Access

Z.ai unveils cheaper, advanced AI model GLM-4.5

Chinese AI startup Z.ai, formerly Zhipu, is increasing pressure on global competitors with its latest model, GLM-4.5. The company has adopted an aggressive open-source strategy to attract developers. Anyone can download and use the model without licensing fees or platform restrictions.

GLM-4.5 is designed with agentic AI, breaking tasks into smaller components for improved performance. By approaching problems step by step, the model delivers more accurate and efficient outcomes. Z.ai aims to stand out through both technical sophistication and affordability.

CEO Zhang Peng says the model runs on only eight Nvidia H20 chips, while DeepSeek’s model needs sixteen. Nvidia developed the H20 to comply with US export controls aimed at China. Reducing chip demand significantly lowers the model’s operational footprint.

Zhang said the company has enough computing power and is not seeking further hardware now. Z.ai plans to charge 11 cents per million input tokens, undercutting DeepSeek R1’s 14 cents. Output tokens will cost 28 cents per million, compared to DeepSeek’s 2.19 dollars.

Such pricing could reshape large language model deployment expectations, especially in resource-limited environments. High costs have long been a barrier to broader AI adoption. Z.ai appears to be positioning itself as a more accessible alternative.

Founded in 2019, Z.ai has raised more than 1.5 billion dollars from investors including Alibaba, Tencent, and Qiming Venture Partners. It has grown quickly from a research-focused lab to one of China’s most prominent AI contenders. A public listing in Greater China is reportedly being prepared.

OpenAI recently named Zhipu among the Chinese firms it considers strategically significant in global AI development. US authorities responded by restricting American companies from working with Z.ai. The startup has nonetheless continued to expand its model lineup and partnerships.

Chinese firms increasingly invest in open-source models, often with domestic hardware compatibility in mind. Moonshot, another Alibaba-backed company, released the Kimi K2 model. Kimi K2 has received praise for its performance in coding and mathematical tasks.

Tencent has joined the race with its HunyuanWorld-1.0 model, which is built to generate immersive 3D environments. The HunyuanWorld-1.0 can accelerate game development, virtual reality design, and simulation work. Cutting-edge features are being paired with highly efficient architectures.

Alibaba also introduced its Qwen3-Coder model to assist in code generation and debugging. Such AI tools are seeing increasing use in software engineering and education. Chinese developers are positioning themselves to compete with Western offerings such as OpenAI’s Codex and Anthropic’s Claude.

The momentum within China’s AI sector is accelerating despite geopolitical and trade restrictions. A clear shift is underway from imitation to innovation, with local startups advancing independent research. Many models are trained on China-specific datasets to optimise relevance and performance.

Z.ai’s strategy combines cost reduction, efficient chip use, and broad availability. The company can build community trust and encourage ecosystem growth by open-sourcing its tools. At the same time, pricing undercuts major rivals and could disrupt the market.

Global AI development is increasingly decentralised, with Chinese firms no longer just playing catch-up. Large-scale funding and state support are helping to close gaps in hardware and training infrastructure. Z.ai is one of several firms pushing toward greater technological autonomy.

Open-source AI development is also helping Chinese companies win favour with developers outside their borders. Many international teams are experimenting with Chinese models to diversify risk and reduce reliance on US tech. Z.ai’s GLM-4.5 is among the models gaining traction globally.

By offering a powerful, lightweight, and affordable model, Z.ai is setting a new benchmark in the industry. The combination of technical refinement and strategic pricing draws attention from investors and users. A new era of AI competition is emerging.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Allianz breach affects most US customers

Allianz Life has confirmed a major cyber breach that exposed sensitive data from most of its 1.4 million customers in North America.

The attack was traced back to 16 July, when a threat actor accessed a third-party cloud system using social engineering tactics.

The cybersecurity breach affected a customer relationship management platform but did not compromise the company’s core network or policy systems.

Allianz Life acted swiftly by notifying the FBI and other regulators, including the attorney general’s office in Maine.

Those impacted are offered two years of credit monitoring and identity theft protection. The company has begun contacting affected individuals but declined to reveal the full number involved due to an ongoing investigation.

No other Allianz subsidiaries were affected by the breach. Allianz Life employs around 2,000 staff in the US and remains a key player within the global insurer’s North American operations.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Huawei challenges Nvidia with AI super server

Huawei has unveiled its most powerful AI server, the CloudMatrix 384, to challenge Nvidia’s grip on the high-performance AI infrastructure market.

The system, launched at the World AI Conference in Shanghai, uses 384 Ascend 910C chips, significantly outnumbering Nvidia’s 72 B200 GPUs in the GB200 NVL72.

Although Nvidia’s GPUs remain more powerful individually, Huawei’s design relies on stacking and high-speed chip interconnection to boost overall performance.

The company claims the CloudMatrix 384 can deliver 300 petaflops of computing power, well above Nvidia’s 180 petaflops, though it consumes nearly four times more energy.

The US recently reversed its ban on Nvidia’s H20 chip exports to China, seeking to curb Huawei’s momentum. However, ongoing reports of smuggled Nvidia GPUs raise doubts over the effectiveness of these restrictions.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

New AI model, Aeneas, assists historians in interpreting Roman inscriptions

Thanks to AI, historians studying ancient Rome now have a powerful new tool.

A research team, including scholars from Google DeepMind and the University of Nottingham, developed a generative AI model called Aeneas that can help interpret damaged Latin inscriptions by estimating their location and date and suggesting likely missing text.

Each year, roughly 1,500 new Latin inscriptions are unearthed, ranging from imperial decrees to everyday graffiti. These inscriptions, written by ancient Romans across all social classes, offer rare, first-hand insights into daily life, language, and society.

Yet many of them are incomplete or difficult to contextualise. Traditionally, scholars must compare each inscription against hundreds of others manually — a process described as laborious and requiring exceptional expertise.

Aeneas, trained on over 170,000 Latin texts, can now predict when and where an inscription was written across the Roman Empire’s 62 provinces. In one test case, it analysed the famous Res Gestae Divi Augusti, narrowing down the date to the same two options long debated by historians.

Aeneas significantly improved research outcomes when used alongside human expertise instead of replacing it, helping scholars piece together history more efficiently than ever.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Viasat launches global IoT satellite service

Viasat has unveiled a new global connectivity service designed to improve satellite-powered internet of things (IoT) communication, even in remote environments. The new offering, IoT Nano, supports industries like agriculture, mining, transport with reliable, low-data and low-power two-way messaging.

The service builds on Orbcomm’s upgraded OGx platform, delivering faster message speeds, greater data capacity and improved energy efficiency. It maintains compatibility with older systems while allowing for advanced use cases through larger messages and reduced power needs.

Executives at Viasat and Orbcomm believe IoT Nano opens up new opportunities by combining wider satellite coverage with smarter, more frequent data delivery. The service is part of Viasat’s broader effort to expand its scalable and energy-efficient satellite IoT portfolio.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Meta forms AI powerhouse by appointing Shengjia Zhao as chief scientist

Meta has appointed former OpenAI researcher Shengjia Zhao as Chief Scientist of its newly formed AI division, Meta Superintelligence Labs (MSL).

Zhao, known for his pivotal role in developing ChatGPT, GPT-4, and OpenAI’s first reasoning model, o1, will lead MSL’s research agenda under Alexandr Wang, the former CEO of Scale AI.

Mark Zuckerberg confirmed Zhao’s appointment, saying he had been leading scientific efforts from the start and co-founded the lab.

Meta has aggressively recruited top AI talent to build out MSL, including senior researchers from OpenAI, DeepMind, Apple, Anthropic, and its FAIR lab. Zhao’s presence helps balance the leadership team, as Wang lacks a formal research background.

Meta has reportedly offered massive compensation packages to lure experts, with Zuckerberg even contacting candidates personally and hosting them at his Lake Tahoe estate. MSL will focus on frontier AI, especially reasoning models, in which Meta currently trails competitors.

By 2026, MSL will gain access to Meta’s massive 1-gigawatt Prometheus cloud cluster in Ohio, designed to power large-scale AI training.

The investment and Meta’s parallel FAIR lab, led by Yann LeCun, signal the company’s multi-pronged strategy to catch up with OpenAI and Google in advanced AI research.

The collaboration dynamics between MSL, FAIR, and Meta’s generative AI unit remain unclear, but the company now boasts one of the strongest AI research teams in the industry.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

UN urges global rules for AI to prevent inequality

According to Doreen Bogdan-Martin, head of the UN’s International Telecommunications Union, the world must urgently adopt a unified approach to AI regulation.

She warned that fragmented national strategies could deepen global inequalities and risk leaving billions excluded from the AI revolution.

Bogdan-Martin stressed that only a global framework can ensure AI benefits all of humanity instead of worsening digital divides.

With 85% of countries lacking national AI strategies and 2.6 billion people still offline, she argued that a coordinated effort is essential to bridge access gaps and prevent AI from becoming a tool that advances inequality rather than opportunity.

ITU chief highlighted the growing divide between regulatory models — from the EU’s strict governance and China’s centralised control to the US’s new deregulatory push under Donald Trump.

She avoided direct criticism of the US strategy but called for dialogue between all regions instead of fragmented policymaking.

Despite the rapid advances of AI in sectors like healthcare, agriculture and education, Bogdan-Martin warned that progress must be inclusive. She also urged more substantial efforts to bring women into AI and tech leadership, pointing to the continued gender imbalance in the sector.

As the first woman to lead ITU, she said her role was not just about achievement but setting a precedent for future generations.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

UK enforces age checks to block harmful online content for children

The United Kingdom has introduced new age verification laws to prevent children from accessing harmful online content, marking a significant shift in digital child protection.

The measures, enforced by media regulator Ofcom, require websites and apps to implement strict age checks such as facial recognition and credit card verification.

Around 6,000 pornography websites have already agreed to the new regulations, which stem from the 2023 Online Safety Act. The rules also target content related to suicide, self-harm, eating disorders and online violence, instead of just focusing on pornography.

Companies failing to comply risk fines of up to £18 million or 10% of global revenue, and senior executives could face criminal charges if they ignore Ofcom’s directives.

Technology Secretary Peter Kyle described the move as a turning point, saying children will now experience a ‘different internet for the first time’.

Ofcom data shows that around 500,000 children aged eight to fourteen encountered online pornography in just one month, highlighting the urgency of the reforms. Campaigners, including the NSPCC, called the new rules a ‘milestone’, though they warned loopholes could remain.

The UK government is also exploring further restrictions, including a potential daily two-hour time limit on social media use for under-16s. Kyle has promised more announcements soon, as Britain moves to hold tech platforms accountable instead of leaving children exposed to harmful content online.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Agentic AI forces rethink of cloud infrastructure

Cybersecurity experts warn that reliance on traditional firewalls and legacy VPNs may pose greater risks than protection. These outdated tools often lack timely updates, making them prime entry points for cyber attackers exploiting AI-powered techniques.

Many businesses depend on ageing infrastructure, unaware that unpatched VPNs and web servers expose them to significant cybersecurity threats. Experts urge companies to abandon these legacy systems and modernise their defences with more adaptive, zero-trust models.

Meanwhile, OpenAI’s reported plans for a productivity suite challenge Microsoft’s dominance, promising simpler interfaces powered by generative AI. The shift could reshape daily workflows by integrating document creation directly with AI tools.

Agentic AI, which performs autonomous tasks without human oversight, also redefines enterprise IT demands. Experts believe traditional cloud tools cannot support such complex systems, prompting calls to rethink cloud strategies for more tailored, resilient platforms.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

The US push for AI dominance through openness

In a bold move to maintain its edge in the global AI race—especially against China—the United States has unveiled a sweeping AI Action Plan with 103 recommendations. At its core lies an intriguing paradox: the push for open-source AI, typically associated with collaboration and transparency, is now being positioned as a strategic weapon.

As Jovan Kurbalija points out, this plan marks a turning point where open-weight models are framed not just as tools of innovation, but as instruments of geopolitical influence, with the US aiming to seed the global AI ecosystem with American-built systems rooted in ‘national values.’

The plan champions Silicon Valley by curbing regulations, limiting federal scrutiny, and shielding tech giants from legal liability—potentially reinforcing monopolies. It also underlines a national security-first mentality, urging aggressive safeguards against foreign misuse of AI, cyber threats, and misinformation. Notably, it proposes DARPA-led initiatives to unravel the inner workings of large language models, acknowledging that even their creators often can’t fully explain how these systems function.

Internationally, the plan takes a competitive, rather than cooperative, stance. Allies are expected to align with US export controls and values, while multilateral forums like the UN and OECD are dismissed as bureaucratic and misaligned. That bifurcation risks alienating global partners—particularly the EU, which favours heavy AI regulation—while increasing pressure on countries like India and Japan to choose sides in the US–China tech rivalry.

Despite its combative framing, the strategy also nods to inclusion and workforce development, calling for tax-free employer-sponsored AI training, investment in apprenticeships, and growing military academic hubs. Still, as Kurbalija warns, the promise of AI openness may clash with the plan’s underlying nationalistic thrust—raising questions about whether it truly aims to democratise AI, or merely dominate it.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!