scx.ai logo

Strategy Blog

Cheap Intelligence, Expensive Everything Else

The State of AI Report 2026 says the cost of a fixed unit of capability is falling roughly 47 percent per quarter. SCX.ai CEO David Keane on why the cheap intelligence era makes power, efficiency and jurisdiction matter more, not less.

By David Keane, Founder and CEO, SCX.ai6 min read

The State of AI Report 2026 landed this week. Nathan Benaich and Air Street Capital have published it for a decade now, and it has become the closest thing the AI industry has to an annual physical. This year's edition is full of headlines you have probably seen. The three-way frontier race. AI models doing AI research. Predictions about agents cloning themselves out of their sandboxes in 2027.

One number stopped me cold. It reframes every conversation I have about AI infrastructure, and I want to walk through why.

Intelligence Is Collapsing in Price

Citing Epoch AI data, the report estimates that the cost of achieving a fixed score on capability benchmarks has fallen roughly 47 percent per quarter. Run that out over a year and a fixed unit of intelligence is getting about 13 times cheaper annually.

Let me put that in context. Moore's Law, the most famous cost collapse in the history of technology, gave us a doubling of transistor density roughly every two years. This is a 13x annual fall in the cost of a fixed unit of capability.

The model capability that was premium, scarcity-priced and frontier-only last year is a commodity this year.

Now here is the part most people get wrong. When the core input of an industry gets radically cheaper, consumption does not fall. It explodes. Economists call this the Jevons paradox. When steam engines got more efficient, Britain burned more coal. When intelligence gets 13 times cheaper per year, everything gets a brain. Every document, every workflow, every customer interaction, every line of code.

Die Trying, or Live to Serve

The report has a line about the frontier race that founders will be quoting for a year: "Either you die trying to reach the frontier, or you live long enough to serve inference."

The frontier has consolidated into a three-way race between Anthropic, OpenAI and Google. Staying on that treadmill costs billions per lab per year, and the moment you slip, you are not a frontier lab anymore.

Meanwhile, OpenAI and Anthropic together have passed a combined 100 billion dollars in annualized revenue run rate. That money is not coming from training models. It is coming from serving them.

The economics of this industry have officially flipped. Training is a capital expense you hope pays back. Inference is a business. Every query, every agent step, every generated token is a transaction. The labs that live long enough to serve inference are the ones turning the cheap intelligence era into revenue.

Power Is the Wall

If intelligence is getting 13 times cheaper per year, what is not?

Look at what happened this week while everyone was reading the report. Google signed a 20-year deal with Constellation Energy to buy 890 megawatts of upgraded nuclear power for its AI data centres, funding 4.3 billion dollars of upgrades across eleven nuclear units. That is a hyperscaler buying its own slice of a reactor fleet because the public grid cannot keep up with it.

The numbers around this are staggering. Global AI data centre power demand is estimated around 132 gigawatts in 2026, comparable to the entire electricity production capacity of Japan. Hyperscaler capital expenditure is projected around 697 billion dollars this year, per J.P. Morgan estimates cited in the press. In Washington, bipartisan draft legislation now proposes making data centre operators shoulder the cost of the grid upgrades their facilities require.

The report says the same thing from the research side: the growth bottleneck has shifted from GPU availability to power availability, grid capacity, and local consent.

Read those two trends together. The model gets 13 times cheaper per year. The megawatt does not. When the constraint moves from the chip to the grid, the economics of AI infrastructure change completely. Efficiency stops being an environmental talking point and becomes the difference between having capacity and not having it.

What This Means for Australia

Every one of those cheap, intelligent tokens still has to run somewhere. The Jevons explosion means Australia's token consumption grows the same way everyone else's does. The question is who serves it, on whose hardware, under whose law, and capturing whose value.

If the answer is Virginia, then the cheap intelligence era is a story about Australian productivity gains paying foreign rents. The value, the jobs, the data gravity and the operational know-how all accrue offshore. And as global competition for capacity heats up, Australians wear the power and price consequences without owning the productive layer.

The report also notes what is happening everywhere else: nations are aggressively funding sovereign AI infrastructure to ensure domestic capability. Australia has had the debate. In the past month alone, security agencies, the union movement, the research community and industry have all landed on the same conclusion: this country needs onshore AI capability.

What Australia does not need is its own frontier lab. That race is already run, at a price no Australian company or government should want to pay. What Australia needs is an inference economy. Serving layers, on efficient hardware, inside our own grid and our own jurisdiction, capturing value every time something gets a brain.

The Bet We Made

This is why SCX.ai exists, and why we designed it the way we did.

We bet on inference, not frontier training, from day one. Inference is where AI becomes real for an economy. It is every query, every agent step, every workflow that touches a customer or a citizen.

We bet on efficient silicon. Our AI factories run SambaNova dataflow hardware that uses up to 75 percent less electricity than comparable GPU clusters and essentially no water cooling, inside existing commercial data centres in Sydney. In a world where power is the wall, that is not a nice-to-have. It is the only way to add serious AI capacity to this country without waiting years for new generation and transmission.

And we bet on jurisdiction. Tokens served onshore, under Australian law, with the value, the jobs and the security posture that come with it.

Intelligence is getting 13 times cheaper per year. The electricity, the water, the land and the laws under which it all runs are not. That is the layer where the next decade of AI value gets built.

It is the layer Australia can still own.

David Keane is the Founder and CEO of SCX.ai, Australia's sovereign AI infrastructure company.

Sources

Related Topics

State of AI 2026inference economicsJevons paradoxAI infrastructureenergysovereign AIAustraliaSCX.ai
Cheap Intelligence, Expensive Everything Else