Signal
from the wire.

Nepal Telecom first promised a 5G trial in 2021. Five years and several missed deadlines later, the country is still waiting on a spectrum auction the regulator hasn't been cleared to hold.

The holdup isn't technical. Nepal's telecommunications regulator has forwarded its 5G spectrum auction file to the Ministry of Communication for a policy-level decision, with frequency bands, pricing and distribution already worked out — the auction is simply waiting on a political sign-off. Proposed bands include the 3.5GHz and 700MHz ranges, priced at roughly Rs 40 lakh per MHz per year for 3.5GHz and Rs 1.35 crore per MHz for the unpaired 700MHz band.

Behind the delays sits a geopolitical knot. Nepal Telecom's 4G buildout relied on two Chinese contractors, and persistent speculation links the years of stalling to outside pressure over the use of Huawei and ZTE equipment in next-generation infrastructure — an allegation the operator has consistently denied.

"The 5G launch will proceed only after 4G gets wider and better."

Government messaging has settled on sequencing rather than urgency. Communications Minister Bikram Timilsina has said 5G will only roll out once 4G coverage and quality improve, while still describing a 5G launch as a government priority with frequency planning underway. Operators echo the caution: with voice revenue declining and OTT apps eating into margins, both Nepal Telecom and rival Ncell have questioned whether a 5G build justifies the return given the country's small device and data-usage base. For now, the most concrete recent progress is humbler than a flagship 5G launch — Nepal Telecom has been pushing 4G into rural ground, including new coverage along the Makalu Base Camp trekking route.

June's biggest theme wasn't a single model release — it was every major lab quietly shipping the plumbing that lets AI act on its own, not just answer questions.

Google led with consumer-facing agents that run continuously in the background. The company launched 24/7 search agents that monitor the web for things users care about, such as event tickets or real estate listings, and notify them when new information appears, alongside expanded Gemini capabilities including voice calling for local bookings and AI-powered shopping assistance. At its Build conference, Microsoft introduced seven new proprietary MAI models built for enterprise workloads and agentic workflows, marking a deliberate move toward less dependence on outside model providers.

Policy moved in step with product. A new U.S. executive order tasks the National Security Agency with building a classified benchmarking protocol to identify high-risk "covered frontier models," and sets up a voluntary scheme giving federal agencies a 30-day early look at risky systems before public release. Industry commentary increasingly frames this shift in blunt terms: agentic systems are moving from chat to actually finishing tasks across research, coding, support and commerce.

"The industry is trying to make agent behavior boring — and boring means trusted enough to become part of work."

The caution flag came from Anthropic's own infrastructure. A major outage affecting Claude AI, its developer console, the API, and the Claude Code execution engine left engineers unable to run automated agent workflows for hours, prompting renewed advice that teams building on a single cloud AI provider need resilient, multi-model routing rather than a single point of failure. As agents move from answering questions to executing terminal commands and editing files, sandboxing and runtime guardrails are quickly becoming as important as the model itself.

Data centers chasing AI training capacity are quietly squeezing the same memory supply that ordinary consumer electronics depend on — and budget smartphones are feeling it first.

The scale of the shift is stark. High-bandwidth memory demand from AI data centers is expected to consume 20% of total wafer capacity by the end of 2026, up from just 2% — and that growth is squeezing production of the standard DDR and LPDDR memory used in everyday devices, already pushing up prices on sub-$100 smartphones critical to markets in Africa and South Asia.

That price pressure is showing up directly in the budgets of the companies building AI infrastructure. Microsoft now expects total 2026 capital expenditure to reach $190 billion, with $25 billion of that increase attributable specifically to surging memory and storage component prices driven by AI infrastructure demand. Despite the spending, returns remain an open debate on Wall Street: the company's AI services have generated a fraction of what's gone into building them.

"The standalone equivalent revenue would approach $50 billion."

Custom silicon is becoming one of the clearest profit stories of the buildout. Amazon's custom chip business, spanning its Graviton processors, Trainium AI training chips and Nitro security chips, has surpassed a $20 billion annual run rate and is growing more than 100% year over year, with multi-year commitments from OpenAI, Anthropic, Meta and Uber, making it one of the top three datacenter chip businesses globally. The pattern across the industry is consistent: whoever controls memory and custom silicon supply, not just model weights, is increasingly setting the pace of the AI race.

Noam Shazeer helped write the paper that made every modern AI model possible. In June, he left Google DeepMind for OpenAI — again — and the move reads as a proxy war for where AI architecture research happens next.

The departure followed an expensive history. Shazeer, co-author of the foundational 2017 paper "Attention Is All You Need" that introduced the Transformer architecture underpinning every major AI model today, announced he was leaving Google DeepMind to join OpenAI as Lead for Architecture Research — less than 22 months after Google paid roughly $2.7 billion to bring him back from the chatbot startup Character.AI, which he had co-founded after first leaving Google in 2021. Markets shrugged: Alphabet shares actually closed up on the news, suggesting investors see the company's revenue scale and compute commitments as outweighing the loss of any single researcher.

"A hire he had wanted since the very beginning of OpenAI."

The move lands amid a broader reshuffling of who's actually winning the assistant market. A widely cited industry report found ChatGPT's share of the global AI assistant market fell to 46.4%, the first time it has dropped below half, while Google's Gemini rose to 27.7% and Anthropic's Claude reached 10.3% — though Claude posted the highest paid-subscription conversion rate in the field, at 13% of its users. Open-weight competition is rising too, with a fourth major open frontier model release landing in roughly a month, evidence that frontier-scale training techniques have spread well beyond the handful of companies that once monopolized them.

Talent, market share and open-source proliferation are now moving together — a sign that the next phase of the AI race may be decided less by any single launch event and more by which labs can keep their best architecture researchers in the building.