Meta Platforms plans to expand its Muse personal agent and charge businesses for transactions it handles, CEO Mark Zuckerberg (pictured) stated on the company’s Q3 earnings call.
He said more than 10 million people in the US and Canada have set up Muse in a month and a half, and about 25% use it on any given day
Zuckerberg stated Muse had a “clear product-market fit” and that Meta will focus on attracting more users, increasing revenue and optimising compute costs.
“We expect to grow Muse to serve billions of people,” he said, calling the goal as “both our mission and our strategy”.
The Muse line began in April with Muse Spark, the first model from Meta’s Superintelligence Labs. This month Meta followed with Muse Wonder and Muse Wonder Full.
Shopping is Muse’s most common use, with more than a quarter of users shopping through the agent.
Zuckerberg said Muse “should be free” for most people, with Meta taking a fee on transactions it facilitates for businesses. Subscriptions would cover access to the most advanced models or extra tokens, and business owners could also subscribe directly.
“We have been discussing our vision for personal superintelligence for more than a year now,” he said. “We want every person [to] have an extremely capable personal agent that understands your goals and can help you achieve them,” he said.
Meta Q3 metrics Meta’s quarterly revenue rose 26% year over year to $64.6 billion.
Costs climbed 68% to $51.6 billion, including a $9.89 billion charge from an August agreement with 52 attorneys general over teen mental health claims.
Operating income fell 37% to $13 billion and the operating margin halved to 20%. Net income rose to $9.9 billion from $2.7 billion a year earlier, when a one-time $15.9 billion tax charge reduced profit.
Capital spending totalled $34.9 billion, with Meta forecasting $135 billion to $145 billion for 2026.
Its apps reach 3.64 billion people a day, up 3% from a year ago but just below the 3.65 billion analysts expected, according to StreetAccount. Zuckerberg said Threads has overtaken X in monthly users.
Reality Labs, which makes the Quest headsets and the Ray-Ban and Oakley AI glasses, lost $4.1 billion on $443 million of revenue, compared to a $4.4 billion loss a year ago.
Meta CFO Susan Li said the company expects Reality Labs to lose $18 billion to $19 billion in 2026, down slightly from $19.2 billion in 2025.
Also yesterday, Reuters reported a New Mexico judge ordered Meta to pay the state $25 billion after a jury found it misled consumers in a case tied to the Cambridge Analytica scandal. The tech giant said it will appeal.
Companies are announcing enormous new data centers to power the AI boom. Some promise facilities bigger than anything the industry has built before.
There's just one problem: Most of these projects may never happen.
Bernstein Research analyzed thousands of planned US data centers and concluded that only about 35% of the announced capacity is likely to get built. That's roughly 170 gigawatts of data center capacity out of a total of almost 500 gigawatts.
The rest? A significant chunk of it belongs to a category the industry has a wonderfully dismissive name for: "Bragawatts." These are ambitious announcements, often from newcomers with little experience building data centers, that have a slim chance of becoming reality.
"Press releases are cheap…concrete is not," Bernstein analysts wrote in a research note.
Bernstein began its analysis with more than 7,000 data center projects across North America, drawing on information from Aterio, a company that tracks construction through satellite imagery, permits, and other records. The analysts narrowed that to more than 5,000 projects (partly by excluding thosealready canceled).
Then they assessed the chances of these remaining projects being completed. The analysts considered factors such as whether construction had started, whether electricity was available, and whether the companies involved had successfully built data centers before.
The results revealed a striking divide.
The 'bragawatt brigade'
Kevin O'Leary.
Daniel Delgado Jr./Disney via Getty Images
Experienced operators including Amazon, Google, Microsoft, and Meta landed at the top. Bernstein gave these companies perfect scores of 5 out of 5 for their development credentials. Other data center specialists, such as Digital Realty, Equinix, and CoreSite, got top marks, too.
O'Leary Digital, associated with celebrity investor Kevin O'Leary, ended up at the bottom,with a credibility score of 0.35 out of 5, according to Bernstein. His company has announced 15 gigawatts of data center projects, and Bernstein estimates that only 489 megawatts, or about 3%, will actually materialize.
Although the term "bragawatts" has been doing the rounds in the data center industry lately, it's unclear who coined it, Bernstein analyst Madison Rezaei told me.
"I wish I could take credit for that," she said. "I believe I originally heard it from an Equinix executive, but I don't think he owned it either."
She said that whilesome ambitious newcomers could succeed, they haven't yet demonstrated they can deliver.
"It's not that all of these guys are sort of terrible and not going to deliver, but we just have lower confidence because we don't know them," Rezaei said.
O'Leary Digital responds
O'Leary Digital disputed Bernstein's assessment, saying the research doesn't account for substantial work that occurs before construction begins.
While acknowledging that O'Leary Digital is a young company, CEO Paul Palandjian said the team has decades of experience building major infrastructure projects and that its proposed 15 gigawatts of capacity would be built in stages over several years.
The company said it has secured binding agreements covering about 11,000 acres in Utah and 13,000 acres in Alberta, obtained a 30-year development agreement in Utah, and made progress on planning, engineering, and permits at both sites. It also said the projects would generate their own electricity.
The company said its early development work is funded, additional capital is committed to reaching key milestones, and experienced engineering and construction partners are involved.
"We would rather be judged on milestones than on a model," Palandjian said.
'An announcement in search of a project'
Opponents of the proposed La Osa data center attend a Pinal County Board of Supervisors meeting in Florence, Arizona, on August 26.
USA TODAY Network via Reuters Connect
Other low scorers in the "bragawatt" brigade included Vermaland and Joule Capital Partners, according to Bernstein analysis based on Aterio data.
Joule has announced 50 data center projects with a combined capacity of 11.7 gigawatts. Bernstein estimates that only about 11% of that capacity will actually get built. Almost all of Joule's planned capacity remains at the announcement stage, with little evidence of construction progress or of major customers committing to use the facilities, Bernstein wrote in a July note to investors.
The analysts gave Joule a credibility score of 2.05 out of 5, placing it among the lowest-ranked developers in its database.
"An announcement in search of a project," the analysts wrote.
Vermaland has announced plans for a massive data center project in Arizona. Bernstein estimates that only 4% of its proposed 2.2 gigawatts of capacity will get built. The company is primarily a landowner, not a data center developer. Its proposed $33 billion La Osa project near Eloy had already been sharply scaled back amid local opposition and without construction approval. In late August, regulators formally rejectedthe project. Even before that, the analysts gave Vermaland a credibility score of 0.75 out of 5.
Joule and Vermaland didn't respond to emails seeking comment this week.
A yawning gap for data centers
Sergio Toro, CEO of Aterio, sees similar patterns in his firm's voluminous, detailed data.
Aterio's tracking shows that roughly 30 to 40 gigawatts of new data center capacity get announced each month. Yet only about 7 gigawatts begin construction each month, while 2 to 3 gigawatts actually start operating.
This leaves a yawning gap.
Many announced projects get withdrawn before construction begins — and before real money has been spent.
"They just make the announcement, but they didn't get their approvals. Or they make multiple filings in multiple locations to see which one gets selected," Toro told me. "That's why we track with satellite if there's real movement on the ground or not."
Indeed, Bernstein analysts wrote recently that the amount of "stranded" data center capacity has surged. These are projects that are either canceled, delayed, or not approved.
The AI boom's three tiers
Toro sees three broad groups competing to build AI infrastructure.
At the top are the established technology giants, including Amazon, Microsoft, and Google. They've built data centers for decades and generally deliver on their commitments. Most of their projects will get done, he said.
In the middle are "neoclouds," newer companies such as CoreWeave that rent out computing power specifically for AI. Although they're more vulnerable to financing problems and delays, many are making solid progress, especially because some of these players already have access to electricity.
"We see them advancing quite well," Toro said. "Sometimes we see some delays, but it's not something significant or any kind of big capacity that is at risk."
Then come the newcomers making extraordinary promises. That's where most delays and cancellations happen, Toro said.
Even Amazon needs a backup plan
A technician works at an Amazon Web Services AI data center in New Carlisle, Indiana.
Noah Berger/Getty Images via Amazon Web Services
Not every project that disappears was a fantasy, though. Even a top player like Amazon struggles sometimes to secure electricity and other supplies on time for these huge projects.
The biggest, most experienced tech companies deliberately pursue more potential data center capacity than they need because they know some developments will fail or get delayed.
Major AI companies were committing to three or four times the space and electricity they needed over the following two years, Jeffrey Moerdler, a data center lawyer, estimated during an interview with Bernstein analysts.
His reasoning: Of data center projects that haven't started construction, roughly a third will arrive on schedule, another third will be delayed, and the remainder won't happen.
Rezaei and Toro estimate that the top players don't over-commit that much — more like double the capacity they need at most.
The strategy provides a "Plan B," as Toro put it, in case of delays involving electricity, equipment, permits, local opposition, or other factors.
Texas has a ghost problem
The Decker Creek Power Station in Austin.
Brandon Bell/Getty Images
This helps explain another extraordinary number.
Data center companies have requested connections for roughly 450 gigawatts of electricity demand on Texas' main power grid, more than five times the grid's roughly 85 gigawatts of existing capacity, according to Toro and others.
Many of those requests are placeholders rather than firm construction plans. Toro calls them "ghost projects."
These speculative projects clog the waiting lists for electricity connections, making the electricity shortage look a lot worse than it actually is.
"Developers routinely spam projects across several sites, and many are speculative builds with no customer yet," Ryan McEntush, a partner at venture capital firm Andreessen Horowitz, wrote in a recent blog. "Often these new developers have never even plugged in a GPU, let alone a power plant."
Texas has paused applications to review and weed out projects that aren't serious. Toro, from Aterio, expects this to favor the established players at the expense of the bragawatts.
The real AI boom
None of this means the AI infrastructure boom is imaginary.
Aterio tracks thousands of buildings, including major projects already under construction. Bernstein estimates that nearly 80% of the capacity the largest technology companies are directly developing is credible.
And when speculative projects disappear, Rezaei doesn't expect a wave of unpaid debt. Many will fail before building begins, the stage when developers borrow serious money through project financing.
Some of those missing data centers were never really coming in the first place.
INTERVIEW: Nokia chief technology and AI officer Pallavi Mahajan and Nvidia SVP of telco Ronnie Vasishta told the Mobile World Live Podcast their AI-RAN partnership is showing tangible progress, as they announced close to 20 operators are lined up for lab and commercial trials of the technology.
In an exclusive interview to outline progress of their efforts, Vasishta and Mahajan highlighted the fact it had only been 12 months since the two companies announced a joint AI-RAN partnership, which included a $1 billion investment from Nvidia into the Finnish vendor.
A year on, today (21 October) at Nvidia’s GTC 2026 event in Berlin the duo announced almost 20 operators globally have committed to AI-RAN collaboration, including 10 from EMEA. Mahajan said these operators are at different stages, lined up for both lab and commercial trials.
In 2027 trials are expected from BT Group, Deutsche Telekom and Elisa, with the latter preparing for commercial deployment.
In the US, T-Mobile has completed what was claimed as the world’s first AI-RAN call on a live network, while in France Orange is introducing new voice and intelligence services for enterprises based on the technology.
Speed of light Vasishta told MWL it was hard to overhype development based on the progress it had seen over the past year, and the number of trials and deployment is only set to grow.
“A year ago, or two years ago for sure, I think people would have scoffed at the idea. Nokia talking about operator commercialisation with AI-RAN, I think, that would have been perhaps seen as somewhat sceptical.”
Mahajan concurred, describing the pace of development as occurring “at the speed of light” with the game changer centred on being able to outline the true “value of what AI brings in terms of making RAN efficient” such as significant spectrum efficiency gains.
“In the next few years, the distinction between the networks and computing infrastructure will just blur,” she said, inevitably opening up a host of operator opportunities across applications beyond 5G and 6G.
Vasishta added Nvidia had started on the journey of software-defined RAN eight years ago and the inclusion of Nokia in 2025 had accelerated this dramatically.
“We see all operators around the world are created differently. They service different markets, they have different DNA, but overall, all of them have gravitated towards this concept now of AI RAN.”
Amazon outlined plans to invest more than €16 billion in its German operations in 2027 to expand its AI infrastructure, cloud computing and satellite connectivity.
The move adds to more than €100 billion Amazon invested in Germany since 2010 as it races against rivals Google and Microsoft to build out its AI services and cloud infrastructure in the nation.
David Zapolsky, chief global affairs and legal officer, described Germany as a centre of innovation and one of Europe’s most attractive investment destinations.
“We want German businesses, communities and creators to have the infrastructure and tools they need to compete and grow.”
The fresh investment includes €4.8 billion to expand Amazon Web Services’ Frankfurt cloud region and AI infrastructure to meet increasing customer demand.
It plans to bolster connectivity through its Amazon Leo satellite network to provide high-speed internet across the country.
Amazon Leo has secured 24 slots with European launch provider Arianespace.
The company plans to also broaden its logistics footprint and create new content for its Prime Video platform.
T-Mobile US is testing AI to allocate radio resources on its 5G-Advanced network, with Ericsson trial results showing the biggest gains during peak traffic.
The operator is trialling AI-Managed Beamforming on live commercial traffic at scale. Chief network officer Ankur Kapoor (pictured) outlined the results in a blog post. The work focuses on sounding reference signals (SRS), short transmissions from phones that help the network measure the path between a device and a cell site and direct beams more accurately.
Due to those resources being limited, they have traditionally been allocated under fixed rules. T-Mobile instead uses a supervised-learning model trained on its network data and running on Ericsson’s RAN compute and silicon to predict where demand is rising and redirect resources accordingly.
Kapoor explained training the model on T-Mobile’s data gives it an advantage on the operator’s network over networks which do not contribute their own tuning data.
The results came from newly activated sites in New York City, but the blog did not say how many were included or give a baseline for the trial.
T-Mobile reported the biggest gains during peak hours, including about a 50% increase in advanced beamforming using full-band SRS, a 16% improvement in downlink user throughput, 12% better spectral efficiency and a 10% increase in downlink data volume.
In May, the companies reported close to 10% better spectral efficiency from a separate AI-native scheduler with link adaptation on live traffic.
Kapoor said the technology suits T-Mobile’s network because beamforming works particularly well in TDD spectrum, where uplink and downlink share the same band, and the operator has a leading position in midband TDD.
He linked the trial to rising network demand, citing Ericsson’s June 2026 Mobility Report, which reported global mobile data traffic growth of 22% year-on-year. Video accounted for about 75% of traffic and 5G subscriptions topped 3.1 billion.
Jaco du Plooy, SVP and head of the T-Mobile customer unit at Ericsson Americas, stated the trial marked another step toward networks which respond more intelligently to demand, enabled by Ericsson’s silicon.
Kapoor said 6G networks will need to sense, learn and respond to support applications such as distributed AI, robotics, vision AI and other forms of physical AI, describing T-Mobile’s nationwide 5G-Advanced network as its proving ground.
Nokia agreed a deal with SoftBank Corp to conduct an AI-RAN field trial in commercial traffic conditions in 2027, progressing efforts to validate the technology.
The Finnish vendor stated the trial in Japan would be an important step in SoftBank’s AI-RAN ambitions by assessing the approach in a live network environment.
For Nokia, the trial also advances its AI-RAN commercialisation roadmap by shifting the technology from laboratory development to real-world deployment.
Nokia explained it will combine its AI-native anyRAN cloud RAN software on SoftBank’s core AI-RAN platform AITRAS, along with Nvidia’s ARC platform.
Through a combination of the platforms, Nokia stated AI-RAN would bring radio and AI workloads on to a common platform, helping operators “improve infrastructure utilisation, acceleration innovation and unlock new revenue opportunities”.
Nokia and SoftBank already have an AI-RAN deal in place, backed by an Nvidia accelerated computing node.
SoftBank SVP and CTO Ryuji Wakikawa said the trial is a significant milestone in the transition of AI-RAN from concept to execution.
“We are excited to begin this next phase of validation with Nokia, paving the way for a new era of networks that can deliver both connectivity and AI services.”
Pallavi Mahajan, chief technology and AI officer at Nokia, pointed to the potential for AI-RAN to “unlock new levels of spectral efficiency while providing the distributed intelligence needed to support physical AI”.
Google's cloud division competes with Amazon and Microsoft for enterprise customers.
Bloomberg/Getty Images
Multiple banking apps in the UK, including Revolut, faced disruptions on Tuesday morning.
Google Cloud experienced issues with its London-based cloud region and has since said services are coming back.
Revolut said in a statement that customers' money was safe while it investigated the disruption.
Google said it had largely resolved a cloud outage that appeared to knock out a number of online banking services.
The company said in an update at around 10 a.m. ET on Tuesday that the problem had been mitigated for the most part, with "a very small number" of customers still experiencing issues.
Google had said the outage had been caused by a "power event" at its europe-west2-b zone, which is based in London.
A handful of banking apps and services, including Halifax and Revolut, reported issues on Tuesday that coincided with Google Cloud's disruptions. Some companies said that the Google Cloud outage was the cause, while others attributed it to an unnamed third-party provider.
Users of some affected banking apps said in social media posts early Tuesday that they couldn't make card payments, log in to the app, or transfer money.
"We can confirm that all our customers' money and data are safe. We've been working to fix the issue since it was identified this morning and apologize to anyone affected and for any inconvenience caused," Revolut said in a statement earlier in the day.The outage was due to Google Cloud, a person familiar with the matter said.
"We're aware some customers are having issues with our app and online. We're sorry about this and are working hard to fix it," Halifax posted on X.
Emma, a popular UK budgeting app, also reported outages on Tuesday. "Emma is currently down due to a Google Cloud outage affecting services in London since 7 a.m. This is a technical outage, not a security issue," the company said on X.
Some took to X to share how the outage was disrupting their lives. One person said they'd had to cancel their beauty appointments in Japan, and another said their card was declined at the airport and they couldn't reach the support team.
Emma is currently down due to a Google Cloud outage affecting services in London since 7am. This is a technical outage, not a security issue. We're monitoring closely and will update here as soon as the app is back. Sorry for the disruption 🙏
The outage appears to have affected non-finance companies, too, according to Downdetector, a site that tracks online service disruptions.
Earlier on Tuesday, an X account for DPD, a delivery service, said in a post that its website was affected, and the web-hosting platform SiteGround said the outage has hit its UK data center. SiteGround said in its post that its engineers were working with Google Cloud engineers to address the issue.
The UK's courts and tribunal service said that it was experiencing issues with its cloud-based video tools for remote hearings on Tuesday. "We're working with our technology partners to restore service," an account for the government service wrote in an X post.
The Google Cloud outage comes as the tech giant has competed with Amazon and Microsoft to win business from enterprises, startups, and governments, particularly amid the AI boom. Reliability is a big promise built into these deals, as outages can have costly consequences for customers.
Google is not alone in being affected by outages, though: Amazon, Microsoft, and others have suffered downtime this year.
Elon Musk (pictured) revealed plans to give a SpaceX share to every child in the US next year, making the pledge a day after AI company Anthropic outlined plans to dedicate part of its equity to public benefit initiatives.
“Next year, I will give one of my SpaceX shares to every kid in America,” Musk posted on social media platform X, adding: “This will be the first of many.”
He provided no details on how the shares would be distributed or how the plan would be funded. According to Census Bureau data, around 73 million children live in the US. This would put the potential value of the giveaway at approximately $10.7 billion based on SpaceX’s closing share price of $146.03 on 16 October.
A regulatory filing made in August showed the billionaire held a 48.4% stake in SpaceX as of 30 June. The company went public earlier this year, raising $75 billion in a record-breaking IPO which briefly made the tech mogul the world’s first trillionaire.
Public BenefitFund Separately, Anthropic announced on Sunday (18 October) that it would commit 5% of its shares to a new Public Benefit Fund, with the potential for the allocation to grow as the company’s valuation increases.
The AI player stated the fund could receive an additional 2% of its shares for every 50% increase in its valuation above a baseline yet to be disclosed. Additional allocations would be made at the board’s discretion.
Anthropic stated the fund would give it “increased flexibility and scale” to pursue its mission, including “large donations to traditional non-profits or direct distributions to specific populations”.
One potential focus is AI-driven disruption to employment. The company stated that, although the technology could generate “enormous economic value”, it could “cause instability for many workers”. It has already committed $200 million to research and initiatives addressing the issue, “but given the pace of AI progress, bolder approaches may be needed”.
“The Fund will have the versatility to pursue them.”
Anthropic, which has confidentially filed for an IPO, stated further details on the fund’s governance and valuation thresholds would follow in the coming months.
stc Group seized the initiative on AI-native networking, setting a target of employing the technology directly in its core infrastructure by 2030.
The operator believes placing the technology at the heart of its infrastructure is essential to keep pace with growing demand for AI services and the resulting changes to connectivity requirements.
stc detailed its ambitions and approach in a white paper released during the Network X event in Austria last week in which it argued AI is essential to handle the complexity of modern telecoms networks.
In a statement publicising the white paper, stc explained the evolution to 5G-Advanced is a principal driver of increased operational complexity. It also pointed to “AI-enabled calling and immersive communications” as factors necessitating the technology.
stc argued operators are set to become essential providers of connectivity as AI services take-off, stating agents and automated systems “need networks to coordinate communications and computing resources”.
It believes an AI-native network is essential in this regard because of the greater levels of automated decision making and resource allocation the technology should afford.
stc is also seeking to improve network maintenance and reliability.
Microsoft reportedly withdrew from a deal to lease computing capacity at AI infrastructure player Nscale’s planned data centre in the UK due to delays related to power supply, adding to mounting concerns around the status of the country’s AI projects.
According to The Telegraph, the US technology giant pulled the plug on the mooted data centre to be constructed in Loughton, Essex, which was originally expected to receive a grid connection by the end of 2026.
However, it now faces a wait until the following decade for supply.
“Our relationship with Microsoft remains strong and long-term, and this unfortunate London-wide grid issue has no impact on our other contracts,” an Nscale spokesperson told Reuters. The pair are still collaborating across other projects, including developments in Norway, Portugal and the US.
The company reportedly added it remained “committed to both the Loughton project and the UK” and was exploring alternative power sources.
Announced last year, the Loughton development was backed by a £2.5 billion investment commitment from Nscale and formed part of a wider UK push to attract AI infrastructure spending. Microsoft had planned to lease the site’s entire 50MW capacity and deploy an AI supercomputer equipped with around 23,000 Nvidia GB300 chips.
The setback revives questions about the delivery of the UK’s headline AI investment pledges. A Guardian investigation in March raised concerns over the status of major announced projects, including the Loughton site, which was still being used as a scaffolding yard when visited by reporters in February. Nscale subsequently said it had purchased the site in September 2025 and was targeting delivery in 2027.
Power availability remains a wider constraint for UK data centre development, as growing demand from AI workloads puts pressure on grid infrastructure.
Nscale, which is backed by Nvidia, filed for a US IPO last month and raised $3.4 billion in pre-IPO financing. The UK-based company named Microsoft and Anthropic as expected major customers in its prospectus.
Global telecom employment fell 2.1% year-on-year in Q2 2026, extending a decline dating to 2016, according to a report by MTN Consulting.
The report’s author, Matt Walker, chief analyst at MTN Consulting, said the panel of 145 operators employed about 4.2 million people, down from 4.3 million a year earlier. The 2.1% fall exceeded the decade-long annual average of 1.7%.
The job cuts did not consistently reduce labour costs or establish a direct link between AI adoption, smaller workforces and wider margins. Among 72 sizable operators, MTN Consulting found no measurable relationship between headcount changes and trailing four-quarter EBIT margins.
It also found no consistent link between AI-related news and restructuring. Telefónica cut its headcount by 19.3%, the steepest reduction in the industry, yet had no tagged AI-automation news, while SK Telecom drew the most AI coverage and reduced headcount by 6.2%.
The findings were less positive than those in other industry surveys. Nvidia’s fourth annual State of AI in Telecommunications report stated about 65% of respondents reported AI-driven revenue gains or annual cost reductions above 5%.
MTN Consulting found fewer than 20% of global operators increased revenue by more than 5% a year, while fewer than 20% cut operating expenses by over 5%. Only BSNL, Iliad and MTN Group achieved both, representing less than 2% of the global market.
MTN Consulting said the gap suggested survey responses reflected industry optimism rather than audited operating results.
Labour costs per employee across the global panel rose 6.8% year-on-year to about $62,500. Higher compensation per worker more than offset savings from job cuts, increasing the industry’s total labour bill.
Asia was the only region to add workers during the quarter, with headcount up 0.6%, largely due to hiring by operators including KDDI. Europe posted the largest decline at 3.4%, partly reflecting Telefonica’s 19.3% reduction. Employment fell 2.8% in the Middle East and Africa and 1.2% in the Americas.
Operators who cut staff were more likely to improve operating margins, but better performance was not assured. Event studies dating to 2011 found they outperformed operators which did not cut staff in about 57% to 67% of cases.
MTN Consulting advised operators, vendors and investors to assess workforce changes and automation returns using detailed financial data rather than broad technology claims.
OpenAI CFO Sarah Friar reportedly warned businesses using low-cost Chinese AI models could face cybersecurity risks and said Western alternatives offered greater protection.
Speaking to Bloomberg at the International AI Summit in Dublin, Friar urged companies considering Chinese models to assess the safety and security risks.
OpenAI and rival Anthropic face growing competition from open-weight AI models, particularly those developed in China. Users can download and run them on their own hardware, often at a lower cost than leading US alternatives, with comparable performance on some tasks.
Chinese releases accounted for the largest share of generative AI model downloads on hosting platform Hugging Face in August, reported the outlet.
OpenAI and Anthropic mainly offer closed models which cannot be downloaded or modified. Both companies have argued open alternatives pose greater risks.
Anthropic head of international Guillaume Princen voiced similar concerns about open-weight models at the event.
“I can’t talk for open models. But by definition, there’s more risk with them,” he said.
The comments follow several cases of advanced AI systems behaving unexpectedly during cybersecurity tests.
In July, OpenAI disclosed its models had breached systems operated by AI start-up Hugging Face during an authorised cybersecurity evaluation.
Anthropic has also disclosed examples of model misbehaviour, including a Claude model submitting a false tip related to a homicide investigation, Bloomberg reported.
Friar nevertheless cautioned US AI companies against slowing releases.
“I do think that we need to pace it but not slow it down so much that we get this imbalance on who has access to the tools that potentially, from a cyber perspective, create real mission-critical issues in the world,” she said.
Katti plans exit Also yesterday, Sachin Katti, OpenAI’s infrastructure lead, stated on LinkedIn he plans to leave his position on 15 November but will continue to support the company as an advisor.
“After an incredible year at OpenAI, I’ve decided to move on to build something new,” he stated.
Prior to joining OpenAI in November 2025, he held various technology roles at Intel, including Chief Technology and AI Officer.
CNBC reported Uday Ruddarraju, who has been serving as OpenAI’s CTO of compute, will replace Katti and lead the company’s overall compute organisation.
Google DeepMind co-founder Demis Hassabis cautioned that recent AI-driven cyberattacks were “early warning shots”, as more serious incidents could emerge over time unless companies and governments stepped up efforts to secure AI systems.
Speaking at the International AI Summit in Dublin, Hassabis said the industry needs to adopt safety practices used in sectors such as biology and nuclear technology, calling for secure, isolated sandboxes to test AI agents before they are connected to the internet, alongside stronger cybersecurity measures.
“We’ve got to kind of asymmetrically power up our defensive side to keep ahead of the attackers,” he said, arguing banks and governments could deploy AI to identify and fix software vulnerabilities before systems are widely released.
He also renewed calls for an international standards body to oversee AI safety, but clarified any framework must be flexible enough to keep pace with technological developments without stifling innovation.
Microsoft president and CEO Brad Smith also advocated for stronger safeguards, describing 2026 as a “sobering year” for AI safety. “No company or technology should be above the law,” he said, adding that regulation needed to be “sensible” and “balanced” while accounting for innovation.
Smith added that companies should be prepared to slow development when necessary, citing Microsoft’s decision to move 34,000 engineers from product development to cybersecurity two and a half years ago. “It was painful, but it was the right thing to do.”
“This allergy that some in the industry have to regulation is frankly something we need to get over,” he quipped.
European AI research Separately, Google DeepMind and Google Cloud announced plans to provide up to 200 European scientific researchers with access to their frontier AI models, alongside $3 million in compute credits.
The initiative will support the European Union’s AI for Science agenda through the Resource for AI Science in Europe (RAISE), which aims to address computing resource gaps faced by researchers.
The companies will provide researchers with access to AI tools including AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve and AlphaEarth Foundations. Google said the programme would equip researchers to tackle complex scientific challenges, including predicting extreme weather patterns and mapping molecular structures.
Telus Digital is now a deployment partner for Germany-founded AI customer service specialist Parloa, expanding its enterprise services portfolio with support for the latter’s AI agents.
Under the agreement, businesses will buy the platform directly from Parloa and use Telus Digital for integration, change management, managed services, and other professional work. Telus Digital will also introduce Parloa’s technology to existing and prospective clients.
Parloa’s agents handle authentication, bookings, payments, and order enquiries, transferring conversations to a human when personal assistance is needed.
The company cited its own research to illustrate the shortcomings of existing customer-service tools. Its study of 10,000 enterprise websites and 4,000 chat interactions found fewer than 9% of chatbot conversations reached a resolution, while only about 10% of escalation attempts connected users with a person.
Separately, its Consumer Patience Index, a survey of 1,001 US consumers conducted by Propeller Insights, found nearly a third would switch brands to avoid being put on hold.
“Consumer expectations are high, and brand loyalty is fragile,” stated Parloa CRO Chris Silver. “They need to have their issue solved, the first time, on their preferred channel and with no unnecessary delays.”
The agreement expands Parloa’s delivery capacity in North America and Europe. Its platform runs on Microsoft Azure and supports multiple AI models, including OpenAI’s GPT series. The company stated it complies with GDPR and meets ISO 27001 and SOC 2 standards.
Telus Digital global SVP of growth Will Mayo framed the partnership around cost, arguing high volumes of repetitive contact-centre interactions can increase the cost to serve without improving customer experience. He said Parloa’s agents could handle routine requests and transfer more complicated cases to a person.
Telus Digital stated Parloa will complement rather than replace its existing tools because no single product is likely to cover every client requirement. Its portfolio already includes conversational AI combining voice, generative AI and agentic automation.
Backed by parent company Telus, the digital services division operates with more than 83,000 employees across over 35 countries.
Nvidia CEO Jensen Huang (pictured) reportedly outlined plans to invest $2 billion across Europe to ramp adoption of AI technology and the company’s products.
Bloomberg reported Huang revealed the plan yesterday (13 October) during an event in Paris, which is part of a tour including stops in London and Berlin.
Huang said the fresh investments, to be announced in the next week, are on top of $8 billion the company previously committed to the continent, according to the news agency.
He also announced new funding for French start-ups Doctolib, Genesis AI, Pigment and Entalpic.
Nvidia’s founder also promoted open-source models, arguing their lower costs and greater flexibility could help Europe close the gap with the US on advanced AI.
The chipmaker is a major partner of French company Mistral AI, which develops open-weight AI models. It has invested in Mistral and supplies accelerators for its data centres.
“Open models are important as they can be customised for languages and for your own sovereignty,” he said, according to Bloomberg.
Nvidia’s GTC Berlin conference is schedule for 20-22 October.
India’s Adani Enterprises Limited and Apple supplier Jabil signed a memorandum of understanding (MoU) to form a 50:50 joint venture, tasked with manufacturing AI and data centre equipment in the country.
The JV is the next step on from a deal struck by the duo in June 2026 to form a strategic alliance to develop an AI and data centre infrastructure manufacturing ecosystem.
Once established, the pair stated the venture will supply hyperscalers, neoclouds, data centre operators and enterprises, serving accelerating data centre demand.
The companies cited industry forecasts which project India’s AI capacity will reach between 5 gigawatts to 8 gigawatts by 2030 and more than $50 billion in planned hyperscale spending across data centre, cloud and AI ecosystems.
In total, the pair peg the global market opportunity at more than $3 trillion over the next seven years, driven by investment in AI computing.
The duo will look to cash in on this demand by providing a manufacturing base in the country to serve domestic demand and support exports.
Adani Group, which runs the enterprise division, has made a $100 billion commitment to develop 5 gigawatts of green-energy powered, hyperscale AI-ready data centres by 2035, and the proposed venture adds hardware manufacturing to the group’s presence across infrastructure, energy and logistics.
Jabil makes components for Apple and will use the JV to expand its engineering footprint across India. It reported revenue of $36 billion in fiscal 2026.
Gautam Adani, chairman of Adani Group, said “the AI revolution will demand far more than just breakthroughs in algoritms” and the proposed JV would “help build the physical backbone of the global AI economy”.
Meta Platforms CEO Mark Zuckerberg (pictured) reportedly claimed the company had reached the “frontier” of AI, as it unveiled two new models designed to handle complex personal and professional tasks.
Following hefty investments in computing infrastructure and research, the technology giant said Muse Wonder and Muse Wonder Full AI mark “our most capable models to date”.
The models can handle tasks ranging from coding and document analysis to interpreting legal information, personal finance and health data.
Among key features is Active Memory, enabling the retention of relevant user information across conversations. The agents are also equipped with Active Focus to help them continue pursuing longer tasks, adapt to new information, recover from errors and ask users for clarification before taking important actions.
Meta claimed its evaluations showed strong results across benchmarks covering agent performance, memory and coding, adding the advanced Muse Wonder Full version performs better on unfamiliar tasks.
Built responsibly Unlike its earlier open-source strategy, Meta is offering Muse Wonder through its own developer and subscription channels in a direct challenge to AI chatbots from rivals including OpenAI and Anthropic. According to the Financial Times (FT), Meta said it would charge companies from between 20 cents per million tokens for basic access to as much as $200 a month for the more powerful Muse Wonder Full.
Speaking to the publication on the launch in which he made the comment about frontier AI, Zuckerberg added the new models would enable “much more sophisticated behaviours that are at the frontier of what is being done in the AI field”.
Meta claims Muse Wonder surpasses Anthropic’s Claude Opus 5.5 and Fable 5.1 and OpenAI’s GPT-6 Astra on certain technical measures.
Zuckerberg added that reaching this point required “a large amount of compute and investment” alongside sustained research. He told the FT that further model development and spending was under way, stating the company was “already working on the next step up”.
The chief also played down mounting concerns about increasingly capable AI systems escaping human control, declaring: “I’m much closer to the belief that AI is a technology that can be engineered with normal guardrails and built responsibly.”
Meta said its new models were trained to prevent unauthorised disclosure of private information and reject requests to facilitate harmful actions, while retaining agentic capabilities for defensive cybersecurity work.
Microsoft took the wraps off the third-generation of its in-house AI chip, as part of a plan to reduce its reliance on buying processors from Nvidia.
The Maia 300 chip will be used for the tech giant’s multi-gigawatt deployments in 2027, making it the primary inference system for Microsoft products and services. Reuters reported it plans to potentially deploy more than 1 million of the chips next year.
According to a blog post from EVP Scott Guthrie, the company expects Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems without naming Nvidia.
The Maia 300 is built as a single system, with software, silicon, physical rack infrastructure, networking, and data centres co-designed to operate as a single platform.
Guthrie noted a 96-chip rack offers seven times the rack-scale token throughput of Microsoft’s Maia 200. He said the new chip is “designed with an inference-first architecture optimized for long-context, memory-intensive and decode-heavy AI workloads”.
The Maia 200 chip powers the company’s production workloads, including Microsoft 365 Copilot, as well as OpenAI models running on Microsoft’s Azure cloud computing service.
Indian IT firm HCLTech struck an agreement with Anthropic to establish a separate business unit designed to accelerate enterprise adoption of agentic AI, with a focus on safety, cybersecurity and governance.
HCLTech stated its services team will build industry solutions and offerings powered by Anthropic’s Claude, Cowork and Claude Code, while its HCLSoftware business unit will embed these capabilities into select enterprise products.
The company will further establish a dedicated “Anthropic Business Unit and Centre of Excellence” to accelerate use of Claude in enterprise environments, providing secure sandbox environments to enable development.
Clients will also be aided by the partnership with the adoption of AI-native engineering practices, training workforces and building organisational capabilities to operate agentic AI safety at scale.
As what is known as an Independent Software Vendor (ISV) partner, HCLSoftware will be given access to advanced models. It cited select products across cybersecurity, automation, business process application and application development as services that will benefit from advanced Claude capabilities.
Claude will also be used across IT support and for digital employee services, workflows supported by policy guardrails, exception handling and human oversight for high-impact changes.
CEO and managing director at HCLTech, C Vijayakumar, said few technology companies can combine the scale of a global systems integrator with the product depth of an ISV.
“That gives us a rare advantage to securely engineer and deploy Agentic AI at scale and deliver the business outcomes that matter most to our clients.”
HCLTech’s deal with Anthropic follows an agreement worth $1.1 billion in July with a major unnamed European company to deploy AI services and boost enterprise services for the business.
Nvidia is reportedly in talks to either acquire or increase its investment in open-source start-up Reflection AI, as it seeks to strengthen its position in open-weight AI models and challenge lower cost Chinese rivals.
Sources told the Financial Times (FT) discussions remain at an early stage and could lead to a full acquisition, further equity investment or a deal covering chips and computing capacity. An acqui-hire, under which Nvidia would recruit staff and license technology, is another option.
A deal could apparently be reached in the coming weeks, though the terms under discussion remain unclear.
According to the report, Nvidia has already invested $800 million in Reflection AI, which was valued at $25 billion in a March funding round. Founded in 2024 by former Google DeepMind researchers, the start-up develops AI models and tools for software development.
Reflection launched its first open-weight model, Beam, last week, targeting coding and agentic tasks as it looks to compete with Chinese models such as DeepSeek and Kimi.
Open-weight models allow businesses and developers to customise AI systems and run them on their own infrastructure, potentially cutting costs compared with closed services such as OpenAI’s ChatGPT and Anthropic’s Claude.
Although Nvidia has its own Nemotron family of open-weight models, FT said it lacks a frontier-scale model of its own. In an essay published in July, CEO Jensen Huang called for a domestic open-weight “ecosystem” across the US, arguing it was essential for “creating opportunities for innovation and prosperity across the country”. “In fact, openness may be one of the most important paths to AI safety and security,” he argued, adding “relying solely on closed models is not inherently safe” as “they can be breached, misused, or fail in ways that outsiders cannot detect”.
Reflection AI has also secured substantial computing capacity from SpaceX. The deal, struck in June, gives the start-up access to Nvidia’s GB300 chips at SpaceX’s Colossus 2 facility in Memphis, Tennessee, for $150 million per month through 2029.
Login
