The money moving through AI infrastructure right now is unlike anything the industry has seen. Nvidia is making a “substantial investment” — reported by Bloomberg as $5 billion — into Ilya Sutskever’s Safe Superintelligence, a lab with no products, no revenue, and a valuation implied at $32 billion, in exchange for deep visibility into its research and lock-in on the Vera Rubin platform. Samsung and Broadcom, meanwhile, have signed a non-binding MOU with projected spend exceeding $200 billion over five years, bundling 2nm foundry, HBM (including HBM4), and advanced packaging under one roof. When a single chip framework and a single equity investment together approach a quarter-trillion dollars, the market is no longer pricing AI infrastructure as a growth sector. It’s pricing it as the foundation of the next decade’s economy.

That scale is producing a second-order problem. The gap between ambition and readiness is widening faster than anyone is closing it. Tata Communications’ data is blunt — 77% of leaders call AI a board-level priority, but only 29% report network infrastructure capable of supporting AI at scale, and an IDC study finds 88% of AI proofs-of-concept fail to reach wide-scale deployment. The network was designed for user-to-application traffic at 100-500ms latency. Tata Communications finds that AI workloads generate 50x to 100x the east-west traffic; latency-sensitive AI workloads demand sub-10ms, and they fail differently — a sixteen-second answer to a visual query is worthless in a way a slow file download never was. The infrastructure conversation is shifting from “can we build enough capacity” to “can we make what we build actually work for the workload.”

On the telco side, the narrative is evolving from aspiration to asset. Orange just posted record results and announced a €3 billion data center platform for sovereign AI in France, part of a broader European consolidation wave that’s beginning to look less like survival and more like strategy. NTN, meanwhile, is maturing past the “satellites can reach phones” proof point toward the harder question of industrial scale — who can manufacture, launch, and operate enough orbital capacity at a sustainable cost to turn technically viable D2D into a durable business. The satellite industry booked $303 billion in revenue last year and surpassed 10 million broadband subscriptions, but subscriber growth (62%) is dramatically outpacing revenue growth (16%), which tells you where the pricing pressure is.

And then there’s the uplink question. The industry’s case for an AI-driven radio supercycle rests on ratios, shipment forecasts, and per-task multipliers — not on published cell-level measurements from any operator’s live network. The one real-world study on record, from Signals Research Group, found modest uplink impact from current AI applications and flagged connection behavior as potentially more consequential than payload volume. Nandlall’s framework — show the burst correlation factor, the cellular duty cycle, the session failure rates — is exactly the standard the investment case should have met before it became a talking point. The industry doesn’t need less investment in uplink. It needs better evidence for why.

Read on for more of today’s top news.