May has been eventful.
Chilean startups with real traction, Claude moving into Wall Street in ways ChatGPT still cannot replicate, and a couple of infrastructure and security signals the market is not discussing enough.
Here are five things worth keeping on your radar this week — especially if you work in technology, telecommunications, or network operations.
1️⃣ Startups using AI to solve real problems — not to run demos
There are two kinds of AI startups in Latin America right now.
Those that built a GPT wrapper and called it a product. And those that identified a concrete industrial problem and used AI to attack it with proprietary data.
The second group is the one worth following.
Bruna is a Chilean startup that uses AI to forecast mineral quality and optimize production plans in mining. In simple terms: instead of a technician manually testing samples to decide whether a batch is ready to process, the model analyzes input data and predicts the outcome before the process starts. Less waste, more consistency, less waiting time.
SearchBrand.ai, incorporated in July 2025, is the region’s first platform that lets brands and SMBs measure and optimize their presence in results generated by AI models such as ChatGPT and Google Gemini. The problem it solves is new but concrete: if someone asks ChatGPT “what is the best dental clinic in Santiago” and your clinic does not appear, you do not exist for that user — just like on Google ten years ago, but without tools to manage it.
The difference between an AI startup that scales and one that dies in the pilot is not the model it uses. It is whether it has access to proprietary data for the problem it is solving.
Sources: Bruna — mining and predictive AI, University of Chile — Madeleine Valderrama, CEO of Bruna AI, SearchBrand.ai — AEO/GEO platform, El Ecosistema Startup — AEO in conversational AI
2️⃣ What Claude does in finance that ChatGPT still does not
On the Chatbot Arena leaderboard for May 2026, GPT-5.5, Claude Opus 4.7 Thinking, and Gemini 3.1 Pro are separated by a single Elo point. For practical purposes, all three are frontier models. The gap in general capability is marginal.
The real difference is where each one is betting as a platform.
On May 5, Anthropic gathered banks, funds, and asset managers in New York to launch Claude for Financial Services: ten ready-to-use AI agents, Microsoft 365 integration, and governed access to data from Moody’s, FactSet, Morningstar, S&P Global, and others.
The most concrete example from the event: JP Morgan CEO Jamie Dimon built, in twenty minutes from a blank sheet, a live analysis dashboard on Treasury asset swaps and bid-ask spreads. Work that normally takes a team of junior analysts half a day.
Three of the ten agents target equity research directly: one reviews earnings calls and updates financial models automatically, another generates daily briefings on coverage lists, and a third turns target-company data into a valuation Excel file ready for review.
Anthropic is explicit: agents produce drafts for qualified human review. They do not execute trades or write directly into accounting records. But that does not reduce the impact — it cuts professional time from hours to minutes.
Anthropic’s bet is not to replace the analyst. It is to replace the three hours the analyst loses preparing context before they can think.
Sources: Anthropic — agents for financial services, CFO.com — Claude’s expansion across corporate finance, ADSLZone — Chatbot Arena+ May 2026
3️⃣ Network traffic changed shape — and most architectures do not know it yet
For decades, network design had a fixed premise: traffic flows downward. Users request, the network delivers. Little upstream, lots of downstream, with predictable peaks in human hours.
The numbers coming out this year show that is no longer true.
In 2025, automated traffic grew 23.5% year over year — eight times faster than human traffic, which grew just 3.1%. And within that automated traffic, traffic generated by AI agents grew 7,851% in a single year. That is not a typo.
Cloudflare, which processes tens of millions of HTTP requests per second across more than 125 countries, reported that in Q1 2026 nearly one in three requests on the internet is already a bot — and 22% of that bot traffic is AI crawlers, the fastest-growing category.
To put it simply: before, a company downloaded data from the internet and barely sent anything back. Now, that same company’s AI agents are constantly sending, querying, and syncing data across multiple systems at once — no schedules, no predictable peaks, in every direction. Networks were designed for the first pattern, not the second.
The problem is not total network capacity. It is that assumptions about how traffic flows — direction, time of day, pattern — are ceasing to be valid.
What these numbers mean in practice: teams sizing network capacity today from historical human-traffic patterns are working with assumptions that no longer apply. This is not a future problem. It is a configuration problem already in production.
Sources: HUMAN Security — State of AI Traffic 2026, Cloudflare Radar Q1 2026 analysis
4️⃣ 5G has 2.8 billion connections — and operators still do not know how to charge for it
Imagine building an eight-lane highway, charging the same toll as the two-lane road you replaced, and then wondering why you cannot recover the investment. That is essentially what is happening to the sector with 5G.
According to the Ericsson Mobility Report from November 2025, 5G already represents nearly 2.8 billion subscriptions — more than 30% of all mobile connections worldwide — yet connectivity revenue growth has been modest: roughly $1.3 trillion in global connectivity revenue in 2025, up about 4% year over year, far below the 5G infrastructure investment curve of the past five years.
The underlying problem is structural. Only about one quarter of operator groups have deployed 5G standalone at scale — the architecture that actually enables network slicing and ultra-low-latency applications. The rest have 5G in name but still run on 4G logic underneath.
A concrete example: network slicing lets an operator sell a hospital a guaranteed “slice” of the network exclusively for its connected medical devices — with contractually guaranteed latency and availability at a premium price. That is real monetization. Without standalone architecture, it does not exist. It is only marketing.
The exit the market is starting to explore is selling network capacity as AI infrastructure — inference as a service, edge computing, guaranteed connectivity for critical applications. In short: stop selling gigabytes and start selling outcomes.
5G did not fail. The assumption that the same 4G business models would work on a radically different network did.
Sources: Ericsson Mobility Report — 5G subscriptions, GSMA — 5G Standalone full steam ahead, GSMA Intelligence — Global Mobile Trends 2026, Omdia via Telecompaper — connectivity revenue 2025
5️⃣ The security risk AI introduced in telecom — and nobody budgeted for
AI is making networks smarter. At the same time it is creating a category of risk the sector did not have before: autonomous failure at scale.
The clearest case last year was NTT DOCOMO in Japan. On January 2, 2025, a DDoS attack caused network congestion that left about 90 million subscribers struggling to access key services for roughly 12 hours. It was not a sophisticated hack. It was volume — enough traffic to overwhelm automated network management systems that responded incorrectly because the data they were processing no longer reflected the real state of the network.
That is exactly the new risk: AI systems that manage networks can take “confidently incorrect” actions when underlying data is manipulated or noisy, causing massive outages — not because the system fails, but because it acts correctly on wrong information.
The scale of the IoT problem is just as concrete. In November 2025, Cloudflare recorded a 31.4 terabit per second DDoS attack — the largest on record — attributed to the Aisuru/Kimwolf botnet, which compromised consumer and industrial IoT devices worldwide. Meanwhile, connected households faced an average of 29 attack attempts per day in 2025 — nearly triple 2024. The average cost of an IoT security incident is around $330,000; in healthcare environments, it exceeds $7 million. IoT malware attacks grew 124% year over year in 2025.
For those of us in network operations, the analogy is direct: it is the same logic as a poorly tested change taking down production — except now an agent executes the change automatically at 3 AM without anyone approving it.
AI is not inherently insecure. We are automating network decisions at scale before we have the validation controls that autonomy requires.
Sources: NTT DOCOMO — official incident notice, Infosecurity — DDoS against DOCOMO, Cloudflare — record 31.4 Tbps attack, Bitdefender and NETGEAR — connected home 2025, DeXpose — IoT statistics 2025
🧭 Five signals, one common conclusion
The architecture decisions being made today — which data feeds an agent, how the network is sized, whether 5G is a product or infrastructure, what controls exist before automating a change — will define who has real operational capacity in 2028 and who is racing to catch up.
May did not bring a single disruptive announcement. It brought a pattern: AI is no longer a software layer on top of what exists. It is rewriting business models, traffic flows, and risk surfaces at the same time.
Which of these five signals resonates most in your role today? If you have a different read, I would like to hear it in the comments.
✍️ Claudio from ViaMind
“Dare to imagine, create, and transform.”
Also available in Spanish: Tendencias IA & Telecom – Mayo 2026.
FEATURED POSTS
ABOUT ME
ViaMind Ecosystem
- ViaMind — The brand hub for ideas, technology and strategy.
- NeuraPRO — Practical software for managing real operations.