#47 - Sovereignty Is an Inference Problem

Last week, in A Hell of a Kill Switch, I ended on this line: own some weights. The point was reliability at the org level: control your most critical dependency so no government can switch it off overnight. However, when talking about sovereignty at a national level (I live in Australia after all), sovereignty doesn't mean what you think it means.

Australia’s National AI Plan names “sovereign digital infrastructure” as a strategic priority. But when you zoom into the fine print, this is what it is: attract the global hyperscalers to build capacity onshore. The key win so far is a $7 billion campus at Eastern Creek: NEXTDC (ASX-listed) builds, owns and operates it, with OpenAI as the anchor tenant. What lands in 2027 is a large-scale GPU supercluster, liquid-cooled and water-efficient, and on paper it passes the “sovereign AI infrastructure” test: SOCI-aligned, Five Eyes, the lot.

However there’s something missing. While NEXTDC owns and runs the cluster on Australian soil, it doesn’t make commitments on tenants beyond OpenAI. The tenancy matters as much as the building since what runs inside is OpenAI’s model, under a US export regime. If Commerce issues a directive tomorrow, that model can disappear just as easily as Fable 5 did a few weeks ago. Owning the data centre doesn’t fix this.

A frontier model hosted in Sydney is still vulnerable to international policies. That's a tenancy, not a sovereignty.

Should Australia train its own foundation model? Maybe. Personally I would love to see that, but it’s the wrong thing to focus on right now.

You don’t need to train the frontier to be sovereign, you need to be able to run it independently of foreign politics. Those are different capabilities, and only the latter is achievable in the short term. Training a competitive foundation model is a multi-billion-dollar arms race against labs with a half-decade head start and a compounding data-and-talent advantage. Running a frontier-class open-weight model on hardware you control is a much simpler investment decision.

Open weight vs Open source: Open weight means you get the pre-trained weights of the model to run locally, but the training data and the full training recipe stay private. Open source provides total transparency, making the entire recipe available for audit, reproduction, and modification.

That doesn’t mean this stuff is easy. While self-hosting at scale takes real capital, real ops and real talent, model inference is a known engineering problem. Serious shops are moving in that direction with small language models for the workloads that don’t need the frontier, and open weights under evaluation for the ones that do.

This is my point about sovereignty: enough compute you control to run weights you can inspect because you can’t guarantee a foreign lab keeps serving you. Open weight models offer a compelling alternative.

While Australia races to lock in twenty-year commitments to host this stuff, the companies selling it are uncertain about their own demand.

In May, xAI leased Colossus 1, the supercomputer it built to train Grok, to Anthropic for $1.25 billion a month (via SpaceX, which holds the lease). By one analysis that single contract throws off $5-6 billion a year, almost exactly xAI’s annualised net loss: the model business breaks even by renting out the cluster built to train it. Weeks later it signed Google too, pushing the combined rent past $2 billion a month. xAI’s compute is worth more leased to its competitors than pointed at its own model, which moves it into the real estate business.

Then, on the 1st of July, Meta. Bloomberg reported it’s standing up a cloud business to rent out AI compute it can’t use itself, and the stock jumped 10% on the news. This from the company that spent $72 billion on CAPEX last year and guided to as much as $145 billion this year. Some analysts now think Meta may have to reduce or even pause further data-centre investment because it already has more capacity than it has demand.

These are related events with a different root cause: xAI is a losing-model problem. Grok lost, so the rational move is to become the landlord. But Meta is a demand problem: it simply built more than the market wants and is scrambling to sublet the excess. Anthropic renting all that compute is the counter-argument here as its inference demand is very real. But the overall message holds: the companies who know most about where AI demand is are more confident as landlords than as labs.

When the people selling the picks and shovels start subletting the mine, ask what they know about the gold.

And it's not only excess capacity. The day after Meta stood up its cloud business, Zuckerberg himself said agentic AI is progressing slower than expected, while, at the same time, promising the real payoff was three to six months out, the way it always is.

Softening demand and excess capacity aren’t a reason to sit tight and wait for the frontier to lower its prices, though. Quite the opposite: the models you can already run yourself are about to get cheaper.

Open-weight models don’t care about the demand hedge, and their cost is already lower for reasons that have nothing to do with the excess capacity. Z.ai’s GLM-5.2 posts frontier-class coding numbers at a fraction of the API cost because it’s cheaper to serve: a sparse mixture-of-experts (MoE) design activates only a fraction of its parameters per token, and a permissive licence means no proprietary margin on top, or you skip the API and run it yourself.

And as the law of supply and demand tells us, the excess GPUs will end up adding a second discount on top: that compute is draining from the labs with disappointing results, and you can rent from the same pool, so the GPUs you’d point at open weights get cheaper even while Anthropic’s demand stays real.

And GLM-5.2 might be the least of it. If DeepSeek was the proof of concept, this looks like the open weights’ ChatGPT moment: a downloadable model going toe-to-toe with the frontier, and a signal to everyone watching that the ceiling is now reachable. Mistral has been circling the same territory from Europe and I expect that others will join. Once the weights are on your infra, nobody else holds the kill-switch.

And yes, the best open weights today are mostly Chinese: DeepSeek, GLM, Qwen. That sounds like the opposite of sovereignty until you remember the whole argument: the model is distributed for free under a permissive licence: you don’t need to run it in China, so you get to control its availability. Provenance becomes a selection criterion (screen it, run inference onshore, keep the weights open), and incentivises other jurisdictions to follow suit, maybe even Australia.

I’ve started making the switch myself. Liv4All, the assistant platform I run, is running GLM-5.2 (with non-China inference) for a subset of its production agents. The trick is having a good eval suite so that evaluating and selecting new models can be done quickly (and I keep Claude as the fallback).

The National AI Plan isn’t missing, but it’s not specific enough: “Attract the hyperscalers onshore” is a starting point, not the end-game. The move is two-sided.

On the supply side: require the compute we’re subsidising to stock a provenance-screened panel of open-weight models (Mistral, GLM-5.2, DeepSeek) run with onshore inference, alongside the frontier labs. And on the demand side: since a supply mandate doesn’t do any good against a hyperscaler’s real demand, the private sector has to generate it. Every Australian enterprise that routes real workloads to open weights is the demand that will make sovereign AI possible. The government can stock the shelf but only buyers justify keeping it stocked.

Plus, it doesn’t even need new buildouts. NEXTDC already picked its anchor tenant (OpenAI) but there’s plenty of room for more. Firmus is building three campuses across northern Tasmania (St Leonards, Wesley Vale, a 288MW factory at Bell Bay) and no anchor announced yet. That’s the decision point, right now: write open-weight availability into who moves in. Those 288 megawatts are worth being choosy about: power, water and grid capacity are finite, so burning them hosting someone else’s tenancy might not be the best long-term strategy.

So, yes, host the hyperscalers, take the jobs and the tax. Just don’t mistake the postcode for sovereignty. We spent this whole cycle worried that AI would make us dependent on the machines. The machines were never the risk. Not owning the weights was.