There is no shortage of AI news this month. Apple is reportedly opening iOS to third-party models. Anthropic shipped an agent “dreaming” technique. Google is leaking Gemini Omni ahead of I/O. OpenAI and Anthropic are buying their way into consulting. AMD just printed the strongest quarter in its history. Intel more than tripled this year. The Pentagon, the EU, and the White House all have AI safety announcements teed up.
Almost none of it actually moves the needle for an operator running a 15-year-old BSS stack in Riyadh, Karachi, or Lagos. Some of it does — and it is not the parts the headlines lead with.
Below are the five stories from May 2026 that I would put in front of an operator CTO this week. Two of them are this week, because FutureNet World London is on Thursday and Friday, and the agenda there will set the tone for half the BSS conversations happening in the second half of the year.
1. The IDC FutureNet panel quietly named the real bottleneck
On April 23, IDC ran a panel at FutureNet World London titled “Unlocking New Revenue Opportunities by Monetizing AI and Digital Infrastructure.” The panellists were senior operator executives from Vodafone, Orange Wholesale, Telekom Srbija, and Tallence AG.
The headline finding was not about revenue. It was this:
“Agentic AI remains one of the most talked-about areas in the market, but many enterprise pilots are still failing to reach production. The problem is not only technical capability. It is trust. For agentic AI to be backed with an SLA, enterprises need confidence that decisions are bounded, explainable, and auditable. In regulated and mission-critical environments, free-form reasoning is not enough. Determinism matters. Governance matters. Standards alignment matters.”
I have been writing variations of this paragraph in client memos for six months. It is genuinely useful to see it on an IDC stage with four operator CxOs nodding along. The implication for anyone building toward agentic BSS is the same one I keep landing on: the model is not the constraint. The constraint is whether the agent’s decisions can be wrapped in something a regulator, an auditor, or a CFO will sign.
If you are an operator currently being pitched “agentic BSS” by a vendor, the right question this quarter is not “what can it do?” It is “what does its decision audit trail look like, and can I show that audit trail to the PTA, TDRA, or NDMO without rewriting the integration?”
That question kills 80% of the demos.
2. NVIDIA released a 30B open-weight telco model. The interesting part is what it costs to actually use
Ahead of MWC, NVIDIA and AdaptKey AI released an open-source 30-billion-parameter Nemotron-based “Large Telco Model” (LTM), along with a Tech Mahindra co-authored guide on fine-tuning it for NOC reasoning workflows. The release went out through GSMA’s new Open Telco AI initiative.
This is genuinely significant. It means an operator no longer has to wait for a vendor to wrap GPT-class capability into a BSS-flavoured connector. The base model is available, the licensing is open, and on-prem deployment is supported.
It also means almost nothing in practice for an operator that has not already cleaned up its data.
To fine-tune a 30B reasoning model on your NOC traces, you need NOC traces — clean, structured, post-incident reasoning chains that look like the way your senior engineers actually think. Most operators do not have this. They have free-text closure notes, half-completed templates, and a long tail of “see attached email.” Tech Mahindra’s published guide is clear about this: the fine-tuning step is the easy part. Producing the structured reasoning traces is the work.
The same logic applies on the BSS side, where it would be even harder. A “Large BSS Model” fine-tuned on order fallout decisions, dunning workflows, and rating exceptions would be enormously valuable. Nobody has the training data sitting on a clean partition.
The honest read: the open model is a tool. The tool does not absorb your operational debt. If you cannot describe your top 50 BSS incident types in structured form today, the open model is not going to do it for you.
3. Microsoft published the case studies, and the gap between Network Ops and BSS Ops is now visible
Microsoft’s NOA (Network Operations Autonomy) framework had a quiet but substantive update around MWC. The new case studies they published with Vodafone, AT&T, Telefónica, Etisalat (e&), and Far EasTone are worth reading as a set.
The number that jumped out at me: Far EasTone reports that roughly 60% of NOC operations are AI-assisted, executing about 10,500 operational tasks per month — alarm correlation, ticket closure, root-cause analysis — with average response times around 16 seconds. Vodafone’s transport-network deployment with Microsoft is autonomously managing more than 65% of fibre-break field dispatches.
Those are real numbers. They are also network operations numbers, not BSS numbers.
This is the gap that I think will define 2026 inside operator IT shops:
- Network Operations Centre (NOC): clean telemetry, deterministic state machines, observable physical outcomes, no regulator on the alarm bus. Agentic AI moves fast here.
- Business Support Systems (BSS): 30-year-old data models, multi-vendor product catalogues, regulatory dependencies on every billing decision, no canonical state machine for an order. Agentic AI moves slowly here, when it moves at all.
The same operator that proudly reports “60% NOC automation” will quietly tell you their order fallout queue is still being worked manually by a team in a third country. The architectural reasons are not mysterious. The lesson for anyone designing a 2026 AI roadmap is to stop benchmarking BSS readiness against the NOC’s progress. Different problem, different stack, different timeline.
4. The CPU and memory squeeze is now an operator procurement problem
This is the boring story, and it is the one nobody on the AI side wants to talk about.
In its Q1 FY2026 results on May 5, AMD reported $10.3B in revenue, with data centre revenue at $5.8B — up 57% year-on-year. CEO Lisa Su now projects the server CPU TAM growing more than 35% annually, reaching over $120B by 2030. Intel had its best stock month on record in April, and Micron’s CEO told CNBC in March that key customers are receiving only “50% to two-thirds” of their actual memory requirements.
Why does this matter for an operator?
Because if you read the Microsoft, Amdocs, and NVIDIA case studies carefully, you will notice that the production deployments running today are almost all on hyperscaler infrastructure. The on-prem story — which is the only story that works for an operator processing regulated subscriber data under NDMO, TDRA, or PTA constraints — is bottlenecked at the inference hardware layer.
An operator IT director planning agentic BSS deployment for 2027 needs to start the procurement conversation now. The 12-week hardware lead times of 2024 are 6+ months for enterprise-grade inference hardware in 2026, and memory allocation is being rationed.
This is the kind of operational reality that does not show up in any AI vendor’s pitch deck. It will show up in the project plan, retrospectively, after the deployment slips by two quarters.
5. MCP at 97 million installs means the connector layer is no longer the bet
The quieter announcement that I think will age best: Anthropic’s Model Context Protocol crossed 97 million installs in March 2026, and every major AI provider now ships MCP-compatible tooling.
For an SI building BSS connectors — or an operator IT team designing the abstraction layer between their stack and whatever AI ecosystem they end up running — this is the protocol war ending without a fight. MCP has become the default. Building to it is a safe bet.
Practically, this means:
- The integration layer between AI agents and operational systems is converging on a single standard.
- The “agent ↔ tool” wiring problem is solved. The “agent ↔ BSS API” problem is not.
- The hard work moves up the stack. It is no longer about choosing the right protocol. It is about exposing the right BSS capabilities through the protocol with the right governance.
The fact that 80% of telecom press releases this year will still mention “agentic AI integration” without specifying which integration is a feature of the press release, not the technology. The protocol is settled. The hard part is what you connect.
What I would actually do this week
If you are running an operator IT shop or carrying an AI quota at an SI bidding into one, the FutureNet week is a reasonable moment to ask three questions of your AI programme leads:
- For every agentic AI claim in your roadmap, can you describe the audit trail in one paragraph? If not, you are building for a demo, not for a regulator.
- For every “we will fine-tune on our data” plan, can someone produce 100 cleanly structured incident traces today? If not, the open models are not your answer.
- For every on-prem inference component you are planning, has the hardware procurement timeline been validated against current lead times? If not, the project plan is fiction.
None of these are AI questions. They are operational discipline questions. That is what I take from May 2026 in telecom AI: the deployment-quality conversation has finally caught up with the model-capability conversation.
The headlines will keep being about models. The wins, this year, will be about everything else.
Sources for this post: NVIDIA Nemotron LTM announcement (March 2026); Microsoft NOA case studies (MWC 2026); IDC FutureNet London panel summary (April 23, 2026); AMD Q1 FY2026 earnings (May 5, 2026); CNBC chip market coverage (May 8, 2026); Anthropic MCP install metrics (March 2026, reported via Crescendo AI). All numerical claims are sourced; some figures, especially install counts and percentage automation rates, are vendor-reported and should be read accordingly.