TelecomTV dropped the inaugural AI-Native Telco Index right before this week’s AI-Native Telco Forum in Düsseldorf—276 pages scoring 56 operators on 20 indicators, full report available to ANTA members. One rule I love: “announce” does not mean “executed.” Commitments count as commitments; deployments count only with published proof they’re running. Nine operators earned AI Vanguard status. The sharpest finding: SK Telecom, the industry’s most celebrated AI-native brand, landed a tier below because much of its work was still in planning at the June cutoff—though the index expects its momentum to put it near the top by 2027. Grade your own AI program the same way: receipts, not press releases.
Amazon Web Services (AWS) will deploy 2 million additional NVIDIA GPUs in 2027 and 2028, on top of the million-plus already coming in 2026. The reason: demand is going through the roof. That’s right. The most sophisticated capacity planner on earth underestimated this market. If your board is still weighing a GPU buildout, this is the closing bell. You cannot win a hardware race against someone buying 3 million units on an 18-month refresh cycle. Sell them the connectivity and edge they need and skip the silicon. It’s too hard to keep up.
Singtel’s investor day showed the most disciplined GPU play in telecom: S$300 million on a 6MW sovereign AI cluster, fully pre-sold, with tranche two waiting until 80% is contracted. I respect the rigor. But look at Singtel’s own numbers: payback lands in early 2030, and the gear needs a refresh every four to five years. The return window and the obsolescence window are the same timeframe. Meanwhile, the wins Singtel actually reported—lower churn and faster fault recovery—came from software. That’s the tell about where AI value will likely live for operators.
OpenAI shipped GPT-6 Astra, and company president Greg Brockman declared it the start of the AGI era. Who cares about the label. Here’s what matters for telco: independent benchmarks say Astra’s raw intelligence is flat versus its predecessor. The leap is in the hands, not the brains. This model doesn’t advise, it operates: computer use, multi-step workflows, real software engineering. It’s also OpenAI’s most cyber-capable model yet—restricted at launch and run past the government before release. (Yikes.) Tokens cost 2.5x, and agents that finish work are worth it—as long as your systems can tell them what they’re allowed to finish.
Salesforce and Anthropic announced Claudeforce, and the CRM’s stock jumped as Benioff and Amodei jointly told the “SaaSpocalypse” crowd to knock it off, because they’re creating something that’s “bigger than the sum of its parts” (or so they hope!). Study the division of labor: Claude brings the reasoning, and Salesforce keeps the data, business logic, valid actions, and governance. Salesforce conceded the interface and rented the model to protect the layer that compounds: context. The problem? It still leaves Salesforce as a big, dumb CRUD. Will this move help a dying SaaS company? Regardless, Claudeforce is a harbinger of what’s to come for all enterprise software, especially telco software.
Verizon went all-in with Google Cloud, deploying Gemini across customer care (it’s already handling most inbound calls and chats), marketing, security, and an autonomous network framework. Smart move skipping the GPU vanity project and riding a hyperscaler instead. But the foundation is a multi-year consolidation of legacy data lakes into one bigger, better-marketed lake. A single source of truth for data still can’t tell an agent which actions are valid. Couple this with Verizon’s effort to hand-build that context layer, and maybe it’s onto something. My only beef: the operator should give itself a head start with a pre-built telco ontology like the Totogi Ontology (because roughly 90% of that context is identical among operators).
An IBM and TM Forum study says 73% of network executives have roadmaps to autonomous operations. Just 6% run Level 4 today. The industry’s prescription: clean the data, then keep humans validating what the machines decide. That fixes hygiene, not architecture. Clean data is still 400 systems with 400 definitions of “customer.” A human double-checking every automated action uses humans as a quality gate. It’s not autonomy. Autonomous loops need a layer that makes invalid actions impossible to attempt. Until that verifier exists, your Level 4 roadmap is a Level 2 reality with better slideware.
Telstra published the external investigation of its July outage, and Vicki Brady deserves credit for the transparency. The findings: ownership of network timing was never defined, the architecture evolved without end-to-end oversight, expertise was too thin, and documentation fell short. Every item on that list is a knowledge failure, and every operator has systems like this: critical, undocumented, understood by three people. This is exactly what an ontology fixes. It gives that knowledge a permanent home where every system, every team, and every agent can find it. The Totogi Ontology already encodes it for telco. Nobody should learn how their network works from an outage report.
Fierce’s Elizabeth Coyne asked the fair question: what if AI is good for telcos but bad for the 300,000 people working at America’s big three carriers? My honest answer: AI will cost jobs, and it will be hardest on workers who stand still. But these people are not just costs to cut. They hold the context your AI needs: every billing exception and eligibility quirk lives in their heads. The workers who encode what they know move up, not out. Leaders, you decide the ratio. Fund the reskilling. The experts you’ll need for the AI era are already on your payroll.
