Computerworld.com [Hacking News]

While delivering record Q4 revenue, Apple’s financial results revealed a $33 billion cost associated with memory price inflation, representing the true cost of the growing bubble of AI data center investments. 

The results also showed something that threatens those same data center investments — a sea change in how enterprises are approaching AI. Inference costs, security, and data governance concerns may finally be dampening enthusiasm for cloud-based AI services.

Enterprises go local

Apple revealed several major enterprise clients who are turning to its products as solutions for local inference. The company told us that Gap Inc. is running AI loss-prevention models entirely on Mac minis to eliminate cloud token costs, while world-famous auction house Christie’s is using Mac Studio to run local pricing estimate models.

Apple’s wider enterprise wins matter too. It told analysts and investors that as well as purchasing 25,000 iPads for pilots and mechanics, American Airlines is deploying over 37,000 iPhone 17 devices to flight attendants for in-flight purchases with Tap-to-Pay. In education, it revealed a huge deal with the Department of Education in the Philippines, which will deploy 170,000 MacBook Neos to teachers and staff.

All these deployments help the company build up its mindshare, which should translate into future customer success and help it stay relevant in fast-changing times. But it’s really Apple’s pivot into AI that is tempting business users to make these purchases. 

The hardware and the economics

It’s not trivial that the company’s M-series and A-series processors are highly performant. It matters that you can patch together four Macs to create a powerful on-premises AI system more than capable of handling even highly complex AI calculations. 

“From Mac to iPhone to iPad, we are proud that our products are the best platforms in the industry for AI for both first-party and third-party applications, and we’re excited about the opportunities that creates,” CEO John Ternus said.

Look at it this way: If a Mac costs a few thousand bucks but can deliver the equivalent of tens of thousands of dollars of inferencing tokens during its lifespan, then it’s a bargain, not a cost. What Apple confirmed is that at least some enterprise purchasers are paying attention to this.

Monetizing both ends

Of course, the other side to AI is the need to build it, so it also matters that when it comes to the deployment of its AI APIs from within applications created by third-party developers, Apple is providing a lot more clarity. It is giving developers building apps with under two million downloads direct access to its powerful server-side LLMs completely free of token costs, it said.

It has also introduced a three-tiered system for developers needing to use its Private Cloud Compute infrastructure. The company is also putting a little more flesh on its plans to charge end user customers for heavy access to its server-side queries through iCloud+. 

What that means is that while delivering hardware capable of running on-device AI, Apple is also building an end user market for AI inferencing the Apple way, and equipping developers with the tools they need to attract end users to make use of its inferencing capacity in the first place. It is monetizing both ends of the story, while also enabling all parties to build success.

That’s not just about building and maintaining AI relevance for Apple. It’s also about building up new business opportunity. Ternus pretty much confirmed this, telling analysts during the Q4 financial call, “in the beginning, iCloud tiers were largely about storage, but we’ve been adding more capabilities and features to it.”

Siri agentic AI

Ternus also talked up Siri AI. “With Siri AI, we really have the kind of the Siri that we always wanted,” he said. “It’s primarily request-based today, but as we talked about at WWDC, it’s built on a modern orchestrator-driven architecture that gives us real room to expand its capabilities over time. As far as agentic capabilities, we think these are promising tools for users. We just want to make sure that users are aware of what they’re giving access to.”

Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon.

“Zoom Got It Done” is the tagline for the company’s new Zoom Ahead campaign, with the focus being how Zoom’s AI-first capabilities take conversations to completion. This one features Kate McKinnon and builds on their “I Use Zoom!” spot with Bowen Yang that kicked off this consumer-friendly campaign late last year.

For the record, I’m not a fan of their comedic approach here, but the ads hold your attention, and it’s great seeing SNL favorites playing over-the-top corporate tropes. Am just not so sure the messaging gets it done.

Wearing my Marketing hat, I’ll note that B2C marketing campaigns from tech companies are really hard to do well, especially when the audience is not the economic buyer. The Zoom ads are aimed at workers – end users – while the buyers – IT and CEOs – are caricatured as the worker’s worst nightmare.

This is a topic I’d love to explore further, but I need to stay on script and tie their Got It Done tagline to the main topic, namely my review of the recent Zoomtopia event. While the event is mainly for customers and partners, I was part of the small cohort of analysts who attended as well.

Since Zoomtopia caters primarily to those economic buyers, it wasn’t surprising that the Got It Done message was the opening theme for the event, capped by a local choir that went full gospel to celebrate the uplifting power of AI to get work done and relieve us of so many tedious tasks that undermine our productivity. Am not sure if everyday workers get that euphoria when using Zoom, but that’s the aspiration – nothing wrong with that, amen.

Choir singing the praises of ‘Got It Done’

Jon Arnold

More importantly, I think Zoom is trying to communicate how AI is transforming their platform into something bigger than video calls and meetings. On that count, I think Zoomtopia succeeded, perhaps in ways that most people don’t realize. In that context (pun intended), I’ve distilled my takeaways into three big ideas.

1. Unified Context Layer

This was the big story at the event, so let’s start there.

I’m never the first analyst to post about industry events, but I’m going with flow here, as this was the lede most of my peers went with for their write-ups. What I haven’t seen so far, though, is connecting the dots to the last time most of our group was at a Zoom event. That would be Perspectives, their analyst-only event, and back in April we got the deep dive on their go-to-market vision, “from conversation to completion”.

The overall reception was positive, and it marked a new chapter in Zoom’s evolution to align more closely with AI’s transformative potential and get beyond being a meetings platform. Building on the premise that all interactions start with some form of conversation, AI does all the hard work of taking those conversational inputs, identifying next-best actions and closing with intelligent task completion.

At Zoomtopia 2026, the big story was Unified Context Layer, which takes that vision one step further. The missing piece was context, where AI is now better able to understand intent, sentiment, urgency, etc., and connect all that to memory across the entire lifecycle of an interaction. Humans simply cannot do this at scale and in real time, so this is where the AI story becomes next level.

On a more basic level, this is no longer about human-to-human communication, the foundation upon which all UCaaS platforms are built. Communication remains the cornerstone for how work gets done, but now it’s just as much about P2P as it is about P2A, A2P, A2A and B2B. Translation – people need to communicate among themselves, but increasingly it’s also agent-to-agent, bot-to-bot and everything in between. Getting context right is more challenging when AI agents are involved, and that point will tie into my second big theme.

Zoom’s new Unified Context Layer

Zoom

To make all this work, the vision needs to be broader, so now it’s “conversation to context to completion”, and that’s what the Unified Context Layer brings to the Zoom platform. You can read more about it here, and I’ll just add some key points to show why the context piece is so important.

• This new layer takes conversations to completion “by linking conversation history with relevant documents and data”. Nothing really new here – other vendors are doing similar things in an effort to become the central hub through which all workflows go. What is new is how these capabilities are framed to support both actual workers – humans – and AI agents. Not only are these agents doing work and completing tasks on behalf of human workers, but as they become more “agentic”, they’ll be doing these things independently. Definitely not something the founding fathers of UCaaS had in mind.
• To become an enterprise-wide “system of action”, this new layer works across all of Zoom as well as third-party applications. This is necessary for agents to pull context from everywhere and every interaction – humans cannot do this on their own, and when all the integrations are in place, data no longer stays in siloes. Their update notes that with MCP connectors and APIs, Zoom can search across over 40 third-party applications, such as ServiceNow, Workday and MSFT OneDrive. Drawing from such a wide range of sources, workers and the workplace become more productive, and makes Zoom Got It Done more than a marketing slogan.
• It’s worth noting how Zoom identifies two forms of data to understand context at a deeper level – dynamic and static. Dynamic would be human-based, real-time interactions in the workplace – calls, chats, meetings, etc. Since all of these are in digital form now, AI can process them as conversational intelligence inputs to drive task completion. Static data represents what’s on record and documented – also in digital form, so it can easily be mined by AI. Together, this presents a more complete picture for automating workflows.
• The emphasis on AI is hard to overstate, as agentic search capabilities serve to get more complete information – and context – to surface answers faster, so tasks can be completed in less time. For example, one of the new features, My Notes, turns meetings into workflows automatically – something that human workers would have spent a lot of time doing manually. Not only that, but Zoom now has prebuilt templates for specific functions, namely sales, marketing and IT. These workflows can now be created directly in My Notes, without human effort to do this with a workflow builder.
• Another sign of where this is going is the distinction between general and specialized AI agents. Why stop at each worker having one AI agent? Never mind the workers – ZoomMate’s AI Agent Teams now can have a general AI agent manage teams of specialized AI agents, further accelerating the time to task completion. Each agent can be assigned a specific task, so you can now have subagents working in parallel to support the overall task. Humans may think they can multitask effectively – it’s a myth – but this approach is next level. AI magic aside, it’s important to note that all of this task automation is subject to human review, so it’s too early to think about using this to replace human workers. Sidebar – this is exactly what Kate McKinnon was getting so worked up about in the ad – let’s just call it misplaced Zoom Rage, where Zoom gets it done in ways her workers simply cannot do.

Zoom Ahead campaign featuring actress and comedian Kate McKinnon.

Zoom

2. AI update – aka, the Anthropic effect

I may be going out on a limb here, but Anthropic was barely mentioned at Zoomtopia, and it seems to me this is driving a lot of Zoom’s evolution, perhaps more than we realize.

If you don’t know, Zoom took a small stake in the company in 2025, and has since added to its position. Financially, it’s been a huge windfall for Zoom, but that’s really beside the point. Zoom and all the other cloud players in their orbit cannot afford to not be close to all the AI players in Anthropic’s orbit – that’s just how Big Tech works.

For Zoom in particular, this is about helping transform the company into an AI-first platform, and what better way to do that than by leveraging their small stake with one of the top three players creating the AI economy? In that context – there’s that word again – their “system of action” vision makes more sense. Yes, human workers will use these tools to take conversations to completion, but longer term, this system will be about AI taking the action.

I’ll call this the “Anthropic Effect” because Claude is now deeply embedded in everything Zoom touches – collaboration and meetings with UCaaS, customer interactions with CCaaS and more recently, ZRA for sales automation – Zoom Revenue Accelerator. All of these data flows cross Zoom’s platform, and this is where AI agents need to be to tap into the data that takes conversations to completion.

Memories can be short, but AI Companion was a cornerstone for how workers engaged with the Zoom platform. Yes, all platforms evolve, and AI Companion was quietly retired, replaced – augmented, really – by ZoomMate. AI Companion hasn’t really gone away; it’s just been subsumed by something bigger and more agentic. Perhaps to make room for Claude, or at least get hooks into what their system of action is trying to do.

On a technology level, this is perfectly reasonable. Zoom has actually been at the forefront of innovation in our space, and they have succeeded at just about everything they’ve taken on internally. However, I think it’s fair to say that even Zoom cannot replicate what Anthropic, OpenAI, Meta, et al have been doing with AI.

Depending on how you look at things, I think the business rationale for riding with Anthropic is even stronger. Consider the chart below, from App Economy Insights. They publish a newsletter that breaks down how tech companies make money, and Anthropic’s revenues are up 45-fold over the last six quarters.

Of greater interest, as their newsletter reports, Anthropic has almost 12,000 enterprise customers spending $100k+ annually – exactly the kind of customers Zoom covets. To whatever extent these customers rely on Claude, this sure dovetails nicely with Zoom’s value proposition. Going from conversation to context to completion at scale requires a lot of AI horsepower, and the more tightly Claude integrates with ZoomMate, the more successful both companies will be.

Of course, Claude could fall out of favor, and all of this will be moot, and every other cloud platform player is partnering with Anthropic, so nothing is for sure. That aside, though, this looks like a good way for Zoom to build a moat they can defend, especially as AI gains trust and is used to automate more workflows and tasks. That makes for a good growth story to become an AI-first system of action that truly takes them out of the UCaaS milieu that is going sideways at best.

Anthropic’s revenue growth

App Economy Insights

3. New management controls for AI

Finally, I wanted to note this focus, as it says a lot about how Zoom wants to be positioned, and how they want to be perceived by enterprise IT leaders.

This is a core piece of the story in terms of Zoom becoming an AI-first platform and truly evolving beyond video calls, meetings and UCaaS. Their new management controls are coming soon, but this signals where and how Zoom intends to play.

With Anthropic now in their DNA, these are the capabilities Zoom will need to become a centralized system of action for enterprises. Not siloed for workplace productivity or in the contact center, but to be the hub for all communications flows.

Again, this is no longer a UCaaS story – one could argue that Teams holds court there, along with other leading UCaaS players – and Zoom is looking beyond this and wants to be positioned on a higher, more strategic plane.

In this – yes, context – the Zoomtopia updates were about the things that IT really cares about when managing these do-it-all platforms. This is about enterprise-wide policy and governance, not updating permissions to edit transcripts, configuring new endpoints for meeting rooms or updating CRM integrations.

These are the bigger picture controls that IT needs to be comfortable and confident that Zoom can really be this system of action for the entire organization. To illustrate, here are three examples for how Zoom intends to earn their trust.

• AI guardrails. Allows IT to customize account-level rules for what AI agents can use or generate. These rules can be highly specific for a team, job function, line of business, etc., but always subject to human review (HITL) to keep AI in check.
• Deepfake risk detection. Not a new idea, but as agentic AI proliferates, and as personal AI agents start clogging your digital channels, the threats become very real. When synthetic data is paired with bad actors, deepfakes can find their way into workflows and customer interactions, making this an enterprise-wide challenge to address, and not just limited to the contact center.
• Compliance Manager. With AI-driven outputs turning up everywhere, there’s a need for centralized visibility to remain compliant. Different platforms track communications in different ways, and when it comes to archiving, data retention, search, retrieval, etc., Zoom will soon be adding more compliance tools to support IT at an organizational level.

Conclusion

Zoom executive Q&A with analysts

Jon Arnold

During the executive Q&A with analysts, Eric Yuan was asked where Zoom would like to be in the future, and I think his three-word response sums up everything I’ve been trying say here. He envisions Zoom as a “work operating system”. Not a UCaaS platform, not a meetings app, but something bigger, and more of a hub that enterprises can organize workflows around.

Conversations to context to completion is the path for getting there, but this puts them up against hyperscalers who are all-in with AI, and not terribly interested in the communications layer that got Zoom this far in relatively little time.

Is Zoom ready for this? Will their customers be ready to view Zoom this way? With much of this data residing with third-party apps, who will ultimately control all the data needed to be a work operating system? What role will the hyperscalers and AI majors play? How will their working relationship with Anthropic evolve to support this vision?

These are just a few leading questions to consider following Zoomtopia. Prediction markets aside, I’m more inclined to bet on Zoom than against Zoom, as their track record for making moves has been pretty good.

The management team has both vision and confidence – they know what they have, where they’ve been and where they need to be going – and I’ll try to answer those what’s-next questions as things progress. As with everything AI, they will progress, and Zoom just might get it done; in which case Kate McKinnon’s Zoom Rage will turn up to 11. Can’t wait to see that!

This article is published as part of the Foundry Expert Contributor Network.

It’s early days for AI-assisted work, and sentiments are mixed about its benefits: Does AI actually bring value and save employees time, or does it bog them down and add unnecessary activities?

Data from ActivTrak seems to vacillate between those two potential realities: The latest research from the productivity tracking software company says employees who use AI most heavily are working up to 90 minutes longer each day, yet its previous research, published earlier this year, found that they are actually working less.

The discrepancy may be due in part to the timespan and sample sizes of the data collection for the two reports, but it indicates that enterprises should still be wary of vendor claims about AI-related productivity gains. Measuring them remains incredibly difficult.

AI use compared

ActivTrak’s current findings are based on behavioral work data gathered from 9,140 knowledge workers across 467 enterprises over a five month period (April 1 to September 25, 2026). This sample is much smaller than that of its previous report based on 443 million hours of work performed by 163,638 employees across 1,111 organizations over a three year period (January 1, 2023 through December 31, 2025).

According to the latest findings, heavy AI users are working between 8.42 and 9.08 hours per day, compared to 7.58 hours for those not using AI. This contradicts ActivTrak’s previous finding that average workdays had decreased by roughly 9 minutes (a 2% decline) thanks to AI.

On the productivity front, heavy AI users have 22% shorter focus sessions than light users (37.8 minutes versus 48.2), suggesting their work is more fragmented due to AI use; further, they switch apps 15% more frequently.

“Over time, those interruptions add friction to high-value work that requires concentration, critical thinking, and sustained attention,” Sarah Altemus, manager of the ActivTrak Productivity Lab, said in a blog post describing the latest findings.

ActivTrak’s previous report had found that over the period of the earlier study, focus efficiency, the percentage of work time spent in focused, uninterrupted activity, fell to a three year low of 60% with AI use, and average focused session time in 2025 declined by 9% from 2023. However, that same report found that daily productive time increased by 19 minutes, or 5%, to 6 hours and 36 minutes, with AI in the workplace. Furthermore, productive sessions grew by 13%.

The company’s latest report also tracked the productivity differences between users of tools with embedded AI and users of web-based AI tools. Those using embedded AI switched apps 0.6 times an hour, while those using AI on the web switched between 1.7 and 4.2 times per hour. That’s up to 4x more switching than by employees who don’t make use of AI at all, which creates what ActivTrak describes as a “swivel-chair tax” — the additional effort spent by employees due to the movement between apps.

Similarly, in its previous report, ActivTrak had found that this multitasking rose 12%, or 10-plus minutes, with AI.

Interestingly, according to the latest research, light users spend more active work time in core job-specific apps (81.4% compared to 78%).  “In other words, more frequent AI engagement doesn’t necessarily translate into more focused work,” Altemus pointed out.

The larger issue is the way organizations manage the additional capacity that AI may create, Altemus noted. “When employees complete tasks faster, managers may assume they have room for more responsibilities,” she said. Without clear workload boundaries, time savings may be redirected into additional work that can be taxing.

The numbers often don’t tell the full story

ActivTrak has a useful view of how work moves across applications, noted Sanchit Vir Gogia, chief analyst at Greyhound Research, but he also pointed out that it sells the capability that produces these measurements.

Recorded use depends on what the instrumentation can detect, he observed. For example, an employee can use AI at work and still appear in the non-user group, and work on an unmanaged personal device sits outside the measured estate. In addition, an API or a background agent can finish work without producing a desktop session, and passive screen time, an offline meeting captured through a calendar, and paid hours are also different measures. Furthermore, switching apps is easier to observe than switching attention within an application.

Ultimately, he noted, those judging the value of AI need to establish what happened after the first task became faster. “An improvement per paid hour can conceal effort left off the timesheet,” said Gogia. “The [enterprise] board should be able to say whose day improved and whose work made that improvement possible.”

This article originally appeared on CIO.com.

Microsoft ended mainstream Windows 10 support a year ago, but PC upgrades to Windows 11 have ground to a halt, analysts said.

Relief may not come until 2028 or even 2029, as high CPU prices, which analysts call “chipflation,” and high memory prices, widely known as the “RAMpocalypse” or “RAMageddon,” keep PC prices up.

PC prices were already high, and recent steps like Intel raising CPU prices are increasing the base cost of PCs, said Dean McCarron, principal analyst at Mercury Research.

“When prices are higher, demand goes down. It causes a deferred upgrade cycle,” McCarron said.

Microsoft’s Windows revenue from PC makers fell 27% in its latest quarter, missing expectations, the company’s chief financial officer Amy Hood said on the company’s earnings call last week.

Last year’s rush to upgrade from Windows 10 boosted PC demand, but sales have since slowed, Hood said. PC makers, which pay Microsoft for a Windows license on each PC sold, are also still sitting on unsold stock, Hood said.

StatCounter numbers also show a steady number of PC users hanging on to Windows 10 throughout most of 2026. In September, Windows 10 was still running on 27.7% of Windows PCs worldwide, slightly up from 26.45% in February. That coincided with the PC market crash due to high RAM prices, uncertainty around tariffs, and the Iran war.

Windows 11’s share of Windows PCs jumped from about half in October 2025 to 72.57% in February after mainstream Windows 10 support ended, then dipped to 71.5% in September. (Windows 10 extended security updates are available for both businesses — for a hefty fee — and consumers.)

For enterprises, upgrades are more complicated than making decisions based on price alone, said Neil Shah, vice president of research at Counterpoint Research. Companies in a Windows 10 end-of-support replacement cycle face a conundrum: whether to invest in expensive AI PCs, Shah said.

“All these trends are deflating the demand slowing down the replacement cycle,” Shah said.

Buyers are being prudent about what to buy and when, waiting for better deals and lower prices, which won’t happen anytime soon, Shah said.

A rush to buy PCs ahead of memory-driven price increases boosted PC shipments by 3.2% in the first quarter of 2026, according to Counterpoint. But PC shipments declined 4% in the second quarter and 21% in the third quarter.

Commercial PC replacements and Windows migrations provided some stability to the market, but “was insufficient to offset the broader market contraction,” Shah said. “We…expect a 27% YoY decline in Q4 2026, highlighting the tough market.”

Analysts were split on when PC prices will stabilize, but they agreed it won’t be 2027. Memory prices will remain inflated as long as most of the supply is directed to data centers feeding the AI boom, Mercury’s McCarron said.

Intel in its recent earnings call telegraphed that its PC chip business will remain soft through the rest of 2026 and into 2027. Its lowest-cost PC chip, Wildcat Lake, is sold out.

Supply constraints for memory and PC parts may last through 2029, said Steven Dickens, principal analyst at Hyperframe Research.

“PC vendors are combating this through restricted configurations and bundles,” Dickens said.

AI isn’t yet driving PC upgrades, which are mostly following a typical cycle, Dickens said. But high-end machines capable of running LLMs on-device are starting to shift the market.

“We are seeing power user devices for local LLM deployment becoming a material part of the mix and changing the unit price average as a result,” Dickens said.

At some point, enterprises will have to upgrade PCs, and the market will also turn around, McCarron said. If a soft PC market goes on for too long, “it will turn into a larger boom perhaps a year later,” McCarron said.

“It’ll result in a larger PC upgrade cycle down the road,” McCarron said. “It’s like compressing a spring. The longer you hold it down, the harder it snaps back.”

After many months of discussion, Apple has figured out a way to bring Apple Intelligence to China, working with Alibaba’s Qwen model on some features. Apple Intelligence will soon be made available as a developer beta and will ship to Chinese customers in December with iOS 27.3.

China’s Ministry of Industry and Information Technology approved the use of Alibaba AI models with the iPhone in October 2026. Confirming the complexity of bringing these tools to China, Apple also received clearance in July from the Cyberspace Administration of China. Apple watchers are at liberty to reflect that the news comes at the best possible time for the company, which continues to grow smartphone market share in China. 

Apple Intelligence gets behind the iPhone Duo

The promise of integrated AI support in the form of Apple Intelligence can only boost interest in the company’s recently introduced iPhone Duo. Recent Counterpoint Research data claims iPhone sales increased 29% year on year in China in the three weeks following launch of the 18 Pro series devices. Initial stock of the devices sold out in hours. 

China is also the world’s biggest market for foldable devices, which prompted some enterprising individuals to attempt to resell purchased iPhone Duos for far above their store price — scalpers have since reduced prices, because even in iPhone-crazy China, people weren’t willing to pay so much.

While it’s a big step for Apple AI in China, it’s not quite the great leap Apple may have wanted to put in place. 

Siri AI sits this one out

Siri AI is not supported and won’t be “anytime soon,” according to market speculation. While little is said about the reasons for that, it seems probable that political pressure in both the US and China may be partly to blame, as both nations see control of the means of AI production as essential for the future success of their differing ideologies. Apple has had to allow the Chinese government to approve ongoing updates to the suite.

Chinese users can use Alibaba’s Qwen in Siri, writing tools, and visual intelligence, including functions related to text and image understanding. They don’t get Image Playground, but can expect:

• Writing tools
• Clean Up in Photos
• Notification summaries
• Email summaries and smart replies
• Priority Notifications
• Visual Intelligence
• Live Translation
• Text to Siri

Qwen if you want to 

Apple and Alibaba also promise to maintain privacy. Customers will be able to choose if they wish to proceed with making a request through Qwen. Alibaba also said it would not use these requests for model training, though it isn’t yet clear if requests are going to Private Cloud Compute in Apple’s Chinese AI deployment.

Will the introduction of Apple Intelligence in China make a difference? It probably already is, as months of ever stronger speculation mean most Chinese consumers knew Apple AI was coming to their iPhones even as they purchased them before the announcement took place. 

While the initial sales rush on Apple’s new devices among Chinese consumers has slowed, the news may help generate fresh enthusiasm for the iPhone in China. That’s great for Apple, but not such great news for some smaller smartphone vendors, who continue to be hammered by vicious increases in memory and other component prices.

A critical catalyst at a critical time

Speaking almost two years ago when industry observers had anticipated an Apple Intelligence launch in China would take place earlier, Morgan Stanley then suggested the Alibaba partnership would be “a critical catalyst for Apple’s competitive standing in China.” 

Other than painful price increases, not much has changed, and the introduction can only help consolidate the iPhone’s appeal in China. That’s good for Apple’s customers there, and great news for Apple’s investors everywhere else. It may even do something to retain a little US soft power.

In the background, we see the continued economic pressure wrought by fast-increasing memory prices. This has driven smartphone makers to raise prices, which has itself driven a 21% year on year decline in sector sales. 

Apple and Samsung have proved more resilient to these price hikes than others, which means the added appeal generated by the launch of Apple Intelligence in the world’s second-largest economy is likely to make a highly significant difference to its market share in the short term, as well as shoring up share in future replacement cycles. 

Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon.

One of the great things about Apple device management solutions is that the market for them is growing. Another is that because no two enterprises are alike, there is scope to bring different offerings to the market. I spoke with Ben Greiner, co-founder and COO of Smplify, a new-to-market device management vendor, to find out what his company provides to IT.

Smplify, which just announced $1m in pre-seed funding, offers what could be described as a headless management tool: everything it does is available through the API and CLI. In use, that means you can build it and wield it from within the tools you already use, Greiner explained. 

He’s building the company with Preetham Gurram, known in device management circles for his work at MobileIron, SOTI, Automox, Addigy, Scalefusion, and Apple, where he contributed to Apple’s core MDM platforms and zero-touch enrollment at scale.

So, who needs this?

“Device management still assumes a person is watching,” said Greiner. “Admins already automate with scripts and pipelines, and now AI agents are showing up that can act on a fleet at machine speed.”

The challenge for IT is that AI is beginning to take actions, rather than suggest them. That means IT must put systems in place to manage and audit the actions AI takes. “Enterprises that win won’t be the ones with the smartest agents; they’ll be the ones that can say exactly what each agent is allowed to do and prove what it did,” he said.

That’s a fair assessment of how things stand. OneTrust’s 2026 AI-ready governance report found that 87% of organizations are encouraging the use of AI agents, but only 47% have clear governance and controls in place. That’s a problem waiting for a disaster.

A recent report from Gartner predicts that 40% of enterprises will demote or decommission autonomous AI agents in the coming months (“by 2027”) because of governance gaps detected only after incidents take place. Patching those gaps is expected to become a big business.

Who watches the watchmen?

Once agents can submit changes at lightning speed, who can approve them? How widely deployed can those changes be? The risk is “an agent with the right key can wipe thousands of laptops before anyone notices,” Greiner told me.

This is a matter of AI governance and is something I know is in the thoughts of leaders at other device management companies, including Jamf and Fleet.

“AI is happening whether organizations know it or not. That’s the problem. You can try to block it, but that’s very hard to do well. It’s far better to build visibility and governance around it,” Jamf CEO Beth Tschida told me earlier this year. Jamf launched its own AI Governance suite in July, focused on AI tool discovery, policy enforcement, and reporting.

The Smplify approach puts some of these decisions into the management workflow itself. “Preetham calls it the last-mile control plane: the final checkpoint between what an agent decides and what actually happens on a device,” Greiner said.

What are the risks of unmanaged AI? 

While most of the conversation is focused on the risk of data leakage or malware, the other risk is lack of accountability. What if an agent gets bad instructions or hallucinates an inappropriate solution to a problem, and then pushes a destructive change across the fleet? By then it’s too late to figure out why it happened and becomes a potentially costly damage limitation exercise.

Fleet CEO Mike McNeil pointed out that part of the risk lies in the difficulty and cost of recovery after an incident. “I’d frame risk by two things: how much privilege the code runs with and how hard recovery is,” he said, noting the risk of scripts running at root, which can take full control of the machine to the extent that reversion can’t undo what’s done.

Greiner thinks the way to minimize the impact of such damage is to create strong rules and boundaries, and to ensure any risky actions are held back pending human approval. The agent can see a problem and create a solution, but can’t apply that fix until IT gives it the go-ahead. It’s very like how human IT works, only now it’s the machine that must report to the boss. Every change is logged with who asked for it, the intent behind it, and the person or policy that approved it.

Creating the opportunity space

Smplify has built its own platform to help manage devices, including AI management. That platform leans into the native protocols of the platforms it supports, naturally including Apple’s Declarative Device Management (DDM). 

“Admins describe what should be true of their devices, and Smplify keeps it true. If a device drifts from that intent, because a setting changed or an update was missed, the platform notices and brings it back in line, automatically or with an admin’s approval, depending on the rules they’ve set,” Greiner said.

He explained how many IT teams use GitOps to “bring discipline” to device management: configuration lives in GitHub, changes go through pull requests, and there’s a history of what happened. It works, but it adds new layers to the management task. He claims his solution has that discipline (versioned intent, review, and approval, and a full history) built in.

Boundaries, Apple and DDM

In the end, I get the feeling that AI management will embrace some of the ideas put in place with Apple DDM. If you think about it, DDM lets IT define a state and leaves it to the tech to work within that definition, and that same sense of boundary setting may be part of the future for AI in IT. AI will be given agency within limited boundaries.

Greiner pointed out, “Apple has also been giving admins controls over Apple Intelligence features, and I expect that to keep expanding as businesses want a say in which AI runs on company devices. Apple’s privacy stance also suits governed AI well: do things on-device where possible and keep a clear line around what leaves it.”

There is one more thing. Because Smplify is an MDM platform while also being API-first, Greiner explained that software companies can offer native device management inside their own products, under their own name, using his company’s solution. 

Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon.

We talk to chatbots as if those conversations are private. But are they?

Most people are confused. A survey commissioned by DuckDuckGo and published in August asked 1,944 American adults about chatbot privacy, and more than half (53%) did not know whether their conversations were being used to train the bots. (They are.)

Only a quarter of respondents knew that chatbot conversations can be subpoenaed. (They can.)

Among self-described AI enthusiasts, 56% said they had confided to a chatbot something they had never told a close friend, parent, colleague, doctor, or therapist.

That’s especially true on Character.AI, a chatbot service where users can chat and form “relationships” with customizable characters, including ones they create and share with the community. On Character.AI, some users chat about thoughts of suicide and self-harm, and they surrender private details like mental health disclosures, ages, private chats and voice calls, deep fears, and relationship struggles.

Character.AI was all fun and games until earlier this month, when the company filed for bankruptcy and opened a court-supervised sale of its assets.

Nobody knows who will end up with the company, or how much of the data on its millions of account holders will be sold as well.

The most valuable asset for sale by the company may be its LLM, which has been substantially trained by users as they build and interact with characters.

Legally, a company can sell personal data about its users if its privacy policy permits the transfer. Character.AI’s privacy policy is spectacularly vague. It says the company may disclose “certain information” in the company’s sale. Nobody knows what that means.

The US Trustee Office, an arm of the Justice Department, wants the judge to appoint a consumer privacy ombudsman to advise the court on user privacy. The judge could restrict the sale of company “records,” including user data.

But here’s the part most people don’t understand: The Bankruptcy Code has no provision specifically addressing trained models. The legal status of LLMs and their training data has yet to be settled in law. That means it’s entirely possible that the data users fed into Character.AI will be sold to whoever buys it.

Character.AI is just one handy example. The larger risk is all the other AI companies out there.

More surprising facts about LLM privacy

The core question is this: Can personal or proprietary information shared in a chat end up inside an LLM, where anyone might coax it out?

The answer is a resounding “uh, maybe.”

First of all, it depends on the AI chatbot service. The companies vary wildly in how they handle chat data, and the differences boil down to four questions: whether training on chats is on by default, whether human reviewers read them, how long they are kept, and whether they feed advertising.

Major chatbot companies generally try to reassure users about privacy. They publish privacy controls, but their own policies admit that consumer chats may be used for training by default, read by human reviewers, or both, and the fine print shifts from company to company.

Tellingly, some of them offer enterprise or business accounts that handle chat sessions privately. For example, they don’t use chat sessions in training data. That business accounts get extra privacy features sounds to me like an admission that individual accounts aren’t private.

Some companies, such as OpenAI, let users opt out of having their chats tossed into the training data. But OpenAI’s help documentation says that turning off training does not delete the chats themselves.

There are three main ways chatbot user data can be exposed. The first is bugs. In March 2023, a bug lurking in an open-source library called “redis-py” let some ChatGPT users see other users’ chat titles and, possibly, the first message of new conversations. More than 1% of ChatGPT Plus subscribers active in a nine-hour window may have had their payment details exposed to other users: names, email addresses, payment addresses, card types, the last four digits of card numbers and expiration dates.

The second is sharing. Some chatbots offer sharing features, and those features can publish chats to the world, whether on purpose or by accident. Once these sessions are posted, search engines are likely to gobble them up and make them searchable. In 2025, Forbes reported that Google’s results showed more than 370,000 Grok conversations after users clicked “share,” and shared ChatGPT chats turned up in search results until OpenAI killed the option.

The third is the absorption of chat sessions into the training data.

Let’s take one simple example: an email address. People often paste in emails so that a chatbot will write the reply. These emails often contain names, email addresses and other information in the header, the signature, or both.

The question I often ponder is this: If emails addressed to one person get pasted into chats hundreds of times, can a stranger ask the chatbot for that person’s address and actually get it back?

In a 2022 study, GPT-Neo models given only a person’s name recovered the right email address no more than 1.24% of the time, but with the domain supplied, the rate climbed to 16.55%, and to 37.06% after five examples. Other factors, including how often an address appeared in the training data and whether surrounding email text came along with it, pushed the rate higher.

Deleting chat records isn’t a remedy, either. Several state attorneys general said deleting chat records may not undo their influence on trained models. And Common Sense Media’s Robbie Torney claimed that exporting or deleting your data “may not mean that that data is gone from model training.”

Note that people not only enter personal information about themselves during AI chats, but about their co-workers, friends, and relatives.

The main thing standing between your secrets and the world is company practice. We depend on the AI companies to keep our information private.

So what happens when those companies sell their assets?

When AI bubbles burst

Online chatter about the “AI bubble” bursting is pretty much constant.

In China, the AI bubble has already burst, according to Reuters. Over the past few months, Chinese AI stocks have been pummeled.

Reuters reported that Shanghai’s STAR50 index of tech stocks fell 40% from its July high in under four months, erasing about $480 billion in market value, and that the broader CSI300 is down about 20% from its June highs.

It’s not a total collapse, but it does put pressure on US stock prices and also provides an incentive to transfer ownership through sales or mergers (though this hasn’t happened in China yet).

Still, the future is unknowable. Which AI companies will go bust as Character.AI did, get sold off, or get hacked?

As for the chatbots themselves, their inner workings are also unknowable.

That’s why we should all listen to the most paranoid amongst us, and never put sensitive information into chatbots. We just can’t trust AI to keep our secrets.

Everyone wants to know what Anthropic is up to, and hackers are gleefully exploiting the fact.

New research from cybersecurity company Cofense describes a phishing campaign that hooks overeager enterprise users by making them think they are getting an invite to a private invite-only Anthropic preview.

In reality, adversarial actors are trying to steal Google and Microsoft login credentials, exploiting video teleconferencing software GoTo Webinar branding to do so.

“It sells exclusivity,” said Frank Dickson of Dickson Research. “Fear of missing out is the oldest lure in the book.”

Every organization is trying to figure out where AI fits, he noted, and a private preview of Anthropic’s newest technology is exactly the invitation an ambitious employee would want to accept. “The attackers did not pick Anthropic at random,” he said. “They picked whatever is hottest in the news cycle. Next month, it may be another name.”

How attackers lure users with spoofed Anthropic invites

GoTo Webinar is a widely used platform, and its meeting invitation emails are typically not blocked by security tools. The Cofense researchers found that attackers are exploiting this fact by sending what seems to be an invitation email from a legitimate sender; when users click on the link to register for the fake webinar, they are instead sent to a compromised URL.

The email offers what the researchers call a “tantalizing” offer: “This private, invitation-only virtual session brings together a select group of participants to a focused conversation on AI systems, research, and the future of frontier models,” it reads.

A URL shortener circumvents email security tools, and the attackers use subdomains and dashes (-) to hide the malicious URL. The link itself appears to have been generated by Zapier, a legitimate company whose tools are designed to work with Claude, ChatGPT, and other AI compatible with Model Context Protocol (MCP).

The email sender address appears to be customercare@gotowebinar[.]com (GoTo Webinar’s website is actually www.goto.com). The text contains no obvious spelling errors, and the email contains GoTo Webinar branding. However, the researchers noted, the email sender’s name is “Austin from nthropic,” with the “A” missing.

“Although this is a small misspelling, it is still a useful warning sign,” they pointed out.

If the phishing link is clicked, a Cloudflare CAPTCHA page appears, and must be completed before the login prompt is presented. This serves as the first of two checkpoints to prevent security tools from scanning attackers’ websites.

“Because CAPTCHAs are common on legitimate websites, phishing recipients may believe all is well and everything is secure,” the researchers explained.

Once the CAPTCHA is completed, victims are led to a Zoom-spoofing invitation page. If they want to “attend” the fake event, they must register, and they are prompted to enter their work email address. When they click “Continue with Google” or “Continue with Microsoft,” yet another CAPTCHA page appears.

Both CAPTCHA pages seem to be good at detecting and blocking analysis by VPN services, scanners, Tor browsers, and other security tools, the researchers noted. Also, because there are two distinct CAPTCHA pages, such tools may not be able to follow every step in the attack.

“A CAPTCHA was built to keep bots out,” Dickson pointed out. “Here, it keeps the security scanners out and lets the victims in.”

The fake Google and Microsoft login pages differ a bit, however. The spoofed Google page does not capture passwords, only usernames, which the researchers surmise is due to a misconfiguration; the spoofed Microsoft page does collect both usernames and passwords.

But both are “remarkably similar” to their respective legitimate pages, and their subdomains are “tricky.” The fake Google URL is goo[.]gle-access-eventregister[.]icu, while the Microsoft URL closely mimics real Office and Sharepoint domains, using office[.]com-access-sharepoint[.]com.

Once credentials are submitted, neither login offers a confirmation message, nor does it redirect to legitimate vendor pages. But by that point, attackers have already succeeded at exfiltrating credentials. That means they could compromise an account, access the sensitive documents in that account, or use it as an initial attack vector to further infiltrate the enterprise, the researchers explained.

They advised checking if the branding of a message remains consistent from the email to a login page, searching for small inconsistencies such as minor misspellings, and reminding employees that both the URL and domain determine whether a login page is legitimate.

“Although the logos and design of the page may look correct, threat actors are skilled at mimicry,  especially now with the use of AI,” the researchers noted.

It’s not the employees’ fault

The AI industry’s habit of gated previews and waitlists does not help in these situations, Dickson noted; it trains people to expect exactly this kind of invitation.

“Start with the cracks the attackers left showing,” he advised: The sender supposedly came from “nthropic,” the email carried GoTo Webinar branding, the registration spoofed a Zoom meeting page, and the login pages were from Google or Microsoft. That’s three brands in three clicks, he noted.

The fake login domains, like office[.]com-access-sharepoint[.]com, are built to be skimmed, not read, he pointed out. And, while the Google page collected only usernames, which may be a configuration error, the threat actors behind this campaign will likely fix that.

“Training helps, but I would not hang the defense on an employee squinting at a URL at 4:45 on a Friday,” Dickson said; how much a stolen password is worth is a leadership decision, not a test the employee fails.

Phishing-resistant multi-factor authentication (MFA) such as passkeys makes a stolen password far less valuable, he noted. But enterprises ultimately have a short to-do list to allow them to defend themselves from this sort of scam, Dickson said:

• Verify any AI preview invitation through an account team or a known contact.
• Treat branding that changes from email to meeting page to login page as a red flag.
• Roll out phishing-resistant MFA. If someone entered credentials on one of these phishing login pages, reset them and review that account’s activity.

“Bottom line: This campaign uses trusted tools to defeat defenses built on trust,” Dickson said. “The employee should not be the last line of defense.”

This article originally appeared on CSOonline.

Following the release of the new 14-inch M6 MacBook Pro earlier this week, Apple has bought the “touch experience” to its professional laptop series, introducing a $2,899 (and up) MacBook Pro you can use with touch. It’s very fast, very thin, very light, with an impressive PROLED XDR display and up to 24 hours battery life.

A brief tour through the key specs

• Very fast: Available in 14- and 16-inch models, the new MacBook Pro uses Apple’s powerful M5 Pro and M5 Max chips. They support up to 64GB of unified memory (128GB on M5 Max) and up to 16TB of storage. Apple says the M5 Max can run LLMs with hundreds of billions of parameters entirely on-device. 
• Very thin: The 14-inch model is 0.59 inch (1.50 cm) thick, while the 16-inch is 0.63 inch (1.60 cm), so these are only a little thicker than a MacBook Air. Apple says the thinner design is a benefit of the move to an OLED display.
• Very light: The 14-inch weighs 3 pounds, half a pound less than the non-touch equivalent MacBook Pro. The 16-inch weighs 4 pounds, which is 0.7 pounds lighter. 
• Industry-leading PROLED XDR display: A tandem OLED panel with up to 2,000 nits peak brightness (1,000 nits sustained brightness) for both HDR and SDR, and a 10,000,000:1 contrast ratio. Refresh rate is variable from 10Hz up to 120Hz.
• Up to 24 hours of battery life: Apple says this is the maximum and that battery life varies by use and configuration. It also varies by model — the 14-inch yields “just” 22 hours battery life. Fast charging takes the battery to 50% in 30 minutes.

Unpacking those numbers, it’s crystal-clear Apple has not stinted on the performance of this pro Mac, with a price tag to match.

IDC data shows that 66% of all touch notebooks are purchased by enterprises. This is a form factor bought for flexibility for field workers, educators, and professionals who move between typing, tapping, and handwriting.

Touch on a Mac

Which brings me to the “touch experience.” Apple hasn’t replaced the traditional Mac interface with touch; it has instead augmented the core OS with it. Swipe to scroll a web page, pinch to zoom a photo, drag-and-drop slides between presentations, double-tap a window to fill the screen, or flick a window to tile it into place. You use touch along with trackpad gestures and keyboard shortcuts, the idea being that touch augments them.

Apple built this integration into macOS 27. It’s a physical thing. The touch sensor is built into the display panel rather than being layered above it. In use, Apple says it reduces latency, making you feel like you’re touching the content.

Apple

The other point to this integration, according to Apple, is that many existing applications will automatically work with touch as the macOS translates on-screen actions. Apple has also built APIs developers can use to create apps with touch. The company told us Microsoft Word, Excel, and PowerPoint; Adobe Lightroom; DaVinci Resolve; Blender; and Cinema 4D will be touch-enabled later this year.

At this point, Apple is evangelizing how touch can be used by creative professionals for things like playing virtual instruments, drawing and fine control of creative apps like Logic Pro or Final Cut Pro. As I see it, it’s as if the Vision Pro was on a Mac, with touch as the interface to the control plane.

It’s an integrated platform thing

When you touch your MacBook keyboard, your display wiggles a bit, but not with this Mac, as Apple has built a new hinge mechanism to keep it steady. Prior to the event, online speculation claimed this hinge was developed alongside that in the iPhone Duo.

Another borrowing from the iPhone is the inclusion of a Dynamic Island on the Mac. Spotlight, Live Activities, and Siri AI all live right there, and you can interact with them on the Mac.

The PROLED XDR display is worth an article in its own right — as is so much about this Mac. As well as built-in support for touch, it also supports Adobe RGB, P3, and DICOM medical imaging reference modes. It supports genlock for virtual production. It boasts a 10,000,000:1 contrast ratio and a variable refresh rate from 10Hz up to 120Hz. For an even better experience, there’s a nano-texture glass option on the screens, which are 14.2 and 16.3 inches diagonally. Following its first look, Tom’s Guide has already called the display “jaw-dropping.”

You also get what Apple called the best audio system in a laptop (six speakers with an all-new woofer design) and plenty of interconnects.

Apple The touch-screen laptop debate is over

“Apple is entering a segment that is gaining share in a shrinking market. IDC expects touch to reach one in five notebooks by 2030, and Apple has just made sure that buyers think about touch as a must-have,” said Francisco Jeronimo, vice president, worldwide devices at IDC in a note shared with Computerworld.

“When the company that spent 15 years arguing against touch concludes that it belongs on a notebook, the debate about whether touch is a gimmick is over, and the question becomes who builds the best implementation,” he said.

The new Macs are available for order now and will be available from November 6. They aren’t cheap — the 14-inch model costs from $2,899 with the M5 Pro chip and from $4,499 with M5 Max. The 16-inch model costs from $3,399 with M5 Pro and from $4,799 with M5 Max.

Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon.

OpenAI and Anthropic agree that frontier AI systems need stronger oversight, but they are proposing sharply different approaches to stopping models that pose serious risks: Anthropic wants courts empowered to suspend dangerous systems, while OpenAI argues developers should be required to maintain and use effective shutdown controls.

The proposals appear in separate submissions to California Governor Gavin Newsom following his September 18 executive order directing state officials to consider additional safeguards for frontier AI systems. Both companies support independent safety assessments and expanded reporting of incidents involving AI systems that act beyond their intended permissions.

“If the AG concludes, informed by the independent evaluation, that a frontier AI company’s models pose a near-term risk of catastrophic harm, they should be able to seek a court order to suspend a model’s deployment, internal use, or further development,” Anthropic wrote in its October 15 submission.

OpenAI, in its October 20 letter, called for an amendment to California’s Senate Bill 53 (SB 53).

“We support amending SB 53 to require frontier developers to maintain effective shutdown controls and an enforceable duty to use them when unsafe conditions arise, and no lesser intervention would adequately address the risk,” OpenAI wrote.

Sanchit Vir Gogia, chief analyst at Greyhound Research, said the two approaches address complementary needs. “The operator must contain the unsafe activity promptly. Independent review must examine both intervention and inaction,” he said.

Newsom’s executive order, N-9-26, directs California’s Government Operations Agency, in consultation with the Governor’s Office of Emergency Services and national experts, to submit recommendations by November 16 on potential amendments to existing AI safety laws. The order asks officials to examine independent verification inside AI laboratories, independent checks of safety assessments, verified shutdown mechanisms, and expanded reporting of critical safety incidents. It does not itself impose these proposed requirements.

Independent verification

Anthropic wants evaluators to assess developers’ safety claims and publish their own findings.

“We strongly support requiring independent verification of what frontier AI companies report about the risks of their models and whether those risks are sufficiently mitigated,” the company wrote.

It proposed requiring developers to publish risk reports at least every six months, covering serious risks from their models, including internally used models substantially riskier than released systems. Evaluators should receive unredacted reports and access to developers’ most capable models and relevant internal information.

Anthropic also proposed that developers pay into a state-administered pooled fund to finance evaluators, rather than pay them directly. It said a team from the independent evaluation organization METR had received employee-like access on a pilot basis, with further partnerships involving SecureBio and Accenture under development.

OpenAI, in its letter, said it supports independent technical assessments but distinguishes them from compliance audits. Technical assessments would examine evidence behind developers’ safety claims and the effectiveness of safeguards; compliance audits would check whether companies follow their published safety frameworks and legal obligations, it added.

OpenAI recommended annual independent compliance audits and focused technical assessments of serious risks arising during model training, evaluation and internal deployment. It also urged California to work with the National Institute of Standards and Technology to develop common assessment standards.

Gogia said oversight requires more than funding. “Credible assurance must survive an adverse finding,” he said.

Beyond a single switch

OpenAI argued that a single kill switch would not address the different ways AI systems operate and are deployed.

“We recommend using the term ‘shutdown controls’ because it will require the development of multiple controls and coordinated action by different actors to stop AI operations when necessary, rather than a single switch,” the company wrote in the letter.

Interventions could include stopping a training run, isolating an AI agent, revoking tool or network access, or suspending a service. OpenAI said responsibilities should extend to inference providers and downstream deployers according to the operations they control. It also noted that developers cannot reliably switch off independently operated copies of open-weight models running on external systems.

Anthropic similarly argued for more targeted safeguards. “We think a more apt metaphor for what frontier AI companies need to build is a circuit breaker,” it wrote, advocating controls that can contain problems at the level of an agent, customer, deployment environment, or model.

For enterprises, Gogia said, contractual shutdown responsibilities should match the controls each party actually possesses. Agreements should specify intervention authority, incident-reporting cooperation, and who can authorize a restart.

“Every assigned duty needs an executable control,” he said, adding that CIOs should test recovery processes rather than assume switching to an alternative model will preserve safe operations. Both companies also called for broader reporting of incidents involving models that circumvent safeguards or take unauthorized actions.

Otter has added a new Work mode to its AI Chat assistant, expanding the range of tasks it can carry out before and after meetings.

It’s part of a wider drive by Otter — which received $320 million in funding last week — to move beyond meeting transcription and into automating tasks and workflows.

The Work interface, announced Tuesday, is accessible in the AI Chat prompt box, alongside the Chat option. While Chat allows users to ask questions, generate content, and retrieve information, Work is aimed at multi-step actions, Otter said. It’s a similar approach to other AI assistants, such as Claude and ChatGPT, that separate simpler chat functionality from newer agentic capabilities.

Prior to a meeting, Otter users can prompt Work to generate briefing documents based on information collated from previous calls and data from connected applications. After a meeting ends, the AI assistant can draft emails, generate documents — including Word and Notion docs, PDFs, and interactive web pages — and update Salesforce records.

The AI assistant requires human approval before completing tasks, an Otter spokesperson said.

The capabilities expand on functions already available with AI Chat’s existing Advanced mode. Otter will now remove the Express and Advanced options from the AI Chat interface.

Work mode is available now to all paid Otter users. There are no usage limits, Otter said — this differs from the Chat mode, which allocates a certain number of queries each month depending on a customer’s payment plan.

The feature updates reflect how meeting note taking is rapidly moving from data capture to workflow automation, said Irwin Lazar, principal analyst at Metrigy. “That is, vendors are delivering solutions that allow meeting notes to be actionable, based on context both from within meetings, as well as from other connected platforms,” he said.

Lazar pointed to similar functionality emerging in other meeting systems such as Microsoft’s facilitator in Teams and Cisco’s Webex meeting agent.

The vast majority (91%) of business users want meeting note-takers to access and act on context from outside a meeting, according to a global study of 403 organizations in Metrigy’s “Workplace Collaboration & Employee Experience: 2027” report.

In addition to the Work mode, Otter announced that its Live Assist feature is now available to enterprise customers, and will expand access to all paid Otter accounts in the coming weeks. Live Assist, announced in April, provides private guidance to a user during a call, such as suggesting questions and tracking meeting “objectives.”

“I see this kind of widget assist a lot in external-facing customer support roles, but for internal productivity meetings, this is quite unique,” said Thomas Randall, a research director at Infotech.

However he said that Otter faces from other meeting software vendors which are also integrating AI capabilities into their platforms. “So I don’t see Otter becoming a center of work, though its new functionality will be of great assistance to pre-existing users,” said Randall.

Related:

• Cisco brings collaborative agents and Claude into Webex chats
• AI notetakers at work could leave companies at risk for lawsuits

A new study from the Coalition for Language Equity in AI Research shows that AI models require significantly more tokens to process certain languages than others. These differences mean that users of underrepresented languages risk paying more for the same services, reports Rest of World.

An LLM’s tokenizer breaks text down into blocks, or tokens, but it does so much more optimally for languages with abundant digitized data, the researchers explain. They found that among the 20 languages with the largest web presence, less common languages used from 1.5 to 5 times as many tokens as the same prompt in English. When the study was expanded to 200 languages, the difference increased to as much as 17 times in one of the models examined.

Since most LLM providers charge by tokens used, that translates to much higher costs. In addition, more tokens mean longer response times and a smaller amount of information that’s able to fit into the AI models’ context windows.

The researchers describe this as a hidden “tokenizer tax” and warn that it exacerbates digital inequality.

This disparity is not limited to languages with few speakers, the researchers note. “If we sort by speakers instead of web presence, the tax lands on enormous populations: Punjabi (~200 million speakers), Bengali (~266 million), Urdu (~288 million), Burmese (~36 million), Amharic (~36 million). This means hundreds of millions of people billed at multiples of the English rate for the same intelligence,” the study says.

Languages that use writing systems not as commonly found in the AI models’ training data are particularly affected. Examples include Burmese, Tibetan, Khmer, and Lao. Languages that use the Latin alphabet, on the other hand, often have an advantage, even when spoken by relatively few people.

This article was originally published on Computer Sweden.

Related: CIOs grapple with subpar global genAI models

Decades ago, the United States established programs that allow businesses to hire highly skilled foreign workers — the H-1B visa program and a related green card program. They’re designed to make sure we lead the world in technology. Not uncommonly, H-1B workers and green card holders eventually become American citizens.

One of the best-known examples is Microsoft CEO Satya Nadella. He came to the US from India on the green card program, then switched to the H-1B visa program in the early to mid-1990s when working for Microsoft. He ended up becoming a US citizen. In 2014, he was appointed Microsoft CEO and subsequently architected one of the biggest turnarounds in tech history. Under his leadership, Microsoft has become nearly a $4 trillion company.

President Trump, who rode an anti-immigrant message to the White House two times, has both the green card and H-1B programs in his crosshairs — and he’s aiming right at Microsoft.

Earlier this month he claimed Microsoft, Adobe, and several other companies had abused an important part of the green card program called PERM, which helps H-1B visa holders get green cards. He’s suspended their use of it. Microsoft fired back that Trump was off target and the company has done nothing wrong.

Who’s right and who’s wrong? We’ll follow the facts to find out.

Trump’s and Vance’s fact-free charges against Microsoft

A green card allows a foreigner to stay and work in the US permanently. Companies often sponsor people for green cards, although there are other ways someone can get one as well. Once someone has a green card, they can switch to work for a company other than the one that sponsored them. Getting a green card is often a first step for someone to become a US citizen.

An H-1B visa, on the other hand, is temporary and allows a skilled worker to stay in the US for three years, which can be extended to six years.

The two programs are closely related. Not uncommonly, someone on an H-1B visa can get a green card by having the company sponsoring them apply for it through the PERM program.

PERM is what Trump targeted earlier this month. He suspended Microsoft, Adobe, and several other companies from the PERM program so they can’t get their H-1B workers green cards.

Vice President Vance claimed Microsoft had “abused” the system by laying off American workers and replacing them with lower-cost foreign workers. He said in a news conference that there was “no company in the United States, unfortunately, that has abused this system more.” 

Vance contends that Microsoft laid off 6,000 American workers in the past year and replaced them with much less expensive foreign H-1B visa and green card holders. He called the foreign workers “indentured servants.”

Vance concluded, “If you do the math, for every worker that Microsoft laid off, they replaced that worker with one and a half foreign indentured servants. Now that is a scandalous system.”

That’s pretty damning, if true. But as so often is true with the Trump administration, the charges bear no relationship to the truth. For one thing, Vance’s numbers include Microsoft’s use of the H-1B program as well as the green card program.

It’s not just PERM and the green-card program that Trump has been attacking. Since his first election, he’s targeted the H-1B visa program as part of his war on immigrants and immigration. Last year, he imposed a $100,000 fee on companies for each H-1B visa they use, a move which a judge earlier this year ruled was illegal. His administration has appealed the ruling.

In August, the Department of Homeland Security proposed a similar fee plan, with companies having to pay $103,265 to hire an H-1B worker. The administration has also filed proposals to end work authorization for some H-1B workers’ spouses and eliminate the 60-day grace period H-1B workers have now to stay in the US after a job loss so they can look for new work.

Microsoft fights back with facts

After Vance’s charges, Microsoft immediately fired back with a statement that succinctly pointed out how wrong his claims were. The company addressed Vance’s false claims about both the H-1B and PERM programs.

Microsoft noted that “of the approximately 6,000 H-1B visa applications we submitted in the last fiscal year, 80% were to extend or change the status of existing Microsoft employees. These were not to hire new people.”

It went on to say, “the remaining filings for new employees were for individuals already legally in the United States who decided to come work for us.”

Beyond that, Microsoft’s layoffs in the past year had nothing to do with H-1B visas or green cards. Even if Microsoft didn’t have a single H-1B visa worker or green card holder, the layoffs would have gone through.

All that’s not to say the H-1B program is perfect. Some companies do hire foreign workers under the program to perform jobs Americans can do and pay them less. Disney, for example, in 2014 replaced several hundred IT staff members with lower-paid foreign workers who had no unique, highly specialized skills. In fact, the Disney workers trained the foreign workers to do their jobs.

There have also long been charges that under the program, companies that provide IT outsourcing hire foreign workers with no greater skills than US workers and pay them less. The Biden administration issued a series of rules cracking down on that. It’s not clear how effective those rules have been. When Trump suspended Microsoft and Adobe from the PERM program, he also suspended several big IT outsourcing firms. 

Trump honors foreign-born CEOs

Vance’s news conference was followed by a very different Trump news conference several hours later, offering the usual head-spinning sense of unreality so often found with the Trump administration. Trump awarded Nadella the National Medal of Technology and Innovation — an award he called “so well deserved.” No mention of the H-1B or PERM charges against the company, or that Nadella worked for Microsoft on an H-1B visa and with a green card in the 1990s.

At the same press conference, Trump awarded Elon Musk, CEO of Tesla and SpaceX, the National Medal of Science. Other foreign-born honorees included Google co-founder Sergey Brin, Nvidia CEO Jensen Huang, and AMD CEO Lisa Su.

Musk says he’s a former H-1B holder and is a big supporter of the program. He’s written on X, “The reason I’m in America along with so many critical people who built SpaceX, Tesla and hundreds of other companies that made America strong is because of H1B.”

Musk being Musk, there hasn’t been a peep from him about Trump’s latest assault on the program.

The upshot

Nothing could be clearer: Trump’s and Vance’s charges against Microsoft are lies. It’s true that some companies have abused the H-1B program, and possibly PERM, but Microsoft isn’t one of them.

Opera has begun rolling out AI models that run locally in its browser, potentially allowing enterprises to perform certain AI tasks on existing PCs without sending the relevant data to cloud AI providers.

The company said it has introduced support for WebAI, a proposed web standard, through an opt-in early preview in Opera One. The technology allows websites and browser extensions to access AI models running locally in the browser for tasks such as generating text and classifying content.

The WebAI preview supports 64-bit Windows PCs running Intel or AMD processors and Macs powered by Apple silicon. Opera does not specify a requirement for a dedicated neural processing unit, although it says performance depends on the device’s hardware.

The preview does not currently support Windows on Arm, potentially limiting its availability across some newer AI PCs.

The feature is optional, with AI models downloaded only after users enable it. Opera said website-specific permissions allow users to control which sites can access the locally running models.

Opera said its implementation supports models including Google’s Gemma and Apple’s on-device foundation model on compatible Macs. The company said it plans to introduce additional models optimized for different devices as it expands the technology.

Local AI shifts enterprise security risks

Opera’s implementation could have implications beyond its own browsers if the proposed WebAI standard gains wider adoption, potentially bringing locally running AI models to more enterprise users.

The privacy benefits for enterprises could depend on how organizations control access to locally running models, particularly through websites and browser extensions.

Pareekh Jain, CEO of Pareekh Consulting, said processing sensitive information locally could reduce enterprise security risks by limiting the need to send data to external AI providers. However, he warned that the approach could make it harder for IT departments to monitor activity on individual devices.

“It is easier for IT to monitor internet traffic,” Jain said, adding that monitoring activity on individual laptops can be more difficult.

He warned that malicious websites could potentially manipulate local AI models through hidden prompts, while rogue browser extensions could gain access to sensitive information displayed on a user’s screen.

“Local processing reduces risks but does not eliminate them,” Jain said.

Enforcing consistent safeguards across employee devices would become imperative for enterprises, according to Tulika Sheel, senior vice president at Kadence International.

“Enterprises still need controls over which websites and extensions can access local models, how information is used, and how models and browser permissions are managed across the organization,” she said.

Could browser AI reduce the need for new AI PCs?

While WebAI can run on existing computers, its ability to reduce enterprise spending on new AI PCs will depend on whether those machines can deliver acceptable performance for everyday business tasks.

Jain said smaller AI models could handle tasks such as summarization and rewriting on existing PCs, potentially reducing the need for immediate hardware upgrades.

“Complex reasoning and advanced AI workloads will still benefit from dedicated AI hardware or cloud infrastructure,” he said.

WebAI’s support for compatible models from different providers could give developers greater flexibility in choosing which AI systems to use.

Browser-based AI could also reduce enterprise spending on cloud services by lowering API usage and dependence on external AI providers. However, limitations in model capabilities and enterprise integration could restrict its usefulness for more sophisticated applications.

“The future will likely be hybrid, with browser AI handling simpler tasks and cloud AI supporting advanced workloads,” Jain said.

Sheel said enterprises would still need to account for the costs of model updates and integration with existing systems, even if local processing reduces their cloud AI usage.

“Lower usage costs do not automatically mean lower total costs,” she said.

Put on your finest propeller-packin’ thinking caps, my fellow Android-appreciating animals — ’cause I’ve got a brain-bending bit of big-picture thinking about the future of Android and everything around it.

This pondering comes out of my ongoing time living with one of Google’s new Googlebook laptops — the beautifully built Lenovo Googlebook (provided to me as a short-term loaner for editorial assessment). Right now, the Googlebook experience itself is best summed up as an awkward start with plenty of potential. Android in a laptop environment just comes with an awful lot of asterisks, especially when you’re trying to use such a system for serious work. In many ways, it feels like a step down from the more polished and productivity-optimized Chromebook setup it’s ultimately designed to replace.

But while working on a Googlebook and using it as your primary computer still seems like a bit of a square-peg situation, I’ve been struck by one area where the Googlebook OS software at the center of the experience feels almost shockingly natural.

And that’s led me down a rabbit hole of pondering about the bigger-picture implications of this operating system and how it could, at least in theory, one day affect the rest of the Android ecosystem.

Lemme explain.

[Get original Googley insight in your inbox with my free Android Intelligence newsletter — one new thing to try every Friday.]

A Googlebook by any other name

So, at its core, Googlebook OS is essentially Android — right? The software that powers Googlebooks is a specialized version of Android 17 designed to have some ChromeOS-like elements along with phone-connecting powers and a handful of feature additions that help it work better on a big screen.

And the more I use it, the more I keep coming back to a thought that I mentioned in passing in my initial Googlebook analysis — that this is what the Android tablet interface oughta be.

Think about it: Google has never really figured out how it wants an Android tablet to work — or if and how it even wants to promote and support the development of that not-quite-a-phone, not-quite-a-laptop form factor in the first place.

Early on, we saw a completely large-screen-optimized Android version called Honeycomb (a.k.a. Android 3.0 through 3.2), where Google gave us a dramatically reimagined interface designed to take full advantage of the tablet’s added screen space.

width="850" height="534" sizes="auto, (max-width: 850px) 100vw, 850px"> Android’s Honeycomb version featured an interface optimized for tablets — not so different in concept from Googlebook OS today.

JR Raphael, Foundry

It wasn’t long, of course, before Google did its Google Thang™ and gave up on that idea — and that’s to say nothing of the even more ambitious Android tablet plan the company spent millions of dollars to develop and then never brought to fruition at all. By a certain point, it became obvious that Google was for all practical purposes giving up on Android tablets and that touch-oriented Chromebooks were effectively filling that role for informed buyers.

Fast-forward through a few other confusing short-lived pivots — the very brief return to tablet devotion with the ChromeOS-running Pixel Slate, the declaration of being done with tablets that followed, and the about-face and (once more) brief return to tablet devotion with the Pixel Tablet and its admirable but never fully realized and rapidly abandoned tablet-smart-display combo philosophy — and we’re back to a present now where Android tablets are mostly an afterthought, as far as Google and the underlying platform are concerned.

With all of that in mind, using a Googlebook is a bit of a revelation — when you think about the experience as less of a full-fledged laptop replacement and more of a potential future for a revamped Android tablet.

The Googlebook-Android tablet connection

Consider: In many ways, Googlebook OS picks up where Honeycomb left off 15 years ago — as far as the interface itself is concerned. It feels like a familiar Android environment, for anyone who’s used Android on a phone, but with some thoughtful touches that make it more sensible in that larger-screen setting. Everything from having the clock in the upper-left corner alongside a split-apart notification panel and Quick Settings area in the right to having the same desktop-style taskbar we’ve seen on recent Android foldables just feels right when you frame a Googlebook as a tablet and not, first and foremost, a work laptop.

width="1024" height="640" sizes="auto, (max-width: 1024px) 100vw, 1024px">

The Googlebook desktop feels like a natural Android tablet interface.

JR Raphael / Foundry

Plain and simple, it’s Android — only optimized for a larger screen. And, thanks to its ChromeOS-inspired elements, it adds in the advantages of a less restrictive big-screen browser environment and smarter tools for interacting with your phone. The elements that feel awkward for a work laptop don’t matter as much in an Android tablet, which you’re more likely to use as a touch-centric supplement and companion to your workday computer and as a device where productivity is more of an occasional add-on as opposed to the primary purpose.

As a tablet interface, Googlebook OS gives you a souped-up version of the standard Android tablet setup that’d also be much more useful than a traditional tablet when you connect a keyboard. And without the expectation of it completely replacing your computer, it suddenly makes a lot of sense.

So riddle me this: With Android tablets being an afterthought now, anyway, what’s to stop Google from starting over — yes, again — and encouraging companies to create Googlebook tablets next?

It’d allow Google to rebrand and relaunch a category that’s always struggled to catch on and get past its not-so-exceptional early reputation. And for us, as users of these products, it’d give us some overdue options for a larger slate-like Android experience that’d work seamlessly with our phones, be pleasingly familiar and capable, and offer the unique advantage of having a mostly desktop-caliber browser for light work when such needs arise (with or without a keyboard connected) — all with the kind of timely and reliable software support promise any such product should provide.

And — oh, yes — there’s more.

Googlebook, Googlebook, everywhere

Here’s the real mind-blowing next-step possibility to consider: If we do one day have Googlebook tablets — and the Googlebook OS software at the heart of them continues to evolve and become a core focus for Google as time moves on — how much of a leap would it be from there to see things streamlined and reaching a point where we also have Googlebook phones?

It brings me back, fittingly enough, to a question I posed over a decade ago — back when Android and ChromeOS were coming ever-closer together and it looked like that was the path Google was pursuing.

Consider this now, with the notion of Googlebook OS in the mix in place of ChromeOS:

If all Android apps can eventually run on ChromeOS — and if ChromeOS evolves to look more like Android while web apps and Android apps grow increasingly similar in design — would you notice the difference between a phone running Android and a phone running Chrome?

The specific platforms and concepts at the center of this have shifted since we last thought this through in 2015, but the idea itself is the same.

And, once more, there’s no better way to end this contemplation than by saying:

Welcome to the Matrix, my friends. Whoaaa. 

Keep the creative thinking coming with my free Android Intelligence newsletter — one new thing to try every Friday and my experience-enhancing Android Notification Power-Pack as an instant welcome bonus.

Apple’s second day of launches this week saw it follow its tempting iPad mini upgrade by pumping up the processor in two of its most popular Macs, even while it injected artificial intelligence into Creator Studio.

24 hours of portable pro performance: 14-in MacBook Pro

Apple’s 14-inch MacBook Pro is popular but sits in a funny place in Apple’s notebook lineup, with the powerful and popular MacBook Air biting at its heels and the more performant and powerful (though not yet M6) MacBook Pro models sitting above.

We already know what the M6 chip is capable of, because it is already inside the Mac mini (which will be made in the US later this year). We know it delivers better single-core scores on Geekbench than the 10-core M5 MacBook Pro. We also know it beats that last MBP in multi-core. Now that same chip is available in a 14-inch MacBook Pro — a really promising dose of power and performance in a highly portable pro Mac that now packs a mighty 24-hour battery life.

The chip gives you a 12-core CPU, a 12-core GPU with Neural Accelerators in each core, a dual 16-core Neural Engine, and faster (16GB, upgradable up to 32GB) unified memory. What that means is power. And what that means is productivity — 30% faster graphics compared to last year’s model. 

For those upgrading from M1, the performance gains are even more significant, with up to 11x faster graphics and up to 13x faster AI performance. The idea is that you’ll pick up your highly portable pro portable and take it anywhere you need it to be.

Apple

What’s new in Apple’s new portable pro laptop other than the processor?

• Networking has been enhanced with Wi-Fi 7 and Bluetooth 6.
• It will support up to two external 6K displays.
• Liquid Retina XDR display with up to 1,600 nits of peak brightness.
• 12-megapixel Center Stage camera. 
• Support for fast charging.
• HDMI, MagSafe 3, three Thunderbolt 4 slots, SDXC card reader.
• Headphone jack.

Available in Space Black and Silver.

Pre-order from October 28, in stores November 6.

Apple New colors, new processors, the new M6 iMac

You don’t even need me to say it anymore, but the introduction of M6 chips inside the iMac makes the all-in-one desktop even more powerful and performant. This tops out a 12-core CPU, 12-core GPU, and Dual 16-core Neural Engine. 

Those fast processors combined with the fast, 170GB/s memory bandwidth means you’ll get twice the graphics performance of the M4 iMac. The Mac ships with 16GB of unified memory, but you can upgrade up to 32GB if you like spending money.  

Inevitably, the new iMac is faster for, well, everything. You’ll be hitting pro performance comparisons all over again as Apple silicon continues to redefine expectations across the industry, putting pro performance in reach of most people. It’s all about AI at Apple at the moment, so expect 12 times faster performance in some tasks compared to the M1, Apple said. 

Some specifications:

• The iMac retains the 4.5K 24-inch Retina display, 500 nits, and P3 wide color of before.
• It brings Wi-Fi 7, Bluetooth 6, and 2.5Gb Ethernet.
• It will support up to two external 6K displays.
• Up to four Thunderbolt 4 ports.
• 12MP Center Stage camera.
• 2x faster SSD read and write.
• Touch ID.
• Available in blue, green, pink, yellow, purple, orange, and silver.

The new iMac is on pre-sale from October 28 and in stores from November 6. It costs $1,499 with a 10-core CPU, 10-core GPU, and 16GB of RAM, or $1,699 with a 12-core CPU and 12-core GPU (16GB of RAM). 

Apple Apple leans into AI for creative customers

Of course, once you have put together the world’s leading mass-market AI ecosystem, you have to do something to show that power off. Apple has done this with a selection of interesting updates to its Creator Studio suite of creative tools. They’ll ship in November.

• Final Cut Pro for Mac now includes a large MotionVFX collection consisting of thousands of visual effects, transitions, and motion graphics tools and built-in Pixelmator Pro editing. Apple has also filled out its Content Hub with a bunch of new video assets that can be snagged from inside Final Cut Pro.

• Logic Pro gains AI, an enhanced Stem Splitter that now separates drums, a redesigned sound Browser, and a new intelligent assistant powered by a combination of Apple-built models and OpenAI.

• Pixelmator Pro introduces Generate Fill, which lets creators transform selected areas of an image using natural language.

• Keynote, Pages, and Numbers introduce new AI features in beta, including assistants that understand document context and new Model Context Protocol (MCP) support for workflow automation. They add stock video in the Content Hub, new premium themes and templates, and that intelligent assistant powered by a combination of Apple-built models and OpenAI.

Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon.

Microsoft is offering voluntary exit packages to about 60,000 employees, roughly 27% of its workforce, including staff on teams building Copilot and other AI products, as the company looks for ways to fund its AI investments.

Amy Coleman, executive vice president and chief people officer, announced the Voluntary Exit Offer (VEO) in an internal memo on Monday, which Microsoft republished on its official blog.

“This offer gives eligible employees who may be considering a future outside the company a new choice to make that transition now with support,” Coleman wrote in the memo. It also gives participating organizations “another way to adapt their workforce to changing needs,” she wrote in the blog.

The offer covers employees at level 70 and below in participating teams in the US, Canada, India, and the UK, according to the memo. “Some organizations are providing the Voluntary Exit Offer, and some are not,” Coleman wrote. She added that participation depends on each business’s “priorities, customer commitments and workforce needs.”

However, her memo did not specify how many employees are eligible for this offer. According to a Reuters report, the move will impact about 60,000 employees globally, or about a quarter of the global employee base.

Participating teams include AI experiences, Copilot, Xbox, and Windows + Devices, Reuters reported, citing a person familiar with the matter. Marketing, sales, commercial business, and security teams, along with some data center staff, are excluded, the report added.

Employees have until early December to decide, and those who accept will leave at the end of the year, the report added. The memo describes the offer as time-limited and says it includes financial support and career planning resources. However, it does not give the terms of the packages.

From retirement to exit

The VEO builds on what Microsoft “learned earlier this year through the Voluntary Retirement Program,” Coleman wrote. That program, introduced in April, was aimed at long-serving US employees.

Hood said in April that the retirement program would carry roughly $900 million in one-time costs. In July, she said its expenses had come in lower than expected.

The new offer is not tied to age or years of service, and it extends beyond the US. “The Voluntary Exit Offer is new for us. We’ll learn from it,” Coleman wrote in the blog post.

A smaller company by design

Microsoft has said its workforce will shrink this fiscal year. “We expect headcount will decrease year-over-year,” CFO Amy Hood had said on the company’s April 29 earnings call, referring to fiscal 2027, which began July 1. On the same call, she said the company was changing how it operates “to increase our pace and agility.”

Microsoft employed about 223,000 people full-time as of June 30, 2026, down from 228,000 a year earlier, according to its annual report filed with the US Securities and Exchange Commission. Hood said on the company’s July 29 earnings call that total headcount had declined 2% year over year.

AI spending and the staffing question

The workforce changes coincide with a sharp rise in AI infrastructure spending. Microsoft spent $41 billion on capital expenditures in the quarter ended June 30, Hood said on the July earnings call, and she guided to more than $50 billion for the September quarter.

Microsoft has also urged customers to rethink how they fund AI. On the April earnings call, CEO Satya Nadella said IT budgets would be reshaped by business outcomes and “reallocation from other line items on the income statement, like OpEx.”

Sanchit Vir Gogia, chief analyst at Greyhound Research, cautioned against reading Microsoft’s moves as a template.

“Microsoft’s internal AI use does not establish substitution of particular roles,” he said. He advised enterprises to measure completed work at an agreed quality level across representative roles and difficult exceptions before changing staffing, and to keep experienced staff available while doing so.

“Workforce savings earn credibility through the capability that remains,” Gogia said.

Account coverage in question

Whether customer-facing teams take part is an open question for buyers. Coleman wrote in the memo that participation varies based on each business’s “priorities, customer commitments and workforce needs.”

Microsoft’s annual report counts 89,000 employees in operations, a category that includes product support, consulting services and data center operations. Another 43,000 work in sales and marketing. Microsoft has not said whether staff in either group are eligible.

“The practical test is whether customer-critical capability remains available throughout the transition,” Gogia said. He recommended that CIOs confirm the accountable account owner and a qualified backup. They should then test the successor on an unresolved exception and compare support cases against a recent baseline.

At the same time, Microsoft is adding people to customer engagements. Nadella said on the July earnings call that the company had launched Microsoft Frontier Co., which will embed 6,000 industry and engineering experts with customers.

Gogia said delivery commitments need checking at a granular level. “Aggregate headcount does not identify the expertise required by an Azure migration or a Dynamics implementation,” he said. He advised verifying Azure dependencies by subscription and region, reconfirming Dynamics implementation specialists, and distinguishing previews from contractual milestones on product roadmaps. Microsoft did not immediately respond to a request for comment.

OpenAI has launched GPT-6.1 Terra, a new reasoning model that it says delivers near-frontier performance on visual reasoning, professional task execution, and computer-use benchmarks without the premium price tag. Replacing GPT-5.6 Terra, the model is aimed at organizations seeking better accuracy and automation capabilities for large-scale operational workloads.

Claiming it to be a step up from GPT-5.6 Terra at 50% lower price, OpenAI said its new mid-tier model is designed to handle everyday professional work such as drafting proposals, interpreting technical drawings, filling out forms, and editing videos.

“We’re introducing GPT-6.1 Terra, a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing,” the company noted.

GPT-6.1 Terra targets developers and enterprises looking to automate everyday knowledge work at scale without paying premium prices for frontier AI models, said Pareekh Jain, CEO at EIIRTrend & Pareekh Consulting. “Enterprises need AI models that balance intelligence, speed and cost. Frontier models are expensive for routine tasks, while smaller models may lack sufficient capabilities. Terra addresses this gap, competing with Anthropic’s Claude Sonnet and Google’s Gemini Flash models.”

Better performance at lower cost

The new reasoning model sits between OpenAI’s lower-cost GPT-6 Luna and its higher-capability GPT-6.1 Sol and GPT-6 Astra models.

As OpenAI has positioned computer use as a central capability of GPT-6.1 Terra, it is backing those claims with a series of benchmark scores. On Agents’ Last Exam, a benchmark of professional tasks, at max effort, GPT-6.1 Terra scored 60.0% at a cost of $1.01 per task, compared to 50.7% for GPT-5.6 Terra ($3.58), 58.6% for GPT-6.1 Sol ($2.00), and 59.3% for GPT-6 Astra ($7.24). In comparison, Claude Opus 5.5 scored 63.2% at max effort, costing $8.82 per task.

Visual reasoning results on the MMMU-Pro benchmark show GPT-6.1 Terra outperforming its predecessor, GPT-5.6 Terra, while nearly matching top-tier frontier models like GPT-6 Astra and GPT-6.1 Sol across every reasoning configuration, at a fraction of the cost. At max effort, GPT-6.1 Terra achieved an 85.5% score at $0.012 per task, compared to GPT-5.6 Terra’s 81.7% at $0.079. Within the current family lineup, GPT-6 Luna scored 79.9% ($0.0021 per task), GPT-6.1 Sol hit 86.3% ($0.013 per task), and flagship GPT-6 Astra led at 86.7% ($0.065 per task). OpenAI did not include competitor model comparisons for this specific benchmark.

On the OSWorld 2.1 benchmark, which evaluates how AI agents navigate desktop applications to perform complex multi-step tasks, GPT-6.1 Terra scored 74.2% at max effort, costing $1.19 per task. That marks a substantial upgrade over GPT-5.6 Terra’s 57.1% score at $4.76 per task. However, it still trails Anthropic’s Claude Sonnet 5.5, which led the category at 83.9% with a cost of $5.74 per task.

OpenAI also claims to have invested heavily in reducing hallucinations and factual errors produced, and as per the internal benchmark of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, and at maximum effort, the error rate fell by 18 percentage points to 4.3%, nearly matching the rates for GPT-6 Astra and GPT-6.1 Sol. On cybersecurity, GPT-6.1 Terra meets the Critical threshold under OpenAI’s Preparedness Framework and retains the safeguards developed for GPT-6 Astra.

“Terra is ideal for high-volume tasks such as invoice processing, document analysis, application automation and routine coding. It improves on GPT-5.6 Terra in speed and accuracy, while GPT-6.1 Sol is better suited for complex reasoning, research, and software engineering,” noted Jain.

However, Jain cautions that technical benchmarks alone are insufficient. Enterprises should test Terra on their own workloads, measuring task completion rates, accuracy, execution time, cost per workflow, human intervention, and compliance with security requirements.

Economics, not just pricing

GPT-6.1 Terra is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API as gpt-6.1-terra.

OpenAI has priced GPT-6.1 Terra at $1 per million input tokens, $5 per million output tokens, and $0.05 per million cached input tokens, positioning it as a lower-cost alternative to its high-tier models. By comparison, entry-level GPT-6 Luna costs $0.10 for inputs, $0.50 for outputs, and $0.01 for cached inputs per million tokens. GPT-6.1 Sol sits at $2 for inputs, $10 for outputs, and $0.10 for cached inputs, while flagship GPT-6 Astra commands $10 for inputs, $50 for outputs, and $1 for cached inputs per million tokens.

Claude Sonnet 5.5, on the other hand, is priced at $2 per million input tokens and $10 per million output tokens, and cached costs $0.10 per million tokens.

“Cheaper inference improves one component of agent economics. Terra’s base rates are half Sol’s, but review and integration still require funding,” Sanchit Vir Gogia, chief analyst at Greyhound Research. “Token processing is fuel with a higher operating cost. A CFO relieved by cheaper fuel needs to establish what happens to acceptance and review demand. In the illustrative coding workload, 5.4 additional review seconds erase the inference saving. At scale, the relevant purchase is completed work that remains correct and authorized.”

The article originally appeared on InfoWorld.

AI is getting expensive, and companies are looking to cut token costs. The solution: move more AI work to PCs.

The PC won’t replace the cloud. Preliminary AI work can be done on a PC, and tougher work — which today’s PCs can’t handle — can go to the cloud.

“We’ll always need to use the most advanced cloud-based frontier models for certain tasks, but… we don’t need to use them for everything,” wrote TECHnalysis Research Insights president Bob O’Donnell in a LinkedIn post.

That’s the vision Microsoft calls “hybrid intelligence,” in which Windows makes AI always available. The cloud models run the hardest tasks, and less demanding tasks run locally on the PC. For Windows machines, at least at this point, that means  Copilot+ AI PCs.

“The future of AI is not about having just one model. It’s about having the right model at the right moment — frontier capability when you need it, smaller, faster models when you want speed,” said Pavan Davuluri, head of Windows and Devices at Microsoft, at an Oct. 7 PC launch event.

Hybrid intelligence will be everywhere, blurring the line between AI in the cloud and the PC, said Microsoft CEO Satya Nadella at the event.

“I just want to use my computing power wherever it is in sort of a seamless flow of what I’m trying to get done,” he said.

Agentic AI PCs

Microsoft recognizes that hefty cloud token bills are making enterprises hesitate to scale AI. Its answer: offload AI tasks that don’t need cloud models to AI PCs.

About 10% to 15% of AI workloads will run on PCs and 40% to 45% in the cloud in the next three to four years, said analyst Jack Gold of J. Gold Associates.

“Companies will need AI-accelerated PCs, and at some point we’ll likely see a renewed emphasis on buying newer machines,” Gold said.

Enter agentic AI PCs, which can run agents, generate video, write code and chat with advanced genAI models on a PC. If the PC can’t do the job, it is sent to the cloud.

“A year ago, such capabilities were exclusive to frontier cloud models. Now you can have these capabilities running on your PC,” Microsoft’s Davuluri said.

Agentic AI PCs are equipped with powerful GPUs that are distant cousins of the GPUs in data centers. Nvidia’s RTX Spark superchip — which has a Blackwell GPU — powers many of them, including Microsoft’s new Surface Laptop Ultra.

Unlike older AI PCs, the new agentic PCs can run full AI models and have enough compute power to run them fast, said Gerardo Delgado, senior director of product management at Nvidia. They give users and enterprises a path to advance AI projects.

“To us, an AI PC is the PC that will… run those local models so that you can actually keep the level of intelligence and enjoy the economic benefits,” Delgado said.

Local, by design

HP’s newly announced ZBook Ultra G3a mobile workstation has an AMD GPU that generates free tokens and cuts cloud bills, said Brian Allen, manager for Global Z workstation products at HP.

“It can be on-prem. They don’t have to worry about confidential information going to cloud, or tracking tokens and getting $1,000 charges a week,” Allen said.

The ZBook Ultra G3a packs Perplexity’s Portable Computer service, which has agents that can link a local model to frontier models in the cloud.

For example, an agentic PC can handle basic AI-based design work in Autodesk’s Revit 3D modeling software without an internet connection. For more advanced design work, Perplexity can advance it to one of 19 frontier models in the cloud, whichever fits best.

Designers will likely offload general tasks to frontier models in the cloud and keep proprietary work local, Allen said.

HP links to Perplexity’s tools in the Windows taskbar. Perplexity has MCP connectors for Revit, and users can just type in their desire.

“I can do contextual typing of, ‘hey, build this for me,’ without having to know the details of that,” Allen said. It takes years to learn the intricacies of Revit, but “I can really learn how to do Revit through agentic AI in a matter of hours,” he said.

Coding pest control

Microsoft demonstrated how a local model could save token costs typically incurred in the cloud.

A team couldn’t reproduce a software bug, so GitHub Copilot picked GPT-6 Luna in the cloud to analyze the code. The cloud model then handed off building and testing to local sub-agents running MAI Code 1.1 Flash on a PC to reproduce the bug.

The local work was free. One sub-agent fed 1.6 million tokens of code and context to the local model, which generated 10,500 tokens in response, a GitHub product manager said during the demonstration.

To make that possible, Microsoft compressed the 130-billion-parameter MAI Code 1.1 Flash model so it can run on a PC.

“We have taken that model to three bits of precision, reducing its size by nearly 80% while preserving coding quality,” Davuluri said.

The demonstration used GitHub’s HydraFusion feature, which moves coding tasks between local models and the cloud. Microsoft is bringing the feature to GitHub’s Copilot app for Windows Copilot+ PCs.

Many challenges remain

Using agentic AI on PCs isn’t as simple as installing models or software, said analyst Gold. Companies will also need to develop and customize models for such PCs, depending on applications in enterprise environments.

Open-source AI models, which can be customized, on those machines could “significantly reduce AI costs and increase ROI,” Gold said.

Also, the first agentic AI PCs hitting the market are expensive. Microsoft’s Surface Laptop Ultra starts at around $2,600, reaching $6,000 with all the bells and whistles.

A cheaper laptop can do the same job using cloud AI models, and the Surface could quickly go out of date as new models arrive, wrote senior reporter Mark Hachman at sister publication PCWorld in a product review.

Analysts expect PC prices to remain elevated until the end of 2027, partially due to high component costs, the Iran war, and an uncertain economic environment.

A hybrid AI environment is the way forward, but using AI on a PC or the cloud isn’t exactly seamless yet, wrote Tom Mainelli, group vice president at IDC, in a LinkedIn post.

“For much of this to waterfall down to the average user, many different things will need to happen across component pricing, model evolution, agentic use cases, and more,” Mainelli wrote.

HP is still trying to figure out a market for agentic AI PCs. It’s a long learning curve for customers on how to use it, Allen said.

“They’re in the ‘this PC is too good to be true’ phase… ‘let me benchmark it on my own RAG model and see how it really works,'” Allen said.

Apple has responded to recent reports complaining that iPhone Duo displays have become damaged in normal use. The company is “aware” of the issue and suggests any of the “small number” of customers who encounter the problem contact AppleCare for a replacement.

Apple acknowledges the issue

While the matter has already been widely reported citing unsubstantiated claims on social media, this is the first credible confirmation of the problem. In making it, Apple took pains to explain the problem does not impact all iPhone Duo units, just a small subset of them. “We are aware of an issue impacting the inner display on a small number of iPhone Duo units,” the company said. 

Reports say that cracks appear in the same place on the iPhone Duo: along the center of the hinge, often followed by a bright line of stuck pixels. Several owners report the cracks appearing after only a handful of openings.

What Apple will do now

I imagine that what is happening now is that Apple is gathering intelligence as damaged units are returned to AppleCare. It will seek to understand the nature of the problem, which it will then attempt to trace through its manufacturing process in order to identify the source. Speculatively, the cause could be something as simple as a misaligned tool on the production line, though the fact that the problem doesn’t impact every iPhone Duo suggests it is not necessarily inherent to the design of the product.

If I’m right and Apple has identified the problem as a manufacturing issue, then Apple representatives will be in crisis contact with their suppliers to track the problem down and resolve it. That’s going to mean a little lost production while factory production lines are checked and verified, and whatever problem is causing these faults is hopefully resolved.

Production challenges and teardowns

Either way, Apple is unlikely to be in position to clarify the nature of the problem immediately pending further investigation of returned devices. Unless, that is, it is connected to one of the production challenges it faced when it began manufacturing the device. We’ve seen multiple reports claiming the yield of good devices was quite low. Was this the problem? Apple hasn’t said.

iFixit’s iPhone Duo teardown speculates that a hinge or other component getting stuck when the device folds may be the cause. “The damage we’re seeing in photos isn’t right where the display cable passes through the hinge, but slightly lower, at a point where several edges meet and change how the thin display is supported as it bends,” they wrote. They also said the iPhone Duo is the most repairable folding smartphone they’ve seen so far.

What happens next? 

That’s up to Apple, but with the company preparing to launch the iPhone Duo in 28 more nations at the end of the week, its response needs to be public, visible, and credible. 

“We all remember when Samsung experienced problems with the original Galaxy Fold in 2019, but it also showed that a fast and transparent response can contain the damage to a category,” said IDC VP Worldwide Devices Francisco Jeronimo in a note shared with Computerworld.

Offering a free replacement for affected devices is the right first step. “At $1,999, Apple is positioning the Duo as a premium alternative to established foldables, and how it handles this first-generation issue will shape buyer confidence in the category well beyond its own sales,” Jeronimo said. 

“Apple has a loyal customer base and a retail network that can swap devices on the spot, but with a further 28 markets launching on 30 October, it needs to clarify its warranty position quickly,” he said.

A first firefight for Ternus

However, while Apple has confirmed there is a problem impacting some devices, the vast majority of the millions of iPhone Duo smartphones now in use don’t seem to be succumbing to it. Nonetheless, Apple’s new product-focused leader, CEO John Ternus, now has a firefight on his hands.

Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon. Apple confirms iPhone Duo problem, what will it do next?