We are not the AI referee: Picking winners, not fights

One new model a week
It is hard to overstate how quickly AI is moving. OpenAI, Anthropic, Google, xAI, Meta and Chinese labs such as DeepSeek, Alibaba and Moonshot are releasing new models roughly weekly - about 35 this year alone - each typically better, faster and cheaper than the last. Competition has turned what was a novelty two or three years ago into a real tool for writing, analysis, coding and research.

The backlash is building
But speed brings risks, and the builders themselves are sounding the alarm. This month, the heads of Anthropic, OpenAI, Google DeepMind and xAI backed calls to slow development after AI agents acted outside their instructions, including one that coordinated to hack a startup. Anthropic's Dario Amodei has called for a global slowdown and independent auditors inside the labs.

It has also hit home. In June, an OpenAI agent researching Medicare spending breached a statistics portal, and it took almost three months to tell Canberra. That disclosure gap matters more than the breach itself. A taskforce is now reviewing reporting, enforcement and cyber protection, with agent liability and prompt incident reporting firming as priorities ahead of AI-specific laws due from 2027. That sharpens the social licence question for every data centre application, just as several of the country's largest projects, with OpenAI and Anthropic as partners, await government clearance.

The second concern is physical. Data centres consume enormous amounts of power, water and land. Victoria's data centre energy use is forecast to quadruple by 2030, Tasmanian developers have struggled to win a social licence, and this week Goodman Group scrapped a A$1.2bn, 90MW data centre in North Sydney after public backlash.

Governments are responding. Canberra is moving to enforce power and water standards, and Victoria's new strategy bans drinking water for cooling, requires a 150-metre buffer from homes and makes operators supply their own renewable energy. As the Victorian government put it, data centres must "power themselves, protect homes and pay their way".

Our job is to pick winners, not referee
These debates deserve serious answers, but it is not a fund manager's job to decide whether AI is regulated enough. Our job is to pick winners and avoid losers. That means treating regulation, community opposition and safety concerns as investment risks to understand and price in.

Tighter water rules change the economics. A mandated buffer changes where you can build. A liability regime changes who pays when an agent misbehaves, and which partners a regulator is comfortable with. None of this kills the opportunity; it forces you to be pickier. Operators with power, water and approvals already locked in become more valuable when new entrants face higher hurdles. Scarcity is a moat, and first-mover advantages are strengthening by the day.

Earnings we can see, not blue sky
Being selective also means valuation discipline. AI has produced extraordinary stories built on five- or ten-year forecasts of what might be earned if everything goes right. At this pace, long-range forecasts are less reliable, not more. A leading model can be obsolete in weeks, a chip spec quickly surpassed and, as this week showed, the policy mood can flip overnight.

We prefer earnings we can see: contracted revenue, visible order books, real customers paying today. Higher valuations mean more risk, not more excitement, particularly for companies not yet profitable. When a theme moves this quickly, the ability to see the earnings is worth more than the size of the promise.

Australia: The underappreciated AI trade
Australia remains one of the cheapest, most underappreciated ways to play AI - not through hyperscalers or chipmakers, but through infrastructure. HSBC puts Australia's data centre pipeline at A$97 billion, about 3.4 per cent of GDP. As PIMCO's Adam Bowe put it, "at the end of the day, this is just real-estate investment in Australia... we rent the space out."

That exposure runs through data centre operators and landlords, industrial property, electrical and construction contractors, and power businesses. The real earnings are in the picks and shovels.

The build-out is getting pricier and facing labour shortages. NRW Group's (ASX: NWH) electrical contractor Fredon just won a $115m contract on a 432-megawatt facility. NWH trades on 17x earnings, well below contractor peers on more than 20x, while growing EPS at more than 15 per cent.

Southern Cross Electrical Engineering (ASX: SXE) provides electrical fit-out, power distribution and backup systems, and is onsite at NEXTDC's S3 and DigiCo's SYD1. Data centres drive much of its EPS growth of more than 50 per cent, justifying its 23x multiple. Genusplus Group (ASX: GNP) designs, builds and maintains transmission and grid connections. With electricity consumption forecast to nearly double by 2050, GNP trades on 17x with its FY27 tender pipeline up 50 per cent year on year.

But playing the theme hasn't been easy. Listed data centre operators have traded poorly over the past year because of the model itself: success consumes capital. The more contracts you win, the more you must build. NEXTDC (ASX: NXT) is the example. A record 250-megawatt increase in commitments triggered a A$1.5bn equity raise in April at a steep discount; FY27 capex is forecast at around A$5.5 billion, and a A$1.1bn convertible note followed this month. NXT now trades more than 10 per cent below the raise price, as rising bond yields and a reassessment of long, cash-hungry build-outs make capital dearer.

Which brings us to Firmus. The Tasmanian-born developer is preparing what could be one of the largest floats in ASX history, reportedly raising around A$7 billion at a valuation of up to A$50 billion, with a listing planned for late October. It claims more than 900 megawatts of contracted capacity and counts OpenAI and Nvidia as partners - a vote of confidence in Australia as an AI infrastructure destination. But its prospectus lands in Canberra's toughest mood yet towards US AI companies, and the same capital arithmetic applies: who funds the next gigawatt, and at what cost?

We are not here to cheerlead, nor to throw shade on a new business the market is crying out for. Firmus has forecasts, not a track record. It is a trust-me story, and we're show-me investors. Show-me stories earn their multiple; trust-me stories need a discount. If the numbers stack up, we'll be happy buyers. If not, we'll wait for the right price. Either way, the AI theme is bigger than one company, and we are genuinely excited by the opportunities it is creating.