Groq to Launch 4.5 MW Inference Site in Sydney with Equinix

Groq is opening its first site in the Asia-Pacific region. The company will deploy a 4.5 MW facility in an Equinix data center in Sydney, and it claims the site will deliver inference compute up to 5 times faster and cheaper than traditional GPUs and hyperscaler clouds. The press release is dated November 17, 2025. The launch continues Groq's expansion after rollouts in the US and Europe. Groq named Canva as one of its Australian customers.

What exactly is being built

Groq is betting on its own LPU chip and the GroqCloud service. The company was founded in 2016 and positions itself as infrastructure specifically for inference, meaning running already-trained models rather than training them. In Sydney, the site will operate together with Equinix, which provides the space, power and network connectivity. The key parameters from the announcement:

• Facility capacity: 4.5 MW.
• Location: an Equinix data center in Sydney, Australia.
• Claimed advantage: up to 5x in speed and cost relative to GPUs and hyperscalers.
• Connectivity: Equinix Fabric, a software-defined interconnection service.
• Target audience: enterprise customers and the public sector in Australia.
• Named customer: Canva, which has more than 260 million users.

Treat the "5x" figure as a marketing claim. The release does not disclose the comparison conditions (model, batch size, context length, GPU configuration), and there are no independent benchmarks for this site.

Why Groq needs Sydney

For inference, distance to the user directly affects latency. When a request to a model from Australia travels to California or Europe, hundreds of milliseconds of network transit are added to the generation time. For chatbots, voice agents and interactive editor tools, and Canva is exactly that kind of product, the difference is noticeable. A local site removes that portion of the delay.

The second reason is data sovereignty. The release explicitly mentions full control, compliance and data sovereignty. Government agencies and regulated industries, such as finance and healthcare, often cannot send data abroad. If inference runs inside the country, these customers get a formal argument for buying.

The scale, however, is modest. By the standards of modern AI campuses, 4.5 MW is small: large hyperscaler projects are measured in hundreds of megawatts. Groq's wording about "one of the largest" fast-inference sites in the country applies specifically to the Australian market and to a narrow class of workloads. For a region where available AI capacity is scarce, that is enough for the first commercial customers, but not enough to push clouds out of the mass market.

The launch also shows Groq's growth model: rather than building its own buildings, it rents space from colocation operators such as Equinix. This is faster and cheaper in capital costs, although the availability of power and cooling at a given location still limits the pace. Jonathan Ross, founder and CEO, framed the logic this way: the world lacks the compute for everyone to be able to build AI.

Who can use this in practice

Teams in Australia and neighboring countries should measure latency and per-token cost on their own models and compare them with their current provider before believing the "up to 5x" figure. The test is most useful where response speed matters: voice assistants, agent chains with many sequential calls, and interactive features in products. Developers in the public sector and regulated industries should ask what guarantees on data location and certification Groq will provide in Sydney. Everyone else can simply watch for whether public pricing, a list of available models and launch timelines appear in the region. The announcement does not name these details.