Firmus & the AI industry - Did Ron Burgundy say it best?

“Well, that escalated quickly.” Will Ferrell’s Ron Burgundy described a street brawl in Anchorman, but the line does a fair job of summing up the AI industry since we last wrote about it. In November 2025, we published a Livewire article, ‘AI Factories – Are They Built on Firm(us) Ground?’, in which we noted that if you mention the letters ‘AI’ to anyone, they will either roll their eyes or start salivating with interest. Fast forward to today, and that still holds. What has changed, and changed fast, is everything around it.

In this content piece, we will attempt to cover the topics below:

• What is an AI factory and how does it differ from a data centre?
• What are the Firmus operations and their associated environmental impact?
• The capital costs of this AI buildout and the explosion in token usage;
• How is the current pricing environment defined?
• What are the returns being generated as we see them today and going forward?
• The circular nature of the financing underpinning the industry; and
• The concentration risk of the associated demand.

AI is certainly dominating headlines around the world, but the blanket AI statements from mid-2025 predicting substantial changes such as 50% of all white-collar jobs to be lost within 5 years have subsided. These have been replaced with a focus on areas such as power usage, risks associated with agentic AI, the debate around open-source vs closed-source models and, of course, the potential for blockbuster global IPOs of both Anthropic and OpenAI. In addition to all of this, the public microscope both globally and certainly within Australia has turned its focus to the topic of data centres & AI factories.

While certainly not alone, one company at the forefront of the Australian media landscape is Firmus Grid Limited (Firmus). Much has been written about Firmus recently, and it is widely understood that the company may be close to undertaking an ASX listing.

The size of such a listing may look insignificant compared to SpaceX Inc. (NDQ: SPCX); however, if a Firmus listing does occur, it will likely be one of the largest domestic IPOs ever. This all remains speculative at this stage. The facts of the recent Firmus capital-raising journey, as we understand them, are below. Irrespective of your views on the company, the size and scale of the recent capital-raising history have been remarkable.

A key part of our investment process at NAOS is spending considerable time understanding the dynamics that underpin the industries in which any potential investment operates.

Below we run through the fundamental concepts underpinning the AI industry as well as Firmus specifically. Ultimately, several of these dynamics will significantly influence Firmus' success (or otherwise) and that of many other companies globally.

SECTION 1 – A 101 REFRESHER ON THE BUILDING BLOCKS OF AI

What Is an AI Token?

Beneath the surface of every AI application, algorithms churn through data in their own language, based on a vocabulary of ‘tokens’. Tokens are tiny units of data created by breaking down larger chunks of information.

AI models process tokens to learn the relationships between them and unlock capabilities such as prediction, generation, and reasoning. Tokenisation is a crucial step in preparing data for further processing.

An AI model processes these input tokens, generates its response as tokens and then translates it to the user’s expected format. The faster tokens can be processed, the faster models can learn and respond. Importantly, tokens are used both when you ask a question and when the AI answers, making it a two-way process.

As AI becomes more advanced and the use of agentic AI has entered the fray, the number of tokens required for AI processes and responses has increased exponentially. As a rule of thumb, 1 token = 3-4 characters in English; ~1 token = 0.75 words. Some AI queries now involve millions of tokens.

AI Training Vs. AI Inference

There are two fundamental phases of an AI model – training and inferencing.

Phase 1 = AI training:

• Training is the process of teaching an AI model by feeding it large datasets to recognise patterns and learn.
• The AI training phase requires massive computational power, most notably silicon chips in the form of NVIDIA GPUs, for this process to occur effectively.
• Training a model from an algorithm often spans days to even months and incurs substantial costs. Furthermore, the computing power necessary to train new iterations of a frontier model (e.g. ChatGPT-6 Astra) is rapidly increasing.

Phase 2 = AI Inferencing:

• This phase is where the trained AI model (from phase 1) applies its learned knowledge to new, real-world data to produce output, such as predictions, classifications, or decisions.
• For every time a model is trained, it may run millions upon millions of inferences before it’s ever trained again. Every time a model guesses the next word, it performs inference. Inference is the true high-volume activity in this new age of generative AI applications, and the stage in which end users typically interact with AI.

At the time of our last article, the story of AI had mostly been about the AI training phase. Whilst that hasn’t slowed down, the narrative has shifted dramatically over the past 6 months. Most of the world’s tokens are no longer used to build AI models. Instead, most AI workloads today have shifted to phase 2: AI inference.

What Is an AI Factory?

To understand AI factories, it helps to break down data centres into two main parts: the physical "shell" (like the secure building and power connections) and the internal IT resources (compute power and networking). In this space, some companies focus only on building and leasing the physical infrastructure as landlords, others specialise in designing and running the internal tech, and some handle both.

Firmus, for example, constructs and operates purpose-built AI factories. These are next-generation data centres optimised specifically for AI workloads and are fully fungible between AI training and inference. Traditional data centres, by contrast, are often inefficient for AI workloads. They're designed for general computing tasks like cloud hosting, email, or web traffic, not the intense, specialised demands of AI processing.

An AI Factory is purpose-built to:

• Power AI;
• Host thousands of high-performance silicon chips such as those graphic processing units (GPUs) produced by NVIDIA;
• Deliver efficient power and cooling; and
• Operate at ultra-high energy density and uptime.

In essence, it is a new type of data centre that houses a massive cluster of GPUs, acting as a ‘manufacturing plant’ for producing AI tokens used to train and run AI models. AI factories enable faster, more cost-effective AI development and iteration compared to traditional data centres, which aren’t built for these intensive AI workloads.

The builders of AI factories can be split into two distinct categories, being:

1. Neoclouds - specialised, AI-first cloud providers built from the ground up to deliver high-performance GPU compute.
2. Hyperscalers – the massive global cloud providers who are also developing their own AI factories.

SECTION 2 – THE OPERATIONS OF FIRMUS

Offshore Operations – Singapore

In June 2023, Firmus partnered with ST Telemedia Global Data Centres (STT) to build an AI factory within the existing STT Singapore data centres. Firmus owns and operates the internal IT network & compute, while STT owns the physical data centre. STT is a major data centre company in Singapore and globally, with >95 data centres. Temasek Holdings, Singapore's major sovereign wealth fund, owns it. Firmus has deployed ~4,000 NVIDIA GPUs across two of these STT data centre sites.

Firmus generates revenue and earnings from large-scale enterprise and government customers across these Singapore operations. These Singapore facilities received the Asia-Pacific Data Centre Project of the Year award for their advanced design, energy efficiency and cost savings (~30% lower than the current status quo) through the use of liquid cooling, meaning NVIDIA GPUs are housed within a cooling liquid to improve power efficiency and lower cost. This is despite Singapore's equatorial climate (i.e., hot, so requiring significant cooling).

Australian Operations – Tasmania

Project Southgate, the company's Australian AI infrastructure program, is anchored by an AI factory in Launceston, Tasmania, which is currently under construction. Firmus built, designed, owns, and operates the project, unlike the Singaporean operations, where Firmus owns and operates the internal AI factory but not the data centre shell itself.

Firmus’ Project Southgate Launceston site is under construction. Note the power substation adjacent to the site at the top of the picture (circled).

While Launceston was the first of the Project Southgate AI factories built in Tasmania, it is not intended to be the last. In late August, Firmus received council approval for its proposed AI factory on the former Gunns pulp mill site in the Bell Bay Advanced Manufacturing Zone in Northern Tasmania. This project at 288MW of IT load would be the largest of the three Tasmanian sites Firmus is currently targeting and would make the company one of, if not the largest, users of electricity in Tasmania.

Firmus’ Project Southgate Bell Bay (George Town) site visualisation.

The two inputs that typically determine whether a site of this scale proceeds (or otherwise) are the availability of both power and water. For Firmus, water use is designed to be minimal, with the company stating that their GPU cooling systems draw water on only around 10 days a year, on days above 26°C, for annual consumption equivalent to roughly 50 Australian households (roughly equivalent to a restaurant), or up to 99% less than a traditional data centre. Power is the open item, with the company website noting the project is "subject to finalisation of energy supply arrangements that are currently being negotiated". The third Tasmanian site is situated in Wesley Vale, which, if approved and developed, would be materially smaller than the above-mentioned Tasmanian locations.

Australian Operations – South Australia

Based on energy usage, South Australia is the largest piece of the pie that Firmus has announced to date. The company has two AI factories planned in regional South Australia: Tailem Bend (~90 km from Adelaide) and Stirling North, near Port Augusta. Firmus sizes Tailem Bend at approximately 1GW of IT load at full build-out, supported by around 1.2GW of planned power connection capacity, and Stirling North at approximately 1.3GW of IT load. From the outside looking in, both of these projects appear to be in their infancy, certainly relative to other Firmus projects, and both are subject to planning approval.

Australian Operations – CDC Strategic Alliance

In October 2025, Firmus announced a strategic partnership with CDC Data Centres (CDC) and NVIDIA to expand Project Southgate to mainland Australia. This strategic partnership is intended to enable Firmus to take capacity within CDC's physical data centre infrastructure. The first of these Firmus/CDC deployments is at the Brooklyn Campus data centre in Melbourne. In August, Firmus said that the site has been installed, commissioned and handed over, and is running a customer's AI models. This site contains 18,400 NVIDIA GB300 GPUs and is under a multi-billion-dollar, long-term contract with a leading global technology company. If it hasn’t already, we expect this site to generate material revenue in the very short term.

On building and owning the physical infrastructure vs deploying within CDC’s footprint, there are two ways to look at this.

1. Owning the entire operation and vertically integrating is an opportunity to capture more of the margin, which would otherwise be paid away to a landlord such as CDC. Furthermore, by taking a full-stack approach, an AI factory can retain control over power, cooling, and how the GPUs are ultimately deployed. This can lead to higher efficiencies, which translate into lower cost of output. That control is also where the operational learnings sit, and those learnings carry across into each subsequent site.
2. The offset is time and capital, being that a full build is measured in years and requires the balance sheet to match, whereas taking capacity within an existing data centre facility, such as CDC, can be operational in months and converts demand into revenue whilst the self-builds are still under construction.

On that basis, we would expect Firmus to continue running both approaches across their portfolio going forward.

Australian Operations – Energy & Water

In 2026, data centres have consistently been on the front page. In the court of public opinion, there is no doubt it has taken a decidedly negative turn. In the US, the world's most advanced data centre market, you could describe some data centres as close to horror stories for local communities because of the negative externalities associated with their existence (i.e., noise pollution, energy consumption, etc.). New York has a moratorium on large data centres, while many local municipalities, cities, and counties nationwide have implemented local prohibitions on new builds.

On the domestic front, regulation is clearly trying to catch up quickly to avoid those horror stories happening here. The Australian Standards for AI attempt to provide a single national framework for how data centres and AI factories are developed and operated (in terms of energy and water use) whilst also safeguarding national security and preserving sovereign capability. This article is not intended to provide an opinion on the data centre debate, but we believe that a national set of standards is obviously a positive step forward.

Regarding Firmus specifically, the company published an Australian Energy Policy and an Australian Water Stewardship Policy in June. The energy policy is built on seven commitments, quoted from the company:

1. We build efficiently;
2. We switch off when prices spike;
3. We invest to strengthen the grid;
4. We back and build more renewables than we use;
5. We pay fair market price for energy;
6. We pay to upgrade the transmission network; and
7. We're here to support the coal transition.

Regarding the above-mentioned commitment 4, it is worth understanding why it is credible in an Australian context. The National Electricity Market is a single interconnected grid spanning QLD, NSW, ACT, VIC, SA & TAS (the latter via the Basslink interconnector), so new renewable capacity contracted anywhere in that footprint can be used elsewhere. Generally, our market isn't geographically bound to match renewables to project locations.

The compute housed within data centres/AI factories sees electricity become heat, and that heat has to go somewhere. The more conventional way to move it, and in most climates the least energy-intensive, is to have the heat move through water and then the process culminates with water evaporating. This is a very simplistic description of complex engineering, but because of the water evaporation involved, the industry has drawn attention for its water consumption.

The World Economic Forum puts conventional consumption at up to 25.5 million litres a year per MW of compute, which on a 100MW site equates to ~2.5 billion litres a year, every year, for as long as it operates. Closer to home, Greater Western Water assessed 19 data centre applications last year in Melbourne's west seeking almost 20 gigalitres (20 billion litres) a year between them, roughly what 330,000 residents use.

A closed-loop water usage system has a dramatically different consumption profile. Rather than evaporating water to shed heat, a fixed volume of water circulates inside the building between the server halls and the chillers. The heat component is expelled to the outside air, and the same water goes around again for another loop. CDC Data Centres has been doing this for years, noting that it needs to “fill the system only once and then circulate it continuously for the life of the data centre”. CDC claimed in 2024 that around five gigalitres a year was being saved across the 13 data centres they then operated in Australia and New Zealand, relative to conventional evaporative cooling. Below is a very simple diagram showing how water is used.

For Firmus, the company’s water policy prioritises dry-mode cooling, site-specific water planning, using recycled or non-potable water where appropriate, and transparent performance reporting. To date, Firmus has deployed direct-to-chip liquid cooling (rather than cooling the air around the servers, the coolant is put in contact with the GPU itself) at their Launceston project currently under development, while in Singapore they have used immersion cooling, where GPUs are submerged and operational within a liquid to reduce heat. Because the heat leaves the building in liquid form rather than air, it can then be rejected through dry coolers with minimal water evaporation.

Further, the current leading-edge NVIDIA Vera Rubin generation GPUs are the first to be 100% liquid-cooled (i.e., direct-to-chip cooling). Paired with a sealed loop and dry coolers, NVIDIA claims this design reduces on-site water consumption to near zero, bar the hottest 1% of the year in warmer climates. The caveat is that this only addresses water consumption at the AI factory site; but the benefit is one that anyone deploying the technology can potentially realise. The design choice Firmus and CDC have made appears to be emerging as the industry default and should be a step in the right direction to reduce water consumption within the industry.

Offshore Operations – Batam, Indonesia

In June 2026, Firmus announced a 170,000-GPU, 360MW IT load AI factory campus in Batam, Indonesia (an island ~20 km from Singapore). For comparison, that is ~5x the 36,000 GPUs going into Launceston and ~9x the ~18,400 at CDC Melbourne.

The Firmus Batam AI factory is being developed in partnership with DayOne, a Singapore-based data centre company which has a portfolio of ~30 data centres across APAC and Europe. The company announced a USD 2 billion funding round in January 2026. Coatue Management LLC, which is a shareholder in Firmus, is also a major shareholder in DayOne. As with Firmus' operations in Melbourne and Singapore, they are not the landlord. DayOne owns and provides the data centre shell, while Firmus owns and operates the IT infrastructure inside it. This may differ slightly from the partnership with CDC: DayOne is building this site new and specifically for Firmus, so one would assume it is co-designed to suit both parties’ requirements. This project is currently under construction and is expected to go live in stages during 2027 and 2028.

What is particularly interesting about this Batam project is its commercial structure: a compute partnership with NVIDIA running through 2034, under which Firmus commits to procure NVIDIA GPUs and NVIDIA-powered cloud services, with NVIDIA earning standard product revenue plus a share of cloud revenue above a certain ‘floor’ price. At or below this ‘floor’ price, NVIDIA will guarantee to pay for the NVIDIA GPUs and associated services, meaning this is essentially a customer backstop. While the future is unknown, there is a reasonably solid likelihood that the NVIDIA customer backstop will never be needed, but only time will tell. Based on customer commitments, Firmus expects USD $25–30 billion in committed offtake agreements over the first six years, and our understanding is that this project is sold out for at least an initial portion of this six-year term.

Firmus/NVIDIA Batam Project Explanation

170,000 NVIDIA GPUs · US$27.5bn of expected revenue over six years · about US$4.6bn a year

Target customers are AI-native companies, enterprises and independent software vendors, rather than a single hyperscale tenant. This customer type is part of the AI landscape, which has seen explosive growth over the past 12 months, underpinned by significant demand for AI compute to power their respective business models.

Offshore Operations – Malaysia

The most recent development came in September when Firmus announced a multi-year strategic partnership with OpenAI, under which OpenAI will contract dedicated AI compute capacity from two Firmus AI factory sites in Malaysia. OpenAI becomes an anchor customer of Firmus. Details are light on at the time of writing but, working backwards, it appears to be a ~385MW IT load deal across two AI factories which are yet to be built.

Malaysia has rapidly emerged as Southeast Asia’s fastest-growing data centre hub due to a combination of Singapore’s land and energy constraints, an explosion in AI demand and an apparent aggressive push by the Malaysian government to develop this industry. If the Indonesian project is any indication, we wouldn’t expect Firmus to build the physical infrastructure in Malaysia. Interestingly, DayOne, which has existing operations in Malaysia, recently purchased a ~78-acre parcel of land in Sepang.

Operations – Summary

Either under development or live, Firmus has AI factories in Australia, Singapore, Indonesia and Malaysia. While only two are operational today (Singapore & Melbourne), the remaining sites, currently under development, are expected to be up and running over the next 24 months.

SECTION 3 – INDUSTRY ANALYSIS

As we know, AI doesn't happen without the commodity underpinning it: AI tokens. The AI sector faces a stark imbalance between surging demand and limited supply, not just for AI tokens, but also for the underlying infrastructure needed to produce them.

Since our last article, the already strong demand signal has, if anything, only gotten louder. For context, Alphabet Inc’s (NDQ: GOOG) monthly AI token processing has gone from ~9.7 trillion tokens in May 2024 to more than 3.2 quadrillion tokens (by the way, which looks like this… 3,200,000,000,000,000) in May 2026, an increase of over 300x in two years[1]. Another way to look at this is in the table below, which shows Meta Platforms Inc. (NDQ: META) usage in a one-month period earlier in CY26. Admittedly, this occurred during peak tokenmaxxing.

Despite replacing tokenmaxxing with a more responsible approach to token usage, AI agents have become much more sophisticated, even since April. This means token usage will continue to rise over time. Goldman Sachs forecasts token consumption will grow ~24x between 2026 and 2030 to ~120 quadrillion (yes, this number 120,000,000,000,000,000) tokens per month as agentic AI becomes the norm[2]. What we also find telling is that industry-wide token processing has consistently blown through forecasts, including forecasts that had already been revised sharply upward. Brookfield, the world’s largest infrastructure manager, forecasts ~USD$2 trillion in global spending on AI factory development by decade's end.

“The greatest shortcoming of the human race is man’s inability to understand the exponential function.” – Albert Bartlett

These tokens must be generated, processed, and monetised, requiring substantial global capital investment. As widely reported, big tech companies are investing heavily in AI capital expenditures to build the data centres and infrastructure needed to power an emerging, generational technology shift and meet soaring demand for AI services.

To put some numbers around this, the scale of that commitment has stepped up considerably over the past 12 months. The five largest US hyperscalers (Amazon.com Inc, Microsoft Corp, Alphabet Inc, Oracle Corp and Meta Platforms Inc) are collectively committing an estimated US$700-800 billion in capex in CY26, of which approximately 75% is directed toward AI-related infrastructure, up from roughly US$410 billion in CY25. Alphabet alone raised its CY26 capex guidance to US$195-205 billion in its June quarter results, and consensus now sees aggregate hyperscaler capex above US$1 trillion in CY27.

These hyperscale capex graphs seem commonplace these days, but honestly, every time we see them, they seem to get bigger in the expected years, and these numbers are hard to put into any sort of historical context. The flip side is a downward spiral in these hyperscaler companies’ free cash flow. Previously generating extreme free cash flow in a relatively capital-light manner, this has transformed into a major investment phase. If you rewound, say, 5 years, who in their wildest dreams would’ve thought we’d see Alphabet producing negative free cash flow?

Alternative asset manager Apollo Global compared this current infrastructure boom to the closest spoken-about recent comparison, the fibre buildout for the internet in the 1990s. As the graph below shows, we are approaching uncharted waters…

Bringing this closer to home, this global capex cycle is now flowing through to Australia at a scale we have simply not seen before. National deployable data centre capacity is expected to more than double from circa 1.35 gigawatts (GW) in 2024 to more than 3.1 GW by 2030, requiring an estimated A$26 billion of new construction. These figures have likely only gone upwards since they were published by the Federal Government earlier in 2026.

As we touched on in Section 2, the constraint which increasingly dictates the pace of this buildout is access to power. Modern AI workloads have driven a step-change in power density, from circa 3 kilowatts (kW) per rack five years ago to over 70 kW today, with next-generation NVIDIA GPUs pushing toward 250 kW per rack. The Australian grid was never designed for this.

From a NAOS perspective, the industry tailwinds are strong, underpinned by compelling demand fundamentals, namely the rapidly increasing consumption of AI. That alone does not guarantee success at Firmus (or any other company in this space), nor an appropriate return on investment. In our previous article, we discussed Firmus's competitive positioning, Australia's sovereign advantages, and the strategic advantages of Firmus's relationship with NVIDIA. We will not rehash this here, but in our view, the competitive positioning has not weakened in recent history. In fact, recent data centre deals by CDC and Nextdc Ltd (ASX: NXT), as well as NVIDIA's recently announced AI Infrastructure partnership in Australia, continue to highlight the global attractiveness of Australian terra firma for AI and cloud infrastructure.

The partnership between Firmus and NVIDIA, whereby NVIDIA serves as GPU supplier, operational collaborator, and equity holder, mirrors successful models in other regions. Top-tier operators similar to Firmus appear to have thrived under this structure. We believe this model succeeds only when Firmus (i.e. David) delivers clear value to NVIDIA (i.e. Goliath), creating mutual commercial benefits. As the saying goes, a picture can say a thousand words.

NVIDIA CEO Jensen Huang keynote presentation NVIDIA GTC Taipei (1 Jun 2026)

In our view, the operating performance of these peers during CY26 has moved the debate from “is the AI demand real?” to “can it actually be funded and delivered?” In the June 2026 quarter, the ‘poster child’ for neoclouds, CoreWeave Inc. (NDQ: CRWV), generated revenue of US$2.6 billion, up 112% year-on-year, with adjusted EBITDA of US$1.51 billion (a ~59% margin) and a revenue backlog of approximately US$104 billion. Despite this, CoreWeave’s quarterly net interest expense alone was ~$640 million. Bringing this back to Firmus, what this says to us is that in this industry, the cost and structure of capital is as important as, if not more than, the demand backdrop. At least, over the medium term.

Two major risks behind these AI factory business models are the duration of the initial customer contracts and the pricing achievable for older-generation GPUs once those initial contracts end. Encouragingly, listed peers have consistently pointed to two favourable trends: lengthening initial contract durations and more stable spot-market pricing for older GPUs than many had anticipated. Both are important for generating strong long-term returns on capital, and we explore these two topics below.

SECTION 4 – DO THE RETURNS STACK UP?

Where Is Compute Actually Priced Today?

Twelve months ago, the widely held fear was that a glut of new supply would eventually deflate compute pricing and strand the previous generation of GPUs. Despite this fear, observable data has gone the other way. Many now predict this demand-over-supply imbalance will remain in place until at least the end of the decade.

Industry-leading research house SemiAnalysis, which now publishes a daily H100 [NVIDIA’s GPU model released in late 2022] spot and contract pricing index, has the one-year H100 contract rate rising almost ~60% from a low of US$1.70 per GPU-hour in October 2025 to US$2.80 by July 2026. For a GPU released nearly four years ago, and one which is technically well inferior to current models, that is not the pricing profile of a technically obsolete asset.

The more interesting development is a two-tier pricing market emerging by contract duration. On its August 2026 quarterly call, listed neocloud company Nebius Group N.V. (NDQ: NBIS) disclosed some interesting economics in the current market environment. Long-dated contracts averaging more than US$1 billion each yield US$20–25 million per MW with upfront payments covering 50–60% of the associated capex, whereas shorter-duration capacity of up to six months is being negotiated in the US$40–50+ million per megawatt range. The company also said it could have sold its entire 2027 capacity today and chose not to.

The largest observable data point on that short-dated premium is the SpaceX Inc. (NDQ: SPCX) arrangement with Anthropic, under which Anthropic pays US$1.25 billion per month for capacity across the Colossus campuses through May 2029, or circa US$45 billion in total, on approximately 300MW of leased capacity and terminable by either party on 90 days’ notice. On those disclosed figures, the implied rate is circa US$50 million per megawatt per annum, at or above the top of the Nebius short-duration range, and Google has since contracted at US$920 million per month for a smaller allocation of roughly 110,000 GPUs.

It would be natural to expect the spot price for an asset to be higher than a long-term contracted price when you have useful-life depreciation to account for. So, if GPU pricing for shorter-duration usage is being priced at a significant premium to longer-duration assets, does that mean fewer and fewer customers are willing to sign longer-term deals for fear of being stuck with out-of-date technology? The industry appears to be seeing things differently, as shown below in recent industry commentary. Not only are older generations maintaining their pricing for longer, but they are also being contracted for longer.

“24 months ago, those [our new contract lengths] were, call it, 3-year contracts. 12 months ago, there were 4-year contracts. I would say now, in our $66.8 billion of backlog that we have, that is 5-year weighted contracts, right, with some contracts in there extending up to 6 years… The other aspect that we’re seeing that’s so strong is [pricing] on older generation infrastructure.” – Brannin McBee, Chief Development Officer, CoreWeave Inc. (4 March 2026)

Is a Return on Investment Actually Being Generated?

There is no doubt this is the trillion(s) dollar question at the forefront of everything and everyone involved in this industry. Will the AI capex boom deliver strong future returns on investment (ROI)?

As a baseline refresher from our previous article, we highlighted how AI factories differ from traditional data centres in the timeframe to reach full contracted capacity. AI training works best when all available computing power is used at once, rather than gradually. This means customers pay for the full capacity upfront, generating revenue and earnings for AI factories much faster than traditional data centres. Faster revenue realisation provides greater certainty for AI factory funding structures and a shorter time to generate investment returns.

As the saying goes, history doesn’t repeat, but it rhymes, so let’s look at history first to see what happened.

Consider Amazon Web Services (AWS), the cloud division of Amazon.com Inc. (NDQ: AMZN). It took nearly a decade from its inception to turn a profit, yet today it is the profit engine of the entire group, contributing over 60% of Amazon’s operating income. We have chosen a point in time at the end of CY24 to measure what we will call the peak of ‘traditional cloud computing’ as a standalone. The reason we have chosen this point in time is that it is largely before the capital expenditure (CAPEX) explosion that has occurred since in CY25 and CY26 due to the onset of all things AI. As at the end of CY24, AWS was generating ~27% after-tax returns on incremental capital. Incremental returns in that ballpark are more than acceptable, particularly for a business as large as AWS. Neither Google Cloud nor Microsoft Azure discloses divisional data at the same granularity, so we have not included them in this analysis.

Amazon.com Inc (NDQ: AMZN) — Amazon Web Services Divisional Financials & Returns

Since the onset of the substantial increases in CAPEX over the past couple of years, as reported in its June 2026 quarter, AWS generated revenue of US$42.2 billion, up ~37% year-on-year, which was its fastest growth rate in 18 quarters and the fifth consecutive quarter of acceleration, on an annualised run rate of approximately US$169 billion. Operating income rose 64% to US$16.6 billion, lifting the divisional operating margin to 39.4%. Importantly, margins expanded, not contracted, while the division spent at record levels, with the company guiding for total group CAPEX of circa US$200 billion for CY26.

“Even at that [capital expenditure] amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027, too. In fact, the demand we already have for 2028 is striking.” – Andy Jassy, CEO, Amazon.com Inc.

Accelerating revenue, an expanding margin and record reinvestment all happening at once is, in our view, one of the most useful data points we have on whether AI infrastructure capital earns a return. Amazon isn’t alone. Both Google Cloud and Microsoft Azure are also seeing such trends in their quarterly financials. That doesn't mean it is bulletproof. In fact, far from it. The outcome, deeming success or otherwise, of all of this investment may not be known for a few years to come, or if it is strong now, that does not mean it will stay strong in the future. 

The point we are highlighting here is that despite significant upfront capital spend, adequate to strong returns are achievable, and the two best disclosed datasets we have access to both show divisional margins expanding rather than eroding as AI workloads scale up. If we add to this some industry rhetoric we are starting to hear about general payback periods, which, if true, are quite remarkable and make deploying such significant billions into this endeavour far more palatable than feared.

“Second, our payback period on AI servers in aggregate is less than 2 years and on our own silicon is half that. So, we have a strong payback period. And the majority of our infrastructure contracts, the total contract value of the infrastructure is long-term committed 5-year contracts.” – Thomas Kurian, CEO, Google Cloud

The other critical factor in the ROI for AI investments is the ability to repurpose GPUs used for AI training for AI inferencing and across various specific use cases with little to no additional cost. By definition, they are fungible assets. If GPUs can serve "two lives", their returns could extend well beyond the typical ~5-6-year depreciation period, potentially to something close to 7-10 years, with the later years being highly profitable. Not all users require the best technology. Think about the iPhone as an example. Does everyone need the iPhone 18 Pro? Some people might, but many others will happily buy older models, or even second-hand phones refurbished after their original owner traded them in. Current NVIDIA H100 pricing suggests this may play out within the GPU ecosystem, even as newer GPUs deliver exponential performance gains.

"NVIDIA compute is fungible, durable and highly rentable. It is a productive, revenue-generating asset." – Jensen Huang, CEO, NVIDIA Corp.

So, What Do the Returns Look Like?

Below, we have put together a generic example based on the profitability of a single NVIDIA GB300 GPU. For context, this is the NVIDIA 2025-released GPU and has since been replaced as their most technically advanced GPU by the NVIDIA VR200 GPU. These are our internal assumptions only, and we kept them intentionally simple. We have no insight into how accurate they are, but we feel comfortable they are at least directionally right.

Our assumptions are below. All figures are in USD. We assume the GPU is 100% utilised throughout its lifetime (8 years), and in this model, we assume a $3.50/hr price with a 50% drop after a 5-year initial contract length.

The model summary is that if the asset has useful life beyond the original contractual period, it can become fully depreciated and debt-free, generating very strong free cash flow margins. Our simplistic assumption is a 50% reduction in price that then flatlines years 6-8. In that time period, it would make sense to discount price, even quite significantly, as it is all ‘cream’, so to speak.

Theoretical financial returns for an NVIDIA GB300 GPU

A lot depends on the price that is either re-contracted or achieved in the spot market after the 5-year initial period. We acknowledge that, given the current pricing environment, the 50% discount seems conservative (i.e., a very large discount to price in), but we are still a long way from knowing how that will play out in reality.

Looking at it from an asset-level NPV perspective, which we believe is the most simplistic way to look at it, using all the same assumptions as previously, whilst also using the below, we can see that, at a GPU level, the payback is 4 years and a ~24.9% IRR is generated. For context, recall the quote previously mentioned in the article from Google Cloud claiming a current 2-year payback period. 

We acknowledge these numbers are simplistic and, at the end of the day, just theoretical. As with any economic market, pricing will depend on supply and demand, but whilst the demand > supply dynamic remains heavily skewed, we believe the economics remain solid for at least the medium term.

“People are demanding tens of thousands of Vera Rubin GPUs now. So, we see demand today as unlimited. Nothing is unlimited in this world. But for now, is the demand in all the visible perspective is much higher than anybody can serve. For how long it will go, we were talking about 18 months. Now I think we can easily talk about 24 months, maybe more. But today is the situation that AI is working. AI creates real value in real life. It needs to be served. We just, as an industry, we can't build as fast. That's why there’s pricing growth.” – Arkady Volozh, CEO, Nebius Group N.V.

What About the Economics of AI From an Enterprise Perspective?

This article is not intended to be a deep dive on the SaaSpocalypse debate, or whether AI ultimately eats the software industry or expands it, as that discussion is well covered elsewhere. For the context of this article, we're asking a narrower question: are the companies spending their budgets on AI-related products and services actually getting something back for it? Every token sold, every GPU leased, and every megawatt contracted eventually has to be paid for by a customer who has run the cost-benefit analysis and concluded that it stacks up today or will stack up in the future.

Of the recent transcripts we have read, e-commerce powerhouse Shopify Inc. (NDQ: SHOP) provided interesting insights on this. Where a shopper begins their journey within an LLM, such as ChatGPT, rather than through a traditional search engine, Shopify notes the customer is 2.5x more likely to land directly on a product page rather than somewhere more general on the merchant's website, and that shorter path is converting at roughly 80% higher rates. While there are winners and losers in this shift (which we won’t discuss here), this is an incredible improvement in customer conversion with tangible value.

On the cost side, Shopify's subscription gross margin is now absorbing rising LLM model and token costs. Furthermore, the company is managing the cost profile by using the best-of-the-best AI models where needed, while augmenting them with older models and open-source models with minimal trade-off in quality.

At the other end of the spectrum are companies like Valvoline Inc. (NYSE: VVV), which are seeing internal efficiencies, and this has a long way to play out, but it seems a long way from the 50% of all white-collar jobs to be lost within 5 years narrative we mentioned in our opening remarks.

“I mean we are not an industry that is going to be transformed with AI, but there is efficiency in software development and marketing and then in just more in the G&A side. So, will it be a step-change improvement? No…” – Lori Flees, CEO, Valvoline Inc.

Taken together, we believe the customer ROI is hard to characterise in one blanket statement. The quantified wins show up in revenue growth (as seen in our hyperscaler revenue growth chart). On the cost side, it looks more like steady, incremental savings in G&A and development than a step change. Closer to home, we have seen some companies make drastic reductions in their staff numbers, WiseTech Global Ltd (ASX: WTC) as an example, but this is not widespread at this point in time. We note that every company is different, and it would be fair to say a clearer picture of ROI from the enterprise perspective has yet to play out, but signs are there; it is promising. An interesting quote from Nebius recently perhaps shapes the direction of travel…

“[We] believe, ultimately, enterprise will represent the majority of the opportunity in the market." – Marc Boroditsky, Chief Revenue Officer, Nebius Group N.V.

SECTION 5 – RISKS

In our previous article, we highlighted a number of risks. We will not rehash those risks here, but they remain live. The risk we want to dive deeper into below is the circular financing nature of this industry. We would be naive not to consider this risk in significant detail, as the capital decisions of a few major companies could change and have wide-reaching ramifications for the entire industry. Furthermore, the industry is now so intertwined that it could have domino effects if something significantly negative occurs.

Risk 1 – The Circular Financing of the AI Industry

The mechanics of the circularity concern are simple enough to describe. Capital from the ‘Fort Knox’ style balance sheet of NVIDIA is invested into various companies which, not always but in a number of examples, are also customers buying GPUs and other products/solutions from NVIDIA.

The list of transactions which fit that description is now very long. NVIDIA has taken equity positions in OpenAI, xAI, Mistral, Nebius, Nscale, CoreWeave and, as we know, also in Firmus. Those same companies sit amongst the largest purchasers of NVIDIA GPUs in the world. Adding to this, NVIDIA can also be the customer of those same GPUs and/or a backstop provider. In addition, NVIDIA has opted to vertically integrate to a certain degree and has invested US$5 billion in Intel Corp (NDQ: INTC), alongside investments in Hugging Face and Groq. It is easy to understand how many have described NVIDIA as the Central Bank of AI.

This phenomenon, however, is not restricted to just NVIDIA. OpenAI has separately committed hundreds of billions of dollars in compute spend across Microsoft, Oracle, Amazon, and CoreWeave, funded largely by capital raised from investors who, in many cases, also sell into the same ecosystem. Anthropic has followed a similar path. In his research paper The AI Circular Economy: Systemic Risk, Vendor Financing, and the Keystone Problem, Rahil Solanki estimates this AI circular economy now exceeds USD $1.4 trillion in committed future spending circulating among roughly the same eight or nine balance sheets. Given his research is from April 2026, it is probably higher than that now.

The International Monetary Fund (IMF) has been explicit about this. In its April 2026 Global Financial Stability Report, the Fund listed “circular financing structures that may artificially amplify reported revenues and valuations” as one channel through which AI revenue expectations could become misaligned with underlying fundamentals, drawing a direct comparison to the fibre and data centre build-out of the dot-com era. That same report also dedicated a box to hyperscaler balance sheets and the risk of GPU obsolescence.

Two things need to be said in the counter-positioning argument to the circular financing economy underpinning AI. First, vendor financing is neither new nor, of itself, improper. Suppliers have supported the growth of their own customers for as long as capital goods have existed, and where the supplier has the best information on end demand, it is arguably the most efficient provider of that capital. Second, a meaningful portion of the money moving around this loop is entirely real. Alphabet’s token volumes, AWS’ accelerating revenue on an expanding margin, and the enterprise adoption we walked through in Section 4 are not accounting constructs.

The genuine risk here, in our opinion, is not that impropriety is occurring, but rather that the money is concentrated in so very few hands. That same dynamic can make this circular economy loop spin faster on the way up, but it can just as quickly spin in reverse. The closest historical analogue is the telecommunications equipment cycle of 1999 to 2001, when the likes of Lucent and Nortel lent their customers the money to buy their own equipment. Demand looked robust right up until the financing stopped, at which point revenue, receivables and collateral values all deteriorated simultaneously, for the simple reason that they were the same dollar viewed from three different angles.

What makes the present cycle worth watching closely is not that any individual transaction is unsound, but that the industry has become so tightly intertwined that the failure (or potentially even a strategic change) of a single keystone counterparty would impair the demand assumptions, the revenue forecasts and the asset values of everybody else, all at once. Adding another dynamic to this industry is that many/most of the AI native companies are in a cash-burn phase, despite their substantial valuations.

The demand profile underpinning this industry is also hard to argue against. Whilst demand remains incredibly buoyant, if this is translating into adequate returns on invested capital, the circular economy can hold its course. At least in the medium term, this seems plausible. On a longer-term view, one would have to assume supply-demand equilibrium will occur at some stage. And then what happens?

So Where Does Firmus Sit in All of This?

It would be remiss of us to write about the above and then not acknowledge that Firmus sits directly inside this loop. As we mentioned in this article, NVIDIA is a Firmus equity holder, GPU supplier, GPU customer, operational partner, and financing backstop in the case of Batam, Indonesia.

There are two ways to read that, and we think both are legitimate. The bearish read is the obvious one: a portion of Firmus' funding, cost base, and downside protection trace back to a single counterparty, which makes the independence of the demand signal a little opaque. We believe the market demand dynamics probably do nullify this to a reasonable degree. The more bullish perspective is that NVIDIA has finite GPU allocation, an effectively unlimited choice of partners, and no commercial reason whatsoever to put its own balance sheet behind an operator it does not believe can execute. Capital is abundant in this industry at present; hence, being chosen consistently as a partner is itself a signal.

Many take comfort in the guarantor and the shape of the rest of the book. NVIDIA’s balance sheet is about as strong as any corporate balance sheet on the planet. That is a materially different proposition to a vendor financing arrangement written by a heavily leveraged equipment manufacturer in the year 2000. Beyond the Firmus Batam project, the Melbourne CDC site is contracted to an unnamed global technology company (now confirmed as Meta), the Singapore facilities service enterprise and government customers, and the Tasmanian and South Australian projects are being developed into a domestic market with a sovereign-capability argument attached. The Malaysian agreement with OpenAI, by contrast, is direct exposure to precisely the counterparty the circularity debate is about. That is not a criticism of the contract, simply an observation that one needs to factor in an assessment of OpenAI's ability to pay rather than the headline megawatts.

Risk 2 – The Concentration of Demand, And What Happens If the Funding Tap Is Turned Off

The second risk is that, for all of the very large numbers referenced throughout this article, the demand side of the AI industry currently rests on a remarkably small number of cheque writers.

Five US hyperscalers account for the overwhelming majority of global AI capital expenditure. Below them is a second tier of AI-native buyers, including OpenAI, Anthropic, xAI, and a handful of smaller companies such as Cursor, Perplexity, and Thinking Machines. At this stage, certainly the 80/20 rule applies – this is not a broad customer base. It is highly concentrated, and concentration can cut both ways.

The starkest framing we have seen is as follows. The four largest US cloud providers now carry a revenue backlog of about USD $2.1 trillion, and roughly half of that figure is attributable to just two counterparties: OpenAI and Anthropic. Both are growing at rates without any real precedent in corporate history. Latest reports also show both are deeply loss-making.

Hyperscaler capital expenditure cannot compound at the current rate indefinitely. At some point this level of growth will normalise, whether by choice, through board-level discipline, or because shareholders eventually lose patience with negative free cash flow. Separately, the private funding that sustains the frontier labs and the AI-natives is a function of sentiment and open capital markets, and both are cyclical. So, what actually happens if the tap is turned down?

The details then become particularly important in any risk assessment. The balance sheets of contracted counterparties, along with pricing and contract term, are all important aspects to consider, the first of those most of all. Based on what has been disclosed, we believe Firmus screens better than most on that measure, though not perfectly, and we expect that to remain the central debate around the company if/when it undertakes an IPO process and beyond.

Conclusion

Industry demand suggests heavy reinvestment is likely to continue over the next decade, but that does not preclude value creation in the meantime. What it does mean is that, through the build phase, current free cash flow may not be the best indicator of what current and future returns could look like.

At the end of the day, only time will tell whether we are in one of the greatest financial bubbles of all time or whether we are simply in the build-out phase of the infrastructure underpinning the 5^th industrial revolution. Having spent considerable time studying this industry recently, our honest conclusion is that both statements can be true at the same time. There is no doubt the technology is real, the demand is real, and the revenue is real. The returns generated by the best-positioned operators appear to be firming up, though they are hard to measure precisely at this stage. It is equally true that the funding structures supporting all of it are increasingly circular, that the demand sits in very few hands, and that the quantum of capital being committed against a still unproven end-user monetisation model is without historical precedent. As with any investment, deciphering an industry, choosing those more likely to benefit, and avoiding those less likely to is easier said than done. The strength of the foundations underpinning any particular investment is as important as ever in the world of AI.

Despite this, we at NAOS believe that following the demand signals can often provide a good barometer for at least the medium, if not the longer, term health of any industry. In the AI industry, consensus signals suggest demand will remain over the coming years, and the largest players appear to have the financial wherewithal and strategic intent to underpin it, given they are highly incentivised to see it succeed. As Napoleon Bonaparte put it, victory belongs to those with the most perseverance, but as Mark Zuckerberg puts it…

“There’s definitely a possibility, at least empirically, based on past large infrastructure buildouts and how they led to bubbles, that something like that would happen here… If we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously. But what I’d say is I actually think the risk is higher on the other side [opportunity cost of not spending those dollars].” - Mark Zuckerberg, CEO, Meta Platforms Inc.

NAOS Asset Management is a specialist fund manager providing genuine, concentrated exposure to quality private & public emerging companies.

Understanding an industry and identifying the businesses best placed within it is central to how we invest, and this article reflects that process. We would welcome the chance to discuss it with you in person at our annual Investor Roadshow, held in cities across Australia throughout October & November. To register, visit naos.com.au/events.