There is a story being told about AI that goes roughly like this. Artificial intelligence will reshape the economy. The most valuable companies of the next twenty years will be the ones that build, train, and deploy intelligence at scale. The physical economy is becoming a commodity layer beneath an intelligent one. The future is software, and software is eating the world.
It is half right.
AI is reshaping the economy. The companies that build and deploy intelligence will absorb extraordinary value. But the conclusion that follows from those two facts, that the physical world becomes less valuable as intelligence becomes more abundant, is wrong. The opposite is true. As intelligence is commoditised and labour markets are reorganised at unprecedented scale, the systems that intelligence acts on, the food, the energy, the materials, the manufactured goods, the physical infrastructure, become more valuable, not less.
That is the inversion at the heart of the MAD thesis. And it is the inversion most allocators have not yet priced in.
Start with the labour-market shift
Begin with the most concrete observation, the one that does not require any aggressive assumption about how fast artificial general intelligence arrives. Even on the most conservative consensus, the economic deployment of AI over the next decade will produce the largest labour-market reorganisation since the urbanisation that accompanied industrialisation. You do not have to believe the most aggressive timeline. The moderate consensus alone is enough to drive everything that follows.
What does that reorganisation actually do to the demand for the things humans need? The standard story says the answer is "less of everything, because productivity rises and economic output is more efficient". The accurate answer is the opposite. When the economy reorganises at this scale, demand for the basics, food, energy, health, water, the systems that keep populations alive, does not fall. It rises. People still eat. Buildings still need power. Health systems are required by larger ageing populations, not smaller ones. The physical systems that sustain life become more important in absolute terms even as they become a smaller share of nominal GDP.
That is the central paradox of the moment. The most abundant input the economy is producing, intelligence, is being deployed largely in service of itself. AI building AI infrastructure. AI investing in AI startups. AI optimising AI advertising. Meanwhile, the systems that have to keep functioning under the pressure of the reorganisation, the systems the world actually depends on, are receiving a fraction of the capital that flows into the abundant input. That is the misallocation the next decade is going to correct.
The commoditisation of intelligence
The other half of the story sits on the supply side of intelligence itself. The economics of AI are unusual. Foundation-model capability is increasing on a curve that has no historical precedent. Compute costs per unit of capability are falling. Open-source alternatives to proprietary models are catching up with closed leaders within twelve to twenty-four months of release. Specialised models that previously required research-lab capability now run on consumer-grade hardware.
The practical consequence is that intelligence as a service is following the path of every previous infrastructure technology: from scarce and expensive to abundant and cheap. The first electric utilities were extraordinary monopolies. Within a generation, electricity was a commodity. The first internet service providers had pricing power. Within a generation, bandwidth was a commodity. AI is on the same trajectory, on a faster timeline.
When intelligence itself becomes a commodity input, like electricity or bandwidth, the software products built on top of it lose their defensibility. An AI model that costs $100 million to train today will cost $1 million in three years and be open source in five. Every SaaS product, every AI-native startup, every platform built on proprietary intelligence is vulnerable to the next model iteration. The moats drain faster than they fill.
This does not mean all software companies die. It means the venture model of backing dozens of software bets and hoping for one or two unicorns becomes structurally less reliable. Hit rates drop. Exit multiples compress. Holding periods extend. The category itself is being disrupted by the technology it depends on.
Investing in AI application companies today carries echoes of investing in early web portal companies in 1998. Some will survive. Most will be absorbed by the foundation-model providers themselves, who keep adding capability and turning yesterday's startup product into next year's default feature. Look at the list of AI application darlings that defined the venture narrative two years ago and check how many have had their value proposition substantially absorbed by the next foundation-model release. The pattern is consistent. The pattern is also accelerating.
What gets more valuable when intelligence gets cheap
When a critical input becomes cheap, the things it acts on become more valuable. This has been true of every infrastructure transition in modern economic history.
Cheap electricity made factories more valuable, not less. The factories existed before the grid. What electricity did was make them an order of magnitude more productive. The same was true of cheap bandwidth and the businesses it served. Logistics, retail, financial services, all became more valuable, not less, as the connectivity layer became free. The companies that built and operated the underlying physical systems were the durable winners. The infrastructure they ran on top of was the commodity input.
The same dynamic now applies to intelligence. The cheaper intelligence becomes, the more valuable the physical systems it makes more efficient become. A precision agriculture business with access to commodity AI is dramatically more productive than the same business without it. A waste-conversion plant with continuous machine-learning optimisation is more productive than one without. A distributed energy network with intelligent dispatch is more useful than one without. A precision-fermentation business with AI-accelerated process design moves to market faster than one without.
In every case, the AI is the input. The physical system is the output. As the input gets cheaper, the output, the operating physical infrastructure, gets more valuable.
AI is accelerating change across every sector. But it doesn't replace the systems humanity depends on. Food. Energy. Health. Water. Waste. These systems are under pressure and must be rebuilt. This is where value is moving.
From MAD Group platform thesis, 2026
What this means for venture capital
The implication for capital is direct. The companies that will absorb the most value over the next twenty years will not be primarily software businesses. They will be physical, biological, and infrastructure businesses, in food, energy, health, water, materials, and waste, that use commodity AI as a multiplier on their underlying operating system.
They will not look like the companies that drove the last decade's venture returns. They will look like industrial businesses with technology multipliers. They will have physical capital requirements. They will operate in regulated industries. They will scale through deployment of physical assets, not through user acquisition. Their margins will improve as they operate, not as they raise. They will generate cash earlier, distribute earlier, and exit through a wider range of acquirers than the typical venture portfolio.
They will need a different kind of capital than the typical venture portfolio receives. Capital that is patient enough to ride longer development curves. Structured enough to generate income during deployment rather than only at exit. Aligned enough to keep founders building rather than always selling the next round. The 80/20 model that combines structured growth capital with equity participation is built for this profile, because the profile is real and the standard venture model does not serve it.
The other side of the argument
There is a serious case to be made on the other side. Foundation-model providers are absorbing extraordinary value. The largest of them are now among the most valuable companies in history. The compute infrastructure they require, GPUs, data centres, energy, is the largest single capital build-out in human history. Some forecasts put it at $6 to $8 trillion globally by 2030. That capital is going somewhere, and a portion of it will produce extraordinary returns.
The physical AI infrastructure thesis is real. Capital allocated to compute, networking, cooling, and power for the AI build-out will compound. The energy transition itself is partly a function of AI infrastructure demand. The metals, materials, and manufacturing required to build that infrastructure constitute a real-economy thesis with serious tailwinds.
But almost all of that thesis is already priced in by hyperscaler equity, large-cap industrial equity, and the existing infrastructure asset class. It is not where venture capital adds the differentiated return. The differentiated return for venture capital sits in the smaller, faster, more entrepreneurial companies that use the infrastructure rather than build it. Companies that take commodity intelligence and apply it to a real-world problem with operating-cash economics. The MAD platform is built around that thesis.
A wider context
There is also a wider context that is worth naming, even briefly. The AI build-out is not happening in isolation. It is happening at the same time as a measurable repricing of capital, a structural rotation away from speculative software toward real-economy categories, and a moment in which institutional investors are rethinking what their capital should do.
Mark Falzon has written elsewhere, in a separate piece on the broader political and economic context, about the way the AI build-out is being financed largely from public capital pools, pensions, sovereign funds, endowments, while the upside is concentrated in a small number of private companies. The systemic question that raises is one for governments and policy bodies to answer. The investment question that follows is also worth naming. If much of the AI build-out itself is being financed by public capital flowing into a handful of mega-cap private and public companies, then the differentiated return for private investors does not sit there. It sits in the next layer down: the operating businesses that use commodity AI to make real-world systems work better.
That is where the MAD platform is positioned.
What follows from this
A few things follow from the inversion described in this article.
First, exposure to AI-native software companies is more vulnerable than it looks, because the moat assumption underneath those companies is being eroded by the foundation-model layer they depend on.
Second, exposure to physical, biological, and infrastructure businesses that use AI as a commodity input is less vulnerable than it looks, because those businesses become more productive as the AI input becomes cheaper.
Third, the venture model that worked for the last decade is structurally less reliable for the next. The combination of compressed software margins, longer holding periods, and a wider range of physical and infrastructure outcomes requires capital instruments that are better matched to the underlying business profile. Structured capital plus equity. Income during deployment plus upside on growth. Multi-vehicle architecture rather than single-thesis funds.
Fourth, the right place to be allocating venture capital now is to companies that operate physical systems, in the sectors the world depends on, with AI as the input rather than the output. That is the MAD thesis. It is also why MAD is positioned the way it is, structured the way it is, and built around the engines it has.
AI is making the physical world more valuable, not less. The capital architecture that recognises this will compound. The architecture that does not will be the one that gets repriced.
Read more about the thesis in the MAD book in our Books library. Wholesale-qualified investors interested in the Information Memorandum for MAD Fund 1 are welcome to enter the Investor Room.
Information for wholesale clients only. This paper is general commentary and does not constitute financial, tax, legal, investment, or other professional advice. It does not take into account the objectives, financial situation, or needs of any person. It does not constitute an offer of securities or an invitation to subscribe. Any investment opportunity referenced is offered privately and only to wholesale clients as defined under sections 761G and 708(8) of the Corporations Act 2001 (Cth), under separate offer documentation. Past performance is not a reliable indicator of future performance and capital is at risk. Tax positions referenced are based on the manager's understanding of the ESVCLP regime at the date of publication; legislation may change. Prospective investors should obtain their own independent financial, legal and tax advice before making any investment decision. Nothing on this page should be relied on as a substitute for the Information Memorandum and Partnership Deed, available on request to wholesale clients via the Investor Room.