Scale Across: Marvell Just Connected the Dots I’ve Been Drawing All Year
Three Threads, One Shape
This week, Marvell’s CEO stood on stage with Jensen Huang at Computex and named something the industry hadn’t fully said out loud: AI clusters have outgrown the buildings that house them.
For the last few pieces I’ve written here, I’ve been circling different corners of the same shift without naming the center of it. Marvell just named it for me.
Go back through what I’ve covered recently and a pattern starts to show.
I wrote about why dark fiber has become the infrastructure play of the decade — because hyperscalers no longer want to lease capacity, they want to own the glass, and the routes worth owning take years to build and are running out. That’s a story about scarcity at the physical layer.
I wrote about the hyperscaler land grab — how Amazon, Google, Microsoft, and Meta have stopped being carriers’ customers and started being carriers’ competitors, buying up dark fiber, submarine capacity, and data center real estate at a pace the industry hasn’t priced in yet. That’s a story about who controls the map.
And I wrote about the battery replacement wave now sweeping through US data centers and tower portfolios — a multi-billion-dollar buildout tied directly to the fact that US hyperscaler capex is on pace to blow past $600B in 2026, with new gigawatt-scale campuses rising across Texas, Virginia, Indiana, and Louisiana. That’s a story about how much physical infrastructure this buildout actually requires, on the ground, inside the US.
Three different stories. Same underlying fact: AI infrastructure has stopped fitting inside a single building, a single network, or a single vendor relationship. Marvell just gave that fact a name — “scale across” — and put a number on it: a business they expect to double to $1B a year by 2028, built entirely on linking data centers together over high-speed optical connections.
What “Scale Across” Actually Requires
Here’s the part the chip-and-optics headlines tend to skip: the technology is the easy half.
Linking data centers — campuses miles apart, often built by different operators, on different timelines, sometimes running on different carriers’ infrastructure — isn’t something a single vendor solves by shipping a faster transceiver. It requires aligning multiple parties around one accountable system: shared standards, shared SLAs, a commercial framework that lets competitors cooperate where it serves the end customer, and someone willing to own the outcome end-to-end instead of pointing at the handoff.
That’s not a new problem to me. It’s the exact problem I built a career around solving — designing the multi-carrier frameworks that let fragmented infrastructure behave like one continuous system for hyperscaler-grade customers. I’ve done it inside a single market and across a border. The mechanics don’t change; only the map does.
Where This Goes Next — Inside the US
What Marvell is describing isn’t a future state. It’s already happening across the same US corridors I wrote about in the battery piece — the same Texas, Virginia, Indiana, and Louisiana campuses that are driving record hyperscaler capex are the ones that will need to be wired together as “scale across” infrastructure over the next three years.
That’s where the opportunity sits now: not in any single building, and not at any single border, but in the connective layer between US facilities that’s about to become as strategically important as the buildings themselves. The operators and carriers who build the frameworks to deliver that — cleanly, at scale, as one accountable system — will be the ones hyperscalers turn to first. The ones still treating it as someone else’s problem at the handoff will be the ones left explaining what happened.
I’ve spent my career building exactly that kind of system. Marvell just told the market where it’s needed next.
— Edgar Mosti, senior telecom executive, 25+ years in yperscaler/carrier sales, strategic alliances, and infrastructure programs across the US and the Americas.