Everyone Wants a Piece of AI. Few Are Looking at the Layer Beneath It.

AI is one of the defining investment themes of this decade. Capital is flowing into model developers, cloud platforms, data centers and the semiconductor supply chain.

That naturally raises a difficult question: if you believe AI adoption will keep expanding, do you have to bet on which AI company wins?

Not necessarily.

Every AI product ultimately depends on physical infrastructure. Models have to be trained. Applications have to run inference. Data has to move through servers, networks and power systems. Behind the software sits a layer of productive hardware that companies either own or rent.

GPUs and other AI accelerators are part of that layer.

The investment thesis is therefore broader than any single model developer or chip generation. The question is who owns the infrastructure that increasingly large parts of the AI economy depend on.

Compute Isn't Just a Training Story

AI training gets most of the attention because the largest training runs require enormous clusters of specialized hardware. But the economics of compute do not stop when a model finishes training.

Inference is the recurring work required every time a deployed model answers a prompt, generates an image, analyzes a document, controls a system or makes a decision.

McKinsey projects that inference will represent more than half of AI compute by 2030 and roughly 30–40% of total data-center demand. That matters because inference is tied to continued product usage rather than a one-time training cycle.

The demand base is also broader than generative AI. Accelerated computing is used across scientific research, drug discovery, simulation, rendering, engineering and other workloads where parallel processing can materially reduce computation time.

That does not mean every workload runs on the same GPU, or that demand will rise in a straight line. Hardware improves, software becomes more efficient and workloads shift between architectures.

What matters is the direction of the infrastructure requirement: more digital intelligence requires physical compute somewhere.

The Numbers Behind the Infrastructure Buildout

The scale of the buildout is easier to see in data-center demand than in speculative market-size forecasts.

McKinsey's 2026 data-center model estimates global demand rising from about 82.3 GW in 2025 to about 219 GW by 2030. In the same analysis, AI inference demand grows from 20.9 GW to 93.3 GW, while AI training demand rises from 23.1 GW to 62.2 GW.

The energy system is seeing the same pressure. The International Energy Agency's updated outlook projects global data-center electricity consumption rising from about 485 TWh in 2025 to around 950 TWh in 2030. Electricity consumption from AI-focused data centers is projected to roughly triple over that period.

NVIDIA's own results offer another view of the same buildout. In its second fiscal quarter of 2027, the company reported $89.0 billion in Data Center revenue, up 117% from a year earlier.

None of those figures guarantees a return for an individual GPU owner. They do show that AI infrastructure is becoming a large and capital-intensive part of the technology economy.

That distinction matters. The thesis is not that every GPU will always be scarce, or that today's rental rates will persist indefinitely. The thesis is that demand for compute is creating a substantial market for infrastructure that can perform useful work.

Owning the Infrastructure Layer

Most public-market AI exposure comes through companies: semiconductor manufacturers, hyperscalers, software businesses or thematic funds.

Those investments can make sense, but they are company-level bets. Their outcomes depend on valuation, management, competition, margins, capital allocation and many other factors beyond demand for compute itself.

Owning a claim on productive compute infrastructure is a different exposure.

The basic economic loop is straightforward:

Capital acquires hardware.

The hardware is deployed.

Customers pay to use compute capacity.

Operating costs are deducted.

What remains is the economic income produced by the asset.

The important words are "productive" and "income."

A GPU does not generate value merely because it exists. Utilization matters. Rental pricing matters. Hosting, power, networking, maintenance and hardware renewal matter. A poorly utilized machine can be a poor investment even in a rapidly growing AI market.

That is why the quality of the operator, the workload pipeline and the economics of the specific hardware matter just as much as the broader AI narrative.

The point is not to eliminate risk. It is to make the source of the economics easier to understand.

Instead of asking only, "Which AI company will win?" an infrastructure investor can ask, "Who is using the asset, what are they paying, what does it cost to operate, and how is that cash flow allocated?"

Those are questions infrastructure investors have asked for decades.

From Physical Hardware to Fractional Ownership

There is still a practical problem: high-performance compute infrastructure is expensive to buy and operate directly.

A single investor may not want to purchase hardware, arrange hosting, manage uptime, source workloads and handle maintenance. At the same time, operators looking to expand productive fleets need capital.

Fractional ownership can connect those two sides.

Instead of requiring one investor to own an entire cluster, ownership can be divided into smaller claims. Tokenization can make those claims easier to record, transfer and settle on-chain.

But tokenization does not create the underlying economics.

The hardware still needs to perform useful work. Customers still need to pay for that work. Costs still need to be accounted for. Ownership records and settlement are only valuable if they remain connected to the productive asset underneath them.

This is the distinction between tokenizing a narrative and tokenizing an asset.

Where IX Finance Fits

IX Finance is building around that second model.

The goal is to connect productive compute infrastructure with an on-chain ownership and settlement layer, while keeping the financial representation tied to the underlying asset and its operating economics.

On the current IX-CORE testnet, users can explore fractional ownership of GPU compute infrastructure on Base. Settlement is in test tokens, and the figures shown are modeled from live market rates for demonstration. Nothing on the current testnet is a live financial product.

That distinction is important to us. We would rather be precise about what exists today than blur the line between a working test environment and production revenue.

The longer-term thesis is simple:

Compute infrastructure performs work.

Customers pay for that work.

Ownership can be divided.

The resulting economics can be accounted for and, where the required legal and technical infrastructure exists, settled on-chain.

We have written separately about how on-chain GPU ownership works and where GPU yield actually comes from. Both lead back to the same principle: the token is not the source of value. The productive infrastructure is.

A Different Way to Think About AI Exposure

AI investing is usually framed as a search for winners.

Which model company?

Which cloud platform?

Which chipmaker?

Which application?

Those are reasonable questions. They are not the only ones.

There is also an infrastructure layer underneath the competition: chips, servers, power, cooling, networking and data centers that have to exist for AI products to run at all.

That layer will change quickly. Today's leading GPU will not remain the leading GPU forever. Rental rates will move. Utilization will vary. New accelerators will compete with old ones. Operators will have to keep upgrading their fleets.

That is precisely why the durable thesis is not one machine.

It is productive compute capacity.

For investors looking at AI from the infrastructure side, that may be the more useful question: not which company captures the most attention, but who owns the machines doing the work.