The Next Offset

Bob Work was sixty-two years old when he stood at the lectern at the Ronald Reagan Presidential Library in Simi Valley on the afternoon of December 14, 2015, and tried to explain to a room full of generals and defense reporters why the American military needed to start thinking like a chess centaur.

Centaurs, in chess, were a thing he had picked up reading. After Garry Kasparov was beaten by IBM’s Deep Blue in 1997, a small subculture of grandmasters had begun pairing themselves with off-the-shelf chess engines and entering tournaments as hybrid players. The engines could see deeper than any human; the humans could see further. Working together, they repeatedly beat both unaided humans and unaided machines. In 2005 a freestyle tournament had been won by two amateurs running three commercial PCs. They had beaten grandmasters with supercomputers. The future, as Work read it, did not belong to the smartest human or the smartest machine. It belonged to the team that figured out how to play together.

He thought it applied to war.

Work had served twenty-seven years in the Marine Corps before retiring as a colonel and crossing into the policy world. He talked like an artilleryman. He drew diagrams on whiteboards with the heel of his hand and walked audiences through them as if briefing a fire-direction center. He had spent the years between the Marines and the Pentagon at the Center for Strategic and Budgetary Assessments and at CNAS, the Washington think tank he ran in 2013 and 2014. He had written a long string of monographs about Chinese and Russian military modernization, arguing in language that grew more pointed each year that the United States was sleepwalking into a strategic crisis. They found a steady readership inside the Office of the Secretary of Defense. In February 2014, Obama nominated him to be the thirty-second Deputy Secretary of Defense. The portfolio that mattered most, the one nobody else wanted, was sitting on his desk the day he walked into the building.

It had a name his predecessors would have recognized. Secretary Chuck Hagel had stood at the same Reagan Library a year earlier, on November 15, 2014, and announced what he called the Defense Innovation Initiative. Reporters in the room, working from briefing papers Work’s staff had prepared, had quickly attached to it a different name. They were calling it the Third Offset.

The phrase was a deliberate echo. The First Offset had been Eisenhower’s nuclear answer to Soviet conventional mass in the 1950s. The Second, Perry and Brown and Marshall’s 1977 bet on precision strike, stealth, and silicon, had been vindicated in the Gulf War of 1991. Each had been an answer to a specific strategic problem and rested on a specific technological wager. The Hagel-Work bet was that a third was now necessary, because the second had run its course.

Hagel laid out the diagnosis plainly. Russia and China were closing the conventional gap. They were fielding their own precision-guided munitions, building their own reconnaissance-strike networks, investing in the same suite of technologies the United States had used to bury the Warsaw Pact, from a starting position no longer thirty years behind. The Eurasian air-defense bubbles around Kaliningrad and the South China Sea, what Pentagon staffers had taken to calling A2/AD for anti-access and area-denial, were dense enough that an American carrier or a non-stealth strike package might no longer be able to operate inside them. The technologies that had once made the United States a magic weapons power had become, in Hagel’s phrasing, a global commodity. Everyone could buy precision. Everyone could buy networking. The advantage that had vindicated Perry’s bet had been worn smooth by three decades of proliferation.

Hagel announced that the Pentagon would launch a new Long-Range Research and Development Planning Program, named for the program that had produced the Second Offset, and direct it at five technology baskets: robotics, autonomous systems, miniaturization, big data, and advanced manufacturing. He named Work as the official in charge. Nine days later, on November 24, 2014, the White House asked Hagel for his resignation.

The political circumstances of Hagel’s departure had nothing to do with the offset. The new secretary, Ashton Carter, sworn in in February 2015, was a physicist who had spent the previous administration as deputy secretary and already understood, in a way Hagel had only intuited, what the Pentagon’s relationship with Silicon Valley had become. He kept Work in place. He kept the agenda intact. And he embraced a piece Work had been pushing in classified rooms for months: that the most decisive technology in the new bet was not robotics or networking in general but a particular kind of computing called machine learning, a kind that depended on a particular kind of silicon, and that neither the silicon nor the talent any longer lived inside the Pentagon’s traditional supply chain.

In Seoul, in March 2016, in a ballroom on the upper floor of the Four Seasons Hotel, the world watched the second piece of the bet land. AlphaGo, a program built by a small London laboratory called DeepMind that Google had quietly bought two years earlier, defeated the South Korean Go player Lee Sedol four games to one over a five-day match. Go had been the standing benchmark for whether machine learning was approaching general intelligence. Most experts had said the milestone was at least a decade away. AlphaGo cleared it on a system that combined deep neural networks of a kind that had not existed in serious form before 2012 with a back end of forty-eight custom Google chips called tensor processing units, put into operation only the previous year. In Beijing, the match drew more than two hundred and eighty million viewers, roughly a fifth of the world’s population watching one chess board. Sixteen months later, in July 2017, the Chinese State Council issued a national plan committing Beijing to global AI leadership by 2030. Kai-Fu Lee would later call the AlphaGo match China’s Sputnik moment. The phrase stuck because it described what had actually happened.

Work had been watching the match from his Pentagon office on the third floor of the E-ring. He did not need a Sputnik metaphor. He had been making the argument inside his own building since 2014, and the chess match had simply made it visible. The sudden capability of deep neural networks was not a normal incremental improvement. It was a phase change. A network that had taken a graduate student a month to train in 2010 could now be trained on a cluster of commodity GPUs in a day. Models that had failed at object recognition five years earlier were beating human radiologists at certain medical-imaging tasks. What had made all of it possible was not a clever new algorithm. It was the chip. The graphics processing unit, invented in the mid-1990s to render polygons for video games, had been repurposed in 2012 by two University of Toronto students, Alex Krizhevsky and Ilya Sutskever, to train an image-recognition network called AlexNet on a pair of consumer Nvidia cards in a bedroom at Krizhevsky’s parents’ house. AlexNet won that year’s ImageNet competition by a margin that broke the field. Every important advance in machine learning since had been a continuation of that lineage. It ran on Nvidia silicon, designed in California, manufactured at TSMC in Taiwan.

This was the part Work had to keep quiet in public, because the Department of Defense did not announce dependencies it did not yet know how to fix. In private he was direct. The compute that drove the new AI lived almost entirely on commercial GPUs, the GPUs were designed by a fabless company eighty miles south of his old Marine Corps assignments, and the fabrication of the leading-edge nodes those GPUs required happened in Hsinchu Science Park on the western face of an island China claimed as its own. The Second Offset had attached American military power to American silicon. The Third would attach it to compute. And compute, by 2015, no longer lived where Perry’s generation had left it.

He spelled out the public version at the Reagan Library that December afternoon. Five technology building blocks, he told the room, would constitute the new offset: autonomous deep learning systems, human-machine collaboration, assisted human operations, advanced manned-unmanned combat teaming, and autonomous weapons. The advantage would come not from any one of them but from putting them together inside a coherent battle network. Russia was openly talking about fielding fully roboticized units; Valery Gerasimov, the chief of the Russian general staff, had said as much in print. China was investing heavily in robotics and autonomy. The United States, Work said, was at an inflection point. The advantage could compound or it could erode. It depended entirely on what the Pentagon did next.

What it did next had an institutional face. Inside the Office of the Secretary of Defense, a small unit called the Strategic Capabilities Office, set up by Carter in 2012, was already doing the fast-cycle weapons engineering Work wanted. Its founder and director was Will Roper, a Mississippi-born physicist who had trained at Oxford in string theory and arrived at the Pentagon, by his own description, having no idea what he was supposed to be doing. By mid-2015 SCO was running Perdix, a swarming-drone program in which an F/A-18 dispensed clouds of small autonomous fixed-wing micro-drones that flocked toward designated targets. Carter had given Roper enough budget cover to keep building. Work, taking up the portfolio inside his offset agenda, kept extending it.

In April 2015, on a trip to Silicon Valley months in planning, Carter announced an experimental unit headquartered in Mountain View whose mission was, in his words, to bridge the cultural distance between the Pentagon and the Valley. Defense Innovation Unit Experimental, soon shortened to DIUx, was supposed to do small, fast contracts with startups whose products the Pentagon needed and could not buy through the normal acquisition process. The first version, run by a career Pentagon hand, struggled. Carter rebooted it in May 2016 under Raj Shah, a former F-16 pilot who had cofounded a Palo Alto cybersecurity company and could speak both Pentagon and venture capital without translation. Under Shah, DIUx began moving small dollars to small companies on timelines startups could survive. By 2018, when the unit dropped the experimental tag and became simply the Defense Innovation Unit, contracts were flowing to firms whose names would appear in defense news for the next decade. Anduril, founded in 2017 by Oculus VR cofounder Palmer Luckey, was one of them.

A second institutional face was harder to name in public. On April 26, 2017, two months into the new Trump administration, Work signed a memo establishing the Algorithmic Warfare Cross-Functional Team. Internally and externally it was called Project Maven. Initial budget about seventy million dollars. Its first job was to take the full-motion video pouring back from Predator and Reaper drones over the Middle East and apply machine-learning models to it, so that an analyst could be told, in close to real time, where the trucks and the people and the buildings were in the frame. The Air Force already employed thousands of analysts to watch this video by hand. It was arriving faster than they could watch it. Without machine-learning processing, much was being dumped, unwatched, on archival storage. Maven was the first operational deployment of modern computer vision inside the Department of Defense.

It needed industry. Specifically, labeled training data and engineers who knew how to build deep learning pipelines. The Pentagon’s traditional contractors did not have those engineers. Google did. By the late summer of 2017, a contract worth around nine million dollars had been signed between the Pentagon and Google’s cloud division, then run by Diane Greene. Inside Google it was small. Outside the executive suite, by the spring of 2018, it had stopped being small.

In late February 2018, an internal letter began circulating on Google’s communication systems. By early April, when Gizmodo and the New York Times published its existence, more than three thousand employees had signed. The opening sentence did the damage. We believe, the letter said, that Google should not be in the business of war. It demanded that Google cancel the Maven contract and adopt a clear policy against ever building warfare technology. A second letter from outside academics collected more than a thousand additional signatures. A handful of engineers resigned. On June 1, 2018, Greene told staff Google would not seek to renew the Maven contract when it expired in 2019. A week later, Sundar Pichai, by then chief executive of Alphabet, published a set of AI principles that included the statement that Google would not design or deploy AI for weapons.

Inside the Pentagon, the Google withdrawal landed as a confirmation of what Work and his colleagues had been worrying about for years. The talent and the tools the Third Offset depended on lived inside companies that did not necessarily want to be defense contractors and whose employees would, given the chance, refuse to work on military programs. Project Maven did not die. The contract went to other vendors, eventually to Palantir, and Maven persisted as one of the longest-running AI programs the department ran. The lesson was clear enough. The American military’s path to AI ran through American technology companies, and American technology companies were, in 2018, an unreliable partner. It would shape every Pentagon AI conversation for the next five years.

It also accelerated something Work had been pushing since 2016: consolidating the Pentagon’s scattered AI efforts into a single organization that could speak for the department to industry. On June 27, 2018, Deputy Secretary Patrick Shanahan signed the memo establishing the Joint Artificial Intelligence Center. Lieutenant General John Shanahan, an Air Force three-star who had directed Maven from its founding, took over the JAIC that December and arrived to four volunteers and no money. Eighteen months later it had a hundred and eighty-five staff and a budget around 1.3 billion dollars. Small numbers by Pentagon standards. Enormous by AI-program standards.

While Shanahan was building the JAIC, another institution was assembling the strategic argument for why it mattered, with the kind of public-private mixed authorship Work and Carter had been trying to engineer. The National Security Commission on Artificial Intelligence had been established by Congress in August 2018 inside the John S. McCain National Defense Authorization Act, signed into law on August 13, twelve days before McCain died of brain cancer at his home in Arizona. The commission was an independent body of fifteen, and its co-chairs were the two men best positioned to write the document the political system would actually read. Eric Schmidt, the former Google chief executive whom Carter had named in 2016 to chair the Defense Innovation Board, took the chair. Work, freshly out of the Pentagon, took the vice chair.

The NSCAI worked for two and a half years and issued, on March 1, 2021, a final report running to more than seven hundred and fifty pages. It was, by a comfortable margin, the most ambitious blueprint for American technology policy that the federal government had produced since the early Cold War. Most of its substance landed inside the parts of Washington already paying attention. One conclusion did something different. It moved the argument about chips, which had been a trade story and an industrial story for fifty years, into the center of the American national-security debate, on grounds impossible to ignore.

The argument was structurally simple. Modern artificial intelligence depended on compute. Compute depended on advanced semiconductor manufacturing. The United States no longer manufactured the most advanced semiconductors. Almost all of the leading-edge nodes Nvidia and AMD and Apple and Qualcomm needed to build the chips on which American AI ran were fabricated in two places, both on islands within reach of Chinese military power: TSMC in Taiwan and Samsung in South Korea. Intel, the standing American answer to this question since the early 1970s, had begun slipping behind both in 2018 and 2019 and would, the report observed dryly, fall further behind unless the United States acted with urgency. The relative comfort the Pentagon had enjoyed by staying two generations of process technology ahead of any potential adversary was eroding. The fab that produced the chip mattered as much as the company that designed it. American military power, by 2021, sat downstream of two foreign foundries.

The recommendations followed. The commission urged Congress to pass refundable investment tax credits for domestic leading-edge fabrication, a national strategy for staying two generations ahead of China, a Digital Service Academy, and immigration changes to keep AI PhD graduates in the country. It urged doubling federal non-defense AI research spending each year, from two billion dollars in 2022 to thirty-two billion by 2026. The preface, signed jointly by Schmidt and Work, used language unusually unguarded for a federal commission. China was a peer competitor in AI. The United States was not adequately prepared to compete or to defend against AI-enabled threats. The preface called for action on the scale of a generational mobilization.

DARPA, which had midwifed Assault Breaker forty years earlier, had already begun moving. In June 2017 the Pentagon’s research arm announced the Electronics Resurgence Initiative, a five-year, roughly 1.5-billion-dollar effort directed by William Chappell, head of DARPA’s Microsystems Technology Office. Moore’s Law, the program documents said, was running out of steam at the leading edge. The next era of electronics would be defined by specialization: domain-specific architectures designed for particular workloads, including AI’s. ERI funded new materials, new circuit-design tools, new system architectures, and a generation of chip-design automation aimed at letting small teams design custom silicon at speeds previously available only to the largest fabless companies. In September 2018, DARPA followed with a parallel two-billion-dollar AI Next campaign. The agency that had once funded the integrated circuit was openly preparing for an era in which chip and algorithm advanced together or not at all.

The architecture all of this added up to was clear enough by the end of the Trump administration that the incoming Biden Pentagon could simply pick up where it had been left. Kathleen Hicks, Biden’s nominee for deputy secretary, had served in the Pentagon under Carter and spent the intervening years at CSIS writing about the problems Work was writing about. She was confirmed in February 2021. She inherited the JAIC, which she would soon fold into a new Chief Digital and Artificial Intelligence Office. She inherited DIU and SCO. She inherited a defense industrial base that had spent five years internalizing the proposition that AI and autonomy were not adjuncts to traditional warfare but the form of warfare itself. What she did not have was an organizing program pointed at a single problem.

She announced one in late August 2023, at the National Defense Industrial Association’s annual Emerging Technologies for Defense conference in Washington. The audience was the same kind of room Work had addressed in 2015. The program was Replicator, and its design was a direct descendant of every argument Work had made since 2014. The People’s Liberation Army, Hicks told the room, had spent twenty years building itself into a force whose decisive advantage was mass: more ships, more missiles, more launchers, more soldiers. The American answer, for most of the post-Cold War period, had been platforms that were exquisite, expensive, and few. Replicator would change that. The Pentagon would, within eighteen to twenty-four months, field thousands of attritable autonomous systems across air, sea, and ground. We will counter the PLA’s mass with mass of our own, Hicks said, but ours will be harder to plan for, harder to hit, and harder to beat.

The phrase that mattered, hidden in plain sight, was attritable. An attritable weapon was one cheap enough to be lost without flinching. Cheap enough was a function of the chip inside it. The compute the swarms required, the perception the autonomous systems needed, the coordination across thousands of platforms, all of it ran on the same kind of silicon Work had described in 2015 and the NSCAI had written about in 2021. Replicator was the first defense program in American history sized to the production rate of the commercial chip industry.

By the time Hicks announced it, the connection had stopped being controversial inside the building. American national security depended on AI. American AI depended on compute. American compute depended on chips no American company could fabricate. The chip war Trump had begun to wage against Huawei in 2019 and 2020, and that Biden would extend dramatically against the Chinese AI ecosystem in October 2022, was the same war as the AI war. Each was the other, viewed from a different angle of the supply chain.

This was the conceptual hinge. The Second Offset had tied American military power to American silicon at a time when the United States made the world’s silicon. The Third, by the time it was finished being designed, had tied American military power to the global compute frontier at a time when the United States no longer did. Perry, in 1977, had been able to take the foundation for granted. Work, in 2015, had not. The fabs that built the chips that ran the models that flew the drones the Pentagon now considered decisive lived in places the Pentagon could not control. Closing that gap, by manufacturing or by chokepoint, would become the dominant project of American chip policy for the rest of the decade.

Work, looking back from his post-Pentagon perch at CNAS in 2020, told an interviewer he worried the United States might be losing the AI race for lack of the urgency the moment required. He was old enough to remember what urgency looked like. He had read, in his time on Marshall’s distribution lists in the late 1990s, the unclassified parts of the Second Offset’s archive. He knew what Brown and Perry had spent and built and gambled to get the Gulf War’s silicon onto the airframes in time. He suspected, in the quiet way he had of suspecting things in print, that the next chapter would not be written by the Pentagon at all. It would be written by the foundries.