The question every Firmus investor should be asking

As investors, we like to think we're always looking for multiple ways to win. But sometimes I think the more important question is whether we have multiple ways not to lose.

It reminds me a little of McDonald's. For decades, investors weren't simply backing a company that sold hamburgers. They were also investing in a business that controlled an enormously valuable real estate portfolio, much of which was leased to franchisees. If restaurant sales disappointed, McDonald's still had an underlying asset base and another income source. It wasn't a guarantee against losses, but it gave the business more than one way to create value.

I see an interesting parallel with the artificial intelligence infrastructure boom. Companies like Firmus are investing enormous amounts of capital to build the data centres that will power AI. They're effectively supplying the infrastructure on which other businesses hope to build the next generation of applications.

Microsoft, Amazon and Alphabet are doing that too. But they also own substantial, highly profitable businesses sitting on top of that infrastructure. Google has Search, YouTube and advertising. Microsoft has its productivity and enterprise software businesses. Amazon has retail, advertising, logistics and AWS. All three can sell compute to external customers, but they can also deploy it into their own applications to generate additional revenue or improve productivity. 

Google and Amazon also have another card to play: proprietary silicon that could reduce the cost of delivering that compute. The distinction matters because building AI infrastructure is extraordinarily capital intensive. With a fully equipped gigawatt potentially costing billions, the question isn't simply who can build the most capacity. It's who can generate an attractive return on the capital invested.

That's where we think the hyperscalers have a structural advantage. They have more ways to generate demand for compute, more ways to monetise the intelligence it produces, and, in some cases, more flexibility over the chips that deliver it.

We're not suggesting Firmus or the neoclouds can't succeed. Some undoubtedly will. But our preference is to invest in established businesses extending their existing economic advantages into AI, rather than businesses whose future returns depend more directly on the economics of renting compute capacity. That's why we'd rather leave Firmus to others. We believe hyperscalers have several structural advantages that specialised AI infrastructure providers struggle to replicate.

The $40 billion question

AI infrastructure is extraordinarily capital intensive. Building and equipping a gigawatt of AI data centre capacity can cost around US$40 billion. To put that in perspective, a gigawatt can supply roughly one million Australian homes. It's also close to the 1.21 gigawatts Doc Brown needed to power the flux capacitor to send his DeLorean Back to the Future. Amazon, for example, is expected to bring on 6 gigawatts of additional data centre capacity in 2026. 

Unfortunately, spending US$40 billion per GW on an AI data centre doesn't guarantee you'll be travelling into a profitable future. And that's really the question investors should be asking. With so much capital being deployed, and much of the computing equipment needing replacement within a relatively short period, who is best positioned to generate attractive returns on that investment?

Simply building more capacity isn't enough. We believe the hyperscalers have three particularly important advantages: more ways to generate demand for their infrastructure, more flexibility in choosing the technology used to deliver compute, and more opportunities to capture economic value beyond compute itself. Together, these could materially improve their prospects of generating attractive returns on invested capital.

1. The first-party demand advantage

Perhaps the most underappreciated difference between the hyperscalers and neoclouds is what we call first-party demand for compute. A specialised AI infrastructure provider generally needs customers willing to rent its computing capacity. Its economics depend heavily on securing contracts, maintaining utilisation and earning sufficiently attractive prices over the useful life of its equipment. The hyperscalers have another option.

Consider Alphabet. It can deploy AI infrastructure into Google Search, YouTube, advertising, Gmail, Maps and Gemini, while also selling compute to external customers through Google Cloud. Microsoft has substantial first-party demand through Microsoft 365 Copilot, GitHub Copilot, Dynamics and its broader enterprise software ecosystem. Amazon can deploy AI across its retail operations, advertising business, logistics network, and customer service, while also supplying external customers through AWS.

These companies don't necessarily need to sell every unit of compute to an external customer to earn an economic return. Google might generate additional advertising revenue through better search results. Amazon might lower the cost of fulfilling customer orders. Microsoft might increase software revenue by adding AI capabilities to its productivity applications. In each case, the economic return from the underlying compute may appear somewhere other than reported cloud revenue. This distinction matters enormously.

Imagine two companies each investing US$40 billion in AI infrastructure. One must generate an attractive return primarily by renting computing capacity to external customers. The other can rent that capacity, deploy it in its own profitable applications, or split it between the two as economics change. We would generally prefer the second business's flexibility. Of course, first-party demand is only valuable if the resulting revenue or productivity improvements justify the investment. Running AI internally doesn't automatically create economic value. But the hyperscalers have substantially more opportunities to earn that return.

2. Two types of silicon are better than one

The second advantage is proprietary silicon. Most neoclouds depend predominantly on Nvidia GPUs, which are exceptional products supported by an extraordinarily powerful software ecosystem. The hyperscalers also buy Nvidia GPUs, often in enormous quantities. But Google and Amazon have something many neoclouds lack: a credible alternative.

Google has spent more than a decade developing its Tensor Processing Units, while Amazon has invested heavily in Trainium accelerators and Graviton processors. 

This gives them greater flexibility to allocate workloads according to economics. Where Nvidia GPUs offer the best combination of performance, compatibility and cost, they can use Nvidia. Where their proprietary chips offer superior economics, they can use those instead.

Amazon has stated that Trainium2 delivers approximately 30% better price-performance than comparable GPUs, while Trainium3 offers further improvements. The company believes its custom silicon could eventually save tens of billions of dollars in annual capital expenditure. These are company estimates, not universal performance guarantees, and Nvidia remains an exceptionally formidable competitor. Nevertheless, the implications are significant.

In an industry that requires tens of billions of dollars of investment per GW, reducing the capital needed to deliver a given amount of useful computing power can substantially improve returns on invested capital. Proprietary silicon also reduces dependence on a single supplier and gives the hyperscalers another way to optimise their infrastructure as AI workloads evolve.

A neocloud might run Nvidia GPUs very efficiently. But Google and Amazon can compete on the economics of Nvidia GPUs and their own silicon. That is a valuable additional card to play.

3. The opportunity extends well beyond renting compute

The third advantage may become increasingly important as we move into the agentic era. Today, much of the discussion around AI infrastructure focuses on training models and serving inference requests. But we believe the next phase of AI adoption will involve autonomous agents performing work across enterprise systems.

Imagine asking an AI agent to prepare a financial forecast, reconcile customer invoices or manage part of a supply chain. The agent might need to retrieve documents from Microsoft 365, access customer records in Salesforce, interact with SAP, update financial databases and execute authorised transactions. Simply providing the GPUs required to run the underlying model is only one part of that process.

Enterprises also need data infrastructure, identity management, permissions, cybersecurity, orchestration, audit trails, and systems to verify that agents act within their delegated authority. This is giving rise to what is increasingly described as the agent control plane. Microsoft is developing these capabilities through Agent 365 and its existing identity and security infrastructure. Amazon has introduced Bedrock AgentCore, while Google Cloud is developing its own agent identity, security and governance capabilities.

The idea is straightforward: as enterprises deploy thousands of intelligent digital workers, they will need systems to manage those agents much like they manage human employees. And critically, the controls governing what an agent can access and what actions it can perform must operate independently of the AI model itself. For the hyperscalers, this creates another opportunity to capture economic value.

They can supply the compute, but potentially also the databases, networking, security, orchestration and governance infrastructure surrounding the workload. 

A neocloud might offer cheaper GPU inference, but that doesn't necessarily mean it delivers the lowest total cost for a successful enterprise task. 

Data movement, integration, latency, security and operational complexity also matter. The hyperscalers have spent decades building the infrastructure that enterprises already depend upon. As AI agents integrate into those systems, we believe existing infrastructure could become more valuable, not less.

Multiple ways to win

One reason we favour the hyperscalers is that we don't need to predict precisely how the AI industry will develop. Consider the debate around whether AI will destroy traditional enterprise software. If AI makes writing software dramatically cheaper, companies may increasingly build their own applications rather than buy them from established vendors. But building an application is different from maintaining a reliable, secure and compliant enterprise system. As AI agents gain the ability to execute increasingly consequential actions, the need for permissions, oversight, verification and governance may become more important.

If enterprises continue buying established software, Microsoft benefits from its applications and productivity ecosystem. If enterprises increasingly build their own software and AI agents, Microsoft, Amazon, and Google can benefit by supplying the development tools, computing infrastructure, and databases needed to run those systems. Regardless of which AI models or applications ultimately prevail, all three can supply the infrastructure needed to manage and govern intelligent agents.

This creates multiple ways to win. We don't need to know whether the winning AI model will come from OpenAI, Anthropic, Google, xAI or somewhere else. We also don't need to know exactly how enterprise software will evolve. We believe hyperscalers can capture economic value across several layers of the AI ecosystem.

What about the neoclouds?

None of this means Firmus or other neocloud providers cannot succeed. Specialised operators can possess meaningful advantages in securing power, designing facilities, deploying infrastructure quickly and achieving high utilisation of expensive equipment. Long-term customer contracts can also provide revenue visibility and support attractive project economics. Indeed, certain neoclouds may achieve better returns than hyperscalers on particular infrastructure investments.

An important feature of the emerging AI ecosystem is that enterprise software and governance don't necessarily need to run on the same infrastructure as the underlying model. An organisation might use Microsoft's identity and security systems while sourcing inference from a specialist neocloud. That creates opportunities for both business models.

The hyperscalers' investment programmes also carry considerable risk. Capital expenditure is reaching extraordinary levels, and if AI monetisation disappoints, equipment becomes obsolete faster than expected, or inference prices fall faster than costs, returns could fall short of expectations. The hyperscalers' substantial existing businesses don't eliminate those risks. 

Our preference isn't based on the assumption that specialised AI infrastructure providers are poor businesses. Instead, we believe hyperscalers have broader structural advantages and more ways to generate returns from their investments.

Owning the factories or owning the economics?

We increasingly think of AI infrastructure as a race to construct the factories that will produce intelligence. But the investment opportunity isn't simply about who can build the most factories. It's about who can generate the greatest economic value relative to the capital invested.

The hyperscalers have substantial first-party demand, access to external customers, multiple semiconductor architectures and established software ecosystems through which they can monetise the intelligence their infrastructure produces.

Google and Amazon may produce that intelligence more efficiently with proprietary silicon. Microsoft, Amazon and Google can all potentially capture additional value through the software, databases, security and governance services required to deploy it. In an industry where each gigawatt of capacity can represent approximately US$40 billion of investment, those advantages could prove enormously important.

For Australian investors seeking exposure to the AI infrastructure boom, we believe Microsoft, Amazon and Alphabet offer a more attractive combination of economic opportunities, competitive advantages and potential resilience than a concentrated investment in specialised GPU infrastructure.

We would rather own businesses with more ways to fill their data centres, more choices about the chips inside them and more opportunities to monetise the intelligence they produce. That's why we're happy to leave Firmus to others.