Beyond the Green Grid: Australia's AI Sovereignty Gap Is About Compute, Not Kilowatts

The PM's speech correctly identified the dependency risk: Australia cannot afford to remain the last link in the digital supply chain. But the policy response — energy obligations and AI standards — addresses the infrastructure footprint without addressing infrastructure access. An Australian university researcher or early-stage AI startup faces the same commercial GPU pricing as any other customer of AWS, Google Cloud, or Azure. The AI Standards framework does nothing to change this.

The instrument the framework is missing is a Compute Dividend: a legal obligation on any foreign hyperscaler granted planning approval to build large-scale AI infrastructure in Australia to allocate a fixed percentage of high-performance AI compute — specifically GPU and TPU capacity — to domestic startups, universities, and government research institutes at zero or nominal cost.

No country has enacted this yet. The closest analogues are all publicly funded: the US National AI Research Resource ($140 million, voluntary industry contributions); the UK's Isambard-AI supercomputer (~£300 million, government-funded); and the EU's EuroHPC AI Factories (€1 billion+, member-state funded). These are important programs, but they require governments to spend money to buy what could instead be mandated as a condition of market access.

Australia's leverage position is unusually strong at this precise moment. Microsoft's $3.3 billion and Amazon's $13.2 billion commitments are forward-looking and not yet fully deployed. The government retains regulatory leverage at the planning approval stage. A 2–3% compute allocation obligation applied to declared AI compute capacity as a condition of planning approval would provide meaningful access for CSIRO Data61, university research centres, and early-stage Australian AI companies that currently cannot afford GPU time at commercial rates.

Australia's domestic AI ecosystem is more capable than the digital colony framing implies. CSIRO Data61 operates one of the world's largest applied AI capabilities. Harrison.ai processes over 35% of all chest X-rays in NHS England. Canva's AI suite serves 220 million monthly active users across 190 countries. Atlassian's Rovo agent platform competes directly with Microsoft Copilot. The constraint is not capability — it is capital architecture.

The Early Stage Venture Capital Limited Partnership (ESVCLP) program's $200 million fund cap and $10 million per-company investment ceiling were designed for an earlier era of the Australian startup ecosystem. They are now binding constraints on the ability of domestic fund managers to deploy at scale or attract large limited partners — superannuation funds, sovereign wealth vehicles, and global institutional investors who cannot write meaningful cheques into a $200 million fund structure.

The March 2026 R&D review should deliver three specific ESVCLP reforms: raise the fund size cap from $200 million to $500 million; raise the per-investee ceiling from $10 million to $30 million to accommodate growth rounds; and introduce upfront income tax relief for investors at 30% of invested capital, modelled on the UK's Enterprise Investment Scheme. This last change — moving from exit-stage CGT exemption to entry-stage income tax relief — materially changes the risk calculus for institutional capital considering the asset class, and is the single reform most likely to catalyse new fund formation at scale.

The copyright protections in the AI Standards framework are the right instinct. Requiring that Australian creative works — books, music, journalism — cannot be used to train AI models without creator consent and financial compensation is both ethically correct and globally leading. The Attorney-General's consultation process should be expedited.

But copyright protection creates an Australian economic vacuum if not paired with domestic compute capacity. Protecting Australian content without funding the infrastructure for Australian AI companies to train on it simply ensures that any content deals will accrue to foreign platforms rather than to domestic AI builders. A revenue-based Digital Services Tax is the mechanism that simultaneously generates funding for domestic compute infrastructure and builds a local AI industry that balances the global AI industry's structural tax extraction.

A DST of 2–3% applied to Australian-sourced revenue above a threshold would generate approximately $660 million to $990 million annually against the $32.95 billion aggregate income base established in our FY2022/23 analysis. Ring-fenced to sovereign compute infrastructure — funding a national AI research compute resource, expanding CSIRO Data61's GPU capacity, and subsidising compute access for early-stage AI companies — this revenue directly converts tax reform into sovereignty capability.

There is a structural imbalance at the heart of Australia's current tech policy that the AI Standards framework does nothing to address. The companies being asked to harden Australia's digital infrastructure — its telcos and domestic ICT providers — pay the full 30% corporate tax rate, absorb multi-billion-dollar regulatory imposts, and carry the cost of every public-good obligation the government mandates. The foreign platforms extracting value from that infrastructure pay a blended effective rate of 1.7%.

Our recent analysis of Australia's telco regulatory environment found that the government treats its domestic network operators as a 'magic pinata' — one that can be beaten repeatedly for policy dividends without consequence or tradeoff. In May 2026, ACMA confirmed a combined $7.32 billion spectrum renewal bill for Telstra, Optus, TPG, and NBN Co — a figure that arrived well above the $5.0–6.2 billion range the industry had been told to expect. Simultaneously, the Universal Outdoor Mobile Obligation places binding coverage commitments on carriers against a technology readiness timeline their own engineers say cannot be met before late 2028 at the earliest. Telstra has invested $12.4 billion in its mobile network in the seven years to FY25; the regulatory environment is making the next $12.4 billion harder, not easier, to justify.

The government cannot simultaneously demand that domestic telcos fund national digital resilience and extract billions from those same companies through spectrum pricing and unfunded compliance obligations. Every dollar extracted via regulatory impost is a dollar that cannot be spent on the redundant backhaul, backup timing infrastructure, and terrestrial densification that outage post-mortems keep identifying as under-resourced. The same capital constraints apply to domestic ICT and software firms competing in a market where foreign hyperscalers face no equivalent regulatory burden and, as our tax analysis shows, no equivalent fiscal obligation either.

Australia has seen this film before. Over the twenty years to 2025, it watched its domestic media sector be progressively hollowed out by OTT streaming platforms and global digital advertising networks. These global operators arrived without equivalent regulatory obligations, paid minimal local tax, and faced no equivalent spectrum or licensing costs. The result was not a competitive market. It was a structural transfer of the advertising dollar from organisations that employed Australian journalists and funded Australian content, to platforms headquartered in California that returned little of that value to the Australian economy. By the time policy responses arrived — the News Media Bargaining Code, the streaming content obligations, the pending platform regulation — the structural damage was largely done. The domestic media sector is now a fraction of what it was.

The government might now repeat that mistake at scale in the technology sector. A domestic AI and ICT industry that carries full tax obligations, full regulatory compliance costs, and no access to mandated compute or venture capital incentives commensurate with the opportunity, will not be able to compete for capital with foreign platforms that carry none of those burdens. The outcome will be a domestic tech sector that looks like the domestic media sector circa 2025: hollowed out, acquisition-prone, and structurally dependent on foreign platforms.

The reset required is not deregulation for its own sake. It is the recognition that regulatory obligations, fiscal imposts, and compliance costs are not free — that they are trade-offs with investment — and that a government serious about AI sovereignty must apply those trade-offs consistently. That means: spectrum pricing methodologies that account for network investment capacity, not just licence value in isolation; UOMO timelines tied to technology readiness, not political calendars; and an R&D and venture incentive framework that gives domestic AI companies a structural advantage on their home soil comparable to what foreign platforms enjoy through tax minimisation. The red carpet currently rolled out for hyperscaler investment announcements should extend to the domestic builders too.