OpenAI Just Became Firmus’s Anchor Customer Ahead of a Potential IPO. We Saw This With SB Energy. How Many Valuations Can the Same AI Demand Support?
Nvidia-backed Firmus is valued above $10.5 billion and has signed OpenAI for two Malaysian data centers just as it considers what could become one of Australia's largest IPOs. Only two of Firmus's seven planned AI data centers are operational. Days earlier, SB Energy filed for an IPO with its entire contracted data-center capacity concentrated in OpenAI and SoftBank. The question is no longer whether AI infrastructure demand is real. It is how many companies, projects, countries and valuations can capitalize the same concentrated pool of future demand before the end customer has generated enough economic value to support them all.
There is a pattern emerging across the global AI infrastructure boom.
A data-center company secures a major frontier AI customer.
Contracted capacity jumps.
The company's growth narrative strengthens.
Capital follows.
Valuation rises.
An IPO approaches.
Then the same frontier customer appears somewhere else.
Another provider.
Another country.
Another infrastructure buildout.
Another valuation.
Another capital raise.
And increasingly I find myself asking:
How many different investments are ultimately claims on the same future AI demand?
OpenAI Just Became an Anchor Customer of Firmus
Reuters reported on September 8 that Australian AI infrastructure company Firmus signed a multi-year agreement with OpenAI for dedicated computing capacity at two data centers in Malaysia.
OpenAI becomes an anchor customer.
The agreement increases Firmus's contracted capacity across customers to more than 900 MW.
Firmus was valued above $10.5 billion in its latest fundraising and is backed by investors including:
Nvidia,
Jane Street,
Blackstone,
and Coatue.
And there is another important detail:
The agreement arrives immediately ahead of a rumored Firmus IPO that Reuters says could become one of Australia's largest listings in recent years.
Firmus declined to disclose the value of the OpenAI contract.
The company currently has seven AI data centers across its portfolio.
Only two are operational.
Five remain under development.
That distinction matters.
Because once again investors are being asked to put present value on:
future AI infrastructure.
We Just Saw a Version of This With SB Energy
Days earlier, another AI infrastructure company moved toward public markets.
SB Energy.
Reuters reported that SB Energy could seek a valuation above $50 billion.
But its AI infrastructure economics contained an extraordinary concentration.
Its entire contracted 8.8 GW of data-center capacity was tied to just:
OpenAI
and
SoftBank.
At the time of its IPO filing, SB Energy was not generating operating revenue from data centers.
Its existing revenue came from its legacy energy business.
Yet the company had accumulated a data-center backlog worth hundreds of billions of dollars.
Again:
future capacity.
Future utilization.
Future revenue.
Future AI demand.
Present valuation.
Firmus Is Not SB Energy
That distinction is important.
Firmus already operates data centers in Australia and Singapore.
Its contracted customer portfolio extends beyond OpenAI.
The company has also announced relationships involving other large technology customers and infrastructure partners.
So this is not a claim that Firmus has only one or two customers.
It doesn't.
The pattern is different.
The pattern is that OpenAI has again become an anchor demand source immediately before another major AI infrastructure company approaches public markets.
That deserves attention.
This Is Global AI Valuation Replication
Global AI Valuation Replication
Global AI Valuation Replication occurs when the future compute demand of a relatively small number of frontier AI companies supports valuations, financing and infrastructure development across multiple nominally independent companies, assets and jurisdictions.
Imagine one company expects to require enormous amounts of compute.
That demand supports:
a data-center developer in America,
another in Australia,
another in Malaysia,
an energy company,
a GPU provider,
a power supplier,
a cloud provider,
a networking company,
and eventually multiple IPOs.
Every business is different.
Every contract may be legitimate.
Every asset may ultimately be needed.
But economically, many can be making claims on:
the same underlying growth assumption.
AI demand becomes enormous.
One Customer Can Create Multiple Valuations
Think about OpenAI.
Its future demand can support:
OpenAI's own valuation,
a cloud company's revenue,
Nvidia's chip revenue,
an energy company's backlog,
a data-center operator's contracts,
a utility's projected electricity demand,
a private-credit investment,
an infrastructure fund,
and an IPO valuation.
The same customer appears across the chain.
This does not make the investments illegitimate.
It makes them correlated.
And correlation is the part investors can easily miss.
Different Companies Do Not Necessarily Mean Different Demand
An investor may own:
Nvidia,
an AI data-center company,
an energy developer,
an infrastructure fund,
and a cloud provider.
That portfolio looks diversified.
But suppose:
OpenAI,
Anthropic,
Meta,
Microsoft
and a handful of other hyperscalers represent a large share of the future demand supporting all five investments.
Then the investor owns:
different securities
but potentially the same economic thesis.
That is AI Demand Concentration Risk.
AI Demand Concentration Risk
AI Demand Concentration Risk occurs when apparently diversified AI investments depend disproportionately on spending commitments from the same small group of frontier AI companies or hyperscalers.
That's why customer concentration matters.
Not simply inside one company.
Across the entire ecosystem.
The Same Demand Can Cross Continents
This is where Firmus adds another dimension.
Australia-based company.
Malaysian data centers.
American AI customer.
American GPU provider.
Global institutional investors.
Potential Australian IPO.
That is an astonishingly international capital chain.
AI demand originates in one place.
Infrastructure gets built somewhere else.
Capital comes from somewhere else.
Power is consumed somewhere else.
Public investors may eventually purchase the equity somewhere else.
This is AI demand becoming globally portable.
Call This Cross-Border AI Demand Export
Cross-Border AI Demand Export
Cross-Border AI Demand Export occurs when frontier AI companies project infrastructure requirements into foreign jurisdictions, allowing their expected future compute demand to stimulate local data-center construction, capital formation and asset valuations outside their home country.
There are benefits.
Investment.
Jobs.
Infrastructure.
Technology transfer.
Digital development.
But there are also externalities.
Malaysia Absorbs the Physical Consequence
Reuters reports Malaysia is now Southeast Asia's fastest-growing data-center market.
The Firmus expansion places OpenAI compute in two Malaysian facilities.
But data centers require something financial models sometimes make look deceptively abstract:
physical resources.
Electricity.
Water.
Land.
Grid infrastructure.
Cooling.
Fuel.
Reuters reported separately on September 8 that Malaysian data-center electricity consumption reached 9.3% of total power consumption during extreme August heat.
The normal 2026 average has been around 7%.
And data centers could consume as much as 31% of Peninsular Malaysia's electricity by 2035.
That is extraordinary.
The AI valuation may sit in:
Australia,
America,
or global capital markets.
The electricity demand sits in:
Malaysia.
This Is AI Infrastructure Externalization Risk
AI Infrastructure Externalization Risk
AI Infrastructure Externalization Risk occurs when the economic upside from AI infrastructure is distributed across technology companies and investors while a host jurisdiction absorbs a disproportionate share of the physical costs—including electricity demand, water consumption, grid expansion, land use and environmental pressure.
That does not mean Malaysia receives no benefit.
It clearly can.
But investors need to understand the complete economics.
Someone ultimately has to build:
power plants,
transmission,
water systems,
roads,
and supporting infrastructure.
AI Demand Does Not Arrive Alone
The compute contract may say:
900 MW.
But 900 MW is not simply a financial number.
It is physical load.
Power infrastructure.
Generation.
Cooling.
Backup capacity.
Land.
And potentially water.
So when a global AI company says:
we need more compute,
another country may hear:
we need more electricity.
That distinction matters.
The Valuation Travels Faster Than the Infrastructure
This is another recurring AI failure pattern.
The deal is announced.
Immediately.
The valuation moves.
Immediately.
Investor expectations move.
Immediately.
The IPO narrative improves.
Immediately.
But the underlying infrastructure can require:
years.
Firmus says five of its seven AI data-center sites remain under development.
That is not unusual.
Data centers take time.
But this creates another mismatch:
Contract-to-Capacity Gap
The Contract-to-Capacity Gap occurs when future AI capacity is financially recognized through contracts, valuations or fundraising materially before the physical infrastructure required to deliver that capacity becomes operational.
The contract exists.
The capacity may not.
Yet capital markets can begin valuing the future revenue immediately.
Contracts Are Not Utilization
This distinction matters enormously.
Contracted megawatts tell us that someone agreed to reserve capacity.
They do not automatically tell us:
how much compute will actually be consumed,
how profitable the workloads will be,
how consistently infrastructure will operate,
what the cost of power will become,
or whether end customers will generate sufficient economic value from the resulting AI.
That is where infrastructure demand and economic demand diverge.
We Need Demand Provenance
I have previously written about Demand Provenance.
The question is not simply:
Is there demand?
It is:
Where did the demand originate, who ultimately pays for it, and what economic activity sits at the end of the chain?
For AI infrastructure, the chain might be:
OpenAI contracts capacity
→ Firmus builds data center
→ Nvidia supplies GPUs
→ utility supplies power
→ infrastructure investors provide capital
→ Firmus potentially goes public
→ public investors assign valuation.
But the chain still needs an endpoint.
Someone eventually has to pay OpenAI enough money for AI services to justify the entire structure.
The Final Customer Has to Pay Everyone
This is the part of the AI investment boom that deserves more attention.
OpenAI can buy compute.
Firmus can buy Nvidia chips.
Malaysia can expand electricity generation.
Investors can fund Firmus.
Banks can finance infrastructure.
But ultimately there must be economic activity at the end of the system producing enough value to pay for:
the AI service,
the model,
the GPU,
the data center,
the electricity,
the financing,
and the investor return.
Everyone cannot simply pay the person directly upstream forever.
Eventually:
the end customer must create economic value.
That Is Where Circularity Risk Appears
The AI ecosystem increasingly contains relationships where companies simultaneously act as:
investors,
customers,
suppliers,
financiers,
and strategic partners.
Nvidia backs Firmus.
Firmus buys Nvidia infrastructure.
OpenAI contracts Firmus capacity.
OpenAI needs Nvidia compute.
Investors value Firmus partly based on that contracted demand.
Nothing about those relationships inherently means something improper is happening.
But they make demand harder to interpret.
How much demand reflects:
independent customer economics?
How much reflects:
strategic capacity reservation?
How much reflects:
competitive fear?
How much reflects:
supplier-supported expansion?
How much is financed by participants elsewhere in the same ecosystem?
That is why AI Demand Circularity Risk matters.
Nvidia Appears Across More and More of the Stack
Look at the Firmus structure.
Nvidia is:
an investor in Firmus
and
the provider of the Vera Rubin processors Firmus plans to deploy throughout Asia-Pacific.
That can create powerful alignment.
Nvidia benefits if Firmus expands.
Firmus benefits from Nvidia's technology and endorsement.
OpenAI obtains compute.
Investors see blue-chip counterparties.
Again:
there is nothing inherently improper in that structure.
But investors should understand how many participants derive value when infrastructure expands.
Expansion Itself Can Generate Financial Value
This is the unusual feature of the current AI boom.
Build more data-center capacity:
Nvidia sells more chips.
Energy companies sell more power.
Data-center operators get larger contracted backlogs.
Infrastructure investors deploy more capital.
AI companies secure more compute.
Valuations can rise across the chain.
This can all happen before the final AI customer has demonstrated equivalent economic productivity.
That is important.
This Is AI Infrastructure Reflexivity
AI Infrastructure Reflexivity
AI Infrastructure Reflexivity occurs when expectations of future AI demand stimulate infrastructure investment that generates revenue, contracts and valuations across the AI ecosystem, which in turn reinforces investor confidence in the original demand forecast.
Expected demand creates infrastructure.
Infrastructure spending creates revenue.
Revenue validates the AI boom.
Validation creates more investment.
Investment creates more infrastructure.
The cycle can be productive.
It can also become reflexive.
Now Add IPOs
This is where the pattern becomes especially important.
SB Energy:
AI infrastructure narrative.
OpenAI anchor demand.
Massive backlog.
IPO.
Firmus:
major AI infrastructure expansion.
OpenAI anchor customer.
$10.5+ billion valuation.
Rumored IPO.
Anthropic itself:
IPO preparations.
Multiple infrastructure providers:
financing and expansion.
At some point, public investors become the next source of capital.
That's where private AI expectations begin entering ordinary portfolios.
This Creates Anchor-Customer IPO Amplification Risk
Anchor-Customer IPO Amplification Risk
Anchor-Customer IPO Amplification Risk occurs when signing a marquee frontier AI customer shortly before a financing event or public listing materially strengthens an infrastructure company's valuation narrative even though the economic durability, utilization and diversification of the resulting demand remain unproven.
Again:
this does not mean the contract is artificial.
OpenAI clearly requires extraordinary compute.
The risk concerns:
how much valuation investors attach to the contract.
A Famous Customer Can Become a Financial Signal
There is enormous signaling value in being able to say:
OpenAI is our anchor customer.
Or:
Anthropic is our customer.
Or:
Meta.
Or:
Microsoft.
Investors may infer:
validation,
future revenue,
technical credibility,
and long-term demand.
The customer therefore creates value beyond the contract itself.
It creates:
valuation signaling.
How Much Is the Name Worth?
Imagine two identical data-center businesses.
One announces:
900 MW of contracted capacity.
The other announces:
900 MW, with OpenAI as an anchor customer.
Would markets value them equally?
Probably not.
The OpenAI name creates:
credibility.
Expectation.
Scarcity.
AI exposure.
That means the customer itself becomes part of the valuation mechanism.
This Is Marquee-Customer Valuation Transfer
Marquee-Customer Valuation Transfer
Marquee-Customer Valuation Transfer occurs when the perceived strategic value of a major AI company transfers partially to suppliers and infrastructure partners simply because those companies secure commercial relationships with the frontier firm.
OpenAI's valuation story can therefore help support:
another company's valuation story.
Then that company can potentially enter public markets.
That's how AI valuation can propagate.
The Same Future Gets Monetized Across Multiple Balance Sheets
This may be the most important point.
Future AI demand can become:
OpenAI valuation.
Nvidia revenue.
Firmus contracted capacity.
SB Energy backlog.
Utility demand forecast.
Data-center financing.
Private-credit assets.
Infrastructure-fund returns.
IPO proceeds.
Different balance sheets.
Different countries.
Different investors.
But potentially the same underlying future.
That's AI Valuation Stacking at global scale.
How Many Times Can the Same Future Be Sold?
Not fraudulently.
Financially.
The future can legitimately support multiple businesses.
Every iPhone sale once supported:
Apple,
chipmakers,
telecom companies,
app developers,
component manufacturers,
logistics companies,
and retailers.
Economic ecosystems naturally distribute value.
The question is whether the final AI economy becomes large enough to support the amount of financial value currently being assigned before that economy fully exists.
That is the distinction.
Apple Had the End Customer
This is one reason comparing AI with previous technology revolutions is useful.
Apple sold:
hundreds of millions of devices
to identifiable consumers
who paid identifiable prices.
The revenue came through an observable transaction.
Much of AI infrastructure investment is occurring further upstream.
Companies are building capacity based on expected future:
agent use,
enterprise adoption,
model inference,
AI automation,
and consumer demand.
Those economics are still developing.
The infrastructure race is arriving before the demand picture is completely mature.
The Infrastructure Can Be Real While the ROI Is Wrong
This is crucial.
A data center can be:
real.
The GPU can be:
real.
The power consumption can be:
real.
The contract can be:
real.
And investors can still get the economics wrong.
Because the issue is not whether the physical asset exists.
It is whether its eventual cash flows justify:
the purchase price,
the financing,
and the valuation.
That's an AI Failure
This belongs squarely inside AI Failure Intelligence.
The model does not need to hallucinate.
The agent does not need to escape.
The system does not need to malfunction.
The failure can occur because:
capital markets overestimated how much independent economic demand existed underneath the infrastructure expansion.
That is a financial AI failure.
Now Multiply It Across Countries
America.
Australia.
Malaysia.
Europe.
The Middle East.
Asia.
Countries increasingly want:
AI data centers,
AI factories,
AI sovereignty,
AI investment,
and participation in the next technology platform.
That creates political competition for infrastructure.
Governments may offer:
land,
tax incentives,
power access,
regulatory accommodation,
or infrastructure support.
Frontier AI demand can therefore move internationally and unlock additional capital wherever it lands.
That is another multiplier.
Global AI Infrastructure Arbitrage
Global AI Infrastructure Arbitrage occurs when AI companies and infrastructure developers can move future compute demand across jurisdictions to access favorable combinations of power, land, incentives, capital and regulation while the resulting infrastructure supports valuations across multiple national markets.
Again:
this can be efficient.
It can also obscure where the real economics sit.
Malaysia Is Not Merely Hosting Servers
If Malaysia becomes one of the world's fastest-growing AI data-center markets, it is making a national resource allocation decision.
Electricity used by data centers is electricity that must be:
generated,
transmitted,
and paid for.
If data-center demand rises toward 31% of Peninsular Malaysia's electricity consumption by 2035, then AI infrastructure becomes part of:
energy policy,
industrial policy,
water policy,
and national economic strategy.
That means the downside is not confined to investors.
What If the Future Demand Is Overestimated?
Suppose AI demand ultimately grows more slowly than expected.
The infrastructure remains.
The power system expansion remains.
The financing remains.
The land-use decisions remain.
The public incentives remain.
The country cannot simply say:
undo the data center.
That is Host-Country AI Stranding Risk.
Host-Country AI Stranding Risk
Host-Country AI Stranding Risk occurs when a country commits energy, land, infrastructure or public support to projected AI demand that later fails to generate sufficient utilization, economic output or domestic benefit to justify the resources allocated.
This is why foreign AI investment should be evaluated through more than:
headline investment dollars.
Who Captures the Value?
Malaysia should ask:
How many jobs remain?
How much tax revenue remains?
How much technology transfer occurs?
Who owns the data centers?
Who owns the chips?
Who owns the models?
Who owns the customer?
Where do the profits go?
Who finances the power infrastructure?
Who bears electricity-price consequences?
Who bears water constraints?
Those questions determine whether:
AI infrastructure investment
becomes
AI economic development.
They are not automatically the same thing.
The Global AI Race Is Exporting Demand Before Proving Value
This may be the broader pattern.
Frontier AI companies are projecting future demand outward.
Into:
power systems,
mineral supply chains,
GPU financing,
data centers,
real estate,
utilities,
sovereign policy,
and public markets.
The future AI economy is therefore already reshaping the present physical economy.
That is remarkable.
It is also risky.
Because the infrastructure is being built before we know precisely how valuable the final AI economy becomes.
The Strategic Questions
Investors, governments, boards and infrastructure providers should now ask:
How many infrastructure companies rely materially on the same frontier AI customers?
How diversified is demand across economically independent end users rather than across contracts?
How much OpenAI demand has already been capitalized into third-party valuations?
How much Anthropic demand?
How much Meta and hyperscaler demand?
Do different infrastructure investments actually represent different risk—or repeated exposure to the same customer ecosystem?
How much of Firmus's contracted 900+ MW is operational today?
How much remains dependent on future construction?
How much additional capital is required before the five facilities under development become operational?
What percentage of projected revenue depends on a small number of anchor customers?
How much valuation premium does securing OpenAI add before utilization begins?
Why do major anchor-customer announcements repeatedly appear near major financing or IPO events?
Are public investors being asked to finance the conversion of private AI demand expectations into physical infrastructure?
How much supplier financing or strategic investment comes from companies that themselves benefit from increased infrastructure construction?
How much AI demand is independently generated versus reinforced within the same ecosystem?
What happens to data-center valuations if frontier-model compute requirements become more efficient?
What happens if enterprises demand better ROI before expanding AI spending?
Who absorbs stranded infrastructure if utilization disappoints?
What benefits do host countries receive relative to the electricity, water and infrastructure they commit?
How many apparently diversified AI investments ultimately depend on the same future dollar of AI productivity?
And the largest question:
How many times can the same future AI demand be turned into present valuation before somebody at the end of the chain has to produce the cash flow that pays everyone back?
Strategic Conclusion
The OpenAI-Firmus agreement is NOT evidence of a global investor money heist.
It is evidence of something more useful for investors to understand:
AI demand has become a globally transferable financial asset.
One frontier customer's expected compute demand can support:
a Malaysian data center,
an Australian infrastructure company's valuation,
Nvidia GPU sales,
energy investment,
private equity,
and potentially an IPO.
Then the same customer can appear in:
Ohio,
another data-center provider,
another energy company,
another infrastructure project,
and another capital structure.
The assets are different.
The jurisdictions are different.
The counterparties are different.
But the economic assumption underneath them can be remarkably similar:
AI demand will become enormous.
That is Global AI Valuation Replication.
It creates AI Demand Concentration Risk when multiple investments ultimately depend on a small number of frontier customers.
It creates a Contract-to-Capacity Gap when contracted demand receives financial value before physical infrastructure is operational.
It creates Anchor-Customer IPO Amplification Risk when marquee contracts strengthen valuation narratives near liquidity events.
It creates Marquee-Customer Valuation Transfer when OpenAI's strategic importance increases perceived value elsewhere in the ecosystem.
It creates AI Infrastructure Reflexivity when anticipated demand stimulates construction, construction produces revenue, and that revenue becomes evidence that the original AI-demand forecast was correct.
And when infrastructure crosses borders, it creates AI Infrastructure Externalization Risk and Host-Country AI Stranding Risk.
None requires fake contracts.
None requires fraud.
None requires conspiracy.
That is exactly why these risks matter.
Every individual actor can be behaving rationally.
OpenAI wants compute.
Firmus wants customers.
Nvidia wants chip sales.
Malaysia wants investment.
Private investors want returns.
Public markets want AI exposure.
Each transaction can make sense independently.
And collectively the financial system can still create far more claims on future AI productivity than the eventual AI economy can support.
That is the bird's-eye view investors need.
Because a contract can be real.
A data center can be real.
A GPU can be real.
A customer can be real.
And the valuation can still be wrong.
The question is no longer:
Is AI demand real?
Clearly, some of it is.
The better question is:
How many companies have already been valued as though the same future demand belongs uniquely to them?
I write about AI failure intelligence, ROI exposure, high-stakes decision architecture, and the hidden pathways through which AI incidents become financial and institutional consequences.
Follow me and subscribe to my work if you are responsible for investing in, acquiring, governing, insuring, or protecting strategically important AI systems and need to understand what technical failure can become after it leaves the engineering team.
Comments