SCX.ai and the Mission to Build Australia's Sovereign AI With David Keane | Ep. 17

SCX.ai listed a week before this recording, and on the morning of the record the Sydney Morning Herald ran a front page calling AI the most important sovereignty issue for Australians since the Second World War. At the same time Australian communities are pushing back hard on new hyperscale data centres, and Australian businesses are being told to adopt AI without being told where their data is being processed. This conversation sits directly on that fault line.

About the guest

David Keane is the founder and CEO of SCX.ai, which listed on the ASX in August 2026. He has taken a company public in Australia twice, the first being Bigtincan, which he started in Sydney in 2011, listed in 2017 and ran largely out of Boston for fourteen years. SCX runs an AI inferencing node in Sydney built on SambaNova ASIC hardware rather than NVIDIA GPUs, and it is pointed deliberately at inference rather than training.

What this episode covers

• Why SCX listed on the ASX rather than taking the offshore capital that was readily available, and what strings that money carries

• What inferencing actually is, why it is measured in tokens, and why the industry should be counting tokens per watt instead of GPUs

• The mega trend from GPUs for training to purpose-built ASICs for inferencing, and why SCX chose SambaNova over the alternatives

• Open weights versus proprietary models, what transparency in the weights really buys you, and the DeepSeek Tiananmen Square example

• Emergent capability: why coding was never trained into a large language model, and why the same is now happening with offensive and defensive cyber

• Project Magpie, SCX's Australian fine-tune of an open weights model, and why as much as 30 per cent of a prompt can be wasted on localisation

• Opir, the guardrails classifier model built in wartime Ukraine to detect 996 types of attack, now being brought to Australia through SCX

• Why network latency stopped being Australia's problem, and why thinking latency replaced it

• Watermarking in model outputs, and whether human-made work is about to carry a premium

Who this episode is for

• CIOs, CTOs and IT directors who need to adopt AI without sending sensitive data offshore

• Cybersecurity and risk leaders working out where guardrails, watermarking and emergent model capability leave their controls

• Cloud, platform and infrastructure leaders evaluating inference economics, hosting decisions and service levels for agents

• Founders, investors and operators watching Australia's position in the global AI build-out

Show notes

This episode is supported by NinjaOne.

Follow Naran McClung on LinkedIn here: https://www.linkedin.com/in/Naranmcclung/
Follow Jono Staff on LinkedIn here: https://www.linkedin.com/in/jonathanstaff/

And don't forget to subscribe and rate the show! We love hearing your thoughts and answering questions.

For all enquiries, contact pru@thingsreasons.show