SK Hynix outlined plans to invest KRW54 trillion ($38.3 billion) in two new semiconductor manufacturing plants in South Korea, stepping up production capacity to meet long-term demand for AI memory.

The South Korean chipmaker will invest KRW35.2 trillion in its Yongin Y2 fabrication plant in the Seoul Metropolitan Area, which will produce high-bandwidth memory (HBM) and other next-generation dynamic random access memory (DRAM) products. Meanwhile, KRW19.1 trillion is earmarked for a second facility, Cheongju M17, in central South Korea to manufacture NAND flash memory chips, with the company arguing the city “offers the fastest fab construction timeline”.

The latest investment forms part of a wider manufacturing push. In June, SK Hynix unveiled plans to invest KRW600 trillion in the Yongin Semiconductor Cluster and KRW100 trillion to expand its Cheongju production base.

Yongin Y2 will become the second of four planned fabs in the Yongin Semiconductor Cluster. Construction of the plant is due to begin in July 2027, while work on the Cheongju M17 fab is scheduled to start in February 2027. The company aims to complete all four planned fabrication plants by 2033, bringing forward the original timeline by 12 years.

The latest commitment comes amid strong demand for AI memory requiring large volumes of HBM. SK Hynix explained demand for NAND memory is “surging rapidly”, adding its investment followed “a thorough review of market demand”. Citing forecasts by Omdia, it noted global DRAM and NAND demand is expected to grow at a compound annual rate of 19% through 2030.

The company noted “in the AI era, technological competitiveness alone is not enough”, arguing the “ability to supply the required volume at the exact moment customers need it is the ultimate competitive advantage”.

In June, South Korean President Lee Jae-myung unveiled a long-term AI and semiconductor initiative led by Samsung Electronics and SK Hynix expected to exceed $1 trillion by 2035.

The post SK Hynix earmarks $38B for AI memory appeared first on Mobile World Live.

ByteDance is reportedly in the process of developing an AI model which could rival Anthropic’s Mythos system as part of a bid to establish itself as a leader in China’s AI ecosystem.

Financial Times (FT) reported the TikTok owner is in the early stages of training a model with as many as 10 trillion parameters.

This would put ByteDance’s offering ahead of Anthropic’s cutting-edge Mythos and make it three-times larger than Moonshot Kimi K3, the biggest model to come out of China so far.

While Anthropic does not reveal the size of its models, FT reported the most advanced Mythos 5 system has 8 trillion parameters and its Fable 5 about 5 trillion.

ByteDance’s model is apparently being pre-trained, a stage in development which typically takes three-to-six-months. It would then be fine-tuned and released, provided there are no issues.

Chinese companies are showing their hand with regards to AI model development, with Moonshot and Alibaba highlighting releases in recent weeks.

FT added ByteDance invested more aggressively in AI than other Chinese companies, building data centres and attracting talent in the sector.

Founder Zhang Yiming reportedly told staff he wants the team to target “world-leading model capabilities” during an internal meeting last month.

The post ByteDance plots leading AI model to rival Mythos appeared first on Mobile World Live.

Mobile World Live (MWL) brings you our top three picks of the week as SpaceX CEO Elon Musk outlined Starlink’s global broadband ambitions, Verizon lifted the lid on the network behind the FIFA World Cup 2026 and the World Bank urged developing nations to boost AI adoption.

Musk targets global broadband domination with Starlink

What happened: Musk outlined plans to increase Starlink’s network capacity 100-fold, claiming the satellite broadband business could eventually “deliver a majority of the world’s internet”.

Why it matters:  The tech mogul said the goal hinges on the rollout of an expanded V3 satellite constellation capable of delivering around 10 times the performance of current V2 models. To support the plan, the company asked the US Federal Communications Commission (FCC) for approval to deploy up to 100,000 low-Earth-orbit V3 satellites, supplementing the roughly 9,600 V2 satellites already in orbit. Musk argued the additional capacity would allow Starlink to eventually carry a majority of global internet traffic in countries where the service is authorised to operate, while driving a tenfold increase in Starlink revenue even if revenue per bit falls by 90%. Quilty Space founder and CEO Chris Quilty told MWL that the company’s “bold claim” flouts its “prior statements that Starlink was most appropriate for rural user”. While V3 satellites are expected to provide a “10x boost”, he noted “aggregate network capacity is a misleading metric because capacity is geographically distributed while demand is geographically concentrated”. Earlier this week, SpaceX president Gwynne Shotwell claimed that Starlink Mobile could compete directly with operators including AT&T, Verizon and T-Mobile US using upgraded satellites and additional spectrum.

Musk targets global broadband domination with Starlink

Podcast: Inside Verizon’s World Cup connectivity

What happened: In the latest edition of the MWL podcast, Kavit and Justin sat down with Verizon wireless chief network officer Srini Kalapala, who detailed how the operator delivered connectivity as the official network provider for the FIFA World Cup 2026.

Why it matters: Earlier this month, the operator revealed its network carried almost 500TB of data across all US host stadiums, including 29.2TB during the final between Argentina and Spain. To handle demand, Verizon boosted stadium capacity by three to five times, installed thousands of under-seat antennas, deployed nearly 140 small cells around venues and fan zones, and laid 80,000 miles of fibre to transport broadcast feeds from all World Cup stadiums. In the podcast, Kalapala described the tournament’s “104 games in 39 days across 18 different stadiums in three countries” as “probably one of the largest projects that we worked on.” He explained Verizon had to ensure stadiums comprising 70,000 to 85,000 people could access “good quality mobile connectivity”, while also delivering “the most reliable network” for broadcasters and powering private 5G applications including referee cameras.

Podcast: Inside Verizon’s World Cup connectivity

World Bank urges lower income nations to grasp AI opportunity

What happened: The World Bank called on developing economies to accelerate AI adoption, arguing low-cost, locally adapted tools could accelerate economic development without requiring frontier models or hefty AI infrastructure investments.

Why it matters: In its World Development Report 2026: The Promise of Artificial Intelligence, the organisation said AI “could allow developing countries to do in a decade what might otherwise take a century”, provided governments address gaps in electricity, internet access, digital skills and institutions. World Bank SVP and chief economist Indermit Gill said AI has “thrown developing economies a lifeline”, but warned countries “must hurry” because the technology is “spreading faster and is more context-specific than earlier general-purpose technologies”. He added many high-impact use cases do not require large models or data centres, instead relying on “small, low-cost AI tools” tailored to local conditions. The report highlighted applications across healthcare, agriculture, tax collection and judicial services, while warning AI could also deepen inequality and weaken public trust without appropriate action. Foresight Communications Consultancy CEO Howaida Nadim argued the World Bank’s framework recognises that “greater opportunity lies in adopting proven tools and adapting them to local languages, data, institutions and development priorities”. She added “the next AI divide” might be “between those that successfully adapt AI to their development needs and those that remain primarily consumers of systems built elsewhere”.

World Bank urges lower income nations to grasp AI opportunity

The post The Friday File: Starlink; Verizon; World Bank appeared first on Mobile World Live.

Indosat Ooredoo Hutchison brought in the big guns of its parent, Nokia and Nvidia to help it develop fresh infrastructure to meet the demand for AI computing in Southeast Asia and beyond, with an initial 1GW target.

The Indonesian operator’s Zankore by Indosat play is to employ Nvidia’s DSX AI Factory reference design as part of an ambitious plan to establish the nation as a regional compute hub.

It aims to provide the first 200MW of capacity in H1 2027, running an Nvidia GB300 NVL72 rack. Information from the chip company indicates the set up uses 72 of its Blackwell Ultra GPUs and 36 Arm-architected CPUs to provide additional reasoning capabilities.

Zankore by Indosat is also employing the Nvidia DSX MaxLPS technology suite, which the chip player separately explained provides ways to “maximise token performance per MW within a fixed power budget”.

Put more plainly, the suite is designed to offer the lowest-possible token cost by optimising power use across the line of GPUs to unlock fuelling which might otherwise be stuck in less useful areas.

The operator’s parent Ooredoo Group is the lead investor and platform sponsor, an arrangement noted to bring access to long-term capital and regional scale.

Nokia is providing AI-native network infrastructure.

Zankore by Indosat appointed ex-Ericsson and T-Mobile US executive Ulf Ewaldsson as CEO, along with establishing a board comprising representatives from principal shareholders and independent members.

Meutya Hafid, the Indonesian Minister of Communication and Digital Affairs, said the nation is advantageously placed to provide AI compute capacity throughout Southeast Asia and the Zankore by Indosat move would speed development of sovereign capabilities to the benefit of “businesses and society”.

The post IOH makes 1GW Indonesian AI data centre play appeared first on Mobile World Live.

Google CEO Sundar Pichai (pictured) revealed a major shake-up within its AI division, with DeepMind co-founder and CEO Demis Hassabis stepping back from day-to-day operational management to focus on overarching strategy, AGI and scientific breakthroughs.

Under the new leadership structure, Koray Kavukcuoglu, former CTO of Google DeepMind and chief AI architect, steps up as SVP of Google DeepMind, reporting directly to Pichai.

Kavukcuoglu, who has been with DeepMind for 13 years and previously led foundational projects like WaveNet and Deep Q-Networks, will now directly oversee Gemini model development, frontier AI research and developer ecosystems.

Hassabis transitions into a newly created strategic role as chair of Google DeepMind and chief scientist of Alphabet.

In a note sent to employees, Pichai stated the shift allows Hassabis to dedicate his attention to shaping the long-term future of AGI and scientific applications, alongside his ongoing leadership at drug-discovery subsidiary Isomorphic Labs.

Hassabis expressed confidence in the transition, emphasising  he plans to focus on broad strategic initiatives and high-impact applications of AI in healthcare, particularly accelerating efforts to treat major diseases.

The executive shift comes at a period of rapid product momentum for the internet giant’s AI stack.

According to internal metrics shared by Pichai, the flagship Gemini consumer application has surged past 950 million monthly active users, while the open-weights Gemma model family has surpassed 900 million downloads.

The reorganisation also marks the end of an era for two foundational figures in modern computing infrastructure.

Jeff Dean, Google’s chief scientist and long-time head of Google AI, is departing after 27 years alongside Google senior fellow Sanjay Ghemawat.

The pair are launching independent public benefit company aimed at applying machine learning to scientific and engineering discoveries. Google will serve as a founding investor and cloud infrastructure partner for the new entity, maintaining an active collaborative research framework on advanced systems.

“Jeff Dean is at a moment where he wants to try something new, and we’re excited to support him in that,” Pichai stated.

As competition in model capabilities intensifies across the industry, Google is doubling down on its full-stack approach which spans custom silicon, cloud infrastructure, and consumer applications to seek an edge in the race toward advanced AI.

The post Google DeepMind CEO steps down in exec reshuffle appeared first on Mobile World Live.

Anthropic confirmed it is building an in-house silicon team to design custom chips for its Claude AI model, a move aimed at improving performance and efficiency as demand for its services continues to rise.

In a statement to Mobile World Live (MWL), Anthropic said it “is hiring engineers with experience across the stack”.

The company stated this move will let it co-design hardware and models together, allowing Claude to run faster and more efficiently at the scale its customers need.

Anthropic told MWL the effort is not a shift away from its existing hardware partners. The company described the silicon team as the latest step in its multi-chip approach

It is building a diversified hardware stack with chips from Amazon Web Services (AWS), Google, Nvidia, and AMD, which will “remain central to how we scale”.

Reuters had previously reported Anthropic was exploring building its own chips while The Information stated the company had held talks with Samsung Electronics about possibly manufacturing them.

Business Insider cited a newly posted job listing which describes what Anthropic is looking for, which includes a “custom silicon team” with engineers spanning multiple areas of chip design and verification.

The listing, which advertises a salary range of $320,000 to $485,000, calls for candidates who can point to direct personal contribution in finalising and shipping semiconductor designs.

Building AI chips in-house is expensive and technically demanding, but the benefits include improved performance and less dependence on Nvidia.

Rival moves
Anthropic joins a growing list of AI companies pursuing custom silicon, following moves by OpenAI, Meta, Google, and Amazon.

OpenAI unveiled its first custom-built inference ASIC, Jalapeno, in June, developed in partnership with Broadcom.

In April, Meta Platforms expanded its partnership with Broadcom to jointly develop multiple generations of custom MTIA chips, building on four new generations it had unveiled in March.

Google is also developing a server chip, called Frozen v2, to improve efficiency for its Gemini AI model and separate from its existing TPUs,

During Alphabet’s Q1 earnings call in April, CEO Sundar Pichai stated the company will begin making its tensor processing units available to a select group of customers for their own data centres, a move seen as a potential challenge to Nvidia’s dominance in AI chips.

Also in April, Meta struck a deal with Amazon Web Services to use its Graviton processors for agentic AI workloads.

Amazon also has an agreement in place with Anthropic to use its Trainium chips to train and run its AI models

The post Anthropic plans in-house chip team to power Claude appeared first on Mobile World Live.

Anthropic appointed Mariano-Florentino Cuellar as its first chief global affairs officer, as the Claude AI model developer faces escalating political and regulatory friction.

In the newly-created role, Cuellar (pictured) will report directly to Anthropic president Daniela Amodei.

His primary focus involves managing relationships with US authorities and international government bodies as the business expands internationally.

Cuellar is stepping down from Anthropic’s independent four-person Long-Term Benefit Trust, a body which ensures the company aligns its commercial goals with its public-interest mission.

The trust is now searching for a successor.

Anthropic described Cuellar as a pragmatist and consensus builder suited to navigating government opposition. His credentials include working as a special assistant to former President Barack Obama and a justice with the Supreme court of California.

He also directed the Carnegie Endowment for International Peace think tank while co-chairing a task force on US national security and nuclear proliferation, a domain industry executives often cite as a blueprint for managing advanced AI risk.

Cuellar also jointly led studies which shaped landmark Californian state law SB 53, which introduced whistleblower protections and safety reporting mandates for high-revenue AI developers.

The executive said politicians everywhere are recognising AI governance is at a pivotal inflection point, noting current decisions would determine if technology can be harnessed safely or increase global inequality.

Anthropic is working to mend relations with the US government following several clashes including a military dispute which resulted in the US Department of War blacklisting its technology, a designation the company is challenging in court.

Officials also issued export control directives temporarily banning international sales of its high-tier Mythos 5 and Fable 5 models over national security concerns.

The post Anthropic recruits first global affairs chief appeared first on Mobile World Live.