Azure just crossed $100 billion in annual revenue, up 43% year over year, its fastest growth in four years. And Microsoft still can’t keep up. The company is sitting on a $678 billion revenue backlog and stood up 31 new data centers in a single quarter. Now the catch: this calendar year’s CapEx is expected to hit $175 billion. Microsoft is spending roughly $1.75 for every $1 of Azure revenue and calling it winning. Even the CFO hedged, noting GPU purchases can slow if demand shifts because the silicon is short-lived. If the richest company in the game hasn’t proven the ROI, your sovereign-AI business case definitely hasn’t. Watch the hyperscalers and learn from them before you launch that GPU-as-a-Service offering.
SpaceX’s first earnings report as a public company reads like a Series B startup: $16 billion burned in Q2 against $7.8 billion in revenue, with $18.4 billion of CapEx, mostly AI data centers. Microsoft’s 1.75x looked aggressive; this is 2.4x. To feed it, Musk is coming for mobile—and not just from the sky. Next-gen satellites plus EchoStar’s 65 MHz deliver 100x today’s D2D capacity, and SpaceX President and COO Gwynne Shotwell announced a terrestrial layer: femtocells bolted onto Starlink dishes. Twelve million broadband customers means twelve million rooftops as potential cell sites—no tower leases, no site acquisition. SpaceX subscribers would pay $66 a month to host its RAN. Shotwell said she expects to take “quite a few” Tier 1 USA customers. Consider the gauntlet thrown.
In March, after drone strikes hit AWS sites in the Gulf, I told you your DR needs a DR. Four months later, Bloomberg’s satellite analysis confirms that Iranian strikes damaged two more AWS sites in Bahrain. Data centers are strategic targets now, alongside refineries and power plants. But this doesn’t mean the cloud is unsafe—AWS customers shifted workloads to other regions within hours because there were 39 to choose from. It’s your on-prem DR that’s exposed: it sits in the same country as your primary, and geography has no failover. When war destroyed Zain Sudan’s data centers, Ericsson quoted up to nine months to rebuild; Totogi had charging back up and running on AWS in 18 days. Put your backup in another country and set it up before you need it.
Jensen Huang’s first-ever X post was a lobbying letter: “Open Weights and American AI Leadership,” signed by NVIDIA, Microsoft, Meta, Dell, and dozens of others. Notice who wants open weights: everyone currently in a losing position. Meta trails the frontier, so Llama is free. NVIDIA sells chips either way. The big name still missing? Anthropic, the lab with the model everyone wants. Why give away the advantage when you don’t have to? Plan accordingly: the best model will always be closed and rented. Instead, own what nobody can rent back to you: your learning loop, the context and decisions that compound. Then you can jump models anytime.
AT&T just handed you a free tool: OTel 2.0, a telco-tuned version of Google’s open Gemma model, trained with the GSMA on 400 billion tokens of industry-standard documents and network data. It’s small, it’s open, and it beats frontier models at telco grunt work: interpreting 3GPP specs, parsing telemetry, troubleshooting network issues. Pair it with a routing gateway like AT&T did—cheap model for the narrow tasks, frontier model when you need reasoning—and watch AI costs drop by up to 90%. Just understand what you bought: a model that speaks telco, not one that knows your telco. Download it. Feed it the context you’ve curated, and see what it can do.
Verizon announced a dark fiber deal with Google worth north of $1 billion. It’s also retrofitting stripped-out central offices into edge inference sites—a trial that sold out in 24 hours. Contracted revenue from assets built for a different era is easy business. But the analysts squinted at the details, asking: new construction? Capacity reselling? Nobody would say what the margin is. Dark fiber is the purest landlord trade in telecom—you supply the glass, Google lights it and keeps everything built on top. Telcos have funded two platform revolutions on those terms. Good deal, but this is not a big idea. Monetizing dormant assets pays the bills; it doesn’t compound. Don’t confuse it with an AI strategy.
Light Reading says 160,000 jobs have been wiped out across 20 Tier 1 operators since ChatGPT launched. Sounds seismic—until you zoom out. Telcos employ about 4.4 million people worldwide, so three years of “AI-driven” cuts trimmed under 4% of workers. Basically, attrition. And it bought nothing: Verizon shed 88,000 jobs over a decade and its operating margin still fell, because it ran the old telco with fewer people instead of redesigning the work. Telcos and other large industry companies will continue to shed jobs as AI gets easier and easier to implement at scale, and that’s a good thing. Get in front of it by redesigning every role to be AI-first and you’ll create tons of value. That’s the kind of AI-first telco we’re helping operators build at Totogi.
