This Week in Dual-Use
NEWS
Chinese AI firms form industry alliances
China's AI companies have announced two new industry alliances, aiming to develop a domestic ecosystem to reduce dependence on foreign (ie American) technology.
This follows the Trump administration’s recent reversal of its ban on NVIDIA’s H20 chip exports to China, allowing the company to resume sales. The H20 isn’t the most powerful NVIDIA chip, but it’s engineered to be faster than the H100 for certain LLM inference tasks. So why is Trump allowing China to buy it?
During the Cold War the Soviets were so good at espionage that they had almost completely penetrated the US military industrial complex. They knew about every semiconductor development the Americans made, and copied it all.
That put them behind in a way that they have never recovered from. The KGB was so proficient that the USSR never had to cultivate an innovation ecosystem for itself. The consequences of that are still felt today. Instead of Silicon Valley, Russians have Lubyanka.
Constraints on resources drive innovation. For today, China is behind. But it will catch up. If it exports its chips, America could have near peers that never surpass it.
That puts China in a bind. It is already using the H20 and any attempt to create its own version could potentially take years. Yet if it wants to replace the US in global technology dominance, it will have to do exactly that.
US releases its AI action plan
The White House unveiled its AI policy blueprint, Winning the Race: America’s AI Action Plan. The plan outlines over 90 near-term federal actions intended to reshape the US AI landscape, with strategic implications across the public and private sectors. It reflects a broader shift in US industrial policy, where AI is increasingly viewed as a domain of national competitiveness rather than commercial innovation.
The document is wide-ranging but the part that caught my eye was the emphasis on open-source models. The rationale being that startups can use them flexibly, without being dependent on a closed model provider. This is a conscious attempt to lower the barriers to entry in a market increasingly dominated by a handful of companies.
The top American AI models are not open-source. OpenAI had a plan to be - hence the name - but then walked that back as the various risks around weaponisation, disinformation, copyright infringement etc. became obvious.
But China’s top models are open-source. The White House has obviously done the risk assessment and decided that speed of adoption trumps tight control. If you can't (or won't) share the model weights, you risk ceding ground to someone who will.
Cartels flew drones 60,000 times along US border
An average of 328 Mexican drones are apparently coming within 500 meters of the US border every day. In the second half of 2024 US border guards detected 60,000 drone flights just south of the border and 27,000 unique remote-controlled craft.
Cartels have always been early adopters of new technology. In the 1980s Pablo Escobar’s Medellin cartel used satellite phones and rotating radiotelephones, employing frequency hopping to evade interception. They also pioneered semi-submersibles for drug smuggling, and used counter-surveillance to monitor government communications. I recently reported that they are developing autonomous semi-submersibles.
This is good news for developers of counter-drone systems, who perhaps didn’t originally plan for their product to be dual-use. It also shows how quickly the lessons from Ukraine are being disseminated - 60,000 drone flights in 6 months is a lot. Given the price point, we can expect to see further proliferation of the technology.
FUNDRAISING
Project Q, a German company developing a software for battlefield integration, raised a €7.5m seed round, led by Project A with participation from Expeditions.
Noma Security, an American developer of artificial intelligence security tools, raised a $100m Series B led by Evolution Equity Partners.
Promptfoo, an American developer of an open-source framework for AI red-teaming and security testing, raised an $18.4m Series A led by Insight Partners.
Classiq, an Israeli company creating quantum software development platforms, raised an undisclosed amount in Series C extension funding led by SoftBank Vision Fund 2.
Lumana, an American company developing AI video security solutions, raised a $40m Series A led by Wing Venture Capital.
QuamCore, an Israeli developer of a superconducting quantum processor architecture that packs one million qubits into a single cryostat by embedding digital control logic directly inside the chamber in order to overcome scaling limits, raised a $25m Series A round led by Sentinel Global.
AIR, an Israeli startup offering eVTOL aircraft for uncrewed commercial and contested logistics, piloted personal flight, and defense use, raised$23m in Series A funding led by Entrée Capital.
Overwatch Imaging, an American developer of airborne imaging systems and intelligence automation software, raised $6m in funding led by Squadra Ventures.
Polaris Spaceplanes, a German developer of reusable space vehicles, raised €5.3m in top-up funding to its recent seed round co-led by Capnamic Ventures Bremen and Spacewalk VC.
Quadsat, a Danish developer of precise spectrum intelligence and RF geolocation solutions for satcoms and defense, raised a €5m Series A extension led by Join Capital and North Ventures.
Not strictly dual-use, but worth noting that OpenAI just raised $8.3bn at a $300bn valuation. That makes it the third most valuable private company in the world (after SpaceX and Bytedance, valued at $350bn and $300bn respectively).