Strategy Blog
From Inference Cloud to Agent Cloud
Agents use up to fifteen times the tokens of a chat, Australia's AI inquiry has found 'contested accountability' for them, and Gartner has created a new category to manage them. SCX.ai CEO David Keane on why the inference cloud is no longer a platform, and what an agent cloud has to provide: inference, memory, compute, and someone who takes responsibility.
In October, Elon Musk called Grok Bot the fastest-growing product he had ever seen. Within weeks he was explaining that the cost of serving it was, in his words, unsustainably high relative to revenue.
That is the agent economy in one sentence. Agents are the first AI workload where usage going up is not automatically good news for the company serving them. It is also why I spent this week describing SCX.ai not as an inference cloud, but as an agent cloud.
The distinction matters for every Australian organisation now sizing an agent deployment. And it matters for how this country's infrastructure gets built.
One Assignment, Many Calls
An employee asks a question and gets an answer. An agent gets an assignment and starts working. On screen the difference looks small. On the invoice it is enormous.
IT Brief Australia this week pointed to Anthropic's published account of its multi-agent research system. In that data, agents typically used about four times as many tokens as chat interactions. Multi-agent systems used about fifteen times as many. Those are one company's numbers for one system, not a universal law. But they explain why a budget built on chatbot usage falls over the moment agents arrive.
An agent reviewing a supplier agreement retrieves documents, compares clauses, checks policy, drafts findings, then checks its own work. What looks like one request is dozens of model calls, a growing context window, and a chain of tool use that branches, repeats, and runs in parallel.
The inference bill multiplies. That part is well understood. What is less understood is that the bill is the least of what multiplies.
What an Agent Actually Needs
A chatbot needs one thing: a model that will answer. An agent needs four.
It needs inference, obviously. It needs memory, because an agent that forgets what it did five steps ago is not an agent, it is a goldfish with an API. It needs general-purpose compute, because agents run code, call tools, query databases and move files, and none of that happens on a token. And it needs a jurisdiction, because someone has to be answerable when the agent does something it should not have.
Strip any one of those out and you do not have a platform. You have a component with a marketing budget.
This is the conversation I keep having with customers and partners. In a session with DDN this week I said something I have been saying internally for months: tokens matter less than how they come together in a complete solution. Customers are not buying tokens. They are buying a compliant report, a resolved customer case, a reviewed contract. Tokens are an ingredient. The solution is the product.
The Accountability Gap Is Now Official
Here is the part the industry is only starting to admit. Agents do not just consume more. They act. And when they act, someone has to be responsible.
Two stories this week made that concrete. Australia's AI inquiry released analysis finding broad public support for AI agents but, in the phrase that stayed with me, contested accountability for them. In the same week, a report commissioned in Europe warned that sovereign AI agents could evade every kill switch and threaten national security. Both are saying the same thing from different sides. We are deploying systems that act autonomously before we have agreed who answers for them.
Gartner clearly agrees it is a problem worth a category. This week it published its Magic Quadrant for AI Agent Management Platforms, naming IBM a Leader and Boomi a Visionary. When Gartner creates a quadrant for something, the something has arrived.
My view is that the accountability question is ultimately a jurisdiction question. An agent running on infrastructure where the operator, the data and the logs all sit under one legal framework is an agent that can be governed. An agent running across three jurisdictions with a platform company in the middle is not. It is a vendor dispute waiting to happen.
Efficiency Is the Price of Permission
None of this works if the agent cloud is built the way the last generation of data centres was built.
Australians got a reminder of that this week. A proposed AI data centre at Cooks River in Sydney was scrapped after local outcry. South Australia moved to tighten planning laws for data centres. In the United States, the midterm elections were dominated by a backlash against the industry. The pattern is the same everywhere. Communities will not accept infrastructure that imports power and exports value.
An agent cloud that multiplies inference consumption by fifteen cannot afford to be inefficient. It has to be efficient at the chip level, which is why our SambaNova partnership matters, and why SambaNova's SN50 systems now shipping through Supermicro for exactly this class of workload is good news for the whole ecosystem. It has to be air-cooled and dense, which is how we have built in Sydney. And it has to sit where the customer's data already lives, so the power and the value stay in the same place.
What We Are Building
So when I say SCX.ai is becoming an agent cloud, this is what I mean.
Inference across open-weight models, priced in Australian dollars, with the capacity to sustain agent workloads rather than single prompts. Storage and memory, with partners like DDN, so that an agent's context belongs to the customer and not to us. CPU and general-purpose compute for tool execution, where our work with Intel becomes relevant. And sovereignty, so that when an agent does something it should not have, there is an Australian company that answers for it under Australian law.
The partnerships we advanced this week point in this direction. A cyber capability with Fifth Domain, because the first thing every enterprise wants from agents is that they do not become attack surface. Validation work with DDN on sovereign storage. Developer channels like Cursor, because agents get built where developers already are.
We also opened pay-as-you-go inference this quarter and have passed around 1,500 sign-ups. I mention that not because sign-ups are revenue, but because of what they tell us. The demand for open-weight models running in Australia is real. Turning that demand into something beyond commodity tokens is the work in front of us. An agent cloud is how we do it.
The inference cloud was a place you sent a prompt. The agent cloud is where work gets done. Australia needs one of its own.
David Keane is the Founder and CEO of SCX.ai, Australia's sovereign AI infrastructure company.
Sources
- AI agents will multiply your inference bill. Is your enterprise ready?, IT Brief Australia
- Australia's AI inquiry analysis finds broad support but contested accountability for AI agents, Australian Cyber Security Magazine
- Sovereign AI agents could evade every kill switch and threaten national security, report warns, Euronews
- IBM named a Leader in the 2026 Gartner Magic Quadrant for AI Agent Management Platforms
- SambaNova SN50 Chips to Power Supermicro's New Inference Racks, SambaNova
- Cancelled AI data centre 'victory for people power' after backlash, ABC News
- AI data centre boom sparks backlash as SA moves to tighten planning laws, 7NEWS
- The data centre industry is one of the biggest losers of the US midterms, ABC News
- The middle layer will make or break enterprise AI agents, CIO