Product Launch
Introducing SCX Labs — The Global R&D Engine Behind Australia's Sovereign AI
SCX.ai launches SCX Labs at labs.scx.ai — the global research and development layer behind Australia's first sovereign AI factory. Open, environmentally responsible, vendor-agnostic research turning sovereign infrastructure into measurable Australian capability.
Why SCX Labs, and Why Now
Sovereign AI infrastructure is necessary, but it is not sufficient. A national network of onshore inference nodes only matters if Australian organisations can answer three practical questions about the models running inside it: Are these models good enough? Are they good for us? And is the way we are running them sustainable?
For most of the last decade, the answers to those questions have been outsourced. Benchmarks were defined offshore. Model behaviour was shaped by datasets that knew very little about Australia. Energy and water consumption sat in someone else's annual report.
Today, we are taking that responsibility back. SCX.ai is launching SCX Labs — our global R&D layer — at labs.scx.ai.
SCX Labs is where we research, build, benchmark, and deploy AI systems that are open, environmentally responsible, and vendor-agnostic. It sits behind Australia's first sovereign AI factory and turns infrastructure into measurable, defensible capability.
What SCX Labs Is
SCX Labs is not a marketing site for our datacentre. It is the working surface where our research becomes products, our products become benchmarks, and our benchmarks become evidence Australian organisations can actually rely on.
Three principles shape everything we publish there:
- Open. Methods, evaluation harnesses, and results are documented so customers, partners, and regulators can verify our claims rather than take them on trust.
- Environmentally responsible. Every system we ship is measured against an Intelligence-per-Watt budget, not just raw throughput. Performance that ignores power draw is not performance — it is debt.
- Vendor-agnostic. We compare SCX models against frontier LLMs in the open, and we choose silicon, models, and serving stacks based on outcomes rather than allegiance.
Visitors to labs.scx.ai will find three things on day one: our flagship research releases, a public model comparison surface, and our environmental commitments — backed by real numbers, not adjectives.
SCX ACE — A Fourth Path for Enterprise AI Adaptation
The flagship release at launch is SCX ACE.
For years, enterprises adapting large language models to their own domains have been forced to choose between three paths: prompt engineering (cheap, brittle), retrieval-augmented generation (powerful, but operationally heavy), and fine-tuning (capable, but expensive and slow to iterate).
SCX ACE introduces a fourth path. It builds a compact, explainable playbook of operational rules that models can use at inference time — improving accuracy and behaviour without retraining the underlying model and without rebuilding the retrieval stack from scratch.
The result is a much faster, more affordable way to continuously improve model performance against the metrics that actually matter to a customer, with a level of explainability that fine-tuned models simply cannot offer.
You can read the research release at labs.scx.ai/ace.
Project MAGPiE and the Australia 1k Benchmark
Alongside ACE, SCX Labs is the home of the work the Australian AI community has come to know as Project MAGPiE — Australia's first sovereign LLM trained to think and act like an Australian.
Generic global benchmarks tell you almost nothing about how a model handles Australian regulation, Australian English, Australian institutions, or Australian context. So we built our own.
The Australia 1k benchmark is the first publicly described benchmark designed to address a model's "Australian-ness" through a cultural, linguistic, historical, and media lens. It is not a marketing artefact. It is the same harness we use internally to decide whether a model is fit for Australian customers.
You can read the technical overview of MAGPiE in our multi-turn reasoning evaluation appendix, and explore the Australia 1k methodology directly on Labs.
Compare — Frontier Models in the Open
Sovereign capability claims need to be falsifiable. That is why SCX Labs publishes a public comparison surface at labs.scx.ai/compare, where SCX-trained and SCX-served models are evaluated side-by-side with leading frontier LLMs across reasoning, function calling, structured output, and Australian context tasks.
We will not always be the fastest number on every chart. But our customers will always be able to see — in the open — exactly where SCX sits, on what task, on what hardware, and at what energy cost.
A Different Approach to the Environmental Question
The dominant narrative in global AI infrastructure today is straightforward: more racks, more megawatts, more water. SCX Labs is built on the opposite premise.
Our research and inference workloads run on SambaNova Reconfigurable Dataflow Unit (RDU) systems. The numbers are unusually clean for an industry that often hides behind complexity:
- 10–30 kW per rack average envelope, against traditional high-density GPU racks that routinely exceed 45 kW.
- Air-cooled SambaRack and SambaStack platforms — no liquid or immersion cooling infrastructure required.
- No overhaul needed to deploy into existing Australian data centres, because lower power draw and standard cooling fit within current operational envelopes.
The architectural reason is simple. RDU silicon minimises power-hungry data movement using an assembly-line-style dataflow and a three-tier memory design. Multiple large models — up to trillions of parameters — can remain resident on a single node, with millisecond switching that avoids the expensive reloads that drive hardware bloat in traditional GPU clusters.
For SCX Labs, that translates into two commitments we are willing to put in writing:
- We will publish Intelligence-per-Watt figures alongside every model we ship.
- We will not chase performance numbers that depend on power and water profiles Australia cannot responsibly sustain at national scale.
Why This Matters for Australian Customers
The launch of SCX Labs sharpens the offer to Australian enterprise and government customers in three specific ways.
Evidence, not assertion. Every claim made by SCX.ai about its sovereign AI factory — accuracy, latency, sovereignty, sustainability — now has a public research surface behind it. Procurement teams, security architects, and regulators can read the methodology before they sign the contract.
Faster adaptation cycles. With SCX ACE in production and Project MAGPiE in continuous evaluation, customers can adapt sovereign models to their domain at a fraction of the time and cost of a traditional fine-tune cycle, without giving up explainability.
Onshore by default. All research, training data, and benchmarking workflows that touch customer-relevant signals are run inside Australia, on infrastructure subject to Australian law, in alignment with the Government's hosting certification framework and the Policy for the Responsible Use of AI in Government.
What's Next
SCX Labs is launching with ACE, MAGPiE, the Australia 1k benchmark, and the public Compare surface. Over the coming quarters, we will progressively publish:
- Technical notes on each new SCX model and release.
- Benchmark expansions covering regulated-industry workloads — financial services, healthcare, government, and critical infrastructure.
- Energy and efficiency disclosures, including standardised Intelligence-per-Watt reporting across our national node footprint.
- Open evaluation harnesses that customers and partners can run against their own deployments.
Australia's sovereign AI future will not be built by any single product or any single datacentre. It will be built by an ongoing, public, falsifiable research practice — operationalised at the scale of national infrastructure.
That is what SCX Labs is for.
Explore the work at labs.scx.ai, or learn more about the sovereign AI factory powering it at scx.ai.