What Makes AI Infrastructure Financeable in 2026?
AI infrastructure is no longer bankable on a large MW headline alone. The strongest projects increasingly combine credible power, contracted or anchor demand, and capital that is explicitly available for infrastructure delivery.
AI infrastructure has entered a phase in which access to capital is abundant for some platforms, but delivery capacity is not. The distinction matters. A project can announce hundreds of megawatts, publish an ambitious construction schedule and attract strategic interest, yet still remain far from financeable if power, demand and capital are not aligned.
Three public developments in early September illustrate the point from different angles. Firmus announced that OpenAI will become an anchor customer for dedicated AI compute at two AI Factory sites in Malaysia. Mistral announced a €3 billion Series D and stated that part of the capital will be used to scale compute capacity and expand infrastructure. In Italy, Greenfield and Finsbury Infrastructure disclosed a 36 MW IT data centre project east of Milan, with up to €360 million of investment and 80 MW of secured power capacity reported for the site.
Together, these developments show a useful framework for evaluating AI infrastructure: power, demand and capital must converge. When only one or two are present, the project may still be real, but its delivery risk is materially higher.
The market is moving beyond headline megawatts
Announced megawatts remain useful, but they are no longer enough to measure executable supply. A 100 MW headline may describe a long-term campus concept, a grid application, a funded project or customer-backed capacity. Treating those states as equivalent creates false precision.
PowerlandMap therefore separates reported capacity from development evidence. The same logic underpins our broader approach to data centre supply intelligence: the question is not simply how many megawatts have been announced, but how much capacity has credible evidence behind it and what milestones still stand between the project and service availability.
The recent Firmus announcement is a strong example. Firmus states that OpenAI will contract dedicated AI compute capacity from two Malaysian AI Factory sites and that its total contracted capacity across all customers now exceeds 900 MW. Crucially, that 900 MW figure is portfolio-wide. It should not be reclassified as Malaysian capacity. The public announcement does not disclose the MW allocation of the two Malaysian facilities.
Reported portfolio capacity, site-specific capacity and contracted customer demand are different data points.
Pillar one: credible power
The first bankability test remains power.
The Milan project announced by Greenfield and Finsbury Infrastructure is particularly instructive because two power numbers are reported separately: 36 MW of IT capacity for the proposed data centre and 80 MW of secured power capacity for the site.
Those numbers should not be collapsed into one metric. The 36 MW figure describes the announced IT programme. The 80 MW figure describes the reported power position. The gap may reflect redundancy, future phases, infrastructure design or simply a larger secured envelope than the initial IT build. Without additional public evidence, those interpretations remain hypotheses.
What matters commercially is that the project is not being presented solely as an unpowered development concept. The partners are explicitly highlighting secured power as a development milestone.
That is increasingly important across Europe. Grid queues, reinforcement requirements and long lead times mean that the difference between “power requested”, “power offered”, “power contracted” and “power deliverable by a specific date” can determine whether a site belongs in an investable pipeline or a longer-term watchlist.
For a deeper explanation of this distinction, see What Makes a Data Center Site Power-Ready in 2026? and The Grid Connection Is No Longer the Whole Power Strategy.
Pillar two: contracted or anchor demand
Power without demand can create a valuable powered-land position. It does not automatically create a financeable AI factory.
The second pillar is therefore evidence of customer demand.
Firmus' announcement is unusually clear on this point. OpenAI is identified as an anchor customer under a multi-year strategic partnership for dedicated compute capacity at two Malaysian sites. Firmus also states that the relationship supports additional deployment across its AI Factory platform.
This changes the project evidence: a named global customer has contracted capacity rather than a developer merely forecasting regional AI demand.
Anchor demand can materially improve the underwriting case by reducing a core uncertainty: whether expensive new capacity will find a user at the required economics. But a customer announcement does not prove the exact MW, contract value, power delivery date or construction status of each site. Those remain diligence items.
This is why PowerlandMap treats explicit unsatisfied buyer demand separately from contracted offtake. Once capacity is contracted, it should no longer be presented as an open buyer requirement in a matching dataset.
Pillar three: capital that is actually available for infrastructure
The third pillar is capital.
Mistral's €3 billion Series D is notable because the company explicitly links the raise to compute capacity and infrastructure, not only to model research or software expansion. Mistral states that the financing will help scale compute capacity for training, expand infrastructure and accelerate international growth.
That makes the financing relevant to physical AI infrastructure.
At the same time, it would be incorrect to convert the €3 billion raise into a data-centre CAPEX number. Mistral has not publicly allocated the entire round to buildings, power, cooling or compute hardware. The capital supports multiple uses.
This distinction between funding capacity and project CAPEX is important. A company can have a strong balance sheet and still need project-specific financing, leases, equipment facilities or infrastructure partnerships. Conversely, a project can have committed project finance even if its sponsor has raised relatively little corporate equity.
PowerlandMap's methodology therefore separates transaction and capital events from project-level CAPEX whenever the public evidence does not support a direct allocation.
The bankability triangle
The most useful way to compare projects is to view power, demand and capital as a triangle.
Power + capital, but no demand
This can support speculative development, especially where powered capacity itself has strategic value, but commercialisation risk remains.
Power + demand, but insufficient capital
A strong customer case can still fail if construction, electrical equipment, cooling or compute procurement cannot be funded.
Demand + capital, but uncertain power
Capital and users may be available while grid delivery remains the gating item, forcing phased energisation or alternative power strategies.
Power + demand + capital
This is the strongest combination. It does not remove delivery risk, but it materially improves the probability that announced capacity becomes deliverable capacity.
That is the distinction the market should increasingly focus on.
Reported capacity is not the same as deliverable capacity
The market will continue to produce very large numbers. Some will represent real commitments. Others will represent potential.
The analytical task is not to discount every large announcement. It is to classify it correctly.
A useful capacity stack can include:
- announced capacity
- requested grid capacity
- secured or reserved power
- contracted power
- deliverable power by date
- announced IT capacity
- customer-contracted compute capacity
- operational capacity
Each layer answers a different question.
This is also why “Up to 1 GW” should not be treated as a single supply number. A long-term campus ceiling is not equivalent to the first phase that can be energised, financed and occupied.
What investors and developers should diligence first
Diligence should trace power milestones, land control, permitting, capital commitment and customer status separately. Developers need to identify which missing pillar is blocking the next stage, while end users need to test whether power dates, construction programmes and expansion rights match their deployment schedule.
PowerlandMap's coverage framework is built around these differences rather than a single headline capacity field. The objective is to compare markets and projects on evidence that is relevant to actual deployment.
PowerlandMap view
The following is PowerlandMap's analytical view, not a reported fact: the next phase of AI infrastructure investment will increasingly reward projects that can demonstrate convergence rather than scale alone.
The most valuable development milestones are likely to be those that reduce one of the three principal uncertainties:
1. power becoming contractually and physically deliverable 2. demand becoming contracted rather than forecast 3. capital becoming committed to a defined infrastructure programme
A project that improves one of those dimensions can move materially closer to bankability even if its headline MW number does not change.
That is why market intelligence needs to track events, not just assets.
From market map to investment decision
Investors and developers need a live view of how supply changes as projects progress through power, planning, financing and commercial milestones.
Static project lists age quickly. Development evidence does not.
PowerlandMap combines market intelligence, structured supply and demand data, power-readiness signals and matching to help investors, developers, operators and end users distinguish announced capacity from capacity that is becoming executable.
For teams evaluating a new market, a site pipeline or a portfolio of AI infrastructure projects, the objective should be to answer three questions early:
Is the power credible? Is the demand evidenced? Is the capital aligned with delivery?
If all three answers are strong, the project deserves a different risk classification from one that merely has a large MW headline.
For a live view of supply, power readiness and development evidence, request access or review the PowerlandMap platform.
Sources
*Matthieu Gallego — Founder, PowerlandMap*
The data behind the analysis
Every analysis is grounded in the tracked dataset
Qualified supply, anonymised demand and evidence-based matching sit behind each read — with the source, confidence level and verification date on every record
Related analysis
Reserved Matters: The Milestone Between Planning and Buildable Data Center Supply
Why outline consent, reserved matters, enabling works and construction should never be treated as one category of data center supply — and what recent UK filings reveal about real development progress.
What “Up to 1 GW” Means for Data Center Supply
Gigawatt-scale AI campuses are built in phases. Here is how to separate ultimate capacity, committed development, power readiness and real near-term supply.
The Grid Connection Is No Longer the Whole Power Strategy
Vertiv’s acquisition of UtilityInnovation Group shows why data center value is shifting from nominal megawatts to deliverable, orchestrated and economically sustainable power.