For twenty years, software taught investors to love asset-light businesses. Build the product once. Distribute it globally. Keep capex low. Let margins expand with scale.
AI changes that. Inference needs physical GPUs, physical sites and physical power. Technology companies are suddenly dealing with a question that energy, telecom and infrastructure developers have dealt with for decades: what should sit on the company balance sheet, and what should be financed as infrastructure?
That distinction matters, because not all capital is equal.
This is how Main Street has been built for decades: Asset-light does not mean asset-free
This is where the term is most often misunderstood. A company can depend heavily on physical infrastructure and still run an asset-light corporate model.
This is how Main Street has been built for decades. A power plant, fibre network or toll road is typically not financed as though it were SaaS. The project is placed into a dedicated legal vehicle. The assets sit there. The project contracts sit there. The financing sits there. The revenues flow there.
Lenders then ask a fundamentally different question from a venture investor: can the project's contracted cash flows support the project's capital structure? If they can, the project can increasingly finance itself, through the capital appropriate to its stage and risk: vendor financing or equipment leasing, then infrastructure equity, asset-backed lending, private credit, institutional debt, and ultimately non-recourse project financing.
There can still be sponsor equity, guarantees, completion support and other protections. “Ring-fenced” does not magically make risk disappear. It allocates risk to the capital designed to price it.
The point most people miss
Asset-light does not mean avoiding heavy assets. It means not financing every heavy asset with the most expensive capital available.
Venture equity is designed to fund uncertainty and exponential upside: software, IP, distribution, customer acquisition, network effects. Using that same equity indefinitely to purchase depreciating hardware can be extremely inefficient once the hardware's economics have become separately underwritable.
So the basic rule is: equity buys what compounds; infrastructure capital funds what depreciates.
The software company can own the IP, orchestration, customer relationship and commercial platform. The physical capacity can sit in ring-fenced vehicles funded by investors seeking infrastructure returns. The technology company gains access to much more capacity without continuously returning to shareholders for equity simply to buy more equipment.
Why AI makes this particularly interesting
The physical infrastructure behind AI has characteristics that traditional infrastructure capital can increasingly understand. It produces revenue. It can serve multiple customers. Utilisation can be measured. Capacity can be contracted. Hardware has a secondary market and residual value. And software can improve the productive output of installed hardware over time.
NVIDIA makes this point well with CUDA: an installed GPU does not necessarily produce the same economic output throughout its lifetime. Improvements in software, kernels and inference optimisation can let the same hardware deliver more useful work. That is quite different from assuming a GPU becomes economically worthless when the next generation arrives.
None of this makes GPUs risk-free infrastructure. Technology obsolescence, utilisation, customer concentration, residual value and power costs all still matter. But those are underwritable risks. And that is what creates an asset class.
The market is now moving this way
This is no longer theoretical. There is a fair criticism of some AI financing first: take the NVIDIA and OpenAI model, where a supplier invests in a customer that then uses significant amounts of that capital to purchase the supplier's equipment. It is fair to ask whether part of the economics have become circular. That does not make the underlying demand unreal. But it is also not the best example of what I mean. There are cleaner structures.
Meta's Hyperion project with Blue Owl puts most of the physical infrastructure into a separately capitalised joint venture rather than requiring Meta to own 100% of it directly.
CoreWeave has begun financing AI infrastructure against the assets and associated customer contracts through large non-recourse facilities backed by institutions including Blackstone, with banks such as MUFG, Morgan Stanley, Goldman Sachs and JPMorgan involved.
The common idea is simple: separate the financing of the productive physical asset from the equity financing of the technology company.
And this week, NVIDIA moved much more clearly in this direction. Jensen Huang announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish independent financing platforms intended to mobilise more than $500 billion of third-party capital into AI infrastructure over time.
The headline number is impressive, but the structure is more interesting. NVIDIA provides the AI-factory technology platform. The financing institutions independently underwrite the customer, utilisation, cash flow and residual value. Long-term institutional capital funds the physical infrastructure. That is much closer to traditional infrastructure finance.
NVIDIA even addresses the circular-financing criticism directly: rather than simply financing customers itself, the intention is to bring independent infrastructure investors into the market. NVIDIA may provide limited residual-value support in individual transactions, but the investment decision remains with the capital provider.
The conversation changes from “who has enough equity to buy all these GPUs?” to “can this infrastructure generate sufficiently predictable utilisation and cash flow for specialist capital to finance it?” That is exactly the question infrastructure investors have always asked. As Huang put it: in AI, compute is revenue.
Where the software value comes from
There is another misconception worth addressing. If a business needs hardware, it does not follow that the hardware is where the enterprise value must sit.
Hardware by itself is available to many companies. The differentiated layer is what determines where a workload runs, which hardware executes it, how capacity is pooled, how latency is managed, how utilisation is increased, how sovereignty and policy requirements are enforced, how customers receive an SLA, how heterogeneous infrastructure is abstracted behind one interface, and ultimately how raw compute becomes an inference product.
The infrastructure makes the software credible. The software makes the infrastructure operate as a network. And once that network exists, the company does not necessarily need to organise every asset itself.
Eventually, the more interesting metric becomes less “how many GPUs do you own?” and more “how much inference flows through your platform?” That is where the economics start moving from infrastructure ownership toward infrastructure orchestration.
How we are building CYANIS
This has been the financing architecture behind CYANIS since January 2026. We need physical GPU capacity. There is no pretending otherwise. But the goal was never to fund Europe's AI infrastructure indefinitely from a venture cap table.
CYANIS venture equity funds the software, IP, team, orchestration platform and commercial layer. Physical capacity is to be financed separately through vendor leasing, infrastructure equity and, once the cash flows support it, asset-backed and non-recourse project financing.
The first CYANIS-organised fleet establishes the network: capacity, operating history, utilisation data, customer trust and a dependable service. Over time, we expect the platform to orchestrate more third-party infrastructure than capacity we organise ourselves. That is the destination.
The fleet establishes the network. The software scales it. Infrastructure capital finances it.
Asset-light development is not about pretending infrastructure does not exist. It is about knowing which part of the business you should own, and which part you should finance. This week suggests the rest of the industry is arriving at the same answer.