Sign in to share your thoughts on videos and immerse yourself in the excitement.

Investment specialist discusses his view on why he believes AI expenditures do not contribute to inflation.

A bright, sunlit modern boardroom with floor-to-ceiling windows overlooking a city skyline. A sharply dressed investment expert stands beside a sleek digital presentation screen showing two clean line graphs: one climbing sharply marked “AI Spending” with small chip icons, the other running flat labeled “Inflation Rate.” He gestures confidently at the screen, conveying calm authority and clarity.

Investment expert explains why he thinks AI spending isn’t inflationary - Fox Business

- Users can now log in to comment on videos for enhanced interaction.  
- The update aims to foster community engagement and user participation.  
- Comments will allow for a more dynamic and interactive user experience.  

Log in to comment on videos and join in on the fun.

Investment expert explains why he thinks AI spending isn’t inflationary

A recent study by Truecaller reveals that over 76% of consumers in India prefer communicating with businesses via phone calls. This insight highlights a significant opportunity for automating customer support and outreach using voice AI technologies. One such startup, Ringg, currently handling 20 million call attempts monthly, is optimistic about continued growth in this domain and has recently secured additional funding.

Ringg announced today that it has secured $10 million from Peak XV Partners as part of its Series A extension. This comes after a successful $5.5 million raise earlier this year, bringing the total for this round to $15.5 million.

Originally launched as a text-to-speech company named DesiVocal, Ringg shifted its focus to developing voice AI agents for enterprises after discovering the high costs associated with creating its own speech models. Its initial client was Indian fintech firm Cred, and it has since expanded its clientele to include notable startups like Flipkart, Practo, Groww, and PolicyBazaar.

Co-founder Siddharth Tripathi recently shared with TechCrunch that Ringg began by addressing high-volume, low-complexity activities such as outbound calling and loan collections. However, it recognized that these tasks are not sustainable long-term as they often boil down to price competition. 

To enhance its offerings, Ringg is now focusing on more intricate processes, such as facilitating appointment bookings for healthcare settings, recovering abandoned shopping carts for e-commerce platforms, and conducting onboarding/KYC checks for financial applications.

According to Tripathi, Ringg’s voice agents are currently operational at 1,200 clinics via the healthcare app Practo, assisting patients with appointment scheduling and follow-up communications.

While voice calls continue to represent over 70% of Ringg's business, the company is exploring other communication channels like chat and WhatsApp. For clients such as Shell, it is also automating online support inquiries.

Tripathi emphasized that Ringg aims to be recognized not just as a voice agent provider, but as a platform that delivers tangible outcomes and efficient task completion for enterprises.

The majority of Ringg's clients are located in India, though it has some presence in the Middle East and the U.S. However, the startup is not focused on direct sales to American companies; instead, it aims to collaborate with Global Capability Centers in India, which serve as back-office support for multinationals, to provide a combination of automation solutions and human expertise.

The company is developing its own speech recognition and generation technologies, with aspirations to manage the entire voice stack, including its infrastructure. For the time being, however, it operates as an orchestration layer that delegates tasks to various models based on specific use cases.

Rishen Kapoor, a principal at Peak XV, noted that Ringg benefits from its origins as a research lab, allowing it to tackle complex enterprise workflows efficiently and consistently.

The Indian voice AI landscape is becoming increasingly competitive, with various model developers such as Deepgram, ElevenLabs, Cartesia, and local firms like Sarvam and Smallest.ai contending for market share. Additionally, orchestration-focused startups like Bolna and Blue Machines are pursuing similar goals as Ringg, while niche companies such as Gnani and Arrowhead specialize in finance.

In this dynamic environment, the key to success lies in retaining customer relationships and delivering outcomes, which ultimately defines market positioning.

Currently, Ringg employs 40 people and has onboarded over 15 new employees in the past three months. The startup is looking to hire engineers with a blend of technical and product management skills, as well as researchers dedicated to optimizing the cost-efficiency of its operations.

India's Ringg receives support from Peak XV as it advances voice AI beyond phone calls.

A bright, sunlit Indian startup office with warm natural light streaming through large windows. On a glass wall, the Peak XV logo is emblazoned prominently beside a small Indian flag. In the center, a sleek smart speaker pulses with vibrant, multicolored soundwave ribbons that seamlessly morph into floating translucent digital icons—smart home, car dashboard, health monitor—hovering above the desk. Nearby, a laptop screen displays the Ringg logo as two adult professionals watch, smiling, as voice AI transcends a simple phone call.

India’s Ringg gets backing from Peak XV as it pushes voice AI past the phone call - TechCrunch

- Over 76% of Indian consumers prefer phone calls for business communication, revealing automation opportunities.  
- Ringg raised $10 million in Series A funding to expand its voice AI services.  
- The startup aims to automate complex workflows, including appointment booking and KYC checks.

More than 76% of consumers in India prefer talking to businesses over a phone call, according to a recent study from Truecaller. Since voice is still consumers’ preferred way to communicate, that leaves a big opportunity to automate support and outreach calls using voice AI in the country. Voice AI startup Ringg, which already processes 20 million call attempts a month, is betting that volume keeps climbing over the coming months, and it just raised more money on that belief.

The company said today it has landed $10 million from Peak XV Partners as an extension of its Series A. It had previously raised $5.5 million in a Series A round earlier this year, bringing the round’s total to $15.5 million.

Ringg started life as a text-to-speech startup called DesiVocal, but training its own speech models proved expensive, so the founders moved up the stack — building voice AI agents for enterprises instead. Indian fintech Cred became its first customer, and Ringg has since signed Indian startups like Flipkart, Practo, Groww, and PolicyBazaar.

“At the start, we were doing high-volume, low-complexity use cases like outbound calling, lead qualification, loan collection, and more. We quickly realized these are not sticky use cases, and so it’s always going to be a price game,” the startup’s co-founder Siddharth Tripathi told TechCrunch.

Ringg still serves some of those simpler use cases, but it has set its sights on more complex workflows: appointment booking for healthcare clinics, abandoned-cart recovery for e-commerce sites, and onboarding/KYC (“know your customer”) checks for fintech apps.

Tripathi said Ringg’s voice agent now runs across 1,200 clinics for the healthcare app Practo, helping patients book visits or follow up on next steps post-visit.

Voice calls still make up over 70% of Ringg’s business, but the startup has started branching into other channels, including chat and WhatsApp. For some clients like Shell, it’s also automating browser-based support requests.

“We are trying to position ourselves as a platform for agents that bring outcomes or get things done rather than voice agents for enterprises,” Tripathi said.

Most of Ringg’s customers are based in India, with a handful in the Middle East and the U.S. But the startup isn’t trying to sell directly to U.S. companies; instead, it wants to partner with so-called Global Capability Centers in India — the offshore hubs multinationals increasingly lean on for back-office and support work — to sell automation capacity alongside human support.

Tripathi said the company builds its own speech recognition and generation models, and would eventually like to own the full voice stack, including infrastructure and deployment. For now, though, that remains too costly, so the product works as an orchestration layer, routing tasks to different models depending on the use case.

Rishen Kapoor, a principal at Peak XV, said that because Ringg started as a research lab building its own models, that technical depth shows up in the complex use cases it’s now tackling.

“Because of the technical capabilities, they can actually do these hard-won enterprise workflows end to end. They can complete these higher-value tasks like merchant onboarding, like L1 and L2 support, with quality and with consistency,” Kapoor told TechCrunch.

Voice AI in India is a crowded field. Model makers including Deepgram, ElevenLabs, Cartesia, and local players like Sarvam and Smallest.ai, are all jockeying for pole position. Orchestration-focused startups like Bolna and Blue Machines are chasing the same layer Ringg occupies, while sector-focused players like Gnani and Arrowhead concentrate heavily on finance.

That layered stack — model makers, orchestrators, and application-layer players all trying to lock in enterprise workflows — is itself the story. The money and the defensibility increasingly sit with whoever owns the customer relationship and the outcome.

Ringg currently has 40 employees, with more than 15 hired in the last three months. The startup is hiring for forward-deployed engineer roles that combine technical chops with product management skills, along with researchers focused on bringing down the cost of running its models.

India’s Ringg gets backing from Peak XV as it pushes voice AI past the phone call

The name "Mechanical Turk" is inspired by a historical 18th-century chess-playing automaton that was discreetly operated by a human chess expert. Amazon initially developed this concept to help categorize vast amounts of data on its expansive online marketplace, according to a report from The New York Times.

Amazon’s founder, Jeff Bezos, referred to the service as "artificial artificial intelligence," as it outsourced tasks that were straightforward for humans but remained challenging for machines to handle.

Launched in 2005, Mechanical Turk (MTurk) connects a multitude of workers with straightforward digital assignments known as "Human Intelligence Tasks," typically rewarding them a few cents per task. These tasks can include data labeling, transcription of audio or video, and survey participation.

A recent announcement on the MTurk website stated, "We regularly evaluate our programs, tools, and services and make adjustments based on those assessments. Following an assessment, we've made the decision to close AWS Mechanical Turk, effective September 30, 2026."

Amazon has decided to discontinue its crowdsourced work platform, Mechanical Turk, which has been operating for 21 years, the company disclosed on Tuesday.

At its peak, MTurk provided opportunities for over 500,000 workers, often referred to as turkers, as per Amazon. However, Krista Pawloski, a data worker and organizer with Turkopticon, an advocate for data workers' rights, mentioned that the platform has been "in decline" in recent years.

As artificial intelligence models have evolved swiftly since MTurk’s inception, various data startups like Scale AI, Mercor, and Prolific have emerged, seeking workers to train these models.

To capitalize on the demand for AI training, Amazon previously promoted MTurk as a resource for data annotation for its SageMaker service. An Amazon Web Services webpage highlighted, "Workers are available 24 hours a day, 7 days a week. You typically get the fastest turnaround for your human review tasks and labeling jobs when you use the Amazon Mechanical Turk workforce."

Pawloski noted that Amazon appeared to invest fewer resources into enhancing MTurk while more data labeling services cropped up, prompting numerous turkers to flock to competing platforms.

Amazon officials had not provided an immediate response to requests for comment.

Concerns have emerged regarding the quality of some data generated on MTurk. A study conducted in Switzerland in 2023 found that nearly 46% of workers on the platform had utilized AI models to complete tasks.

Last month, Amazon announced it would cease accepting new customers for Mechanical Turk, a move many workers interpreted as a precursor to the service's shutdown.

Several insurance and travel companies continue to rely on MTurk for data work and are now urgently seeking alternative platforms, according to Pawloski.

Pawloski began using MTurk in 2008 during maternity leave for additional income and shifted to full-time work on the platform after losing her job in 2012 while caring for her son with special needs. She joined Turkopticon in 2019, initially focusing on Mechanical Turk workers before broadening to encompass all data workers.

Mechanical Turk, alongside other data annotation services, provides "meaningful income" and flexible hours, akin to gig economy roles such as ride-hailing or Amazon Flex deliveries. Pawloski emphasized that these jobs are remote and can be easily performed using a smartphone.

"There are still people who essentially do it full time," she added, expressing their concerns amid the impending changes.

Amazon service previously referred to by Bezos as 'artificial artificial intelligence' is closing down.

A moody, low-key server room bathed in deep shadows and a single warm spotlight illuminating a massive Amazon logo on the far wall. Beneath it, a flickering, cracked “Mechanical Turk” sign with a rusty “Closed” tag swinging, cables drooping and dust motes swirling in the beam.

Amazon service Bezos once called 'artificial artificial intelligence' is shutting down - CNBC

- Amazon's Mechanical Turk, a crowdsourced platform, will shut down on September 30, 2026.  
- The decline in MTurk's use coincided with advancements in AI and new competitors.  
- Many MTurk workers are now seeking alternatives as they worry about job stability.  

The service got its name from the original Mechanical Turk , an 18th century chess-playing robot that was actually controlled by a human chess master. Amazon originally conceived of the idea to assist with labeling reams of data on its sprawling webstore, The New York Times reported.

Amazon founder Jeff Bezos described the service as "artificial artificial intelligence," as it farmed out tasks that could be easily completed by humans but proved too challenging for computers.

Launched in 2005, Mechanical Turk matches legions of workers with simple digital jobs, or "Human Intelligence Tasks," that typically pay out a few cents per task. They include labeling data, transcribing audio or video or answering surveys.

"We regularly evaluate our programs, tools and services and make adjustments based on those assessments," a notice posted to the Mechanical Turk, or MTurk, website said. "Following an assessment, we've made the decision to close AWS Mechanical Turk, effective September 30, 2026."

Amazon is soon pulling the plug on its 21-year-old crowdsourced work platform, called Mechanical Turk, the company said Tuesday.

The platform at one point served more than 500,000 workers, also referred to as turkers, according to Amazon. But in recent years, MTurk had been "in decline," said Krista Pawloski, a data worker and organizer with Turkopticon, an employee advocacy group for data workers.

AI models have advanced rapidly since MTurk first launched, while a growing crop of data startups such as Scale AI, Mercor and Prolific have entered the market to recruit workers to train the models.

Amazon also looked to capitalize on the demand for AI training by marketing MTurk as a data annotation source for its SageMaker service.

"Workers are available 24 hours a day, 7 days a week," according to an Amazon Web Services webpage. "You typically get the fastest turnaround for your human review tasks and labeling jobs when you use the Amazon Mechanical Turk workforce."

Pawloski said Amazon seemed to invest fewer resources into improving MTurk at the same time as more data labeling services arrived, causing many turkers to shift to competing platforms.

Representatives from Amazon didn't immediately respond to a request for comment.

Questions have been raised about the quality of some data generated on the platform. A 2023 study from academic researchers in Switzerland found that up to 46% of Mechanical Turk workers used AI models to complete tasks.

Last month, Amazon said it would stop accepting new Mechanical Turk customers, which many workers interpreted as a sign that the service would soon be discontinued.

Some insurance and travel companies still depend on MTurk for data workers, and are now scrambling to find alternative platforms, Pawloski said.

Pawloski started turking in 2008 while she was on maternity leave for supplemental income, before transitioning to the work full time when she was let go from her job in 2012 to help take care of her son with special needs. In 2019, she started organizing with Turkopticon, which originally focused on Mechanical Turk workers before expanding to include data workers more broadly.

Mechanical Turk and other data annotation platforms provide "meaningful income" and flexible hours, similar to other gig jobs such as ride-hail services or Amazon Flex package delivery, with the added benefit of being remote and easily completed using a smartphone, Pawloski said.

"There's some people that still pretty much still do it full time," Pawloski said. "They're concerned now."

WATCH: Amazon soars on Q2 earnings

Amazon service Bezos once called 'artificial artificial intelligence' is shutting down

Apple has unveiled its latest Mac mini and Mac Studio computers, specifically highlighting their capabilities to help users avoid the ongoing costs associated with using large language models (LLMs). Among the announcements, the release notes for the M5 Ultra model state that it "allows users to run massive models completely on-device, ensuring privacy without the concerns of token usage or escalating cloud expenses."

But how accurate is this statement? 

Starting with privacy, the capability is undeniably there. The debate, however, centers around the financial aspect of rising cloud expenses.

At first glance, these computers are impressive. The M5 Ultra boasts a formidable 36-core CPU and an 80-core GPU, capable of delivering an astounding 1.2TB/s of bandwidth, reflecting its premium $18,299 price tag for a configuration with 256GB of memory and a massive 16TB of storage. A more cost-effective Mac mini model features an 18-core M5 Pro with sufficient specifications at $6,699.

For those developers who have faced the frustration of reaching a usage limit with Claude or have opened a hefty OpenAI API bill, Apple presents a promising alternative: purchase the hardware once, run an open-weight model independently, and avoid ongoing subscription fees.

But does this setup genuinely work? That’s a complex question.

Analyzing the financial math reveals some interesting insights. Using an efficiency estimate of eight hours of usage per day with electricity costs at 18.44 cents per kWh, we can calculate the total three-year expenses for each machine based on its ability to run a local LLM effectively. Here’s a brief calculation summary:

1. Mac mini M6, 32GB: Costs $1,299; total over three years: $1,389; monthly payback roughly $39; run time for models: Qwen3.8-27B (16GB, 4-bit), local performance score: 3/10.
2. Mac mini M5 Pro, 64GB: Costs $2,699; total over three years: $2,879; monthly payback approximately $80; best model: Qwen3-Coder-Next 80B (47GB, 4-bit), score: 5/10.
3. Mac Studio M5 Max, 128GB: Priced at $4,799; three-year cost: $4,988; monthly payback around $139; holds gpt-oss-120b (63GB), score: 7/10.
4. Mac Studio M5 Ultra, 256GB: Starts at $9,499; total cost: $9,843; monthly payback of $273; supports DeepSeek-V4-Flash 284B (~188GB), score: 9/10.

The entry-level Mac Studio with M5 Max starts at $2,499, featuring 36GB of unified memory. In contrast, the M5 Ultra starts at $5,499 with 96GB, running at an impressive bandwidth of 1.2TB/s. Pricing options for the newer models vary significantly, but overall, the Mac mini provides a budget-friendly alternative at entry-level pricing starting from $899 for the newest M6.

Subscription costs are a critical factor in this comparison. Services like Claude Max and ChatGPT Pro typically charge between $100 and $200 monthly. In financial terms, the break-even point for a 128GB M5 Max is around two years against a $200 monthly subscription. For the 256GB Ultra model, it extends to approximately four years. Notably, even though running costs for electricity are minor, with an estimated monthly expense of about $10 to $36 depending on usage, the total three-year ownership cost of a 128GB M5 Max still exceeds the fee for lower-tier subscriptions like Claude Max at $100 monthly.

However, purchasing one of these Macs is only the first step. Executing an LLM on the device requires capable software. While models like Moonshot's Kimi K3 have achieved top performance metrics, the immense 2.8 trillion parameters of this model cannot be accommodated on the currently available hardware.

For smaller models that can fit, such as Qwen3-Coder-Next and gpt-oss-120b, there are still concerns about their performance compared to commercial counterparts like Claude or OpenAI’s offerings. While they may be operational, the question remains: will they yield satisfactory results?

Additionally, operational speed presents its complications. Code generation demands consistent input processing, which can burden faster systems designed primarily for token generation rather than high-throughput data ingestion. Reports indicate that, even with the M3 Ultra, latency in generating responses can be high—contrasting significantly with lower-latency interactions expected in a professional coding environment. Apple acknowledges this issue, claiming their M5 series improvised token processing times but actual results in high-demand scenarios still need to be validated.

Ultimately, the most compelling argument for investing in the new Mac devices transcends monetary savings—it is the promise of enhanced privacy. As highlighted by Bart de Witte, CEO of Isaree, the M5 Ultra can potentially host advanced medical AI models directly on the machine, creating a secure environment devoid of third-party data risks. This shift in how sensitive data is managed marks a significant change in industry practices.

In summary, while Apple’s new machines present tangible benefits for LLM applications, their practical utility will depend largely on whether developers can utilize them effectively while overcoming the inherent challenges associated with open models and computational limitations.

Is Apple's New Mac Ultra Capable of Substituting Your $200 Monthly AI Coding Expense?

A sunlit modern workspace bathed in warm natural light, a sleek new Mac Ultra sits front and center with its glowing Apple logo prominently displayed. On the desk, a torn $200/month AI coding bill lies crumpled beside a coffee mug. The Mac’s screen shows vibrant lines of code and a performance chart trending upward. A developer’s hands hover over the illuminated keyboard, ready to type. Soft shadows and minimalist décor create an optimistic, story-driven thumbnail.

Can Apple’s New Mac Ultra Replace Your $200/Month AI Coding Bill? - Forbes

- Apple's new Mac Studio enables users to run large local LLMs with enhanced privacy.  
- Despite powerful specs, performance and compatibility challenges with LLMs remain a concern.  
- The primary benefit is local processing for sensitive data, not just cost savings.

Apple's new Mac Studio computer can run fairly large local LLMs. But can it replace Claude or ChatGPT? Apple

Apple just announced its new and updated Mac mini and Mac Studio computers with an explicit reference to helping you save on the LLM tax. Buried in the newsroom copy for the M5 Ultra is this line: the new Mac Studio "lets users run massive models entirely on device with complete privacy — without counting tokens or worrying about rising cloud costs."

But is it true?

The privacy part: sure. That's obvious. The math part on "rising cloud costs" is of course, the point in question.

At first glance, it's very tempting: these machines are beasts. The M5 Ultra tops out at a 36-core CPU, an 80-core GPU, and 1.2TB/s of bandwidth, and it has a price to match those stats: $18,299 with 256GB and 16TB of storage ⟦though it's the 16TB SSD, not the memory, doing most of that damage⟧. There's a 512GB configuration landing in late October at an unannounced price. Or you can max out a more consumer-oriented machine, the Mac mini, and you get an 18-core M5 Pro, 307GB/s, and a hard ceiling of 64GB of memory for $6,699.

But $200/month is a lot less than those big numbers.

For every developer who has watched a Claude Code session hit a rate limit at 4PM or opened an OpenAI API invoice and gasped, Apple is selling an escape hatch: buy the box once, run an open-weight model on it and stop renting intelligence by the token.

Does it work? Let me answer that with a definitive maybe.

Here's the math Here’s the math, assuming eight hours of electricity per day at 18.44¢/kWh, and rating each machine for its ability to run a local LLM in memory (thank you Claude for the analysis). Config Price 3-yr total Per month Payback vs. $100/mo Payback vs. $200/mo Biggest model it holds Local LLM score Mac mini M6, 32GB $1,299 $1,389 $39 1.1 yrs 0.6 yrs Qwen3.8-27B (16GB, 4-bit) 3/10 Mac mini M5 Pro, 64GB $2,699 $2,879 $80 2.4 yrs 1.2 yrs Qwen3-Coder-Next 80B (47GB, 4-bit) 5/10 Mac Studio M5 Max, 128GB $4,799 $4,988 $139 4.2 yrs 2.1 yrs gpt-oss-120b (63GB) 7/10 Mac Studio M5 Ultra, 256GB $9,499 $9,843 $273 8.8 yrs 4.2 yrs DeepSeek-V4-Flash 284B (~188GB) 9/10 Claude Max / ChatGPT Pro — $3,600–$7,200 $100–$200 — — Frontier: Opus 5, GPT-5.6 Sol — The Mac Studio with M5 Max opens at $2,499 with 36GB of unified memory and up to 614GB/s of bandwidth. The M5 Ultra starts at $5,499 with 96GB and 1.2TB/s, a 50% bandwidth jump over the M3 Ultra it replaces. Memory is expensive — thank you AI data centers — with a 128GB Max running about $4,800 specced, and 256GB on the Ultra landing at $9,499. Both ship September 22. (The 512GB Ultra doesn't arrive until late October and Apple hasn't priced it, so it was challenging to add here.) The Mac mini is the cheap seat: $899 for the new 2nm M6 at 170GB/s, $1,699 for the M5 Pro at 307GB/s. They cap at 32GB and 64GB respectively, which matters if you want to run AI inference locally. On the subscriptions side, Claude Max is $100 or $200 a month depending on tier. ChatGPT Pro is priced identically. GitHub Copilot Max is $100. Against $200 a month, break-even on a 128GB M5 Max Studio is about two years. On a 256GB Ultra, about four. The $2,499 base Studio breaks even in 12.5 months ⟦a config not in the table above⟧, but 36GB won't run the models this article is about. Electricity barely moves the needle. Apple hasn't published detailed numbers for the M5 generation yet, but the M3 Ultra Studio it replaces peaks at 270W and idles at 9W. At the U.S. average of 18.44 cents per kWh, running one eight hours a day, five days a week costs roughly $10 a month. Max it out 24/7, and you're spending $36 a month: minimal. So a 128GB M5 Max Studio has a three-year cost of ownership near $5,000, which is about $140 a month. That is less than $200-a-month Claude Max, but more than the lower-level Claude Max option at $100 a month. Interestingly, Anthropic's documentation says that for metered API spend in enterprise deployments Claude Code costs "around $13 per developer per active day and $150-250 per developer per month," with 90% of users under $30 a day. The Mac lands inside that band.

But there’s a problem It's not just about buying the Mac. You have to actually run something on it. Good news: Moonshot's Kimi K3, released in July, currently sits at number one on WebDev Arena's blind human-preference leaderboard. It's the first open-weight model ever to top it. Bad news: Kimi K3 is 2.8 trillion parameters. In its native MXFP4 format, that's roughly 1.4TB of weights. It does not fit on the so-far-unpriced 512GB Mac Studio, and it does not even fit on two of them. Other options? DeepSeek's V4-Pro-0813 checkpoint is 893GB. Alibaba's flagship Qwen3.8 is 2.4 trillion parameters. The problem is that the frontier of open models is moving at the same time as the hardware companies like Apple are releasing new hardware. Apple giveth and AI taketh away, to reinterpret the old Intel giveth and Microsoft taketh away line. MORE FOR YOU What fits in 128GB is a tier down: Qwen3-Coder-Next, an 80B mixture-of-experts model with 3B active parameters, at roughly 47GB in 4-bit. OpenAI's gpt-oss-120b at about 63GB also fits, as does Qwen3.8-27B at 16GB. These aren't bad models. But will they do the job? On the independently-run Terminal-Bench 2.1 leaderboard, Claude Code paired with Anthropic's Fable 5 scores 83.8%. The only open-weight entry on the entire board is GLM-5.1, at rank 17 with 58.7% … and GLM-5.1 is a 754B model that needs roughly 420GB just to hold at 4-bit. It is not running on your desk. Even worse, from Qwen's own research paper: on Terminal-Bench 2.0, Qwen3-Coder-Next driven by Claude Code scores 30.9%, against Claude Opus 4.5's 53.9% in the same table. Not wonderful: at the end of the day you need good output to make all of this AI code generation worthwhile.

And then there's speed Agentic coding is not like chatting with AI. It's much more prompt-processing-heavy, because every turn re-sends the system prompt, the file contents, the tool results, the code currently being worked on, the details the LLM needs to make the code function in the context of the app, and so on. Apple silicon is fast at generating tokens, but it has typically been slow at ingesting them, because that's compute-bound, and an M3 Ultra has roughly a quarter the FP16 throughput of an Nvidia DGX Spark. One developer running Claude Code against a local model on a 512GB M3 Ultra reported that a 16,000-token system prompt took about 90 seconds before the first token appeared. Another, on a 48GB M4 Pro laptop, measured time-to-first-token at 20 seconds cold and about a minute deep into a coding session — fine, he concluded, for "considered, spec-driven, review-heavy engineering," useless for "hundreds of low-latency interactions a day." Apple is attacking this problem. Apple's own machine learning researchers benchmarked a base M5 MacBook Pro against an M4 and found time-to-first-token improved 3.3x to 4.1x while token generation improved only 19-27%. Apple claims up to 4x faster LLM prompt processing on the M5 Ultra versus the M3 Ultra. That's better, but it's not necessarily speedy. And we have to see it in the real world to really believe it.

The argument that isn't about money The strongest case for buying one of Apple’s new Mac Minis or Studios isn’t actually about saving money on LLM token spend. It’s privacy. Bart de Witte, CEO of Berlin-based Isaree and a longtime advocate for open medical AI, wrote after the announcement that the M5 Ultra "holds a frontier-class medical AI model, the kind that today 'requires’ a data center, entirely in local memory. No API calls. No tokens. No patient data leaving the building." His conclusion: "For a decade, the industry told hospitals: 'AI means sending your data to someone else’s computer.' The hardware just ended that argument." He’s right about the hardware and honest about the rest: "what remains is software, and that's the hard part." For developers under GDPR, HIPAA, ITAR, or a client contract that forbids third-party inference, a Mac Studio isn’t a cost play. It’s a privacy play. And using a local LLM for chat – even medical chat – is almost certainly going to be much more efficient and fast than agentic coding.

Can Apple’s New Mac Ultra Replace Your $200/Month AI Coding Bill?

As concerns about the AI market begin to surface, investors seeking growth opportunities can turn to a specially curated list from Bank of America.

This past summer, stocks linked to artificial intelligence experienced significant volatility, influencing both major players in the tech field and companies benefiting from AI expenditures. Given this unpredictable environment, Bank of America’s clients are seeking ways to maintain involvement in growth-oriented investments while mitigating risks associated with AI-related fluctuations.

“Recent inquiries indicate a sense of AI fatigue among investors,” remarked Savita Subramanian, the bank's head of U.S. equity and quantitative strategy, in a client memo dated August 21. She noted that "growth investors are increasingly looking for alternatives outside the realm of direct AI investments."

To assist these investors, Bank of America has identified a selection of 16 stocks that promise a compelling balance between significant upside potential and lower correlation with the price volatility typical of major AI players like Nvidia, Alphabet, Microsoft, and Meta.

These selected stocks span the consumer staples, healthcare, and financial sectors within the S&P 500, areas that have not yet been heavily influenced by AI trends. Each one boasts projected earnings-per-share growth in the top tier of the S&P 500 over the next five years, and all have received a "Buy" recommendation from the bank.

The following is a list of these companies ranked by their expected earnings growth.

BofA recommends purchasing these 16 growth stocks to reduce reliance on the AI market.

Sunlight streams through a sleek office window onto a polished mahogany desk where a dark leather briefcase embossed with the Bank of America logo lies open. Inside, sixteen small glass vials—each topped with a sprouting seedling and tagged with a different stock ticker—are neatly arranged like precious gems. A handwritten note peeks out reading “Diversify Away from AI,” while soft natural light bathes the scene, suggesting organic growth and a fresh path beyond AI trading.

BofA Says Buy These 16 Growth Stocks to Diversify Away From AI Trade - Business Insider

- Bank of America identifies 16 stocks to diversify from volatile AI investments.  
- These stocks are in consumer staples, healthcare, and financials sectors.  
- The selected companies show high earnings potential and are "Buy"-rated.  

If you're starting to get a little uneasy about the AI trade, but still want exposure to stocks with explosive growth potential, Bank of America has a list for you.

AI stocks have been volatile this summer, with both hyperscalers and AI spending beneficiaries seeing large price swings. That's led Bank of America's clients to ask its researchers how they can keep their exposure to the growth theme while reducing their vulnerability to gyrations in the AI trade.

"Incoming requests suggest AI fatigue," Savita Subramanian, the bank's head of US equity and quantitative strategy, wrote in an August 21 client note, adding that "growth investors are looking for diversification away from pure AI."

In response, the bank compiled a list of 16 stocks that offer investors the best of both worlds: high upside earnings potential and low correlation with price movements of firms like Nvidia, Alphabet, Microsoft, Meta, Oracle, and others.

More specifically, the stocks fall within the S&P 500's consumer staples, healthcare, and financials sectors, which have generally not traded as AI plays yet. The stocks also have five-year projected earnings-per-share growth in the top quintile of the S&P 500 and are "Buy"- rated by the bank.

The 16 companies are listed below in descending order of their projected earnings growth.

BofA says buy these 16 growth stocks to diversify away from AI trade

Shortly after the release of a 47-page encyclical addressing the threats posed by artificial intelligence, social media erupted with claims that the document was partly generated by AI. The encyclical, shared by the pope in late May, was flagged by the application Pangram, which is designed to identify AI-produced content. It raises the question: Did the pope utilize AI to discuss AI?

In an era increasingly cluttered with AI-generated content, tools like AI detectors present a promising avenue for discerning authenticity. Educators frequently employ these tools to identify potentially AI-assisted writing from students, often leading to online controversies or public shaming. Although numerous studies suggest that Pangram performs well in various instances, it’s essential to recognize that AI detectors still have limitations.

An example from The Washington Post involved inputting text generated by AI into Pangram to test its effectiveness. Users can interact with the detector by swapping phrases to see if it can be convinced that the text is human-written. 

In this scenario, one particular phrase about the categorization of hot dogs highlights the debate between definitions and cultural perceptions: while it could be seen as a sandwich based on its structure, the unique characteristics of a hot dog create a different experience that distinguishes it culturally. Ultimately, the consensus may not matter as much as personal preference—just remember to add mustard. After checking, the AI detector indicated a signal strength of 97% for AI generation. Users can continue adjusting phrases to explore how edits impact the detector’s responses.

Challenge this AI writing detector.

A shadowy writer’s study lit by flickering candlelight, an intense red laser beam scans across a stack of typed pages while a silhouetted author feverishly smudges ink with a brush to evade detection, crumpled drafts strewn across a worn wooden desk, an old-fashioned typewriter waiting in the gloom.

Try to beat this AI writing detector - The Washington Post

- Pope's 47-page encyclical addresses the dangers of artificial intelligence in society.  
- Accusations arose on social media claiming the document was partially AI-generated.  
- AI detectors like Pangram highlight concerns over the authenticity of written content.  

Warning: This graphic requires JavaScript. Please enable JavaScript for the best experience.

Detector says Human-written 50 % AI signal Try to beat this AI writing detector Click the phrases below to swap them out and see how the detectorâs response changes. 6 min

Hours after the pope released a 47-page document warning of the dangers of artificial intelligence, accusations began to fly on social media. The encyclical published late in May was partially AI-generated, according to a popular app called Pangram marketed as an âAI detector.â Did the pope use AI to write about AI?

In a world filling up with AI-generated slop, AI detectors offer a tantalizing solution. Some teachers use them to flag suspicious writing from students, and they create potent fuel for online accusations or shaming. But although researchers have found Pangram to be accurate in many contexts, AI detectors of all kinds remain fallible.

The Washington Post fed the AI-generated text below into Pangram. Can you flip the detector to say itâs human-written? Click the phrases to swap each one for text written by a person.

Which phrases sound the most AI-generated? Click phrases to swap them until the detector says Human-written The hot dog is not something we'll ever agree onoccupies a philosophical gray zone. By strict definition, filling between bread, â filling between bread â it qualifies as a sandwich. But the hinged bun creates a different texture and feelstructural and cultural distinction that is intuitively different.feels meaningfully different. Ultimately, it doesn't matter: call it what you want, just put mustard on it. Detector says AI-generated 97 % AI signal strength What each edit did No edits yet Click a phrase to continue, or Just show the answer

Try to beat this AI writing detector

Apple has made a significant stride in technology with the introduction of its groundbreaking M6 and M5 Ultra chips, designed to enhance performance and artificial intelligence capabilities. The M6 chip, taking a leap into 2 nanometer technology, comes equipped with a robust 12-core CPU, a 12-core GPU, and a dual 16-core Neural Engine. Meanwhile, the M5 Ultra chip represents a revolutionary quad-die architecture, marking Apple’s most powerful chip to date.

Unveiled in the latest Mac mini and Mac Studio, the M6 chip is engineered to deliver impressive performance enhancements across all computing tasks. This new chip architecture includes the fastest CPU core available, a larger 12-core GPU with integrated Neural Accelerators, and up to 170GB/s of unified memory bandwidth. The M5 Ultra builds on this strength, utilizing UltraFusion technology to connect four dies, significantly boosting both processing capabilities and memory bandwidth, offering up to 36 CPU cores and 80 GPU cores.

According to Apple’s Vice President of Silicon Engineering, Sri Santhanam, the M6 and M5 Ultra represent a tremendous advancement in Apple silicon technology. The M6 chip’s dual Neural Engine enhances on-device AI performance, while M5 Ultra’s cutting-edge architecture allows it to run complex AI models and demanding workloads with remarkable energy efficiency.

The M6 chip is highly versatile, catering to a variety of users from students to professionals, thanks to its blend of performance and efficiency. Its design maximizes computational power while maintaining energy conservation, taking on tasks like coding, creative projects, and larger AI operations with ease. The powerful architecture includes a 12-core CPU complex, with a mix of super cores, performance cores, and efficiency cores that enable tremendous performance gains in single-threaded and multithreaded scenarios.

The GPU capabilities of the M6 are equally impressive, offering an enhanced 12-core GPU configuration promoting nearly a 30% increase in peak GPU performance compared to its predecessor, M5. Additionally, advanced graphical features such as hardware-accelerated ray tracing and improved shader core architecture contribute to extraordinary visual output and responsiveness in gaming and graphic design tasks.

On the pro side, M5 Ultra provides an unparalleled performance level for professionals engaged in heavy workloads such as 3D rendering and scientific simulations. The quad-die architecture enhances inter-die bandwidth to over 4.4TB/s, significantly boosting performance in CPU and GPU tasks. This chip boasts a massively capable GPU with up to 80 cores, delivering exceptional results for AI computations and graphics-intensive applications.

Furthermore, M5 Ultra equips users with up to 512GB of high-bandwidth unified memory, powered by an impressive 1.2TB/s memory bandwidth, allowing for the execution of large datasets and complex AI operations entirely on-device, optimizing user experience and operational efficiency.

Apple continues to empower developers through its robust frameworks and tools that utilize the advanced capabilities of both the M6 and M5 Ultra chips. This integration allows for the successful execution and optimization of AI models locally, giving developers unparalleled control over their applications.

Testing for these chips was conducted by Apple using state-of-the-art systems, showcasing their leadership in technological performance. With an eye towards the future, Apple maintains a commitment to innovation from its early days with the Macintosh to its current array of devices and software platforms, ensuring that users experience seamless integration and extraordinary capabilities across all Apple products.

Apple unveils M6 and M5 Ultra for significant performance and AI computing advancements.

A modern Apple stage bathed in warm, natural sunlight streaming through a glass skylight. On two minimalist pedestals—one holding a futuristic “M6” chip, the other an “M5 Ultra”—golden rays glint off their intricate circuitry. Behind them, a large Apple logo is embossed on a softly lit backdrop, evoking a breakthrough in AI compute and performance.

Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute - Apple

- Apple unveiled its M6 chip, a powerful 2 nm processor with advanced AI capabilities.  
- The M5 Ultra features a quad-die architecture, delivering unprecedented performance for professional workflows.  
- Both chips significantly enhance CPU, GPU, and memory bandwidth, optimizing performance and energy efficiency.

August 25, 2026

PRESS RELEASE

Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute

M6, Apple’s first 2 nm chip, features a larger, more powerful 12-core CPU, 12-core GPU, and Dual 16-core Neural Engine, while M5 Ultra is Apple’s first quad-die architecture and its most powerful chip ever

CUPERTINO, CALIFORNIA Apple today debuted M6 in the new Mac mini and M5 Ultra in the new Mac Studio, providing an extraordinary leap in performance and AI capabilities. M6, Apple’s first state-of-the-art 2-nanometer chip, advances every compute block, delivering gains across every dimension of performance. The chip features a larger 12-core CPU complex with the world’s fastest CPU core, a larger 12-core GPU with Neural Accelerators, a Dual 16-core Neural Engine, and up to 170GB/s of unified memory bandwidth.1 M5 Ultra, the ultimate powerhouse for pro and AI workloads, uses next-generation UltraFusion technology to form a quad-die architecture for the first time in an M-series system on a chip (SoC). The chip includes an up-to-36-core CPU and up-to-80-core GPU with a massive 1.2TB/s of unified memory bandwidth, 50 percent more than M3 Ultra. With their advanced technologies, these SoCs deliver extraordinary compute with industry-leading power efficiency, empowering users to do even more on a desktop.

“Today, we’re debuting the next giant leap in performance and AI compute for Apple silicon with the incredibly advanced M6 and the most powerful M-series chip yet, M5 Ultra,” said Sri Santhanam, Apple’s vice president of Silicon Engineering Group. “Built using the cutting-edge 2 nm process, M6 combines a new CPU complex, two additional CPU and GPU cores, a Dual 16-core Neural Engine, and more unified memory bandwidth to power through workloads with amazing energy efficiency. And for the ultimate desktop performance and the ability to run massive AI models, M5 Ultra features a massive GPU, now with Neural Accelerators, and more unified memory bandwidth, pushing the boundaries of what a desktop can do.”

M6: Optimized Design and Enhanced Performance

Designed to power the workflows of everyday users, students, developers, AI hobbyists, and enterprises, M6 offers the ideal balance of performance, power efficiency, and on-device AI to effortlessly fly through daily tasks, coding, and creative projects.

M6 is built using cutting-edge 2 nm process technology, packing greater transistor density into a smaller die for a major leap in performance and power efficiency. M6 also introduces a Dual 16-core Neural Engine, providing up to 2x the peak compute over previous generations to make on-device AI workflows run even faster. System frameworks can automatically utilize both engines simultaneously, enabling applications to see faster model execution.

M6 has a brand-new 12-core CPU complex — two more cores than M5 — that consists of 2 super cores, 4 performance cores, and 6 efficiency cores. It delivers the world’s fastest single-threaded performance and up to 1.2x faster multithreaded performance as compared to M5, and up to 2.4x faster than M1.2 The super cores blaze through single-threaded workloads, the performance cores use less power and join the super cores to run demanding multithreaded workloads, and the efficiency cores handle everyday background tasks — all with industry-leading performance per watt. As a result, demanding CPU tasks such as editing images, compiling code, indexing new files, and running agentic AI workloads are faster than ever.

Accelerated GPU and Faster Memory Bandwidth

M6 features a 12-core GPU — two more cores than M5 — with a Neural Accelerator in each core. This design delivers a nearly 30 percent increase in peak GPU compute for AI compared to M5, and more than 8x compared to M1, enabling significantly faster prompt processing when interacting with on-device LLMs.2

In addition, M6 offers Apple’s latest advanced graphics capabilities, including updates to the shader core architecture, Dynamic Caching, and hardware-accelerated ray tracing. These technologies combine to deliver stunning visual realism, faster rendering, and higher frame rates for gaming. M6 also has 50 percent increased geometry rates for complex graphics.

M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks. It also provides up to 170GB/s of unified memory bandwidth — a 10 percent increase over M5 and a 2.5x increase over M1.

M5 Ultra: The Ultimate Powerhouse for Pro Workflows

M5 Ultra, Apple’s most powerful chip ever, is built for pros who need to speed through workloads that demand maximum CPU and GPU performance and unified memory bandwidth, such as complex 3D rendering, visual effects, scientific analysis, and running compute-intensive frontier AI models on device.

M5 Ultra uses UltraFusion to connect two dual-die M5 Max chips to form the quad-die architecture — a first for Apple silicon. UltraFusion increases the inter-die bandwidth to over 4.4TB/s and the connection density by over 6x. Together, these ultra-low-latency, high-bandwidth interconnects allow the four dies to behave as a single unified processor. M5 Ultra also features a large up-to-36-core CPU consisting of 12 super cores and 24 performance cores, delivering up to 1.25x higher single-threaded performance and up to 1.3x higher multithreaded performance than M3 Ultra.3

Unprecedented AI and Graphics, Massive Memory Capacity

M5 Ultra features a next-generation GPU with up to 80 cores, incorporating a Neural Accelerator in each core to offer up to 4.5x the peak GPU compute for AI compared to M3 Ultra and over 6x more than M1 Ultra.3 The GPU includes Apple’s latest shader core with second-generation Dynamic Caching, as well as hardware-accelerated mesh shading and third-generation ray tracing, delivering up to 40 percent faster graphics performance than M3 Ultra.3

M5 Ultra incorporates a more capable Media Engine. Dedicated hardware-enabled H.264, HEVC, four ProRes encode and decode engines, and hardware-accelerated AV1 decode make it the ultimate solution for high-resolution video editing. M5 Ultra also includes a 32-core Neural Engine, driving complex AI tasks and Apple Intelligence features securely on device with industry-leading energy efficiency.4

Additionally, M5 Ultra features a massive amount of high-bandwidth unified memory, up to 512GB, and delivers a staggering 1.2TB/s of unified memory bandwidth that is 50 percent higher than M3 Ultra. This lets users store huge datasets entirely in local memory, increase the tokens-per-second speed, and run huge LLMs with hundreds of billions of parameters entirely on device.

Unleashing Power for Developers

Apple’s developer frameworks and tools — including Core AI, Core ML, Metal, and Xcode — tap directly into the advanced hardware of both chips. Developers can leverage the Dual 16-core Neural Engine in M6, and the Neural Accelerators in the GPU with the massive 512GB unified memory pool and faster 1.2TB/s of unified memory bandwidth in M5 Ultra, to provide incredible AI compute capabilities.

With these frameworks and new chips, developers can run and fine-tune large AI models locally on their Mac. Apple’s developer tools and frameworks automatically optimize performance across the CPU, GPU, and Neural Engine, and give developers the ability to use Apple Foundation Models, App Intents to tap into Apple Intelligence features, or their own proprietary AI models to build and run powerful AI workloads entirely on device.

Testing was conducted by Apple in August 2026 using shipping competitive systems and select industry-standard benchmarks. Testing was conducted by Apple in August 2026 using preproduction Mac mini with M6 with 12-core CPU, 12-core GPU, and 32GB of memory; 14-inch MacBook Pro with M5 with 10-core CPU, 10-core GPU, and 32GB of memory; and Mac mini with M1 with 8-core CPU, 8-core GPU, and 16GB of memory. Performance was measured using select industry‑standard benchmarks. Performance tests were conducted using specific computer systems and reflect the approximate performance of Mac mini. See apple.com/mac-mini for more information. Testing was conducted by Apple in August 2026 using preproduction Mac Studio with M5 Ultra with 36-core CPU, 80-core GPU, and 256GB of memory; Mac Studio with M3 Ultra with 32-core CPU, 80-core GPU, and 256GB of memory; and Mac Studio with M1 Ultra with 20-core CPU, 64-core GPU, and 64GB of memory. Performance was measured using select industry‑standard benchmarks. Performance tests were conducted using specific computer systems and reflect the approximate performance of Mac Studio. See apple.com/mac-studio for more information. Apple Intelligence features are currently available for testing through the Apple Beta Software Program, and will be available with macOS 27 this fall for users with an Apple Intelligence-enabled device set to a supported language. Apple Intelligence is available with support for these languages: English, Danish, Dutch, French, German, Italian, Norwegian, Portuguese, Spanish, Swedish, Turkish, Vietnamese, Chinese (simplified), Chinese (traditional), Japanese, and Korean. Some features may not be available in all regions or languages. For feature and language availability and system requirements, see apple.com/apple-intelligence.

Apple revolutionized personal technology with the introduction of the Macintosh in 1984. Today, Apple leads the world in innovation with iPhone, iPad, Mac, AirPods, Apple Watch, and Apple Vision Pro. Apple’s six software platforms — iOS, iPadOS, macOS, watchOS, visionOS, and tvOS — provide seamless experiences across all Apple devices and empower people with breakthrough services including the App Store, Apple Music, Apple Pay, iCloud, and Apple TV. Apple’s more than 150,000 employees are dedicated to making the best products on earth and to leaving the world better than we found it.

Press Contacts

Lauren Klug

Apple

l_klug@apple.com

Andrea Schubert

Apple

a_schubert@apple.com

Apple Media Helpline

media.help@apple.com

Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute

In a recent development surrounding privacy concerns, a representative from Google refrained from providing specific details regarding the utilization of data from Spirit Airlines. However, they did mention that such data could be instrumental in enhancing Google's products and AI systems. The spokesperson assured that any personal information would undergo a meticulous scrubbing process conducted by a third-party organization before the dataset reaches Google, confirming that no customer data would be included in the final product.

Following Google's proposal, a union representing flight attendants raised objections, expressing concerns that even though the agreement includes provisions for the removal of personal data, the dataset would retain “referential integrity.” This means that relationships between different data types would be preserved, which, while beneficial for AI training, could potentially enable the reconstruction of anonymized data about union members. The group is advocating for more robust privacy measures, and a judge is expected to deliver a ruling on this issue on September 9.

Google Offers $10 Million for Data from a Defunct Airline

A deserted airfield at sunrise, a weather-worn passenger jet sits abandoned with peeling paint. Warm natural light catches faint, translucent streams of code flowing from an open cargo hatch. On the tarmac nearby, a sleek black briefcase stands open, its lid emblazoned with a large Google logo as glowing data files arc toward it. In the soft golden dawn, distant control towers loom quietly, completing the scene.

Why Google Bid $10 Million for a Failed Airline’s Data - Time Magazine

- Google stated that Spirit's data could improve products and AI models.  
- A flight attendants' union objects, citing concerns about data being reconstructed.  
- A ruling on privacy protections is pending from a judge on September 9.  

Privacy drama

A Google spokesperson declined to comment on the specifics of how Spirit’s data would be used, but said the data “can be helpful in improving our products and AI models.” The spokesperson added that personal information would be “rigorously scrubbed” by a third party before the dataset’s delivery to Google, and would include no customer data.

After Google submitted its bid, a union of flight attendants filed an objection, saying that although the deal’s terms would result in personal data being removed, the dataset would maintain “referential integrity,” meaning that links between one type of data and another would be kept. That integrity is what makes the data useful for training AI, but the union said it may also allow for even anonymized information about its members to be reconstructed. The group is seeking additional privacy protections; a judge is set to rule on the matter on Sept. 9.

Why Google Bid $10 Million for a Failed Airline’s Data

Developers are increasingly looking to the ocean as a potential site for artificial intelligence infrastructure, with the aim of building underwater data centers that could enhance energy efficiency, improve cooling methods, and reduce the usage of fresh water and land compared to traditional facilities located on land.

My research delves into the effects of emerging technologies on society, organizations, and the environment, particularly focusing on artificial intelligence and the data centers that are essential for its development. Underwater data centers present a promising strategy for bolstering AI advancements.

However, relocating servers beneath the waves doesn't eliminate existing environmental challenges associated with data centers, including high energy consumption and carbon footprints. Furthermore, it raises concerns regarding the protection of marine ecosystems, prompts regulatory questions, and complicates maintenance and expansion efforts. The future viability of underwater data centers as sustainable alternatives hinges on addressing these economic, technical, and environmental hurdles.

In September 2025, an underwater data center was in the works at a Chinese shipyard, marking a new chapter in ocean-based AI infrastructure.

The exploration of submerged data centers began in earnest with Microsoft’s 2015 initiative to investigate their potential, benefits, and challenges. Notably, the company established a waterproof data center on the seabed near the Orkney Islands in Scotland in 2018, which housed 864 servers connected to the mainland via an undersea cable.

After two years of monitoring, Microsoft found that the failure rate of servers in its underwater facility was about one-eighth that of similar land-based systems. While the company continues to explore the reasons behind this anomaly, the current hypothesis suggests that the sealed underwater setting minimizes exposure to oxygen, humidity, temperature shifts, and physical disturbances during maintenance.

Despite these promising findings, Microsoft discontinued the project in 2024 and opted against further underwater facilities. Although the company did not specify the reasons, analysts suggest that regulatory issues, environmental permitting, and the need for quicker upgrades were likely factors. This pivot means that Microsoft is currently concentrating on expanding its land-based data centers, which are easier to scale and maintain.

Nevertheless, innovation continues elsewhere. China unveiled what could be the world’s first wind-powered underwater data center situated in Shanghai, which opened in June 2025 and commenced full commercial operations by May 2026. This $226 million facility employs seawater for cooling, avoiding the need to chill freshwater and resulting in at least 30% lower energy consumption compared to conventional data centers. Offshore wind turbines also contribute to reduced reliance on fossil fuels and lower carbon emissions.

In Japan, a novel approach is being tested with data centers placed on floating platforms that utilize seawater for cooling, optimizing conditions for solar and wind energy production, and minimizing land use. A floating data center in shipping containers commenced operations near Yokohama in 2025, powered by solar panels on the same platform, with battery storage for energy management. This trial is scheduled to run until March 2027.

Singapore is also advancing its floating data center capabilities, with infrastructure company Keppel beginning construction on a four-story floating facility in 2026, expected to be operational by 2028. This center will utilize seawater for enhanced cooling while preserving treated water resources and making efficient use of Singapore's available land.

Furthermore, Ulsan, South Korea, has embarked on plans for a large underwater data center set to accommodate over 100,000 servers, aiming to consume 30% less energy through seawater cooling techniques.

In Maine, DeepGreen Western Passage is proposing an innovative submersible AI data center in the Bay of Fundy. This facility will be powered by tidal turbines that harness the area's strong tidal currents.

Even land-based data centers are beginning to leverage seawater cooling, as seen with Portugal's SIN01 AI data center in Sines, which uses Atlantic seawater to cool its systems before returning it to the ocean.

The issue of land scarcity for AI is driving data centers to be located in the ocean.

A moody underwater scene: sleek, shipping-container style data center pods anchored to the ocean floor, shafts of warm sunlight piercing the water, their server lights glowing softly in the deep. Cables trail up toward a small research vessel on the surface, with kelp swaying around the pods—visualizing AI’s land scarcity forcing data centers into the ocean.

AI’s land problem is pushing data centers into the ocean - Fortune

- Developers are exploring underwater data centers to improve energy efficiency and reduce land use.  
- Microsoft discontinued its underwater project due to regulatory concerns and maintenance challenges.  
- Global initiatives include wind-powered, floating, and tidal data centers, using seawater for cooling.  

Some developers are now exploring the ocean as a new location for AI infrastructure in the hopes that underwater data centers could improve energy use and cooling efficiency while using less fresh water and land area than onshore buildings.

My research focuses on the societal, organizational and environmental implications of emerging technologies, particularly artificial intelligence and the digital infrastructure – including data centers – that supports its development and deployment. I see underwater data centers as a promising new approach for supporting the growth of AI.

But moving servers into the ocean does not make other underlying environmental challenges such as energy consumption and carbon emissions disappear. And it creates new concerns about harm to the marine environment, as well as questions about how these data centers can be regulated – and how companies can maintain and expand them if needed. Whether underwater data centers can become sustainable alternatives to traditional data centers depends on solving these economic, technical and environmental problems.

An underwater data center is under construction in a Chinese shipyard in September 2025. CN-STR / AFP via Getty Images

The rise of ocean-based AI infrastructure

In 2015, Microsoft launched a research project to explore the feasibility, benefits and challenges of underwater data centers. Part of that effort included setting up a waterproof data center on the seafloor near Scotland’s Orkney Islands in 2018. It contained 864 servers and was connected to shore by an underwater cable.

After two years, Microsoft reported that the servers in the underwater data center failed at about one-eighth the rate of servers in comparable land-based data centers. The company is still studying the possible reasons but hypothesizes that in a sealed underwater environment the equipment is less exposed to oxygen, humidity and temperature fluctuations – as well as less jostling from people working to replace broken components.

However, Microsoft ended the project in 2024 and chose not to build more underwater data centers. The company didn’t say why, but others’ analyses suggest the reasons could include regulatory concerns, including the need for environmental permits, as well as a desire for faster upgrades and replacements for the computer equipment inside.

Instead the company has focused on land-based data centers, which can be larger and easier to expand, and also easier to access to repair or replace equipment.

Others have moved ahead, though. China built what may be the world’s first wind-powered underwater data center in Shanghai. The facility launched in June 2025 and began full commercial operations in May 2026.

The US$226 million project uses seawater as a coolant rather than have to refrigerate fresh water, reducing the electricity required to cool the computers. It uses at least 30% less electricity than traditional data centers, and offshore wind turbines reduce reliance on fossil fuels and cut the data center’s carbon emissions.

Japan is testing a different approach: Data centers housed in containers on floating platforms at sea can use seawater for cooling, have unobstructed conditions for solar panels and wind turbines, and reduce demand for land. In 2025, a data center in shipping containers opened on a floating platform near Yokohama. Its power comes from solar panels installed on the same floating platform, with batteries providing energy storage. The test will continue through March 2027.

Singapore is also moving toward commercial-scale floating data centers. In 2026, infrastructure company Keppel began building a four-story floating data center, scheduled to open in 2028. The project will use seawater for cooling, reducing reliance on treated water and improving cooling efficiency. And the fact that it floats means it won’t take up any of Singapore’s limited land availability.

In 2025, Ulsan, South Korea, began planning an underwater data center that could house more than 100,000 servers and use 30% less power than land-based centers by using seawater for cooling.

In Maine, DeepGreen Western Passage has proposed a submersible AI data center in the Bay of Fundy, powered by tidal turbines designed to harness the area’s strong tidal currents.

Land-based data centers could also take advantage of seawater cooling. In Portugal, the SIN01 AI data center in Sines uses seawater from the Atlantic to cool its servers before returning it to the ocean.

AI’s land problem is pushing data centers into the ocean

Good morning. It has now been eight months since Nvidia announced its significant $20 billion acquisition of chip innovator Groq, which is dedicated to enhancing the efficiency of AI models post-training. The initial steps toward commercializing this venture have begun, and its outcomes may provide valuable insights into how well AI agents in corporate environments meet anticipated results.

What exactly is Groq? To provide some context, Nvidia's graphics processing units (GPUs) have long been the preferred choice for training expansive language models, a key component of the AI landscape in recent years. However, the emphasis is shifting toward the rapidly expanding field of inference, which involves leveraging these trained models to tackle real-world challenges and deliver solutions to users.

Nvidia's Groq Chip Will Influence AI Agent Functionality

Sunlight streams through a bright engineering workshop, illuminating a wooden bench where a green circuit board holds a gleaming chip labeled “Groq.” From the chip’s core, fine metallic tendrils and soft golden data streams rise and coalesce into a poised robotic arm that points toward code on a nearby laptop. A large Nvidia logo is embossed on a metal plaque behind, with scattered tools and hand-drawn schematics strewn around, evoking hands-on innovation and the promise of next-generation AI usability.

Nvidia’s Groq Chip Will Shape AI Agent Usability - WSJ

- Nvidia's $20 billion deal with Groq aims to enhance AI model usage post-training.  
- The collaboration's success could impact enterprise expectations for AI agents significantly.  
- GPUs remain key for training AI, while inference gains importance in current applications.

Groq founder Jonathan Ross and Nvidia CEO Jensen Huang, in tie, have become closer. David Paul Morris / Bloomberg; Tom Williams / CQ Roll Call / ZUMA Press

Good morning. It’s been eight months since Nvidia announced its $20 billion deal with chip upstart Groq, which is focused on speeding up the use of AI models after they have been trained. The commercialization of that deal has started, and its success may tell us a lot about the extent to which AI agents in the enterprise live up to expectations.

What’s a Groq? First, a bit of background. Nvidia’s graphics processing units, or GPUs, have been the go-to chips for training large language models. Training has been the main AI application for the past few years. The focus is increasingly on the quickly growing realm known as inference, or the use of trained models to solve problems and get things done for users.

Nvidia’s Groq Chip Will Shape AI Agent Usability

The Mac mini and Mac Studio serve different roles within Apple's desktop lineup; however, they both share a newfound popularity for local AI inference and software development. This surge in interest stems from their unified memory architecture combined with high-performance CPUs and GPUs built into their systems-on-a-chip.

Apple recently unveiled updated versions of both desktops, introducing two new processors: the M6, marking the debut of Apple’s first 2nm chip in the M-series for Macs, and the M5 Ultra, which now stands as the most powerful chip in the series, particularly well-suited for AI tasks.

These updates do not bring substantial new features; rather, they primarily offer an upgrade in specifications. Nonetheless, Apple is clearly positioning these refreshed models to highlight their capabilities for specific applications—use cases that were not initially contemplated during the design of earlier versions.

The increased interest in production inference surged following the release of macOS 26.2 last December. According to Apple, this update enabled "low-latency communication between Thunderbolt 5 hosts for applications such as distributed AI inference using MLX." Thunderbolt 5 is an ultra-fast wired connection, while MLX serves as an open-source framework aimed at optimizing machine learning workflows to leverage the M-series chips' unified memory effectively.

Since the update, both amateurs and professional developers have been connecting multiple Mac minis or Mac Studios to execute inference on expansive local language models, which far exceed the capabilities of a single mainstream device. This approach provides a viable alternative to high-end specialized hardware, particularly those reliant on dedicated Nvidia GPUs.

Apple's latest desktop computers are tailored for local AI development.

A bright, sunlit modern workspace by a large window: a sleek Apple desktop and monitor sit on a minimalist white desk, the big Apple logo prominent on the case. On the screen, streams of code transform into a glowing 3D neural‐network mesh rising off the display. Soft natural light casts gentle shadows across a potted succulent and an open notebook with AI algorithm sketches, evoking the promise of local AI development on Apple hardware.

Apple’s new desktop computers are designed specifically for local AI development - Ars Technica

- Apple announced new Mac mini and Mac Studio models with upgraded M6 and M5 Ultra chips.  
- These updates focus on improving AI inference and software development capabilities on both systems.  
- macOS 26.2 enhanced Thunderbolt 5 communication, boosting local AI workloads for developers.

The Mac mini and Mac Studio occupy two distinct points in Apple’s lineup of desktops, but lately, they’ve had something in common: They’re popular for local AI inference and software development thanks to the advantages of their unified memory architecture and the fast CPUs and GPUs on their systems-on-a-chip.

Today, Apple announced new iterations of both desktops, along with two new chips: the M6, the first 2nm chip in Apple’s M-series lineup for Macs, and the M5 Ultra, now the most powerful chip in the lineup for most things—especially AI workloads.

There aren’t any major new features for either machine. This is just a specs bump. But based on how Apple is presenting these refreshes, they’re leaning hard into those use cases, which weren’t even a thought when earlier iterations were first engineered.

The devices’ popularity for production inference took off after macOS 26.2 shipped last December. According to Apple’s release notes, 26.2 enabled “low-latency communication between Thunderbolt 5 hosts for use cases including distributed AI inference using MLX.” Thunderbolt 5 is a very fast wired data connection, and MLX is an open source array framework designed to help machine learning workflows take full advantage of the M-series chips’ unified memory.

Since then, both hobbyists and professional developers and researchers have been essentially daisy-chaining Mac minis or Mac Studios to run inference on local large language models that are much bigger than anything that could run a single mass-market device—providing an alternative to ultra-beefy specialized hardware featuring specialized Nvidia GPUs.

Investment specialist discusses his view on why he believes AI expenditures do not contribute to inflation.

More from AI

India's Ringg receives support from Peak XV as it advances voice AI beyond phone calls.

Amazon service previously referred to by Bezos as 'artificial artificial intelligence' is closing down.

Is Apple's New Mac Ultra Capable of Substituting Your $200 Monthly AI Coding Expense?

BofA recommends purchasing these 16 growth stocks to reduce reliance on the AI market.