EIF Speaker Series: ‘Inference Could Shift the AI Compute Mix’ with Paul Golding

The AI infrastructure boom has pushed data centers from a niche real estate story into one of the defining investment themes of the current market cycle. But as demand for compute accelerates, the debate is becoming more complex: what kind of compute will be needed, where will the power come from, and how much of today’s data center buildout is supported by durable fundamentals?

That is where Paul Golding brings a useful lens. As a senior analyst at Macquarie, Golding covers AI infrastructure, data centers, payments and blockchains — sectors that increasingly overlap as compute demand, power availability and digital assets reshape the infrastructure market.

This conversation is part of TheEnergyMag’s EIF Speaker Series, which features speakers attending the Energy Investors Forum and their views on the macro questions defining the next investment cycle across energy, AI infrastructure, data centers and bitcoin mining.

Golding’s answers point to a market where the AI buildout remains supported by strong fundamentals, even as the investment focus may broaden beyond GPUs. He sees inference potentially increasing the relevance of CPU compute, experienced data center operators continuing to find powered land despite grid constraints, and bitcoin tracking a familiar post-halving cycle while benefiting from structural tailwinds from institutional adoption.

Below is the full Q&A with Paul Golding, senior analyst at Macquarie.

Q: Bitcoin is around $60,000, and a lot of retail attention has left the crypto market. Does bitcoin still have a future?

Our research has shown that Bitcoin’s performance across post-halving cycles has been quite consistent. When we overlay the previous cycles, we see a similar pattern emerge each time: an initial period of bullishness, often leading to what might be seen as a point of exuberance, followed by a correction and a period of disillusionment before the next phase of the cycle.

We’ve now seen this pattern play out across three full post-halving cycles, and the current cycle continues to track closely with those historical precedents. Based on that analysis, our view is that Bitcoin is behaving largely as expected.

The macro environment is certainly different this time around, but there are also important tailwinds that didn’t exist in previous cycles. [With] institutional adoption, there’s a general acceptance of Bitcoin potentially representing part of an asset allocation for institutions in general. It’s at a place, sentiment-wise, where it wasn’t in prior cycles. [So] while there may be macro headwinds, there are also meaningful structural tailwinds supporting the asset.

Q: Everyone is focused on AI and the major frontier labs. What’s a less obvious or underappreciated investment theme in AI?

There are a couple of key areas worth watching. I wouldn’t necessarily call them contrarian, but they could be evolutionary themes in the AI space.

One of those is that while the focus is very much on GPUs at the moment and installing accelerator capacity—and we believe that is very likely to continue to be the case based on analysis—as enterprise adoption of AI inference grows, and as models become more pervasive in business and enterprises become more comfortable using agentic workflows, you could see inference become of greater significance from a compute-mix perspective.

Some of our coverage universe would suggest that this could become a significant demand case for CPU compute, as opposed to GPU compute.

To the extent that data center capacity and data center design today are focused around GPU compute—and while that will continue to be highly relevant—we do see a universe where there is some incremental focus that is not yet being fully paid to CPU compute, but perhaps should be, depending on inference uptake. That could pivot over time and is not being focused on as much right now.

Q: What are some names that illustrate that theme?

I can’t speak to names that we don’t cover, but what I can say is that we cover SoftBank Group, ticker 9984 in Tokyo, and one of SoftBank Group’s largest holdings is ARM.

ARM has recently been in the press and also announced through its official communications the development of an AGI-focused CPU design.

Q: Power availability is increasingly seen as a constraint on AI growth. How do you think the industry addresses that bottleneck?

One theme right now is the general issue of power as a constraint.

Power is seen as a constraint on a consensus level across the U.S. market in terms of having enough power availability or interconnect capacity to energize data center sites for training factories and other large-load use cases.

However, across our coverage universe of colocation and vertically integrated cloud players, what we’re seeing is that the platforms with the expertise to engage in load studies, procure sites with a high probability of interconnect approval, navigate the interconnect regulatory landscape, and maintain the relationships needed to bring generation to these sites have created an environment where those with the skill and experience are able to continue adding scaled, powered land sites despite the constraints around generation—or, more specifically, transmission delivery—that we see in the marketplace.

Q: So you think experienced operators can still move the needle despite those constraints?

It’s to say that while the consensus view is that power is very constrained, that doesn’t necessarily mean we won’t continue to see some of these more experienced players do what they’ve been doing very well, which is finding powered land sites at scale where they might not otherwise have been considered obvious opportunities to other parties.

Q: Do you see the energy mix changing? Could geothermal, nuclear, or other sources become a bigger part of the equation?

Yes. We cover some names that are engaged in more unique generation and energization approaches.

That includes “bring your own generation” strategies, whether thermal today or potentially nuclear in the future. We’re also seeing renewables paired with battery energy storage systems to smooth out intermittency.

Overall, we’re seeing a varied mix of approaches that take advantage of geographic positioning against what remains a very strong demand backdrop.

Q: This might be a leading question, but when do you see the AI buildout—or AI bubble—popping? Do you see valuations coming down in the next 6, 12, or 18 months?

We evaluate our covered companies on a fundamentals basis relative to the demand picture that we assess in the marketplace and the supply of powered-land inventory, as well as any prospective or demonstrated capabilities around energizing land and establishing data center infrastructure.

Our recommendations and target prices for our covered companies are on a 12-month horizon. At the moment, over the course of that horizon, what we are seeing is a continued demand picture that outstrips supply.

We continue to see demand for compute outstrip supply of available compute and GPU hours. Our global teams continue to see demand for high-bandwidth memory as a component of accelerators outstrip supply as well.

Overall, we continue to see a bullish backdrop for our coverage universe.

Q: So, strong fundamentals?

Strong fundamentals.

Golding will join other analysts, investors and infrastructure leaders at the Energy Investors Forum, where these questions — how AI data centers get powered, how compute demand evolves, and how bitcoin fits into the next market cycle — will be part of the broader discussion.

Join EIF to hear more from Golding and other speakers on how the next wave of AI, energy and digital infrastructure will be financed, built and valued.