Takeaways
#- Zabelin separated sovereign AI funding into legislation and incentives, direct state allocation and sovereign wealth funds.
- He attributed large platform investments partly to uncertainty about which industry-specific applications would win.
- Perz questioned whether infrastructure spending would translate into the best investment returns.
- Energy efficiency, greater AI usage and regulation complicated the commodity forecast; cybersecurity offered another area of demand.
Jason Perz came to the October 30, 2025 Gold Rush Live with a question about commodities. Countries and companies were committing heavily to AI infrastructure. How much of that spending would flow through to energy and metals?
His guest, PitchBook senior research analyst Dimitri Zabelin, had just published a report on sovereign AI. Zabelin's work examined how national efforts to build and control AI capabilities affected public and private markets. He saw a powerful reason for continued infrastructure investment, but he was less willing to identify the eventual commodity winners.
With Sam Gatlin joining the discussion, they traced the spending from government policy to foundation models, then to the power systems and security tools required to operate them.
Governments were funding AI through different channels
Zabelin divided national AI policy into three broad approaches: legislation and industrial incentives, direct bureaucratic allocation, and sovereign wealth funds.
He associated the first with the United States and Europe, where chip legislation offered grants, credits and tax incentives. China illustrated the second, with state-directed capital targeting strategic technologies. Several Middle Eastern countries illustrated the third through their sovereign funds.
Those differences mattered because governments did not distribute capital in the same way. A company's opportunity depended partly on where it operated and how its technology fit the country's priorities.
Despite the varied funding mechanisms, Zabelin saw the United States attracting a large share of the capital. He pointed to its strength in advanced chips and software, including Nvidia's role in supplying the hardware used to train large models. In his analysis, access to that technology had also become a source of geopolitical bargaining power.
The largest checks favored the underlying platforms
Zabelin distinguished foundation-model platforms from applications built for particular industries. The former supplied a broad capability; the latter adapted it to uses such as legal work or healthcare.
His charts showed fewer deals in the platform layer but much greater capital requirements. Application developers could start with smaller amounts and build on an existing model instead of financing one themselves.
Investors he had interviewed did not know which applications would win. They were more confident that successful applications would depend on a limited set of foundational platforms, helping explain where the largest checks were going.
Perz challenged the implication for returns. Heavy infrastructure spending did not automatically make infrastructure the best investment. He asked whether, as in earlier technology cycles, more attractive businesses might emerge among the applications built afterward.
Zabelin emphasized a difference in the funding base: established companies such as Amazon, Microsoft and Apple had substantial businesses outside AI. He also expected countries' desire to avoid technological dependence to sustain investment. That was his case for continued spending, while Perz's question about which layer would earn the best returns remained open.
Better chips complicated the energy forecast
Perz was watching strength across several commodity groups and leaned toward a near-term inflationary effect from the buildout.
Zabelin saw two competing forces. Training and operating AI required power, but chips designed specifically for inference could use energy differently from the GPUs employed to train models. More efficient hardware could reduce the energy required for a task while also making more usage economically practical.
He did not offer a precise net-demand forecast. He expected energy-related commodities, including uranium, to remain prominent in the investment narrative, but cautioned that price action was already being driven by expectations. Identifying the ultimate corporate winners was harder.
Regulation was another variable. Rules that enabled or delayed infrastructure development could reinforce the rally or interrupt it. Perz agreed that the immediate trading horizon looked more actionable to him than a confident forecast ten or twenty years ahead.
Cybersecurity was the next research question
Zabelin also described work underway on AI and cybersecurity. Greater use of models created additional places to protect data, from training and inference to movement between systems. AI could also make some attacks faster or easier to execute.
He expected customers to favor providers offering a broader suite of protections, rather than assembling many disconnected products. That preference supported his view that consolidation would continue in a fragmented industry.
The opportunity he described therefore extended beyond another data center or chip order. Companies adopting AI would also need to secure it, and the ability to combine useful protections in one platform could influence which security businesses attracted customers and buyers.
Questions and answers
#Questions answered in the original 2025-10-30 episode.
What did sovereign AI mean in this conversation?
It described national efforts to develop or control AI capabilities, supported by government policy and capital. Zabelin examined how those efforts affected companies and investment flows.
Why were foundation-model platforms attracting large investments?
They required substantial capital, and investors expected many future applications to build on them. Zabelin said investors were less certain which individual applications would succeed.
Would more efficient chips reduce total energy demand?
Zabelin did not give a definitive answer. More efficient hardware could lower power use per task, while cheaper and broader usage could offset some of the saving. Regulation and infrastructure deployment also mattered.
Adapted by Stock Market TV Editorial from the original video. Source date: . Article published: .