AI has already proven it works. The question now is where it becomes impossible to remove.
In our manifesto we drew a line between AI that automates workflows and AI that restructures the physical and institutional systems it touches. We called the second category Essential AI, and proposed a simple test for it: try to remove the intelligence from the system. If the system collapses, the AI is essential. If it doesn't, you have a feature.
We invest across three pillars: sovereign infrastructure, physical AI, and biology. Each connects directly to a substrate the world already runs on: energy and compute, industrial and autonomous systems, and human health. These pillars are also interdependent. Sovereign programmes need energy and compute to train the physical AI systems that operate in healthcare, agriculture, and defence. Robotics companies need simulation environments. Bio AI companies need sovereign compute to process genomic datasets under national data sovereignty requirements. Every layer feeds the others. That reinforcing dynamic is what makes this a thesis, not a theme.
Sovereign Infrastructure
Nations are not adopting AI. They are stockpiling the ability to produce it, and the capital tells the story.
Saudi Arabia's HUMAIN secured $1.2B and is acquiring several hundred thousand Nvidia chips over five years. The UAE is deploying $8-10B per year through MGX. France is building a €30-50B data centre campus. The UK committed £18B. The EU's InvestAI initiative pulled in €200B in commitments.
These are procurement programmes that need suppliers. Compute architectures, energy systems, security frameworks, edge deployment platforms. Companies selling into this market don't need product-market fit in the traditional sense. Their market is being created by policy, and their buyers are governments.
Australia is quietly becoming a meaningful node. AWS committed $20B AUD to Australian data centres. Google signed a three-year Defence cloud agreement. The National AI Plan allocated over $460M, with industry bodies estimating $2-4B is needed for sovereign compute.
AUKUS is tying AI infrastructure directly to defence partnerships. In our pipeline, one Australian company is replacing the attention mechanism in large language models with a memory-inspired architecture that eliminates the computational bottleneck and runs on sovereign edge hardware where transformers simply cannot. That's the kind of enabling layer these programmes must acquire.
Physical AI
Travis Kalanick built Atoms in total stealth for eight years. Thousands of employees. Food production, mining, industrial transport. Not a single press release until the technology was ready.
That patience is unusual in venture, but it's becoming the norm in physical AI.
Figure AI reached $39B in under three years. Skild AI raised $1.4B at $14B from SoftBank, Nvidia, and Bezos. 1X and EQT are deploying 10,000 NEO androids across 300+ portfolio companies. At GTC 2026, Jensen Huang announced the GR00T N2 foundation model and partnerships with ABB, FANUC, KUKA, and Universal Robots. The message was clear: every industrial company will become a robotics company.
Physical AI requires fundamentally different companies from AI software. Hardware-software integration. Real-world operational data that cannot be synthesised. Domain expertise measured in decades. You can't assemble these capabilities around an API. You build them over years, in labs and on factory floors, with founders who were working on these problems long before the current hype cycle.
Australia has more of these founders than most people realise. CSIRO Robotics placed second at the DARPA Subterranean Challenge against teams with orders of magnitude more funding. The University of Adelaide's AIML is globally competitive in computer vision. Hardware and robotics topped both funding and deal counts in Australian VC for the first time in Q3 2025. The Defence Trailblazer programme is channelling research into quantum, hypersonics, robotics, and autonomous systems. The pipeline is real.
Bio AI
The World Health Organisation projects a global shortage of 11 million health workers by 2030. Recruitment won't close that gap. AI will.
The science is outrunning the capital. Isomorphic Labs (DeepMind spinout) raised $600M and secured partnerships with Eli Lilly and Novartis worth up to $3B. Their Drug Design Engine doubled AlphaFold 3's performance. Insilico Medicine hit Phase 2a with a fully AI-designed therapy, compressing drug discovery from ten years to 18 months. The FDA published its first comprehensive guidance on AI in drug development in January 2025.
Biology is a domain where you can't fake competence. Regulatory moats compound. The consequences of failure are measured in human lives. That combination creates the deepest defensibility in all of AI and the highest barrier to new entrants. Exactly the dynamic you want at seed.
Australia's research depth in computational biology, agricultural science, and health systems feeds directly into this pillar. Proximity to the fastest-growing healthcare markets in the Asia-Pacific makes it a natural launchpad for Bio AI companies building for global scale.
Why Early Access Is Structural
In 2025, 50% of all venture capital raised went to companies in AI-related fields, making artificial intelligence the leading sector for funding, as it was for the past three years. That concentration is accelerating.
Look at how the strongest Essential AI companies raise. Figure AI's $675M Series B was led by Microsoft and Nvidia. Skild's $1.4B came from SoftBank and Bezos. Isomorphic's $600M came from Thrive Capital and Alphabet. When these companies raise their next rounds, a16z, General Catalyst, and Nvidia's venture arm will take them whole. There will be no allocation left for new investors showing up with a $200M cheque and an interest in the space.
The only way in at growth stage is to have been there at seed. Founders choose their earliest investors with care, because those investors shape the company. Domain expertise, operational support, regulatory navigation, bridges to global capital markets. Access is earned.
And here's the uncomfortable corollary: if a growth-stage Essential AI company is still accepting capital from investors who weren't there at the beginning, it usually means the tier-one funds passed. Late access is an antifilter.
The founders building Essential AI aren't just looking for capital. They're looking for escape velocity. They're domain experts (physicists, roboticists, computational biologists) who were working on these problems long before the current wave. AI multiplied their possibilities by orders of magnitude, but it didn't multiply the hours in a day. Every hour spent on fundraising or investor management is an hour not spent on the science.
We built Unlock Capital for exactly this problem. We syndicate the first cheque so founders get the capital table they need without the overhead of managing it. We open the pathway to US Series A through our co-investor network. And we hand founders the same AI operating platform we built for ourselves: skills, automations, and compounding improvement loops that let a two-person team run like a tier-one operation. The operational drag that kills deep-tech companies between seed and Series A disappears.
Capital, conviction, and a system that compounds both. So they can do the thing that matters most. Build.