Scenario III: AI, Dual-Use & Defence 2026

Present day

The Capital Influx

There has been a massive influx of venture capital into defence and security startups. Funding rounds for defence-focused startups have risen sharply, and this is happening as global military spending on AI accelerates.

629 VC transactions in defence tech in 2024, up from 414 in 2020¹

$29bn in round values in 2025 — nearly triple the 2020 total¹

$38.8bn projected global military AI spending by 2028, from $4.6bn in 2022²

Most VC capital is going into battlefield autonomy,⁵ with Ukraine having demonstrated the potential for drone technology to transform warfare, proving that cheap and scalable autonomous systems can rival traditional platforms.

In parallel, VC investment decisions are moving at breakneck speed, with some firms deploying capital into defence tech within 24 hours without deep due diligence. There is growing concern that the “tech bro” culture of Silicon Valley — which historically breaks rules and skips steps — is now colliding with the Pentagon, potentially eroding democratic checks and balances.

Present day

Four Categories of Use

Broadly speaking, four major categories of use case for AI in defence exist. They carry very different risk profiles — and the capital is not flowing to the safest ones.

I · Battlefield / kill chain autonomy

AI systems that operate on the battlefield — crewless drone swarms, ground vehicles and submarines, AI-assisted fire control systems. Deployed for surveillance and combat operations, they reduce the time between threat detection and response. They may be designed to be capable of independently searching for, selecting and killing a target without human input, though this is not standard practice yet.³

Opportunities

Faster decisions than human reaction times
Reduced risk to personnel through unmanned and semi-autonomous operations
Coordinated swarm operations across air, land and maritime domains
Scalable at lower unit cost than traditional platforms

Risks

Loss of meaningful human control over lethal force, creating gaps in accountability
Adversarial “spoofing” and sensor deception, leading to misidentification
Escalatory dynamics — the speed of autonomous engagement may cause volatile escalation
Unclear legal framework under International Humanitarian Law for autonomous targeting

II · Intelligence, surveillance and reconnaissance

AI applied to the collection and analysis of intelligence — processing satellite imagery, signals intercepts and sensor data. These algorithms aid target identification, tracking, object recognition and situational awareness. A notable example is Palantir and Project Maven, which fuses open-source, signals intelligence and satellite data to identify targets.⁴

Opportunities

Processing of data at scale, leading to increased precision
Reduced analyst burden, freeing specialists for higher-order judgement

Risks

Model bias may produce systematic blind spots or misclassification
Over-reliance on AI outputs reduces capacity to evaluate their accuracy
Data poisoning and adversarial inputs can corrupt intelligence
Privacy and civil liberties risks if capabilities are turned inwards to surveil a country’s own citizens

III · Predictive maintenance and logistics

AI which optimises maintenance and logistics for armies — forecasting component failures before they occur, or optimising supply chains under contested conditions. These applications are the most practically mature application of AI in defence, helping the military operate more efficiently and save resources.⁶

Opportunities

Optimised supply chains under contested and austere operating conditions
Reduced sustainment costs and more efficient use of spare parts inventory
Most concrete near-term ROI, with mature data and well-defined success metrics

Risks

Cybersecurity vulnerabilities, given that logistics networks are high-value targets
Dependency on data quality; poor training data produces flawed predictions
Skills atrophy if AI judgement displaces human expertise

IV · Simulation, wargaming and decision support

AI that helps model adversarial behaviour, stress-test operational plans, and support commanders in complex decision-making, including large-scale scenario generation that allows faster and more accurate campaign analysis.²

Opportunities

More rigorous campaign analysis, stress-testing numerous plans before execution
AI-driven adversary models that can adapt and surprise, improving exercise realism
Real-time decision support that surfaces options and second-order effects
More accurate wargaming that demonstrates the full extent of losses may incentivise leaders to de-escalate and prevent wars altogether

Risks

Scenario bias; AI trained on historical data may not model novel adversary behaviour
Commanders may anchor decisions to AI recommendations, reducing strategic creativity
Overconfidence if simulation outputs are treated as prediction rather than exploration
Classification and data security risks when sensitive order-of-battle data feeds models

Mid 2026

Keeping the Human in the Loop?

Speed and accuracy gains from AI create accountability gaps that existing legal and doctrinal frameworks haven’t been built for. The risks are most acute where the task involves lethal or irreversible decisions; lawyers and ethicists warn that autonomous weapons will make warfare less humane and more volatile.⁶

The need to keep humans in the loop is commonly raised as a way to mitigate these issues.⁷ But even when a human is technically in the loop and manually approving targets, the process is so fast and complex that humans may be deciding “without cognitive clarity” or awareness of the full range of information available.⁶ It is therefore hard to discern when human involvement is meaningful, rather than perfunctory.

While AI may improve decision-making over the status quo, its increased usage may also cause an atrophying of human capabilities to judge and evaluate its output. Research has shown that humans are highly susceptible to automation bias, meaning they tend to over-rely on a machine’s guidance even when it is flawed or incorrect.⁸

Expand: what is data spoofing?

Where malicious actors intentionally manipulate real-time inputs to disable, distort or deceive an AI system’s functionality. For example, an adversary could take a civilian object, such as a lamppost, and print visual patterns on it that would convince an autonomous weapon it is looking at a legitimate military target.

Late 2026

A Vacuum Where Regulation Should Be

A lack of regulation has enabled the pace of developments in this area, which in turn makes it harder for regulators to keep up. The US, which has the power to set norms globally, has shown resistance and even hostility towards regulating AI applications in defence technology — with the Pentagon pressuring firms to remove ethical guardrails on AI tools, and threatening to cut ties with firms like Anthropic for refusing to allow unconstrained use of their models for military targeting.⁹

Against the context of the current US administration and existing geopolitical tensions between the US and China, forecasters see it as highly unlikely that the two powers will formally ratify any kind of agreement specifically addressing the military use of AI in the next couple of years.¹⁰ In fact, many experts expect a NATO member state to use autonomous weapons within the next two decades, citing increased geopolitical competition and the erosion of norms.

2027 → 2028 · Branch point

Where, Then, Do We Go From Here?

Three futures follow. The first is a hot war in which capital compounds into an arms race and then strands. The second contains conflict through precision and contractual guardrails. The third redirects capital toward deterrence entirely. Choose one to read first — all three are below.

Hot war

2028 → 2030

AI Begets AI

Geopolitical tensions continue to rise, and more countries are dragged into conflict. VC investment into defence technology reaches new highs, with investors prioritising speed-to-market over safety testing. The use of AI tools becomes widespread in warfare.

AI on one side begets AI on the other,⁶ and a global arms race materialises without safety guardrails. As armies use AI to identify and hit targets with increasing rapidity, adversaries are forced to turn to AI to keep up. In response to skyrocketing demand, defence tech companies produce massive arsenals of AI-powered military technology.

This introduces a set of misaligned incentives for political and industry leaders, whose financial returns now depend on the continuation and escalation of war rather than deterrence and de-escalation. Due to these vested interests, conflicts become increasingly prone to intensification.

Hot conflict breaks out internationally. The value of non-defence consumer portfolios collapses. Capital flows are channelled into defence, militarising the economy, inflating input costs, and leaving general investments as stranded assets. State and industry converge, and governments enforce strict, nationalistic principles on the tech workforce — denying visas to workers from adversarial countries — dismantling the vibrancy that made the tech industry competitive in the first place.

2031 → 2035

Two Ways Down

From here the path splits again, depending on whether investors and policy-makers make choices that prevent the most extreme existential-level outcomes — while not being enough to prevent catastrophic harm. The choices that leave Scenario A open while preventing Scenario B are narrow:

Avoiding investment in fully autonomous lethal systems. No AI is given autonomous authority over nuclear weapons.
Mandating some level of explainability for high-risk decisions involving escalatory strikes.

Scenario A · Guardrails hold, barely

Large-Scale Civilian and Infrastructure Casualties

Increasing reliance on AI systems for targeting causes civilian casualties to rise dramatically. AI models fail to recognise hostile intent in nuanced edge cases requiring human judgement — an AI cannot distinguish between a combatant with a real gun and a child holding a toy gun, or a wounded soldier attempting to surrender from an active sniper.¹¹

Data spoofing and poisoning, combined with broader cyber attacks, corrupt the inputs targeting AIs rely on. An adversary feeds false sensor readings into the system, causing it to identify its own civilian convoys as hostile, or to ignore real threats. As casualties mount, commanders authorise broader targeting parameters. Legal review becomes cursory, or is waived entirely.

Then a hacking event — perpetrated by an adversarial state or a non-state actor — causes a drone swarm to turn on its own troops and civilians, destroying the standard of living for an entire region overnight. The world realises that no nation’s AI weapons are immune to being used against them. AI has greatly reduced the barrier to entry for catastrophe: where a nuclear weapon requires specialised infrastructure, AI is relatively cheap and accessible.

This serves as the traumatic catalyst for major powers to end global conflict. The global community agrees sweeping treaties banning military AI. The massive influx of capital into defence technology becomes misspent capital, as valuations and business models are wiped out overnight — leaving VCs and LPs holding portfolios of stranded assets and facing irreversible reputational damage, and delivering a systemic shock to a global economy already on its knees.

Scenario B · Guardrails absent

Nuclear War and Extinction

In a rapidly escalating situation, the legal system removes accountability of commanders. With no human accountability for war crimes committed by AI, commanders deploy AI on the battlefield with total impunity.

Humans abdicate increasing amounts of decision-making power to AI. Commanders, suffering from automation bias or fearing they are losing a rapidly developing hot war, give AI full power to respond to and execute strikes. AI systems show a tendency towards unpredictable escalatory behaviour — a King’s College London study found AI models opt for nuclear signalling in 95% of simulated war games¹² — and escalation spirals out of control. Nuclear strikes are authorised by AI systems in an automated decision faster than humans can intervene. An extinction event occurs.

Contained

2028 → 2035

Guardrails Written Into Contracts

Investors realise the severity of the risks in developing defence tech without adequate guardrails. They take action to prioritise cybersecurity, explainability and transparency in the design of the defence tech they invest in, and require portfolio startups to:

Fund independent red teams to perform adversarial training, where dedicated teams of hackers stress-test the system during training and ensure it ignores false data designed to cause misfires in a cybersecurity attack.
Prioritise algorithmic transparency to develop explainable AI systems that can show why a target was chosen — and the limits of the technology — allowing human overseers to audit the decision logic.
Include specific human-in-the-loop requirements in contracts with military procurers, with those clauses flowing down to subcontractors, ensuring humans meaningfully evaluate outputs at the deciding stage rather than merely being present.
Deploy trainers to support military units, teaching end-users how to use the technology and how to evaluate its outputs.

This results in fewer civilian casualties, more precise military targeting, and a lower probability of uncontrolled escalation. Conflicts remain contained. The technology augments its operators rather than replacing them, reducing the harm inflicted by armed conflict rather than multiplying it. VC returns are sustainable, because the technology is deployable without triggering bans or reputational collapse.

Deterrence

2028 → 2030

Rising Public Concern

International bodies, civil society organisations, and military practitioners and experts raise increasing concern around the safety of deploying AI for war, leading to heightened public attention. Citizens alarmed by the prospect of their pensions being used to fund lethal weapons place increasing pressure on asset managers to disclose their policies on investing in defence technology.

Asset owners begin to view the unchecked development of defence technology as a sizable systemic risk. Recognising the significant reputational and legal risks of investing in dual-use technology — and viewing it as necessary to avoid a race to the bottom that would undermine global stability — LPs apply pressure on VCs to enforce more rigorous safety and human rights requirements on their portfolio companies.

2031 → 2033

Peace Tech and Counter-Drone

VCs avoid investing in systems capable of lethal force, diverting capital towards less risky but still profitable forms of military technology compatible with geopolitical stability. Working with civil society and policy experts, investors establish shared frameworks connecting investment activities directly to outcomes like conflict deterrence.

Guided by these frameworks, a category of “peace tech” develops — AI tools that support conflict prevention and peacekeeping, including wargaming and negotiation co-piloting that rely on AI’s probabilistic modelling to simulate the outcome of wars with increased accuracy. This “geopolitical chess” allows world leaders to see so many steps ahead that they recognise conflicts as lose-lose scenarios, incentivising them to de-escalate and prevent wars altogether.

Investors also channel capital towards counter-drone technology, a currently underdeveloped field that leaves cities and infrastructure defenceless against drone swarms. As defensive capabilities reach parity with offensive drone technology, this significantly strengthens deterrence and helps prevent escalation.

When funding AI used for the battlefield, investors mandate compliance-by-design — international law and ethical norms embedded directly into the model’s architecture from day one, rather than retroactively enforced — and require developers to undergo international law and ethics training. VCs also enforce Responsible Use policies requiring KYC due diligence, end-use monitoring, and voluntary human rights assessments, including maintaining red flag lists of countries where the technology might be used to violate international humanitarian law.

2034 → 2035

A New Mode of Deterrence

As geopolitical tensions are appeased, a form of global stability centred on deterrence allows defence companies to continue to grow and profit on annual recurring revenue models — selling software subscriptions for intelligence and targeting — without the economic destruction of a hot war. Governments allocate a portion of GDP to military tech, continuously buying, upgrading and testing equipment without deploying it in conflict.

AI wargaming tools lead to greater international communication, coordination and anticipation, forming new de-escalation protocols that reduce the likelihood of a hot war occurring again. Wargaming also helps discover new ways to pre-empt, avoid and exit violent conflict faster, should war break out again.

These innovations, while rarely deployed on the battlefield, spill over into non-defence ecosystems — benefiting sectors like climate tech, automated mining and commercial logistics. This gives defence technology companies another revenue stream, further divesting their interests from warfare.

Side by side

The Futures Compared

The same variables across all four outcomes — including both branches of the hot war. The investor-relevant difference is not capability but control.

Variable	Hot war · A	Hot war · B	Contained	Deterrence
Investor behaviour	Speed-to-market over safety testing; avoids fully autonomous lethal weapons; mandates explainability for escalatory strikes.	Unconstrained capital deployment; no safety measures mandated; fully autonomous lethal weapons funded.	Funds ISR capabilities; recognises the risk of inadequate guardrails; prioritises cybersecurity, explainability, transparency.	LP pressure enforces safety and human rights requirements; capital diverted to wargaming, negotiation co-piloting and counter-drone; compliance-by-design; KYC and end-use monitoring.
Human control	Eroding — legal review becomes cursory.	Eliminated — commanders deploy AI with impunity.	Human-in-the-loop contractually mandated, flowing down to subcontractors; end-users trained on the systems.	Capital redirected to decision-support technologies where human control is inherent, ensured by explainable and transparent AI.
AI accuracy & security	Fails on nuanced edge cases; weak cybersecurity measures.	Extremely weak: systems given full authority to authorise strikes while prone to false warnings and nuclear escalation.	Improved, due to precision gains.	High — systems adept at predictive logistics and wargaming, use cases that don’t aggravate tensions and help incentivise de-escalation.
Geopolitical state	Hot war, aggravated by the scale of civilian casualties.	Hot war, rapidly escalating out of control.	Conflict, but contained by fewer civilian casualties.	Tensions appeased; a new “cold war” centred on deterrence rather than active conflict.
Point of failure	Mass civilian and infrastructure casualties; a catastrophic cybersecurity event.	Automated nuclear strikes; extinction event.	—	—
Outcome	Catastrophic event triggers sweeping bans; defence valuations wiped out; reputational and economic shock.	Civilisational-scale collapse.	AI augments operators rather than replacing them; lower probability of uncontrolled escalation; VC returns sustainable.	Stability becomes the business model — recurring revenue without a hot war, plus spillover into climate tech, mining and logistics.

Whichever path

Takeaways for Investors

The future direction of AI’s impact is not inevitable. The decision tree remains broad and, at many critical points, is shaped by the decisions of innovators and capital allocators. There are strategic investable opportunities within defence tech that are compatible with peace:

Administrative AI

Predictive maintenance — saving the US Air Force $25m a month by predicting when A-10C warplanes need repair¹¹ — plus logistics optimisation and human resources. These tools can be profitable in industries beyond defence, such as mining.

Wargaming

Rigorous wargaming and negotiation co-piloting can produce meaningful returns by reducing the number of analysts needed and generating novel war plans on the fly, such as DARPA’s SCEPTER programme.¹¹ AI may help commanders recognise conflicts as lose-lose scenarios.

Cybersecurity & red teaming

Military AI relies heavily on data and software, introducing massive vulnerabilities to cyber-intrusion, hacking and data poisoning. A growing need exists for startups securing these systems and identifying deepfakes, disinformation and poisoned datasets before they are ingested by military intelligence.

Counter-drone technology

Defensive capabilities have not kept pace with drone advancements. There is opportunity in counter-drone tech, including directed energy lasers and signal jamming.

Practical actions

Embed the law in the architecture. Ensure international law and ethical norms are built into the AI model’s architecture from day one, and mandate international law and ethics training for AI developers in portfolio companies.
Fund explainability. Back explainable and transparent AI systems to make algorithmic decisions understandable to human operators, countering automation bias. Make the limits of the system clear to end-users.
Put humans in the contract. Encourage portfolio companies to include specific human-in-the-loop requirements in contracts with military procurers, ensuring humans meaningfully evaluate outputs.
Vet the capital. Vet co-investors and LPs to ensure they are not taking “adversarial capital” from shell companies linked to hostile nations.
Require KYC and end-use monitoring. Require portfolio companies to conduct Know Your Customer due diligence, end-use monitoring, and voluntary human rights assessments.
Audit after deployment. Mandate continuous performance tracking, re-evaluation and auditing post-deployment — military AI can be unpredictable in real-world scenarios, and developers must retrain and retest iteratively after launch.
Build the shared standards. Collaborate with other investors and civil society to develop taxonomies and outcome metrics for peace, security and resilience.

References

Sources

All sources below are retained from the original Reframe Venture analysis.

S&P Global (2026), Venture capital investment in defense tech surges while M&A activity slows.
Belfer Center (2025), Code, Command and Conflict (PDF).
United Nations, Lethal autonomous weapon systems, UNODA.
CSIS, What is the Maven Smart System and what does it do?
New Market Pitch, Defense tech funding trends.
The Economist (2024), How AI is changing warfare.
Amoroso & Tamburrini (2020), Autonomous weapons systems and meaningful human control, Current Robotics Reports.
Defense One (2026), Military AI and troops’ judgement.
CNN (2026), Hegseth, Anthropic and military AI.
Forecasting Research Institute, LEAP report, wave 5.
The Economist, How is AI changing warfare? (Inside Defence).
King’s College London, Artificial intelligence under nuclear pressure.