AI boomers, doomers, and why sovereignty and investment now matter more than hype

AI is now expensive, fast moving strategic infrastructure, not a toy, and Australia and our region are still arguing with the AI boomers and the AI doomers while other countries quietly build the machines that will decide who sets the rules.

The “AI boomers” in this debate are the people who see upside everywhere, productivity gains, new drugs, climate models, and a wave of innovation that will supposedly lift all boats. The “AI doomers” are focused on existential risk, runaway systems, and the concentration of power in a few US based firms who control both the models and the cloud infrastructure they run on. Both camps are responding to a real phenomenon, capability is improving on timescales measured in months, and each new frontier system arrives with broader modalities, sharper reasoning, and tighter integration into existing toolchains.

From an Australian vantage point, the more interesting position sits between these extremes. It asks less, “will AI save or doom us?” and more “who owns the compute, who sets the terms of access, and how much are we prepared to pay for capability we do not fully control?” That is a sovereignty and investment question, not a philosophical one.

Who pays for frontier AI, and how do they get their money back?

Frontier AI is not cheap. Training and serving the current generation of large models requires massive capital spend on accelerators, data centres, power, cooling, and networking, and those costs are rising just as demand from enterprises and governments spikes. In practice, three groups are paying the bill:

• US hyperscalers, who treat AI as a way to deepen lock in and sell more cloud.
• Venture investors, who are betting that at least some of today’s model labs will turn into high margin platforms or get acquired.
• Governments, who are increasingly underwriting domestic compute as “national infrastructure”, either directly or via subsidies and joint ventures.

Return on investment is still speculative. The platform players are banking on subscription revenue, usage-based pricing and differentiated enterprise services to recoup multi-billion dollar training runs. Startups are gambling that they can move up the stack and own workflows rather than raw models, capturing recurring software revenue instead of living solely on API calls. Governments, particularly in the EU and parts of Asia, are framing high end compute as a public good, necessary for science, defence, and industrial competitiveness, with “returns” measured in growth and strategic autonomy rather than dividends.

For Australia and the wider Indo Pacific, the danger is obvious. If we rely entirely on someone else’s machines to run our critical workloads, then our ability to govern AI is constrained by their pricing, their export controls, and their policy choices. Sovereign AI risk shows up here as regulatory dependency and commercial dependency rolled into one.

China’s GPU‑free supercomputer and the sovereignty play

China’s new LineShine supercomputer is a concrete example of what it looks like when a country decides that reliance on US GPUs is no longer acceptable. LineShine is built around domestic CPUs rather than Nvidia or AMD accelerators, reportedly packing around 13.79 million cores into dense cabinets, using a proprietary interconnect and KylinOS, and drawing roughly 42 MW of power to reach about 2.198 exaflops on the traditional LINPACK benchmark. It takes the top slot on the TOP500 list for high-performance computing (HPC), displacing US systems like El Capitan and Frontier on that metric.

On mixed precision benchmarks designed to capture AI style workloads, GPU heavy architectures still dominate, and the US retains a lead, which is why LineShine ranks lower on tests like HPL AI even while topping the classic HPC list. But this misses the strategic point. China has shown that it can build and operate an exascale class machine without restricted Western accelerators, using its own chips and software stack end to end. That matters because it reduces vulnerability to export controls and gives Beijing more freedom to decide what gets run, when, and at what scale, whether that is climate modelling, military simulations, or large-scale AI training.

LineShine also reframes the conversation about who pays and why. It is not built to chase consumer AI subscriptions. It is built as national infrastructure, justified as a long-term sovereignty investment rather than a near term product P&L. For countries like Australia, that is the uncomfortable contrast. We talk about sovereign AI and data sovereignty, but our high-end compute is scattered across university clusters, national facilities that are oversubscribed, and commercial clouds headquartered elsewhere.

Bigger isn’t always smarter

There is a tendency in AI marketing to equate “bigger” with “smarter”. The scaling literature tells a more nuanced story. DeepMind’s Chinchilla paper argued that many large language models have been trained on too few tokens relative to their parameter count, and that for a fixed compute budget you get better performance by using more data for a smaller model rather than simply increasing parameters. Follow up work across labs has shown that beyond certain scales, returns diminish rapidly unless you change architecture, objectives, or training regimes in ways that better match your tasks.

Richard Sutton’s “bitter lesson” is relevant here. Over time, general methods that can absorb more compute and more data tend to beat hand engineered cleverness, but even within that paradigm there are more and less efficient ways to spend your FLOPs. Studies of distillation, compression, and retrieval augmented generation demonstrate that relatively modest models, paired with good training and access to external tools and knowledge, can rival or beat much larger dense models for many practical applications.

For investors and policymakers, the takeaway is simple. Paying for “the biggest model” is often a vanity project. The smarter spend is on right sized models, data pipelines, evaluation, and integration that actually deliver value per unit of compute and power. For Australia, that should mean prioritising shared mid-scale capability and smarter use of regional and alliance infrastructure over chasing headline parameter counts we cannot realistically fund or sustain.

An Australian and regional lens on AI sovereignty and investment

From a middle power perspective, the key questions about AI are less metaphysical and more practical.

• Are we building or buying the compute that underpins our science, defence, and critical services?
• Who can turn off the models that our institutions increasingly depend on?
• How exposed are we to other people’s export controls, platform rules, and investment cycles?
• Where, exactly, do we expect the return on our AI spending to come from, and who captures it?

If LineShine represents one end of the spectrum, state backed, domestically controlled exascale infrastructure, then the other end is a public service that quietly runs its core workflows on a single US cloud region, using models whose access conditions are governed by foreign regulators and corporate risk committees. Most Australian organisations sit closer to that second end today.

A sensible regional strategy would combine several moves, more serious investment in shared compute and data infrastructure here at home, deliberate participation in allied initiatives that give us seats at the table rather than just access keys, and a shift in board level thinking from “AI as a feature” to “AI as a dependency that needs to be governed like any other strategic supplier”. Locally, that also means treating questions like “who pays and how much?” as governance questions rather than procurement details. If we are underwriting other people’s machines with public money, we should be clear about what we are getting back, not just in jobs, but in capability, control, and long-term resilience.

The AI boomers and doomers will keep arguing loudly about whether AI will bring us utopia or catastrophe. The more useful conversation, especially for Australia and our region, is quieter and more concrete: what infrastructure we are willing to build, what dependencies we are willing to accept, and whether we intend to be customers or owners in the systems that will shape our future.