SCX.ai CEO: AI inference doesn't need new data centres

SCX.ai's second computing node, which doubles its AI capacity, is on track for later this year, and the company is not waiting for new data centres to be built, co-founder and chief executive David Keane told the Certified Strategic Podcast.

"You use existing buildings. So we're not waiting for new buildings. We're not waiting for power allocation," Keane said in an interview with host Michael Cooney, recorded on 28 October 2026. The ASX-listed AI inference provider had released its first quarterly report on 27 October and announced it had joined SpaceX's Grok partner network.

SCX.ai CEO David Keane on Node 2, sovereign AI and chips built for inference.

Key points from the interview

• Node 2 copies Node 1. Keane said Node 2 is "a copy-paste from Node 1". It doubles SCX's computing capacity, and Keane said it is on track for later this year.

• Air-cooled and spread across sites. Keane said SCX's model is to deploy small blocks of computing at several sites and link them into one cluster. "Our systems are air-cooled," he said.

• Inference only. SCX runs trained AI models for customers and does not sell training capacity to hyperscalers or AI labs. "That's not our business," he said.

• Chips built for inference. SCX uses SambaNova's application-specific chips rather than GPUs. Keane said the chips can be two and a half to five times more efficient at turning a kilowatt of energy into tokens, and four to six times faster.

• 84 models, with AI agents driving growth. Keane said agent workloads can drive a thousand times more use than a chatbot.

• A named customer. Keane said SCX works with Elfworks, a Brisbane-based company that makes an AI tool for Australian accountants.

• Cost. He said SCX can deliver cost savings of up to 80 per cent compared with proprietary AI models.

• Cursor connection. Through the SpaceX network, users of the AI coding tool Cursor can reach SCX's models through a free MCP server, a standard connector between AI tools and outside services.

• Apple developers. SCXSwiftKit, announced in September, connects iPhone, iPad and Mac apps to SCX's hosted models.

• Demand. "I never talk to someone who says, gee, I'm using a lot less AI than I was last month," Keane said.

SCX runs its AI computing on SambaNova chips in rented space at Equinix's SY5 data centre on the Sydney data centre map.

Chapters

• 00:00 Introduction

• 00:40 Why SCX focuses on inference, not training

• 02:30 Who is driving demand

• 03:53 84 models and the rise of AI agents

• 06:02 Why customers want sovereign AI

• 07:22 National security and digital dependence

• 09:43 Why SambaNova chips instead of GPUs

• 11:27 How customers get started

• 12:52 SpaceX, Grok and Cursor

• 15:35 Node 2 and the distributed build model

• 17:21 Why SCX doesn't build big data centres

• 19:04 Node 2 is a copy of Node 1

• 20:05 Data sovereignty and cost

• 21:13 The demand outlook

• 22:09 The risk of slow AI adoption

• 24:30 Customer example: Elfworks

• 25:34 Open and closed models

• 27:06 Air-cooled racks, power and space

• 28:29 What's next for SCX

• 29:46 Three reasons to use SCX