What did Megaport announce in June 2026?

Megaport announced a fully underwritten A$827.3 million entitlement offer on 3 June 2026, priced at A$14.30 per share. The raise funds four new AI inference contracts worth A$458.9 million in total contract value, with US-based technology providers, plus a new A$350 million on-demand GPU Pool. The contracts require A$369.5 million of capital expenditure, mainly on NVIDIA GPUs, and commence in the first half of 2027. Megaport also reported network annual recurring revenue of A$277.7 million, up 25% year on year, and group pro forma ARR of A$662.9 million including its Compute division.

What is Megaport's GPU Pool?

The GPU Pool is a pool of on-demand GPU capacity that Megaport is funding with approximately A$350 million and deploying across its 1,100 connected data centres in 31 countries. Enterprise customers can access it through both contracted and consumption-based pricing. It underpins Megaport's Globally-Distributed AI Inference Cloud, which aims to place GPU compute close to where AI applications run to reduce latency. Deployment runs over 6 to 9 months, with servers reaching optimal utilisation 3 to 6 months after installation.

How is Megaport's strategy different from a neocloud like CoreWeave?

Neoclouds such as CoreWeave, Nebius and Crusoe own large GPU fleets concentrated in a small number of very large campuses, optimised mainly for AI training. Megaport owns GPUs too, through its Latitude.sh acquisition and the new GPU Pool, but distributes them across its existing global interconnection network and targets latency-sensitive inference rather than training. Megaport's differentiator is that it owns both the GPU capacity and the network fabric that connects it, where neoclouds buy connectivity and interconnection providers like Equinix do not own the GPUs.

Megaport raises A$827m to build a distributed AI inference cloud across 1,100 data centres

At a glance

Megaport (ASX:MP1) launched a fully underwritten A$827.3 million entitlement offer on 3 June 2026 at A$14.30 per share, a 13.9% discount to its A$16.61 last close.
The raise funds four new AI inference contracts worth A$458.9 million in total contract value, requiring A$369.5 million of capital expenditure on NVIDIA GPUs, network and storage, with the contracts commencing in the first half of 2027.
A separate A$350 million funds an on-demand GPU Pool, sold on both contracted and consumption-based pricing, to be deployed across Megaport's 1,100 connected data centres in 31 countries over the next 6 to 9 months.
The plan reweights Megaport away from the asset-light network model it built its valuation on, putting company capital at risk on depreciating GPU inventory ahead of signed demand.
Megaport's combination of an owned global interconnect fabric and owned GPU capacity is uncommon. Equinix and Lumen connect to GPU clouds; neoclouds such as CoreWeave and Nebius own GPUs but concentrate them in few locations.

What Megaport announced

Megaport entered a trading halt on 2 June 2026 and disclosed the package the following morning. The company secured four AI inference contracts with United States based technology providers, carrying a combined total contract value of approximately A$458.9 million and requiring roughly A$369.5 million of capital expenditure, predominantly for high-performance NVIDIA GPUs alongside network and storage infrastructure. Those contracts commence in the first half of 2027.

Alongside the contracted work, Megaport committed approximately A$350 million to a new on-demand GPU Pool. The Pool gives enterprise customers access to GPU capacity through both contracted and consumption-based commercial models, and underpins what the company calls a Globally-Distributed AI Inference Cloud spanning its 1,100 connected data centres across 31 countries. Deployment runs over the next 6 to 9 months, with servers expected to reach optimal utilisation 3 to 6 months after they go in.

To fund both, Megaport launched a fully underwritten, accelerated non-renounceable entitlement offer to raise A$827.3 million, one new share for every 3.08 held, priced at A$14.30. That is a 13.9% discount to the A$16.61 at which the stock last traded on 1 June, after a roughly 84% run over the prior month. Megaport CEO Michael Reid framed the rationale directly: "AI inference is becoming a global infrastructure challenge, not simply a GPU problem. As AI adoption accelerates, organisations need seamless access to GPUs, CPUs, storage, and the connectivity that powers them."

The trading update that accompanied the raise shows why the market has been re-rating the stock. Network annual recurring revenue rose 25% year on year in constant-currency terms to A$277.7 million as at April 2026, with net revenue retention of 113%. Including the new strategic contracts, the Compute division now carries pro forma ARR of A$385.2 million, lifting total group pro forma ARR to A$662.9 million. Compute, in other words, is now the larger half of the business on a pro forma basis. FY2026 revenue guidance was tightened to between A$307 million and A$315 million, with EBITDA margin and capital expenditure guidance for the existing business unchanged.

Item	Figure	Detail
Entitlement offer	A$827.3 million	Fully underwritten, 1-for-3.08 at A$14.30, 13.9% discount
New AI contracts	A$458.9 million TCV	Four contracts, US-based AI providers, commence 1H 2027
Contract capex	A$369.5 million	Mainly NVIDIA GPUs, plus network and storage
GPU Pool	A$350 million	On-demand, contracted plus consumption pricing
Footprint	1,100 data centres, 31 countries	Distributed inference across existing network fabric
Network ARR	A$277.7 million	Up 25% YoY constant currency, 113% net revenue retention
Group pro forma ARR	A$662.9 million	Compute A$385.2 million, now the larger segment

Source: Certified Strategic Editorial, Megaport ASX disclosure and trading update, June 2026.

The bet on the model

Until now the company deployed GPU hardware only against a signed customer commitment, typically 24 to 36 months. The GPU Pool inverts that sequence. Megaport is buying NVIDIA silicon speculatively and carrying it as inventory, betting that enterprise burst demand and sovereign inference workloads absorb the capacity at premium rates once it is live.

For a network-as-a-service company that built its valuation on capital-light software margins, carrying depreciating GPUs is a different kind of business. The hedge sits in two places. The four signed contracts provide a revenue floor that de-risks much of the capital before it is committed. And Megaport keeps the clusters when each contract ends, then redeploys them through Latitude.sh, its GPU compute platform, to serve new customers. That extends the revenue life of each asset beyond its first tenant.

The Latitude.sh deal is the part of this story that predates the headline. Megaport bought the bare-metal and GPU provider in November 2025 for approximately A$459 million, funded by a roughly A$200 million equity raise. Latitude.sh then signed three binding contracts worth US$182.9 million with US AI companies in May 2026. Counting those, Megaport now holds more than A$710 million of contracted AI infrastructure revenue. The pivot from asset-light networking to asset-heavy compute began seven months ago.

Why inference, and why distributed

Megaport's wager is specifically on inference rather than training. Training concentrates demand in a small number of very large clusters, where the economics favour hyperscalers and the largest neoclouds. Inference is the opposite shape: it is latency-sensitive, it runs continuously once a model is in production, and it benefits from sitting physically close to the data and the end user. Reid's framing of inference as the bigger infrastructure opportunity of the next decade is a bet that enterprise AI spend is shifting from one-off model builds to always-on serving.

That shape is where Megaport's existing assets matter. The company already operates a software-defined interconnect fabric reaching 1,100 data centres in 31 countries. Placing GPU capacity at nodes across that fabric, and provisioning compute, network and storage through the same software layer, lets Megaport offer inference close to where workloads actually run without the customer stitching together multiple vendors. The strategic claim is that this fills a gap between hyperscaler clouds, which are centralised and general-purpose, and single-location GPU specialists, which are fast but geographically concentrated.

Is the move original?

Within the Australian market, at this scale, Megaport is largely on its own. Globally, two halves of the move are familiar; the combination is not.

The connectivity-to-AI repositioning is well established. Equinix has built its AI strategy around being the neutral interconnection layer, letting customers reach neoclouds, model providers and GPU platforms through Equinix Fabric, and it is explicit that it is not in the business of owning GPUs to resell. Lumen has pursued the same logic through its hyperscaler fibre deals. Cisco has its Secure AI Factory reference designs with NVIDIA. All three connect AI infrastructure; none of them take inventory risk on the silicon.

The GPU-ownership half is equally well established, in the neoclouds. CoreWeave, Nebius and Crusoe own large GPU fleets and sell them as a service, and they are the reference point for GPU-as-a-service economics. Their footprints, though, are concentrated in a small number of very large campuses optimised for training density, not distributed for low-latency inference.

Player type	Owns the GPUs?	Owns the network fabric?	Inference posture
Megaport (MP1)	Yes, via Latitude.sh and the GPU Pool	Yes, 1,100 data centres in 31 countries	Distributed, latency-led
Equinix / Lumen	No, connects to GPU clouds	Yes, interconnection and fibre	Hosts partners' capacity
Neoclouds (CoreWeave, Nebius, Crusoe)	Yes, at scale	No, buys connectivity	Concentrated, training-led
Hyperscalers (AWS, Azure, Google)	Yes	Yes	Centralised, general-purpose

Source: Certified Strategic Editorial, company disclosures and product documentation, June 2026.

What is distinctive about Megaport is the fusion: an operator that owns both the interconnect fabric and the GPU capacity, and uses the former to distribute the latter for inference. As we set out in our neocloud market report, Australia's GPU-first operators have moved from concept to capital deployment quickly, but most have built around owned data centre capacity rather than a global network layer. Megaport is approaching the same demand from the connectivity side.

Who else is on the same track

The closest parallel to watch is Equinix's partnership with inference specialist Groq, which is being rolled out in Australia and Europe and feeds dedicated inference capacity through Equinix's interconnection estate. If that model proves out, other interconnection operators will likely replicate it, which would validate Megaport's distributed-inference thesis while competing for the same enterprise demand. The difference remains ownership: Equinix routes to Groq's hardware, while Megaport carries the hardware on its own balance sheet.

The connectivity layer of the AI stack is becoming an active battleground for inference, and an ASX-listed operator has chosen to compete by owning capacity rather than only carrying traffic. Megaport now joins the operators tracked in the Australia Data Centre Index. The move sits within the broader supercycle we covered in our analysis of the JLL Asia Pacific data centre report, and it reinforces the timing argument in our piece on Australia's window to capture its share of the AI infrastructure boom. It also connects to the demand picture set by hyperscaler commitments landing locally, including the operator relationships flagged in our coverage of the Anthropic Australia MOU.

What to watch

Three checkpoints will show whether the bet is working. The first is GPU Pool utilisation, which Megaport expects to reach optimal levels 3 to 6 months after deployment; the August 2026 full-year results are the first window onto early take-up. The second is contract signings beyond the initial four, which would show the speculative capacity is converting to committed demand rather than sitting idle through depreciation. The third is the cash-flow shape over the next 12 to 18 months, when the A$369.5 million of contracted capex is spent while the four contracts only begin generating revenue from the first half of 2027.