Sovereign Coding Agents: Cursor and SCX.ai, End to End

Last week we signed something I want Australian engineering leaders to notice. SCX.ai is a Cursor Marketplace Partner. Our plugin is live on the Cursor Marketplace, and Cursor has accepted us into the Cursor Partner Network early access program. Cursor, now part of SpaceXAI, is launching its partner network this fall, and SCX.ai will be there as a launch partner.

I have written a lot about the inference flip, the shift from humans asking questions to agents doing work. Nowhere is that shift more visible than in coding. This partnership is our answer to the question it raises: where do your coding agents actually run?

Coding Went Agentic. Sovereignty Didn't Keep Up.

Cursor has over 50,000 businesses on its platform, including 67 percent of the Fortune 500. Over 150 million lines of enterprise code are written per day with Cursor. Those are Cursor's own numbers, and they describe the new normal: agents reviewing code, refactoring, writing tests, and running migrations overnight while the team sleeps.

Here is the part that keeps security officers awake. Every one of those agent loops carries context. Your repository, your internal documentation, your architecture, your customer code. An agentic coding session is not a question and an answer. It is your IP, streaming continuously to wherever the inference happens.

For most of the world that is a cost question. For Australian government agencies and regulated enterprises it is a governance question. Data residency obligations, the US Cloud Act, procurement policy. The tooling went agentic before the sovereignty layer caught up. That is the gap we just closed.

The Workflow, End to End

This is how an Australian team runs Cursor on sovereign rails today.

Your developers work in Cursor as they already do. Nothing about their workflow changes. They install the SCX.ai plugin from the Cursor Marketplace, point it at api.scx.ai with an SCX key, and pick their models: SCX Coder for coding, MAGPiE for reasoning, or the open-weight line-up.

From that moment, every completion, every agent step, and every tool call in the loop resolves on SambaNova ASIC infrastructure in Sydney. The developer sees Cursor. The agent sees an API. The regulator sees Australia.

The controls are the ones our enterprise customers already run: no input or output logging, no training on customer data, immutable audit trails. Configuration maps cleanly to APRA CPS 234, IRAP, and ISO 27001 expectations. The agent works all night. The tokens stay home.

Why This Stack Wins on the Numbers

Microsoft Research measured agentic coding tasks consuming on the order of 1000x more tokens than chat. At that multiplier, per-token economics are not a procurement detail. They decide whether the agent program scales or dies in the budget review.

So here are our numbers. SCX.ai serves inference at roughly 12 cents per million tokens, against 80 cents to a dollar at the major providers. Up to 80 percent lower cost, with transparent per-token billing, capacity reservation, and payment in Australian dollars.

SCX Coder, our fine-tuned coding model, delivers 6x throughput and 41 percent lower latency than the baselines we measured it against. And the whole stack runs on SambaNova dataflow ASICs that are roughly 10x more energy efficient than GPU-based inference. Air-cooled, no water. Node 1 saves 10 MWh of energy and 30,000 litres of water per million reasoning requests.

One more thing that matters to platform teams. SCX Router is a single OpenAI-compatible API across the model catalogue. Your engineers move workloads to sovereign infrastructure without re-platforming. It is a routing decision, not a migration project.

What We Are Building Next

The marketplace plugin is step one. It runs through MCP, which works today. What we are building with the SpaceXAI solutions team is the end state we all want: SCX as a native Cursor model provider. Models selectable directly in the model menu, with requests routing to SCX for sovereignty or speed requirements. Sovereignty becomes a dial your administrators set, per workspace, per project, per data classification.

We are also live as a provider in the Portkey enterprise AI gateway, so teams already routing through Portkey can reach Australian inference the same way. The theme across all of it is simple. Where your agents run should be a policy you control, not a decision someone else made for you.

The Point

Every Cursor seat is a potential sovereign token consumer. For the first time, an Australian engineering team can run a world-class AI coding platform end to end on infrastructure that never leaves home soil. The editor is Cursor. The models are ours and the best open weights available. The tokens are processed in Sydney, under Australian law, at a fraction of the cost.

The inference flip means agents now do the work. Sovereignty means you decide where the work runs. That is the partnership we just signed, and we are just getting started.

Sources

• Cursor
• SCX.ai plugin on the Cursor Marketplace
• Microsoft Research: How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks
• SambaNova × SCX.ai: the world's first RDU sovereign AI cloud
• SCX.ai models and pricing

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