the Naval Research Laboratory (mid-1990s, Navy patent 1998) and refined under DARPA before Tor launched in 2002: Onion routing.
DARPA robotics funding for Boston Dynamics' BigDog and Atlas, with the roughly $200 million across about fifty-seven government contracts treated as an approximate aggregate: Boston Dynamics.
Total Information Awareness (Poindexter's Information Awareness Office) defunded by Congress in 2003 amid privacy backlash, flagged as a civil-liberties controversy: Federation of American Scientists congressional archive (FAS publishes advocacy-adjacent policy material; the underlying congressional records are primary).
Policy Analysis Market denounced July 28, 2003 and cancelled the next day: Policy Analysis Market.
HTV-2 failing both flights (April 22, 2010 and August 11, 2011) at roughly Mach 20, with the second flight's partial stable-flight data: Hypersonic Technology Vehicle 2.
AFRL as the Air Force's in-house science and technology enterprise, formed October 31, 1997 at Wright-Patterson, with a soft workforce figure: AFRL fact sheet, af.mil and Air Force Research Laboratory overview.
AFRL's X-51A Waverider record hypersonic flight on May 1, 2013 (about Mach 5.1, about 210 seconds), a joint AFRL, DARPA, NASA, Boeing, and Pratt & Whitney Rocketdyne effort: AFMC news release and Boeing X-51 Waverider.
The Airborne Laser (YAL-1) destroying two missiles in February 2010 but cancelled in December 2011 after more than $5 billion over roughly sixteen years, primarily a Missile Defense Agency and Air Force program with Boeing: Boeing YAL-1.
GPS as a Defense research win driven mainly by the Air Force, Navy, and the DoD Joint Program Office rather than DARPA as prime: Global Positioning System.
The portfolio and market-underprovision argument, presented as reasoned framing rather than a precise ROI figure: National Academies (NCBI), reducing the time from basic research to innovation.

The high-risk research bet: how DARPA and AFRL are built for a few huge wins and many failures