Strategy Blog
Bragawatts Don't Run Inference
Bernstein found only about 35 percent of the 500 gigawatts of planned US AI data centre capacity is likely to ever be built, and Gartner warns over US$3.5 trillion of AI infrastructure spending could face impairment risk by 2030. SCX.ai CEO David Keane on why the next phase belongs to delivered compute, not announced compute.
The AI industry discovered a new unit of measurement this month. The bragawatt.
Bernstein Research analysed nearly 500 gigawatts of planned AI data centre projects in the United States. Their finding: only about 170 gigawatts, roughly 35 percent, is realistic and likely to actually get built. The rest is press releases. Capacity announcements with no secured power, no land, and no path to concrete. The analysts coined the term bragawatts for exactly this kind of megaproject theatre.
I have been in infrastructure long enough to know how this movie plays out. Announcements are cheap. Transformers are not.
The Money Is Starting to Ask Questions
The part that should focus everyone's attention is that the capital markets are now doing the maths.
Gartner published a warning that over US$3.5 trillion of AI infrastructure spending could be exposed to impairment risk by the end of 2030 unless enterprise AI revenue catches up to the capex. In the same analysis, Gartner predicted that up to 80 percent of independent neocloud providers could merge or cease operations by 2030 under capital pressure.
Meanwhile, cloud tenants have started charging data centre developers late-delivery penalties of more than US$5 million per day when power timelines slip. That is up from around US$1 million. When your penalty clause triples and your grid connection is stuck in a queue behind three hundred other projects, the financing math gets brutal fast.
Even NVIDIA is now looking at repurposing decommissioned nuclear plant sites in Germany, because the existing high-capacity power lines are worth more than the ground they sit on. When the smartest money in AI is scavenging old reactor sites for grid access, the constraint is no longer subtle.
To be clear, I am not an AI bear. Inference demand is real and it is compounding. Agents already make up the majority of token traffic. The problem is not demand. The problem is a chasm between announced capacity and deliverable capacity, and an industry that has been valuing the announcement over the delivery.
Why This Matters More in Australia Than Anywhere Else
Australia does not have gigawatts to waste on theatre. Our grid is tighter, our capital pool is shallower, and our sovereign capability is at stake.
The federal government's consultation on "Getting it right: Building AI infrastructure that works for Australia" closed earlier this month. The direction of travel is unmistakable: mandatory minimum standards for large data centre development covering energy, water, land use and security, with supporting legislation expected in early 2027. The updated Policy for the Responsible Use of AI in Government also now embeds hard data sovereignty requirements across federal agencies.
Read those two together and the message to the market is simple. Australia will not be awarding sovereignty points for megawatts that exist only in an investor deck. Every megawatt we commit as a nation should be a megawatt that actually serves tokens to Australian enterprises and agencies.
That is a feature of our market, not a bug. Scarcity forces discipline.
How We Built SCX.ai for This Exact Moment
We never built SCX.ai on the bigger-is-better theory. We built it on efficiency and delivery.
Node 1 runs today at Equinix SY5 in Sydney. It is air-cooled and ASIC-based, delivering up to 10x the energy efficiency of traditional GPU-based inference without a drop of water for cooling. We measure ourselves on Intelligence per Watt, because in a country where power is the price of admission, tokens per megawatt is the only efficiency metric that matters. Node 2 is targeted for late 2026, and we will announce it the same way we announced Node 1: when it is racked, powered and serving customers.
That discipline is looking less like conservatism and more like foresight. When the bragawatt bills come due, operators whose capacity is live, metered and efficient will still be standing. Operators whose capacity was a press release will be explaining themselves to lenders.
Five Questions to Ask Any AI Infrastructure Provider
If you are an enterprise or government buyer signing a multi-year inference contract, the bragawatt era changes your due diligence. Here is what I would ask.
Is the capacity live today, or announced for a future quarter? Ask for the rack count and the uptime, not the roadmap slide.
What is the energy efficiency in tokens per watt? If a provider cannot answer that with a number, they have never had to.
Does the facility use water cooling? In Australia, water is as contested as power.
Who controls the node, the telemetry and the model weights? Residency without operational control is just a rented postcode.
And finally: does the provider's business model survive if capex markets tighten? A neocloud with a 2030 impairment risk profile is a supply chain risk, not just an investment thesis.
Delivered Compute Wins the Next Phase
The AI buildout is not a bubble in the way sceptics claim. The demand is real, the workloads are real, and the productivity gains are landing in P&L statements, not just demos.
But the financing layer is finally being marked to market. Trillions in planned spending is being stress-tested against actual enterprise revenue, and only the efficient, the delivered and the operationally sovereign will clear the bar.
That is the phase SCX.ai was built for. We are not the biggest announcement in the market. We are one of the few Australian providers whose inference capacity you can use this afternoon, on efficiency-first hardware, on our own soil, under our own laws.
Bragawatts don't run inference. Delivered compute does. That has been the plan all along.
Sources
- Business Insider: AI data centers that may never be built and the rise of 'bragawatts'
- Gartner: Over US$3.5 Trillion of AI Infrastructure Spending Could Be Exposed to Impairment by End of 2030
- Tom's Hardware: NVIDIA eyes converting old nuclear plants into AI data centres
- The Guardian: The datacentre next door, on the frontline of tech's latest David v Goliath battle
- Department of the Prime Minister and Cabinet: Getting it right, Building AI infrastructure that works for Australia
- digital.gov.au: Policy for the Responsible Use of AI in Government
David Keane is the Founder and CEO of SCX.ai, Australia's sovereign AI infrastructure company.