AT&T’s 90% Cost Cut
On August 12, AT&T’s Chief Data and AI Officer Andy Markus made a statement that should have every enterprise CIO paying attention.
OpenAI models already power about 25% of AT&T’s total AI usage. The target? 70% to 80%.
But here’s the real story. By switching from closed, proprietary models to open ones, AT&T cut its costs by 80% to 90% in certain applications. The company processes an average of 45 billion AI tokens a day through a “smart router” that picks the cheapest suitable model for each task.
“We’re not scared of the token future,” Markus said. And the numbers back him up. AT&T built its own customized open model, OTel, and says its AI initiatives have delivered a fivefold return on investment this year.
This is the moment the economics of AI shifted. Not through a new model. Through a new way of buying.
The Router That Just Sold for $7.5 Billion
Speaking of routers: Stripe just acquired OpenRouter for approximately $7.5 billion. Three months ago, investors valued OpenRouter at just $1.3 billion.
What does OpenRouter do? It helps organizations select the best and most cost-efficient AI model for every task. In a world where businesses are drowning in token consumption costs, routers have become one of the hottest areas in enterprise tech.
The math is simple. If AT&T can cut costs by 90% by routing to the right model, the company that builds the best router is worth a lot of money. Stripe just proved that thesis.
AI Agents Now Outnumber Humans on the Internet
This is the one that stopped me.
Cloudflare reported that AI-agent traffic has crossed 50% of network volume. AI agents now generate more internet traffic than humans.
Think about that for a second. Non-human traffic now outpaces human traffic on one of the world’s largest edge networks—a threshold that would have seemed implausible two years ago.
Cloudflare’s stock jumped 5.1% to $293.14 on the news. CEO Matthew Prince cut 20% of the workforce in May to pivot toward an “agentic AI-first operating model”. The bet is paying off. Cloudflare reported Q1 revenue of $639.8 million, up 34% year over year.
This is not a marginal shift. Unlike human web browsing, AI agents execute multi-step workflows—scraping data, calling APIs, running inference, and coordinating across services. Each agent session can generate dozens of requests per second. The request volume compounds.
We are now living in a world where machines talk to machines more than humans talk to machines.
Alibaba’s Open-Source Dominance
While U.S. companies were making headlines, Alibaba was quietly becoming the most downloaded AI model provider in the world.
Alibaba’s open-weight Qwen models have racked up more than 3 billion downloads in six months, blowing past Meta’s 227 million and Google’s 418 million. The milestone caps a run of more than 460 open-sourced Qwen models and 300,000-plus derivatives.
The flagship, Qwen3.8-Max, is a 2.4 trillion-parameter vision-language model trained for long-running coding and knowledge work tasks. It ranks fifth overall and second among open models on Artificial Analysis’ Intelligence Index.
The message is clear: even as Washington restricts chip exports, Chinese open models are winning developer mindshare globally.
Digicel’s Ruby: The AI Agent That Handles 55% of Customer Queries
Let me start with the one that actually made me sit up straighter. Digicel, the Caribbean telecom giant serving 9 million customers across 25 markets, revealed that its AI agent Ruby now resolves more than half of all customer queries without human involvement.
Since launching in Jamaica, Ruby has supported over 535,000 conversations, successfully handling 55% of interactions autonomously. It handles routine tasks—checking balances, activating plans, paying bills, resolving technical issues—while human agents focus on complex cases. Built on a platform developed by Canadian AI firm Ada, Ruby evolved from a knowledge-based tool into an AI-powered assistant available across WhatsApp, voice, web, and the MyDigicel app.
Here’s what caught my attention: Ruby is powered by Anthropic and OpenAI models, with customer data processed on servers in the US, Canada, and Europe. Digicel plans to expand Ruby across additional markets by March 2027.
This matters because it’s a real-world case study in AI-driven customer care at scale. But it also raises questions. Jamaica’s outsourcing sector has already dropped from 60,000 workers to 40,000. As Ruby expands, the workforce implications will be impossible to ignore.
Zong’s 39% Network Throughput Gain
While Digicel focused on customers, Pakistan’s Zong focused on the network itself. The operator reported a 39% improvement in peak-hour network throughput in Islamabad using AI.
The system, described as AI-optimised Air Ran, operates across more than 17,000 sites and 26,000 kilometres of fibre, adjusting capacity and resources in response to traffic patterns during busy periods. Zong is embedding intelligent automation to predict congestion, detect faults before they affect customers, and optimise energy use.
The company is also developing packages featuring AI assistants, cloud storage, and content streaming as part of a strategy combining AI with 5G, cloud computing, and IoT services. Zong’s locally hosted Tier III Intelligent Cloud Computing Centre, with zones in Islamabad and Lahore, is designed to improve access to cloud and AI tools for startups, SMEs, government bodies, and large industries.
SoftBank and Ericsson’s AI-Native RAN Breakthrough
On August 20, SoftBank and Ericsson conducted Japan’s first trial of Ericsson AI in RAN on a 5G commercial network. The AI-native scheduler for link adaptation ran directly within the radio access network in real time.
The results: spectral efficiency improved up to 25%, and downlink user throughput improved up to 50% compared to conventional technology. Across all locations, both metrics improved by about 10% on average.
This isn’t about faster internet—it’s about networks that learn. By applying AI directly within the RAN, operators can adapt to changing radio and traffic conditions, optimize performance, and make better use of existing resources. Conventional link adaptation relied on rule-based algorithms based on offline analysis. Ericsson’s AI-native scheduler runs on baseband equipment based on real-time AI decisions.
This marks an important step toward AI-native mobile networks. The industry is moving from selling bandwidth to selling intelligence.
Huawei’s Xinghe Intelligent Network: Security First
At the Huawei Network Summit 2026 Asia-Pacific in Bali, Huawei unveiled its upgraded Xinghe Intelligent Network Solution under the philosophy of “Secure and Intelligent Connectivity”.
The context: AI-driven cyberattacks have surged by 327%, straining traditional defenses. As AI agents penetrate core business processes, network instability is magnified exponentially.
Huawei’s response is a network built on four pillars: lossless compute to maximize token efficiency, integrated sensing and communications, full-scope security to address AI-driven threats, and high-level network autonomy.
The technology works. Huawei’s Hyper-Converged Fabric achieves a network-wide throughput of over 98% while boosting inference tokens per second by 20%. The upgraded solution covers Xinghe AI Fabric 2.0, Xinghe Intelligent WAN, Xinghe AI Campus, and Xinghe AI Network Security.
The GSMA Warning: Don’t Outsource Your AI Future
The GSMA’s Director of AI Technologies Louis Powell issued a warning that every telecom executive should read: the biggest frontier models still don’t really understand telecom.
The GSMA has spent the past two years benchmarking AI models against telecom-specific tasks, and the results show a stubborn gap. Frontier models aren’t getting much better at parsing 3GPP standards, troubleshooting RAN issues, or handling the domain-specific complexity operators need for autonomous networks.
Powell noted that while deployments in customer service and marketing have taken off, “only about 16% of deployments are on the network”. He also expressed concern that “the supply chain of AI is very, very narrow, and we don’t want everybody outsourcing everything to the hyperscalers”.
The response is clear. The percentage of operators planning to develop AI solutions in-house rose from 37% in 2025 to 42% in 2026, with another 38% co-developing with partners. Open telco AI models could be critical to autonomous networks—and to network security.
The Agent Compiler That’s Changing Enterprise AI
On the enterprise side, a Chinese startup called Duda Technology (独到科技) launched something genuinely innovative: an “Agent Compiler” called DLM-AOP that transforms scattered business knowledge, process rules, and employee experience into AI agent systems that can actually execute complex tasks.
Here’s the problem it solves. Gartner predicts that by 2028, 33% of enterprise software will include Agentic AI capabilities, but warns that by the end of 2027, over 40% of such projects may be cancelled due to cost overruns or unclear value.
DLM-AOP takes a different approach. Instead of starting with prompts and tool calls, it takes real business operations as input—SOPs, historical data, rules, and human experience—and “compiles” them into agents that understand business, connect to real systems, and continuously execute tasks.
The results are striking. In testing, DLM-AOP increased complex workflow execution accuracy from 72.4% to 95.6%, while compressing agent delivery cycles from months to weeks. In one retail project, agents handled over 85% of customer communications, and AI-assisted teams saw a 125% increase in conversion compared to human-only teams.
The Anonymous Model That’s Shaking Up the Industry
And then there’s the one that genuinely surprised me. On August 20, an anonymous model called “Ox Alpha” appeared on OpenRouter’s API with a 1,048,576-token context window and multimodal reasoning across text, image, and video.
A million-token context window means the model can process an entire codebase or several full-length novels in a single prompt. It supports tool and function calling alongside structured JSON output, making it immediately useful for production workflows rather than just chat.
The developer community has gone into full detective mode. Tokenizer fingerprinting and API signature matching point strongly toward China’s Zhipu AI (now Z.ai), with the model likely being a variant of GLM-5.3.
This marks the fifth time an anonymous AI model linked to Chinese labs has appeared on OpenRouter. Previous anonymous drops were eventually claimed by Chinese labs after the preview period ended. The model is currently free for a one-week preview.
The Funding Reality: $100 Billion Chips and $1 Billion Startups
While the models and networks grabbed headlines, the infrastructure behind them told a different story.
Broadcom is reportedly in talks to raise over $60 billion in debt financing for AI chip projects benefiting Anthropic and other customers, with the total package potentially reaching $100 billion. Apollo and Blackstone are in talks to participate.
Velaura AI raised $110 million in a Series A round, crossing a $1 billion valuation. The company’s Titan Core chip platform delivers a 2-4x improvement in performance per watt for AI accelerators—a critical advantage as power consumption becomes the primary constraint on AI infrastructure. As one investor put it, “Physical AI will require a fundamentally different approach to compute centered on extreme power efficiency”.
And Nvidia is reportedly in talks with South Korean AI chip designer Rebellions for potential cooperation, investment, or even acquisition. Rebellions is the centerpiece of South Korea’s “K-Nvidia” national strategy, backed by SK Hynix, Samsung Ventures, and Arm, with a valuation of approximately $2.3 billion. The company focuses on AI inference chips—the larger workload as AI moves from training to mass deployment.
#AI, #AT&T, #Stripe, #OpenRouter, #Cloudflare, #Alibaba, #Qwen, #Huawei, #AIIinfrastructure, #OpenSourceAI, #AgenticAI, #TechNews, #Innovation, #DigitalTransformation