For readers unfamiliar with Megaport, can you explain how its approach to networking differs from traditional telcos?
Megaport is a global Network-as-a-Service platform that “lives” inside the world’s data centers. We have physical infrastructure in roughly a thousand data center locations and stitch those sites together over about 6,000 fiber routes spanning 26 countries. In practice, that means customers can stand up private connections—up to 100 Gbps—between data centers and clouds in about 60 seconds.
Everything is software-orchestrated on top of massive hardware, so connectivity can be spun up or down on demand, and we routinely see tens of thousands of active connections changing continually as customers automate their networks through DevOps.
We began as the world’s largest cloud connector and still are, but the platform has expanded: data-center-to-data-center connectivity, internet services in those facilities, and global WAN. Unlike telcos that are strong in a single country, Megaport holds 26 carrier licenses and operates across providers—about 120 different data center operators. We lease, rather than build, the underlying fiber and focus deeply on inter–data center and data-center-to-cloud paths (not last mile). That specialization enables the 60-second, code-driven provisioning—and, oddly enough, we’re still the only company doing it at this scale.
You mentioned an “AI exchange” and surging AI needs. How is AI-driven demand showing up in your network usage and customer growth?
AI impacts us in three clear ways. First, after a period of post-COVID belt-tightening—2022 into 2024—budgets have reopened and projects are flowing again. Board-level pressure to “do something with AI” pushed CEOs to their CIOs, and CIOs to say, “We need to modernize security, software, and the network.” That pent-up spend benefits core infrastructure, including high-speed connectivity, whether or not a given project ends up being overtly “AI.”
Second, dedicated AI farms—GPU operators and cloud competitors—are appearing everywhere, and enterprises need to move huge datasets to and from them fast. Where a 10 Gbps circuit once sufficed, customers now ask for 100 Gbps to train or run inference for a defined window, then ramp it down. Third, we’re selling directly to the AI providers themselves. Some inference-first companies are rolling out global clouds by shipping their own chips to any data center with space; the long pole is usually the network, which with traditional providers can take one to six months. Megaport turns that into ~60 seconds, so these companies choose locations with “Megaport inside,” light up capacity overnight, and scale globally at a pace that used to take hyperscalers months or years.
You’ve built an AI exchange with dozens of GPU providers. Does your model introduce trade-offs vs. traditional connectivity—and how quickly will enterprises adopt this approach?
Our model is closer to cloud economics for networking: software-driven control on pre-provisioned, global infrastructure, so customers pay for what they use and change it instantly. That doesn’t add new technical constraints for the customer; it removes old ones—namely, waiting months for circuits and being locked into fixed capacity while demand fluctuates. The value becomes obvious when an enterprise wants to burst a 100 Gbps link for backups, training, or interconnection, then spin it back down. It’s also powerful for arbitraging GPU availability and price: if capacity is scarce or costly in one region this month and cheaper elsewhere next month, the network needs to keep up.
Adoption is well underway. We serve on the order of a few thousand large enterprises—the “Fortune 3000,” so to speak. Many began with cloud on-ramps via credit card and then expanded. We haven’t even cross-sold all of them on data-center-to-data-center internet services or the AI exchange; in many cases they don’t yet know what’s possible. The next wave will be driven by inference at scale—SaaS platforms and real-time applications that need sub-second responses—where speed and elasticity are business-critical. As those teams get specific on their AI roadmaps, instant, programmable connectivity will move from “nice-to-have” to default.
Inference is pushing bandwidth and latency higher than ever. Where are the biggest technical infrastructure bottlenecks ahead?
The pain points stack up in familiar ways. First, chips: GPUs, LPUs—supply has struggled to keep up with demand. Second, space and power: if you can secure land with power, capital will find you; but transformers, generators, and the rest of the power train are deeply backlogged, slowing new capacity. Third, data center construction timelines remain a gating factor even as investment surges.
Then there’s the network. Procuring high-speed links from traditional telcos can take three to 18 months, especially for long-haul paths with multiple hops. That’s why customers need optionality: the ability to reach whichever site has capacity today and to pivot tomorrow. We can’t manufacture chips or build the facility for you, but we can neutralize the network wait time and give you Switzerland-style neutrality across operators and locations, so you can place workloads where it makes sense right now.
You’re in a sweet spot but also have to invest heavily. How are you balancing scaling a global AI-ready network with profitability and resilience?
We grew quickly during the “growth at all costs” period, reaching critical mass in revenue and footprint. When the market flipped to profitability in 2022–2023, we pivoted hard, becoming very profitable and adding roughly $100 million to the balance sheet rather than raising capital. That gave us the headroom to reinvest substantially without dipping into reserves.
Our posture now is disciplined: invest significantly in expansion and product while staying cash-flow balanced. We’re not reverting to burn-heavy tactics; instead, we scale where demand is clear (AI farms, inference hubs, sovereign deployments) and continue building out the fabric so customers can reach more places instantly. The objective is durable growth: widen the platform, deepen utilization, and keep the model efficient so we can fund expansion from strength.
Stepping back, how will AI reshape networking and telco more broadly—and what won’t change?
The outcome—moving data from A to B—won’t be disrupted; the “big fat pipes” metaphor still holds whether it’s fiber or satellites. What must change is how networks are delivered: automation has to become the default. Most of the industry still deploys like the copper days—roll a truck, wire the wall, test each hop—hence the months-long lead times. We pre-built the provisioning pain so a click deploys 100 Gbps worldwide in about a minute.
As AI drives arbitrage of compute cost and availability, the only viable response is an automated network that can re-route and resize on demand. You can’t survive three-month delays when the “cheapest/fastest” location changes every quarter. The hard part is automating physical infrastructure at global scale—that’s why few, if any, others do what we do today—but the direction of travel is clear: software-defined control atop pre-provisioned capacity.
Forecast question: will this keep compounding for years, or is there a bubble forming?
No crystal ball, but two forces argue for durability. On training, the spend is led by a handful of giants—xAI, Meta, OpenAI/Microsoft, Google—who can bankroll massive deployments and are already seeing benefits in their core businesses. It’s an arms race they can afford, and even “only” a 50% improvement at the frontier is enormously valuable, justifying more chips and more scale.
On adoption, companies are just figuring out how to embed AI into products. As token costs fall, you’ll see AI stitched into everything from SaaS workflows to voice interfaces, and models will improve—fewer hallucinations, more domain-tuned accuracy. That spans every industry, fast and slow movers alike, so diffusion should be steady. Rather than mass job loss, what we observe is speed: teams run faster with copilots and code assistants, and competition forces everyone onto a quicker treadmill. It’s disruptive, yes—but broadly additive, and it keeps the demand for flexible, instant networking very much alive.
