SECRET · SUBJECT TO REDACTIONFAB·BOOK / CHAPTER 48 / FOLDER OFFSET-3CLASSIFIED IN PRACTICE; DECLASSIFIED IN REPORTING
EYES ONLY
THE THIRD OFFSET — A PROGRAM MAP
PORTFOLIO · OSD/CIO · DARPA · CDAO
Robotics · Autonomy · Compute · 2014 → 2024
How the Pentagon learned to buy its silicon from someone else.
The Second Offset, in 1977, attached American military power to American silicon at a
time when the United States made the world’s silicon. The Third Offset, by the time
it was finished being designed, attached American military power to the global compute
frontier at a time when it no longer did. Below: the public ledger of programs that
the Pentagon spun up between 2014 and 2024 to do something about that.
I
Eisenhower · 1950s
First Offset
Nuclear weapons as the answer to Soviet conventional mass in Europe. Tactical and strategic warheads counterbalanced army-group-level land power.
Wager: Atomic primacy
II
Carter · 1977
Second Offset
Perry, Brown, and Marshall’s bet on precision strike, stealth, and silicon. Vindicated by Iraq, 1991. Anchored to American semiconductor leadership.
Wager: Silicon + sensors
III
Hagel/Work · 2014
Third Offset
Robotics, autonomy, big data, miniaturization, advanced manufacturing. Wagered on AI/compute that lived in commercial GPUs designed in California and fabbed at TSMC in Hsinchu.
Carter’s bridge between the Pentagon and Silicon Valley — small, fast contracts
with technology startups whose products the Department needed and could not buy
through the normal acquisition process. Rebooted under Raj Shah in May 2016. Anduril,
Palantir-adjacent firms, geospatial-intel and signals-intel startups absorbed early.
Raj Shah, ex-F-16 pilot & Palo Alto cybersecurity founder
OPERATIONALCONTRACTING
$400M
cumulative est. through 2023
26 Apr 2017
Maven
Project Maven
Algorithmic Warfare Cross-Functional Team
The first operational deployment of modern computer vision inside DoD.
Took full-motion video from Predator/Reaper drones and applied ML models
so analysts could be told, in near real time, where the trucks and people
and buildings were in the frame. $9M Google contract canceled June 2018
after 3,000+ Google employees signed an internal letter. Contract reassigned
to Palantir. Program persists.
Lt. Gen. Jack Shanahan (USAF), founding director
CONTROVERSIALPERSISTENT
$70M
FY17 initial budget
27 Jun 2018
JAIC
Joint Artificial Intelligence Center
Started: 4 volunteers, no money · absorbed into CDAO Feb 2022
Patrick Shanahan’s memo established the Pentagon’s single point of AI contact for
industry. Lt. Gen. Jack Shanahan took over from Maven in December 2018 and arrived
to four volunteers and zero budget. Eighteen months later: 185 staff, $1.3B
budget. Merged into the Chief Digital and Artificial Intelligence Office in
February 2022 under Kathleen Hicks.
Lt. Gen. Jack Shanahan; merged into CDAO under Craig Martell
OPERATIONAL → CDAO
$1.3B
cumulative through 2021
DOSSIER 02The agency that midwifed the ICDARPA / MTO
June 2017
ERI
Electronics Resurgence Initiative
Five-year program · Microsystems Technology Office
Moore’s Law was running out of steam at the leading edge. The next era
would be defined by specialization — domain-specific architectures designed for
particular workloads, including AI. Funded research on new materials, circuit-design
tools, 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.
William Chappell, MTO director
DARPAFY17 — FY22
$1.5B
5-year, ~$300M/yr
Sept 2018
AI Next
DARPA AI Next
Companion campaign to ERI · the agency that funded the IC, preparing for the era after it
$2 billion announced as a parallel push to ERI. The agency that had once funded the
integrated circuit was now openly preparing for an era in which chip and
algorithm advanced together or not at all. Funded contextual reasoning,
common-sense AI, third-wave reasoning systems, and a slate of programs intended
to do for AI what the LSI program had done for VLSI in the 1970s.
Co-chaired by Eric Schmidt (ex-Google CEO, Defense Innovation Board chair) and
Bob Work (ex-DepSecDef). The report’s argument was structurally simple:
modern AI depends on compute; compute depends on advanced semiconductor
manufacturing; the United States no longer manufactures the most advanced
semiconductors. It recommended doubling federal non-defense AI research
spending each year — $2B in 2022 to $32B by 2026 — refundable investment tax
credits for domestic fabs, and a national strategy to stay two generations ahead
of China at the leading edge.
Eric Schmidt (chair), Bob Work (vice-chair); 15 commissioners
RECOMMENDED→ CHIPS ACT
$32B
non-defense AI R&D / yr by 2026
28 Aug 2023
Replicator
Replicator Initiative
All-domain attritable autonomous systems
Hicks announced at the NDIA Emerging Technologies for Defense conference: within
18–24 months, field thousands of attritable autonomous systems across air,
sea, and ground. The PLA had built itself into a force whose decisive advantage
was mass; Replicator was the American answer.
“We will counter the PLA’s mass with mass of our own, but ours will be
harder to plan for, harder to hit, and harder to beat.”
The first Pentagon program in U.S. history sized to the production rate of the
commercial chip industry.
Kathleen Hicks, Deputy Secretary of Defense
CURRENTFY24+
~$500M
reorganization of existing funds
The compute that drives the new offset does not live in the United States.
Every program above runs, ultimately, on silicon. The accelerator chips that train the
models, the inference parts inside the autonomous systems, the SoCs inside the
attritable hardware — none of them are fabricated on American soil at the leading
edge. The Second Offset attached American military power to American silicon.
The Third Offset attached it to a global compute frontier in which the
leading-edge nodes lived in places the Pentagon could not control. Closing
that gap, by manufacturing or by chokepoint, became the dominant project of
American chip policy for the rest of the decade.
Disclosed program spend
~$5.4B
Across DIU, JAIC/CDAO, Maven, ERI, AI Next, Replicator, 2014–2024.
NSCAI target by 2026
$32B/yr
Non-defense federal AI R&D recommended.
Leading-edge fabs in the U.S.
0 → 1 (TSMC AZ)
Phase-1 production ramp 2024. Two generations behind Hsinchu.