- Megaport closed its Latitude.sh acquisition in November, pairing automated bare-metal compute with its global network-as-a-service fabric
- The company raised A$827M ($594M) and signed four AI contracts to build a Globally-Distributed AI Inference Cloud across 1,100+ data centers
- Beyond inference, Megaport sees surging demand for AI "sandboxing" — running coding agents in isolated micro-VMs to protect proprietary data
For more than a decade, Megaport's job was to connect customers to other companies' clouds. Now the Australian network-as-a-service (NaaS) provider wants to be one, building a Globally-Distributed AI Inference Cloud stretched across the 1,100-plus data centers Megaport connects in 31 countries.
The pivot began in earnest with the November acquisition of Latitude.sh, a Brazilian compute-as-a-service company with presence in 10 countries, and accelerated June 2, when Megaport disclosed four AI infrastructure contracts worth a combined A$458.9 million (US$329 million) and launched an A$827.3 million (US$594 million) capital raise, its second in about 18 months.
From plumbing to compute
Megaport sees the marriage with Latitude as a natural match. Megaport has spent 13 years automating connectivity; Latitude spent its life automating servers. "They orchestrate bare metal, and we orchestrate network, so the two fit together quite well," Cameron Daniel, Megaport's CTO and a founding engineer, told Fierce in an interview at Cisco Live this month.
Matt Simpson, Megaport EVP of channel and another founding employee, agreed. "We built this incredibly large network globally. Acquiring a compute business makes sense when you've got a network," he said.
Most bare-metal providers stand servers up by hand, on a bespoke, per-customer basis. Latitude automates the process to deploy in roughly five seconds, "which is unheard of, and now when you couple that with the network, it's pretty powerful," Simpson said.
Daniel added, "They automate their compute almost exactly the same way we approach automating a network."
Latitude brings more than 7,700 servers across 20 markets and more than 1,150 customers, skewed toward developers — ad tech, gaming, streaming, blockchain nodes and, increasingly, AI. Megaport contributes roughly 3,100 enterprise customers, heavy in financial services, media and healthcare. Latitude's compute users get private, on-demand connectivity into clouds and data centers, and Megaport's enterprises get GPUs and CPUs a 100-gig private link away.
Why inference and why distributed?
Megaport is deliberately not trying to be CoreWeave. In a March interview at the Nvidia GTC conference, CEO Michael Reid compared his company more closely to Vultr than to the GPU hyperscalers training frontier models. CoreWeave-class players have spent "billions and billions of dollars" on chips and data centers, he noted; Megaport is spending much less — hundreds of millions — but spreading that investment across many locations.
That distributed posture is the heart of the bet. Training concentrates GPUs in a few enormous halls. Inference can be spread across many locations closer to end users, cutting latency. The connectivity layer, long treated as an afterthought in AI buildouts, has become the missing ingredient for neoclouds and enterprises scrambling to wire GPUs to their data.
The distributed requirements of AI workloads are emerging as a key competitive advantage for telcos, which have the sprawling networks and infrastructure required for inference at the edge. It's a second chance for telcos to exploit new revenue streams, after losing out to hyperscalers in the cloud. But telcos face barriers to seizing the opportunity, including dirty data, lagging autonomy and organizational inertia.
The bare metal advantage
Bare metal provides a performance advantage, handing customers the whole machine, sidestepping the "noisy neighbor" tax of shared virtual machines. "The economics of bare metal are extremely favorable compared to traditional cloud. The performance you get out of bare metal is significantly higher," said Daniel. Megaport seeks to deliver a bare-metal product that "feels like a cloud to use."
Megaport also offers cloud sovereignty — a fast-growing slice of demand as governments and regulated industries reshape the cloud market. "If you buy compute in London, we'll never move workloads anywhere else. The data never leaves London," Daniel said.
Agentic AI is another application for Megaport servers and connectivity. The company is seeing demand for AI sandboxing — running coding agents such as Claude Code or OpenAI Codex inside isolated micro-VMs rather than on a developer's laptop, so the network controls exactly what proprietary data the agent can touch. "These sandbox products are effectively like a micro virtual machine in cloud somewhere," said Daniel. The category barely existed a few months ago and is now "consuming more than we can even provide," Reid said.
The connective tissue: Cisco and the neoclouds
Megaport's edge runs on Cisco Silicon One, and the two companies have a five-year-old partnership spanning silicon, automation software and joint selling. The practical payoff for enterprises is consolidation: "You can spin up a Cisco SD-WAN or firewall virtual appliance on our network edge, and provide connectivity to the cloud through the Cisco platform. You don't need to swivel-chair between two different platforms," said Simpson.
The fabric reaches 35 neoclouds across colocation hubs including Equinix, CoreSite, CyrusOne and Cologix, Simpson said. That on-ramp role matters because where enterprise inference should actually run is still unsettled, and Megaport profits regardless of who wins. Workloads remain stubbornly scattered, Simpson said. "Workloads are still distributed very extensively" — across hyperscalers, repatriated data centers, bare metal and neoclouds at once.
Megaport's advantage is pairing a global, code-driven NaaS fabric with automated bare metal and GPUs. Its weaknesses are equally real: it is a minnow next to the hyperscalers and CoreWeave on raw GPU firepower, it is exposed to brutal memory and GPU supply chains, and finding high-density AI racks is increasingly difficult. And costs are going up; a server that was once priced at about $15,000 now weighs in at $50,000 in memory alone, Reid said.
"We're subject to the supply chain, like everybody else is, but the market seems willing to pay for it," Daniel said.