AI arms race and implications for Australian technology infrastructure: Jefferies sees durable AI-infrastructure demand supporting selected Australian technology names

Overview

Jefferies summarises an expert call with GlobalData's Bill Rojas on the AI arms race and its implications for Australian technology. The central conclusion is that model competition and falling AI-model prices need not undermine infrastructure demand: rapidly rising token consumption should continue to require substantial chips, power, data-centre capacity and GPU workloads.

Core views

The call's central thesis is that AI infrastructure demand is driven by token consumption rather than the near-term profitability of frontier-model developers. GlobalData's modelling suggests daily global token demand could grow from more than 300 trillion currently to 2-5 quadrillion by 2030, with China already generating roughly 100 trillion tokens daily. More advanced reasoning models can generate an order of magnitude more tokens, so inference demand could exceed expectations. The report says this trajectory supports continued expansion in chips, power and data-centre infrastructure through the end of the decade, with growth potentially moderating after 2032. It cites potential global power needs of approximately 200GW and semiconductor-related capex of about US$950bn in 2026, rising to roughly US$1.4tn by 2030. The discussion distinguishes the competitive position of US frontier labs from the economics of the broader model market. OpenAI and Anthropic are said to retain an advantage in highly complex reasoning, supported by larger R&D budgets and access to leading chips; Bill estimates OpenAI alone spends about US$20bn annually on R&D. China has nevertheless narrowed the performance gap materially over the past 24 months through architecture, efficiency, optimisation and distillation, partly because export restrictions limit access to leading NVIDIA GPUs. Chinese production is described as concentrated around 7nm while lacking EUV capability for sub-approximately-4nm processes, but the report argues that semiconductor design talent and algorithmic innovation can offset part of the hardware constraint. Sanctions are therefore viewed as delaying, rather than permanently stopping, Chinese progress. China's open-weight strategy is presented as a potential pressure point for closed-model economics. The report explains that open-weight models release model parameters, whereas fully open-source models also release training code, data and development processes; most supposedly open-source models today are described as open-weight. Cheaper Chinese models and widespread use of distillation could enable enterprises to use open-weight models for 80-90% of inference tasks, reserving closed frontier models for complex reasoning, mission-critical work and specialised scientific applications. This could reduce the addressable market for OpenAI and Anthropic even as total AI usage rises. Bill expects no meaningful economic returns from frontier-AI investment before 2029, with ROI potentially emerging in 2029-30 but remaining below 10%, as advanced GPUs and growing model complexity keep costs high. The report argues that the demand outlook is particularly supportive for neocloud and data-centre providers over the next three to five years. Neoclouds seek long-term customer commitments to secure utilisation and justify major capital spending; one gigawatt of AI infrastructure can require US$50-60bn of investment. Strong contracted backlogs, financial backing, broad service offerings, workload optimisation and anchor-customer relationships are described as increasingly important advantages as GPU access alone becomes less differentiating. Regional operators may face less direct hyperscaler competition and therefore have stronger local pricing power. Constraints on US data-centre power availability and project approvals could direct incremental activity toward regions including Australia, Malaysia, Indonesia, Thailand, Scandinavia and Northern Europe. For Australian technology, Jefferies says the call reinforces its favourable view of Megaport, NextDC, Infratil through CDC, and Macquarie Technology Group as beneficiaries of durable AI infrastructure demand. Cheaper open-weight models can enable AI start-ups and new business models, which the report says should increase demand for GPU workloads through Megaport's Latitude.sh. Data-centre capacity pipelines at NextDC, CDC and Macquarie Technology Group are positioned to house AI servers as neocloud demand expands. The report separates this infrastructure opportunity from where AI's ultimate economic value may accrue. Enterprise adoption has been rapid, but measurable productivity evidence remains mixed and AI-generated office outputs may provide only incremental benefits. Bill instead sees physical AI—industrial automation, robotics, autonomous systems, transportation, manufacturing, healthcare and intelligent edge computing—as the larger long-term opportunity, potentially exceeding US$100tn over time. Physical AI can affect utilisation, output, costs and productivity more directly, and edge processing is important where cloud latency or connectivity dependence is unacceptable. The report therefore argues that leading long-run winners may be companies embedding AI into industry-specific products, machines and workflows rather than model creators alone. Company valuation frameworks remain specific to the covered names. Infratil is valued through SOTP, including CDC at 32x EV/EBITDA, described as reflecting about 8x EV/contracted EBITDA; One NZ at 7.5x; RHCNZ/Qscan at 13x; Wellington Airport at 16x. Macquarie Technology Group is valued using DCF and SOTP, with an 8.0% WACC, 4.5% risk-free rate, 5.0% equity risk premium, 1.10 beta and 4.5% terminal growth rate. Megaport's 15-year DCF uses an 8.8% WACC and 4% terminal growth, while NextDC's DCF uses a 6.5% WACC and 3.5% terminal growth.

Analysis framework

Jefferies uses an expert-call framework that starts with AI model competition and token-demand drivers, then links token growth to chip, power, neocloud and data-centre requirements. It applies this industry logic to selected Australian and New Zealand infrastructure providers, while retaining company-specific DCF and sum-of-the-parts valuation frameworks and risk discussions.

Methodology notes

• Industry AnalysisSupply-demand framework

  Token consumption as the demand driver for AI chips, energy and data-centre capacity.

  The report links projected growth in daily tokens to higher compute and power requirements, then to demand for AI infrastructure providers.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Transmission from models and enterprise inference workloads to neoclouds, GPU services and data centres.

  Cheaper open-weight models are expected to broaden AI workloads, which the report says increases demand for hosting, networking and data-centre capacity.

• Valuation methodsDCF (Discounted Cash Flow)

  Discounted-cash-flow valuation for MAQ, MP1 and NXT.

  The report values these companies using stated assumptions for WACC, risk-free rate, beta, capital structure and terminal growth.

• Valuation methodsSOTP (Sum-of-the-Parts) Valuation

  Sum-of-the-parts valuation for Infratil and Macquarie Technology Group.

  The approach values separate business assets using assigned EV/EBITDA multiples or asset values, then aggregates them.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Infratil Limited (IFT)

  Covered beneficiary through its CDC data-centre exposure.

  Strengths

  CDC is positioned to serve expanding AI-server demand; its SOTP valuation uses 32x EV/EBITDA, reflecting about 8x EV/contracted EBITDA.

  Weaknesses

  Potential need for additional equity or capex injection.

  Risks

  Construction and contract delays at CDC and Longroad Energy, additional funding needs, and mobile-price competition at One NZ.

• Macquarie Technology Group (MAQ)

  Covered beneficiary with a data-centre capacity pipeline for AI infrastructure.

  Strengths

  The report highlights capacity to meet growing data-centre demand.

  Weaknesses

  Exposure to construction execution and customer concentration.

  Comparison

  The SOTP applies different multiples to CS&G, stabilised data-centre assets, IC3SW and Telecom.

  Risks

  IC3 Super West construction delays, loss of government or enterprise customers, and hosting-price competition.

• Megaport Limited (MP1)

  Covered beneficiary through Latitude.sh GPU workloads and broader AI networking demand.

  Strengths

  Global scale and the potential for more bandwidth demand and value-added services.

  Weaknesses

  Low networking barriers to entry and possible price declines as hardware costs fall.

  Comparison

  Its global scale is described as the main advantage unless major data-centre operators create a virtual cross-connect platform.

  Risks

  Competition, declining pricing, difficulty securing the latest GPU chips, and interest-rate sensitivity.

• NextDC Limited (NXT)

  Covered beneficiary with data-centre capacity to house AI servers.

  Strengths

  Capacity pipeline aligned with rising demand for AI infrastructure.

  Weaknesses

  Capital must be deployed upfront before customer revenue is received.

  Comparison

  Competitors with a lower cost of capital may offer lower prices.

  Risks

  Hyperscaler bargaining power, lumpy contracts and capacity-utilisation gaps, lower-cost competitors, and interest-rate sensitivity.

Key data

• Global daily token demandMore than 300 trillion currently; 2-5 quadrillion by 2030GlobalData modelling cited by the expert; growth is expected to remain strong through 2032.
• China daily token generationRoughly 100 trillionBill Rojas's estimate of current daily AI token generation.
• Potential global power requirementApproximately 200GWPotential capacity requirement associated with 2030 token-demand estimates.
• Global semiconductor-related capexRoughly US$950bn in 2026; approximately US$1.4tn by 2030Bill Rojas's estimate.
• Frontier AI ROIPotentially below 10%Bill does not expect adequate returns before at least 2029; possible ROI emergence is cited for 2029-30.
• Open-weight enterprise use80-90% of use casesBill's expectation for enterprise inference usage.
• AI infrastructure investmentUS$50-60bn per 1GWEstimated capital requirement cited for AI infrastructure.
• Physical AI market opportunityUS$100tn+ over timePotential long-term opportunity cited by the expert.

Impact & implications

The report says model-price compression and competition may weaken the economics of frontier-model developers without reducing infrastructure demand. Instead, broader and more token-intensive AI usage should support regional GPU services, networking and data-centre capacity, with MP1, CDC/IFT, NXT and MAQ identified as relevant beneficiaries. Longer term, the report expects more value creation in physical and edge-AI applications than in foundation models alone.

Risks

• AI safety, liability and cybersecurity risks could become more material as models become more autonomous and capable.
• Open-weight models and distillation could pressure the addressable market and economics of closed frontier models.
• Neocloud and data-centre operators require substantial capital and depend on long-term customer commitments and utilisation.
• Construction delays, customer losses, competitive pricing pressure and higher interest rates are explicit risks for the covered infrastructure companies.
• Enterprise AI productivity outcomes remain mixed, with broad-scale evidence still limited.

What to watch

• The pace of token growth, particularly whether reasoning workloads drive more-than-expected inference demand.
• Data-centre power availability, project approvals and the ability of regional markets to absorb incremental AI infrastructure.
• The development of contracted backlogs and anchor-customer commitments for neocloud and data-centre providers.
• The adoption split between open-weight models and closed frontier models.
• Evidence that physical and edge-AI deployments are delivering measurable productivity, cost and output gains.