Nearly four years into the surge of artificial intelligence, major technology firms continue to make ambitious commitments about what lies ahead. However, this optimism comes at a steep financial cost, leading to significant cash burn.

According to Goldman Sachs, AI expenditure among the prominent tech players is anticipated to hit $765 billion this year, with projections suggesting it could escalate to nearly $1.2 trillion by 2027. Amazon recently increased its capital spending forecast to $220 billion, the highest among its peers within the hyperscaler category.

In a concerning trend, Amazon also registered a negative free cash flow of $7.6 billion over the past year, following Meta's report of a staggering 91% decline in cash generation compared to the previous year. Moreover, Alphabet announced that it experienced negative cash flow for the first time in its history, a significant turn of events for one of the industry's most lucrative entities.

During an earnings call, Alphabet's CFO Anat Ashkenazi conveyed to analysts that free cash flow would continue to face challenges as the firm capitalizes on the burgeoning "AI opportunity."

As the tech earnings season approaches its conclusion—Nvidia is set to announce its results on August 26—it has become evident that investments in AI are significantly impacting financial statements, despite industry leaders persistently emphasizing the prospective advantages of their extensive investments in new data facilities and the associated hardware.

A primary factor contributing to the rising costs is the ongoing memory shortage, driven by an insatiable demand for AI processors that are heavily dependent on memory produced by a limited number of suppliers.

During last week's earnings call, Tesla CEO Elon Musk labeled memory prices as "insane," while Amazon CEO Andy Jassy attributed part of his company's increased capital expenditure guidance to the "inflated price" of memory chips.

In contrast, Apple, which has scaled back its spending relative to its Big Tech counterparts, finds itself particularly vulnerable to the memory shortage since this component is essential for all consumer devices. The company has already implemented price increases on its Macs and iPads, and many analysts predict that iPhone prices will also rise later this year.

Rising memory expenses and decreasing funds: The cost of tech's AI expansion is escalating.

A dimly lit, moody server room: racks of humming machines fade into shadows while a cracked-open vault in the center drips a few coins and crumpled bills onto the floor. Beside it, an oversized stack of glowing memory modules towers under a single flickering red light, a digital price counter above climbing relentlessly. Deep shadows and warm amber highlights emphasize the strain of dwindling cash and soaring memory costs.

Dwindling cash and soaring memory costs: Tech's AI buildout has ballooning price tag - CNBC

- AI spending among major tech firms is expected to reach $765 billion this year.  
- Amazon reported negative cash flow of $7.6 billion over the past 12 months.  
- Rising memory prices are impacting overall costs and leading to product price hikes.

Almost four years into the artificial intelligence boom, the world's biggest tech companies are still making grand promises about the future. The problem is, they're burning through their cash in the process.

AI spending among the megacaps is projected to reach $765 billion this year, before rising to nearly $1.2 trillion in 2027, according to Goldman Sachs . Amazon boosted its capital spending forecast for the year on Thursday to $220 billion, the highest among the four hyperscalers.

Amazon also reported negative free cash flow for the trailing 12 months of $7.6 billion, a day after Meta disclosed a 91% drop in cash generation from a year earlier. Last week, Alphabet said cash flow turned negative for the first time on record, a stunning development for one of the most profitable companies on the planet.

Alphabet finance chief Anat Ashkenazi told analysts on the earnings call that free cash flow will remain under pressure as the company seizes on the "AI opportunity."

With tech earnings season largely wrapping up this week — Nvidia is set to report on Aug. 26 — it's become readily apparent that AI investments are distorting balance sheets, even as industry leaders continue to tout the future benefits of their mammoth bets on new data centers, and the chips and systems that populate them.

One big reason that costs are rising more than previously expected is the memory crunch, caused by insatiable demand for AI processors that rely on memory supplied by a small set of vendors.

Tesla CEO Elon Musk described memory pricing as "insane" on the automaker's earnings call last week, and Amazon CEO Andy Jassy said the "inflated price" of memory chips drove his company's capex guidance higher.

Apple , which is spending far less than its Big Tech peers, is particularly susceptible to the memory crisis because the technology is a key piece of every consumer device. Apple has already raised prices on Macs and iPads, and many analysts expect iPhone price hikes later this year.

Dwindling cash and soaring memory costs: Tech's AI buildout has ballooning price tag

In Pepper Pike, Ohio, Superintendent Lynn Campbell reassured parents in the Orange City School District that the rollout of artificial intelligence (AI) in schools will be implemented thoughtfully and at a measured pace. During a recent meeting of the Orange Board of Education on August 24, Campbell clarified popular misconceptions, stating, “We are not just opening up AI.” He emphasized that at the elementary level, particularly K-5, AI will be introduced in a way that supports teachers in curating their lessons, rather than being an overwhelming presence in the classroom. The deeper integration of AI into high school assignments will come later, as students will have more opportunities to utilize guidance and structure provided through these technologies.

The discussion about AI took place during a review of professional development initiatives for staff that occurred from August 17 to 18, right before students in grades 1-12 returned to school on August 20. Heather Miller, the new director of curriculum and instruction, highlighted the use of the AI platform MagicSchool, which is designed specifically for K-12 education to enhance lesson planning, grading, and classroom interactions. During the same professional development session, staff received insights from Kelly Fishman, the regional director of the Anti-Defamation League (ADL), on discrimination definitions and bias, marking the first of three training sessions for the year.

Looking ahead, Campbell mentioned that the upcoming session in November will delve into the implications of AI and social media as potential drivers of bias and prejudice, and how educators can proactively address these issues in their teaching. “AI brings with it a lot of greatness and a lot of concerns at the same time,” he reflected, expressing enthusiasm for developing staff proficiency with the MagicSchool platform.

Additionally, during these professional development days, staff were introduced to the Marzano framework for High Reliability Schools by Melanie Lewis Magee. The school board recently approved a contract with Marzano to further enhance professional development for educators and administrators.

On a separate note, resident Liz Gillmore voiced her concerns during the board meeting regarding how the district processes eligibility decisions related to a federal law protecting students with disabilities, specifically referencing Section 504. This law is designed to ensure that students with disabilities are not discriminated against and have access to the same educational opportunities as their non-disabled peers. Gillmore emphasized that good academic performance does not necessarily indicate that a student is unaffected by a disability and insisted that students should not be required to struggle before receiving necessary support. 

Gillmore urged the board to foster trust by promoting an environment where parents feel heard and respected, even when disagreements arise. She stressed the importance of open communication and constructive dialogues between families and the school, especially concerning accommodation requests.

In other developments, board President Beth Wilson-Fish announced the recent passing of Anita Tenney, a former teacher in the district who contributed significantly to the community and education. Tenney, a Pepper Pike resident, passed away on August 20 at the age of 76 and was remembered for establishing a successful early Spanish program at Moreland Hills Elementary School.

Lastly, the board approved a one-year contract for Julianne Cochran as an intervention specialist at Moreland Hills Elementary, effective August 17, and accepted the resignation of part-time bus driver Sabeen Chromjak-Harris, effective August 7.

Superintendent: Orange Schools taking initial steps into the realm of AI

A warm sunlit school corridor bathed in natural light, with orange-accented banners reading “Orange Schools” hanging overhead. In the center, a small, friendly robot taking its very first baby steps on polished tile, its mechanical hand outstretched for balance. At the edge of the frame, a lone superintendent silhouette in business attire holds a clipboard, watching proudly as the robot approaches a door labeled “AI Lab.”

Superintendent: Orange Schools entering into world of AI with ‘baby steps’ - Cleveland.com

- AI integration in the Orange City School District will be gradual and teacher-curated.  
- Parent concerns address eligibility decisions for students with disabilities under federal law.  
- The district hired Julianne Cochran as an intervention specialist after accepting a resignation.  

PEPPER PIKE, Ohio – Parents in the Orange City School District can rest assured that integration of artificial intelligence (AI) into the district’s schools will be gradual, Superintendent Lynn Campbell says.

“We are not just opening up AI,” Campbell told the Orange Board of Education Monday (Aug. 24).

“I know there’s a misconception that all the schools are going to use AI.”

At the K-5 level, AI integration is “really about the teacher curating specific things to show kids what it is,” Campbell said.

“It won’t be until high school that they will be pulling more into their assignments what guidance and structure they’re allowed to use,” he said.

“Who knows where it will go?

“We’ve got into it with baby steps, and we’re going to continue that because it impacts everything we do, or it stands to.”

The topic of AI was discussed during a review of the district’s professional development for staff that took place at the start of the school year Aug. 17-18.

Students in grades 1-12 began classes Aug. 20.

Heather Miller, the district’s new director of curriculum and instruction, said on Aug. 18, staff members looked at AI integration with MagicSchool “as another resource for staff.”

MagicSchool is an AI platform built specifically for K-12 schools, teachers and students to simplify lesson planning, grading and classroom activities.

Also on Aug. 18, staff members heard a presentation by Kelly Fishman, regional director of the Anti-Defamation League, which Miller said covered the common definitions of discrimination and addressed bias.

It was the first of three professional development sessions for district staff designed by the ADL for this school year.

Campbell said he believes the second session with the ADL in November will focus more on “how AI and social media are being used as tools and platforms to promote biases and hate and how our staff can get in front of that.”

“AI brings with it a lot of greatness and a lot of concerns at the same time,” he said.

“I’m excited as we continue to develop our staff’s knowledge of MagicSchool.”

On Aug. 17, Miller noted that staff members heard from Melanie Lewis Magee, an associate for Marzano Resources, about Marzano High Reliability Schools.

In 2024, the school board approved a contract with Marzano High Reliability Schools to provide additional professional development for staff members and administrators.

The Marzano framework defines five progressive levels of performance that a school must master to become a “high reliability” school.

Miller said the district received positive feedback from staff on both days of professional development.

Parent addresses law related to disabilities

Also on Monday, Pepper Pike resident Liz Gillmore, a parent in the district, expressed concerns to the board about how eligibility decisions related to a federal law regarding students with disabilities are being made and reviewed within the district.

She was referring to Section 504, a federal civil rights law that protects students with disabilities from discrimination and guarantees them equal access to education.

Governed by the U.S Department of Education under Section 504 of the Rehabilitation Act of 1973, it requires public schools to give disabled students the same opportunities as non-disabled peers.

“Good grades do not mean a student is unaffected by a disability,” Gillmore said.

“In many cases, they reflect extraordinary effort, extensive support and hours of additional work to keep pace with their peers.

“A student should not have to fail before receiving support.”

Gillmore said she’s also concerned that accommodation requests are sometimes viewed by the district’s administration “through the lens of potential advantages on college entrance exams, such as SATs or ACTs.”

“Accommodations do not provide advantage,” she said. “Their purpose is to remove barriers.

“They exist to allow students to demonstrate their knowledge and abilities on equal footing with their peers.”

Gillmore said she was asking the board to consider this question: “How do we build and maintain trust with families in our school?”

“Trust is built when parents believe they have been heard, respected and treated as partners, even when there are disagreements,” she said.

“Every parent should feel comfortable asking questions, seeking clarification, presenting additional information or challenging a decision they believe may affect their child’s educational opportunities.

“Those conversations should not be viewed as adversarial,” she continued.

“They’re often drivers of the very same goals shared by educators and administrators, ensuring that qualifying students receive the support necessary to reach their potential.”

A parent should never feel there are consequences for advocating for their child, Gillmore said.

“They should never feel that asking difficult questions will negatively affect their relationship with the school or create adverse consequences for their student,” she said.

“They should never feel that utilizing the review processes available to them will be viewed as disruptive, rather than constructive.

“As board members, your role is not to decide individual eligibility cases,” she continued.

“Your role is to ensure that the system, processes, culture and those responsible for administration promote fairness, consistency, transparency and trust.”

Gillmore said her request was simple: to review the district’s practices surrounding Section 504 eligibility determinations and appeals, “ensuring that decisions are based on the totality of evidence and that legitimate accommodation requests are evaluated on their merit and not through assumptions about testing advantages.”

Former Orange Schools teacher dies

In addition, board President Beth Wilson-Fish shared that Anita Tenney, a former Orange Schools teacher, died recently.

“I learned of Anita’s passing last week, and it hit our community very hard,” she said.

Tenney, a Pepper Pike resident, died Aug. 20. She was 76.

Wilson-Fish noted that Tenney was an Orange Community Education and Recreation instructor as well as a teacher at Moreland Hills Elementary School.

She credited Tenney with piloting the elementary school’s Spanish program around 2009 or 2010.

“The PTA brought Anita on to kind of try a Spanish program for our kids, beginning in the younger years,” she said.

“Obviously, her work there for that one year proved to be very successful, as the program has continued on.”

Wilson-Fish added that Tenney was a longtime community member and that her two daughters, Rebecca and Rachel, are graduates of Orange High School.

“She’ll be a true legacy here to us at Orange,” she said.

Tenney dedicated much of her professional life to education, with a particular passion for the Spanish language, according to her obituary on the Cleveland Jewish Funerals website.

A native of New York, she earned a master’s degree in Latin American literature from New York University and spent many years tutoring students of all ages in Spanish, the obituary stated.

A funeral service for Tenney was held Monday at Cleveland Jewish Funerals in Warrensville Heights.

New teacher hired

In other action, the board approved a one-year contract for Julianne Cochran, intervention specialist at Moreland Hills Elementary School, effective Aug. 17.

It also accepted the resignation of Sabeen Chromjak-Harris, part-time bus driver, effective Aug. 7.

Superintendent: Orange Schools entering into world of AI with ‘baby steps’

At the 2025 Icons of Porsche Festival, held at The Slab in Dubai on November 22, 2025, Porsche announced a significant five-year partnership with Tata Consultancy Services (TCS), valued at 1.25 billion euros (approximately $1.46 billion), focused on advancing artificial intelligence initiatives.

As part of this agreement, TCS will take over Porsche's IT consulting division, MHP, in a transaction worth 320 million euros. This strategic move is designed to integrate Porsche's deep automotive knowledge with TCS's robust digital and AI expertise. Michael Leiters, Porsche's chairman, emphasized in the announcement that this collaboration aims to enhance the carmaker's innovation, operational efficiency, and competitiveness in an era that increasingly prioritizes data and software in mobility solutions.

This partnership comes on the heels of TCS securing a range of AI-driven business transformation agreements throughout the year. The company reported an annualized revenue of $2.6 billion from AI activities in the June quarter, reflecting a 13.6% increase compared to the previous quarter.

Porsche signs $1.5 billion agreement for AI implementation with India's top IT services company Tata Consultancy.

A sleek silver Porsche sports coupé parked on a gleaming plaza under warm natural light, a modern Tata Consultancy Services office tower rising behind it with a large TCS logo on its glass façade, an Indian flag fluttering atop a nearby flagpole, subtle abstract AI circuitry patterns etched into the marble pavement to hint at the $1.5 billion AI deployment deal.

Porsche inks $1.5 billion deal for AI deployment with India’s largest IT services firm Tata Consultancy - CNBC

- Porsche signs a five-year €1.25 billion AI contract with Tata Consultancy Services.  
- TCS acquires Porsche's IT consulting business, MHP, for €320 million.  
- Partnership enhances Porsche's innovation and competitiveness in the digital mobility landscape.  

General atmosphere at the 2025 Icons of Porsche Festival at The Slab on November 22, 2025 in Dubai, United Arab Emirates.

Porsche has signed a five-year contract worth 1.25 billion euros ($1.46 billion) for artificial intelligence with India's largest IT services firm, Tata Consultancy Services.

As part of the deal, TCS will acquire Porsche's IT consulting business, MHP, for 320 million euros, the company said in a release.

This will combine Porsche's "automotive expertise with TCS's digital and AI capabilities," Michael Leiters, chairman of Porsche, said in the press release. The partnership will boost the European sports-car maker's innovation, efficiency and competitiveness "in an increasingly data- and software-driven world of mobility."

Since the start of the year, TCS has secured multiple AI-led business transformation deals and has said its annualized AI revenues were $2.6 billion in the June quarter, up 13.6% from a quarter earlier.

Porsche inks $1.5 billion deal for AI deployment with India’s largest IT services firm Tata Consultancy

Access your account to share your thoughts on videos and participate in the excitement.

Rep Chip Roy emphasizes the need for an important discussion in Washington regarding AI.

A wood-paneled congressional hearing room bathed in warm natural light, Representative Chip Roy standing at a podium before a single American flag, shadowed senators seated behind him as a large screen glows with the letters “AI,” their serious expressions hinting at a pivotal debate.

There needs to be a serious conversation in Washington about AI: Rep Chip Roy - Fox Business

- Users can now log in to share comments on videos.  
- The platform encourages community engagement through interactive features.  
- Participation enhances the overall user experience and content enjoyment.

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

There needs to be a serious conversation in Washington about AI: Rep Chip Roy

During the World Robot Conference held in Beijing on August 21, 2026, both local attendees and international visitors gathered to explore the latest advancements in robotics.

At the conference, the founder of Unitree highlighted a significant hurdle facing humanoid robots: the acquisition of human-like skills. While artificial intelligence plays a role, the search for a definitive solution continues.

Greetings from Beijing, I'm Evelyn, and I’m here to share insights from the local business landscape as part of The China Connection.

QJ Robots, an innovative startup focused on enabling robots to grasp human skills, has received support from Temasek, as revealed by founder and CEO Haichuan Gao. Their AI framework is designed to enhance robots' ability to anticipate their environment and execute tasks with greater efficiency.

QJ Robots reports impressive earnings of over 100 million yuan (about $15 million) through the application of its technology across more than 100,000 robots, primarily utilized by companies in mainland China. This substantial user base allows the startup to gather extensive data each month, according to Chief Technology Officer Tianren Zhang. However, he emphasized that the current data pool is insufficient for optimal model training.

Both Zhang and Gao earned their PhDs in automation from Tsinghua University, where their three-year-old venture originated within the school's brain computing research center. Their work aligns with the tech sector's increasing emphasis on creating world models, which seek to better replicate the dynamics of the physical environment, as opposed to the predominantly textual focus of large language models like ChatGPT.

To foster AI integration in industry, Beijing is actively encouraging major corporations to enhance their technological capabilities, an opportunity that startups are keen to exploit.

Another Tsinghua graduate, Chao Yu, established the robotics firm Lumos two years ago, with partnerships that include Mitsubishi Electric. Lumos has also initiated a center aimed at amassing training data. Earlier this year, Yu recognized that factories and other enterprises needed a system to translate data and AI directives into actionable tasks. As a result, Lumos launched the NexCore platform last week, designed to interface with various types of robots, not limited to Lumos products.

The competition to create superior autonomous robotics solutions is heating up. Just a year ago, Ace Robotics was founded, with Wang Xiaogang, co-founder of SenseTime, taking on the role of chairman. Wang stressed that a successful humanoid robot must integrate hardware, models, and real-world data scenarios. He cautioned against relying on online videos for training, as they may introduce unrealistic visual effects into the model.

Ace Robotics' offerings include wearable sensors capable of accurately capturing human activity data. By the end of next year, Wang anticipates that the industry will amass enough data to trigger a pivotal moment, similar to the transformative launch of ChatGPT in late 2022 that reshaped business applications of AI.

However, it's uncertain whether the market can afford to wait until then. Following comments from Unitree's founder Wang Xingxing that tempered commercialization expectations for humanoid robots, the company's shares experienced a drop. Meanwhile, visitors at the conference expressed their skepticism about the technology, with many finding the performance of robots—such as those responsible for folding clothes or selling water—slow and awkward. One individual waiting to purchase a beverage remarked in Mandarin that patience was only extended to the robots, asserting, "A human would definitely be scolded if it took this long."

CNBC's The China Connection newsletter: AI firms are hurrying to assist robots in acquiring human skills.

A bright, sunlit tech workshop with warm natural light streaming through tall industrial windows. In the foreground, a humanoid robot sits at a wooden workbench, flipping through a manual titled “Human Skills,” while a human instructor guides its hand holding a paintbrush. Behind them, other robots practice handshakes and cooking with utensils on worktables. On one wall, a large CNBC peacock logo glows on a digital screen, with a small Chinese flag pinned beside it.

CNBC's The China Connection newsletter: Robots need help learning human skills. These AI companies are rushing to meet the demand - CNBC

- Unitree's founder discussed humanoid robots' challenges in learning human skills at the conference.  
- QJ Robots has generated over $15 million by enhancing robots’ predictive tasks with AI.  
- Industry competition is intensifying, with startups focused on improving data-driven robotic functionalities.  

The public and foreign visitors alike visit the World Robot Conference in Beijing on Aug. 21, 2026.

Unitree's founder reminded everyone at the World Robot Conference last week that the biggest challenge for humanoids is learning human skills. AI helps, but is there a secret sauce?

Hi, this is Evelyn, writing to you from Beijing. Welcome to the latest edition of The China Connection — a snapshot of what I'm seeing and hearing from local businesses.

QJ Robots, a startup that's helping robots learn human skills, has been backed by Temasek, founder and CEO Haichuan Gao told me.

The company's AI model focuses on helping robots predict the world around them and perform tasks more efficiently.

QJ Robots claims it's already made more than 100 million yuan ($15 million) from deploying the model to more than 100,000 robots, mostly by companies in mainland China.

That gives the startup a large base from which it collects data every month, Chief Technology Officer Tianren Zhang said. But it's not enough — he said more kinds of data are needed to train the model effectively.

Zhang and CEO Gao both have PhDs in automation from Tsinghua University, where their three-year-old company got its start in the school's brain computing research center.

It's all part of the tech industry's growing focus on world models, which aim to imitate the physical world more closely than ChatGPT-type "large language models" that focus on text generation.

One of Beijing's approaches is to encourage industrial giants to integrate and develop AI capabilities. It's a manufacturing ecosystem edge that local startups are tapping into.

Another Tsinghua alumnus, Chao Yu, launched industrial-focused robotics company Lumos two years ago. Its backers and partners include Mitsubishi Electric.

Lumos has launched its own center for collecting training data. But Yu said that earlier this year, he realized that for factories and other businesses to use more robots effectively, they needed a platform to translate data and AI instruction into task completion. The startup last week officially launched its NexCore system that claims can work with a range of robots, not just those from Lumos.

But the race to build the best autonomous robotics solution is intensifying.

Just one year ago, Ace Robotics launched with SenseTime Co-Founder Wang Xiaogang as the startup's chairman.

An effective humanoid robot solution, he said, must combine hardware, models and data with use-case scenarios. Real-world data is critical, Wang said, as using online videos for training can introduce unrealistic special effects into the model.

Ace Robotics' products include wearable sensors for precisely capturing data from human activity.

By the end of next year, Wang predicts the industry will have gathered enough data for a breakthrough moment, similar to how ChatGPT's launch in late 2022 transformed what businesses could do with AI.

It's less clear whether markets will be willing to wait that long. Unitree's shares tumbled after founder Wang Xingxing sought to lower expectations for humanoid commercialization, while the public visiting conference exhibition booths were often bluntly unimpressed.

Dancing robots aside, machines folding clothes or selling water moved slowly and stiffly. One person waiting to buy a bottle of water noted in Mandarin Chinese that people were only patient because it was a robot.

"A human would definitely be scolded if it took this long."

CNBC's The China Connection newsletter: Robots need help learning human skills. These AI companies are rushing to meet the demand

A recent study has revealed minimal overall changes in employment rates across different sectors, despite variances in how significantly they are affected by artificial intelligence (AI). Researchers examined the impact of AI on employment, finding that for young workers aged 22 to 25, jobs classified as heavily influenced by AI saw an approximate 11 percent decline in jobs since 2022. In contrast, positions considered minimally impacted by AI experienced a 10 percent growth in employment for this demographic during the same timeframe.

The analysis conducted by Brynjolfsson and colleagues indicates that while the overall economic effects of AI on employment appear muted, there is a widening gap specifically for entry-level positions between job sectors with significant AI influence and those without. In these AI-heavy fields, the primary issue seems to stem from diminished hiring rates rather than an uptick in layoffs or job resignations. This trend suggests that for younger workers, the effects of AI are reflected more in decreased job availability rather than lowered wages.

Furthermore, the researchers highlighted that not all jobs facing potential AI transformation are equally affected. Through their Economic Index, Anthropic classifies tasks into two categories: "automative," which fully replace human labor, and "augmentative," which enhance human capabilities in their work. Jobs such as accounting and reception have been identified as more vulnerable to automation, while roles like chief executives and registered nurses frequently employ AI to support human efforts rather than replace them.

The study emphasizes that fields susceptible to AI automation showcase the most significant drops in employment for entry-level roles. The findings indicate a trend where automation-focused applications of AI are displacing human labor, while roles that utilize AI to complement human work tend to maintain stable or even increasing employment figures.

Stanford study reveals AI is most affecting entry-level jobs.

A moody, dimly lit open-plan office at dusk: rows of empty cubicles with cardboard boxes and scattered resumes on desks. In the shadows looms a giant, circuit-patterned robotic hand pointing toward the exit. A few dejected young professionals (adults only) pack their belongings under flickering overhead lights, while a faint Stanford crest is plastered on a report lying on one desk.

AI is hitting entry-level jobs hardest, Stanford study finds - Ars Technica

- Overall employment across industries remains stable despite varying AI impact levels.  
- Employment for young workers in AI-impacted fields dropped 11% since 2022.  
- Jobs involving automation see lower entry-level hiring compared to those with AI augmentation.

When crunching the numbers economy-wide, the researchers found little to no difference in relative overall employment between the jobs judged most and least affected by AI on these metrics. When separating out workers aged 22 to 25, though, the researchers found that, since 2022, employment in the top 40 percent of “AI-impacted” jobs had fallen by about 11 percent. In the 60 percent of jobs with the least AI impact, by contrast, total employment for those young workers grew by 10 percent over the same period.

Brynjolfsson et al Across the entire economy, relative employment effects in more heavily AI-impacted fields are relatively muted. Across the entire economy, relative employment effects in more heavily AI-impacted fields are relatively muted. Brynjolfsson et al Brynjolfsson et al For entry-level workers, though, there is a large and widening gap between employment in fields that are or are not impacted by AI. For entry-level workers, though, there is a large and widening gap between employment in fields that are or are not impacted by AI. Brynjolfsson et al Across the entire economy, relative employment effects in more heavily AI-impacted fields are relatively muted. Brynjolfsson et al For entry-level workers, though, there is a large and widening gap between employment in fields that are or are not impacted by AI. Brynjolfsson et al

Digging deeper into the data, the researchers found that this phenomenon is mainly manifesting itself through lower hiring rates for entry-level workers in AI-impacted fields, rather than increased firings or employees quitting. They also found that the labor market effects among this age group were mostly seen in lower overall employment, rather than reduced pay rates.

But not all jobs that show potential for AI “disruption” are created equal, the researchers found. In its Economic Index, Anthropic differentiates between queries related to tasks that are “automative” (i.e., fully replacing work previously done by a human) or “augmentative” (i.e., helping human workers be more effective at tasks they are still needed for). By this measure, jobs like “accountants and auditors” and “receptionists and information clerks” were among those judged most susceptible to AI automation, while jobs like “chief executive” and “registered nurse” were among those using AI augmentation most often.

Unsurprisingly, jobs where AI automation is prevalent are the ones showing the worst relative employment levels for entry-level workers these days. “The findings are consistent with automation-oriented uses of AI substituting for labor while complementary uses are associated with flat or rising employment,” the researchers write.

AI is hitting entry-level jobs hardest, Stanford study finds

Close to 200 activists connected with President Donald Trump's "Make America Healthy Again" initiative are encouraging him to rethink his initiative aimed at constructing artificial intelligence data centers across the country.

There has been a noticeable bipartisan backlash against these resource-intensive facilities due to their detrimental effects on the environment and public health. Meanwhile, Trump and his supporters have been characterized as taking a villainous stance, insisting that the establishment of new data centers is unavoidable and dismissively suggesting that critics should be ignored.

It's particularly noteworthy that this opposition includes members of the MAHA movement, which has typically endorsed various policies that could jeopardize public health and potentially lead to increased illness among Americans. Clearly, the issue of AI data centers might be where they draw a line.

In a letter sent on Friday, these activists—operating under the banner "United We Eat"—urged Trump to retract an executive order that places coal power at the forefront of energy sources for these data centers.

"The executive order from April 2025 links the rising need for electricity stemming from AI data processing centers to the enhanced support of coal power," the letter stated. "It also tasked federal agencies with pinpointing areas where coal-based infrastructure could sustain AI data centers. This connection warrants immediate reevaluation."

The activists expressed their desire to voice their concerns to maintain the integrity of their stance:

MAHA activists call on Trump to change his stance on AI data centers.

A moody nighttime protest scene: adult activists clutch “Stop AI Data Hubs” placards outside a sprawling, dimly lit data center whose windows glow ominously in warm reds and ambers. In the distance atop the building’s roof, a lone silhouette of Trump stands at a podium. A single US flag billows above them under stormy clouds, casting dramatic shadows across the crowd.

Opinion | MAHA activists urge Trump to reverse course on AI data centers - MS NOW

- Nearly 200 Trump-aligned activists are urging a rethink of AI data centers nationwide.  
- Bipartisan backlash highlights environmental harms and public health impacts of these centers.  
- Activists call for reconsideration of coal power use in data center operations.  

Nearly 200 activists aligned with President Donald Trump’s “Make America Healthy Again” movement are urging him to reconsider his push to build artificial intelligence data centers nationwide.

There’s been growing bipartisan backlash to these resource-sapping centers over their environmental harms and negative impacts on public health. Meanwhile, the president and his allies have basically sounded like comic book villains with their rhetoric about the supposed inevitability of new data centers and the need to “roll over” critics.

So it’s telling that the opposition includes activists in a movement like MAHA that, on many fronts, has advocated for a host of policies that undercut public health and stand to make Americans sicker. Perhaps AI data centers are their red line.

In a letter dated Friday, the activists — under the banner “United We Eat” — pleaded with Trump to reverse course on an executive order that prioritizes the use of coal power to help fuel data centers.

“The April 2025 executive order on coal connected rising electricity demand from artificial intelligence data processing centers to expanded support for coal power,” the activists wrote. “It also directed federal agencies to identify regions where coal-powered infrastructure could support AI data centers. That connection deserves urgent reconsideration.”

The activists said they’re speaking out to avoid the appearance of hypocrisy:

On AI data centers, MAHA activists urge Trump to reverse course

This narrative is derived from a discussion with Chris Domanski, a 41-year-old business management and development director at Renau, a manufacturing plant located in California. Domanski resides in Thousand Oaks with his family. The following content has been edited for brevity and clarity.

According to Domanski, ChatGPT has significantly eased his workload by acting as a study companion for his children.

The motivation to utilize AI was ignited by the progress reports from his daughters, aged 8 and 6, which indicated that one was having difficulty with math while the other faced challenges in reading. While their performances weren't poor, they were not meeting the district's expectations.

Recognizing the need to support their educational growth, Domanski felt uncertain about his ability to effectively teach specific concepts, such as TouchMath, a unique math curriculum he was unfamiliar with.

In an innovative approach, he began taking photos of their assignments and using AI to clarify concepts. By leveraging ChatGPT's voice mode, he found that it could engage his daughters with quizzes, tailor questions to their needs, revisit misunderstood topics, and explain ideas using multiple methods—all without showing signs of frustration.

Initially, Domanski questioned whether the time spent setting up the AI assistance would be worthwhile. However, just five minutes into the process, his daughters were discussing "KPop Demon Hunters," prompting him to make some adjustments to their focus.

Through a mix of experimentation, the AI system became a valuable asset for their educational journey, improving their grades while saving time. Although he is uncertain how it will adapt to increasingly complex schoolwork, the current arrangement seems effective.

Domanski capitalizes on AI to ensure his daughters fully comprehend their study material.

With both girls preparing for weekly math and spelling exams, he routinely informs ChatGPT of their test requirements. By uploading images of their study sheets and instructing the AI to maintain focus amidst the inevitable distractions from his daughters, he crafts a structured learning environment.

He prompts ChatGPT with clear expectations like: "You are a tutor supporting grade schoolers with math and spelling studies. You will receive study guides and will continue to work with the students until they can confidently and correctly answer all questions."

To keep the lessons on target, he provides additional guidance, instructing ChatGPT to express interest in other topics but redirect the discussion back to the studies at hand. This ensures that if his daughters can quickly solve equations, such as determining 9+3, they’ll confidently record their answers during tests.

Chris Domanski works closely with both his children and ChatGPT to optimize their learning process.

Domanski remains present during these sessions to monitor both his daughters' responses and ChatGPT’s input, allowing him to promptly intervene if he notices any misconceptions.

AI has proven to be a resourceful tool in maintaining their attention, adjusting to their immediate thoughts, and repeating concepts multiple times without any feelings of frustration for the children. This adaptability has considerably reduced educational friction in their home.

Since introducing this approach, his daughters’ progress reports have improved from "slightly below satisfactory" to "above satisfactory" levels. While both have shown academic improvement, the real benefit lies in creating a more engaging and less confrontational study experience. Domanski feels more equipped to observe their challenges instead of solely administering questions from worksheets.

Additionally, he incorporates AI to digitize their school handouts into a more manageable calendar system.

Every Monday, Domanski reviews their take-home folders filled with school correspondence, scanning them to extract important dates, deadlines, and necessary supplies. ChatGPT processes this information and organizes it into actionable calendar entries, dramatically simplifying what would typically be a cumbersome task.

His prompt to the AI is straightforward: "You will receive images of take-home folders. Identify key dates and action items that require follow-up, then add relevant tasks to my calendar." ChatGPT efficiently exports these events as calendar files that Domanski can integrate into Outlook.

So far, ChatGPT has functioned reliably, eliminating the stress of last-minute project discoveries. The family now avoids late-night scrambles to prepare for assignments due the next day.

Lastly, the Domanski family enjoys the creativity that AI brings to their lives.

In their artistic endeavors, they often turn to ChatGPT for generating custom images for drawings. Initially asking for character mashups like Disney's Ariel and Elsa, they encountered copyright issues. However, after requesting more generic, open-source alternatives, they found a compromise everyone could enjoy.

Now, several times a week, both children create art featuring their preferred character combinations.

Ultimately, AI does not just create additional free time but rather enhances the quality of their existing time together. They can focus more on projects, artwork, games, or simply enjoying each other's company, minimizing distractions and the need to decipher instructions from scratch.

I created a voice-activated AI tutor for my kids to assist with spelling and math, freeing up time for more fun together.

A cozy living room bathed in warm morning sunlight: an adult sits on a soft couch holding a tablet showing colorful spelling letters and simple math equations, with gentle voice waveforms rising from the screen. Two small empty children’s chairs and backpacks sit nearby on a rug strewn with crayons and worksheets beside an open board game, hinting at playtime off-screen. The adult smiles, turning toward unseen laughter—natural light, warm tones, story of tech freeing up family fun.

I made my kids a voice-mode AI tutor to help them with spelling and math. It saves me time to have more fun with them. - Business Insider

- Chris Domanski uses ChatGPT to assist his daughters with studying math and reading.  
- AI helps manage school-related paperwork, organizing important dates and tasks efficiently.  
- Generative art prompts from ChatGPT encourage creativity, allowing for fun family drawing sessions.  

This as-told-to essay is based on a conversation with Chris Domanski, 41, the director of business management and development at Renau, a manufacturing facility in California. He and his family live in Thousand Oaks, California. This interview has been edited for length and clarity.

More than anything, ChatGPT has saved me time and drama by being a study partner for my kids.

What spurred this were the progress reports that my daughters, ages 8 and 6, got from school last year. I found out that one struggles with math, and one struggles with reading. Not that they're doing poorly, but maybe below where the district thought that they should be.

I knew we could try to spend more time on these areas. But at the end of the day, I thought I might not be the best person to explain a concept like TouchMath, a multi-sensory math curriculum, which I'd never heard of until I had to help the girls with homework.

Domanski's daughters, ages 8 and 6, use AI for homework help and customized drawing prompts. Chris Domanski

I decided to take a picture of their homework and have AI explain some concepts to them. Using voice mode, ChatGPT now quizzes them, varies the questions, revisits what they missed, and explains the same concept several different ways without becoming impatient.

Related video How AI takes a toll on the freelancers who train, test, and refine it

At first, I wasn't sure if the time it took to set this up was worth it. Five minutes in, they were talking to ChatGPT about "KPop Demon Hunters." I needed to revise it a little.

Through trial and error, it became something that really saved time and I think improved their grades. I'm not sure how it's going to look as things get more and more complicated with their schoolwork, but at least for now, it works well.

I programmed an AI tutor to make sure they truly understand the material

Domanski uploads his daughters' homework and has ChatGPT create custom quizzes to test their knowledge. Chris Domanski/ChatGPT

Both my daughters have weekly math and spelling tests at school.

In general, I tell ChatGPT what kind of test my daughters are preparing for, take a picture of the study sheet, and instruct it to stay focused — despite the inevitable interruptions by my daughters.

I prompt ChatGPT by saying something like: "You're a tutor helping grade schoolers study for math and spelling tests. You'll be given images of the study guides, and you'll work with the students until they can confidently and accurately answer all questions."

I also add some guardrails: "You will not be distracted if other topics come up. Express interest in those topics, make them relevant to what you're discussing, but keep it on the task at hand."

Many of the instructions are for ChatGPT to politely and firmly make sure my daughters are comfortable and confident in the material, not engaging just for the sake of engaging. If they're able to quickly say what 9+3 is, they're able to write that on a test.

An example of Domanski's instructions to ChatGPT. Christ Domanski/ChatGPT

I also make sure I'm present to hear both what they're saying and what ChatGPT is saying. If I hear that something is very clearly wrong, I can jump in and steer things back to the task at hand.

Overall, AI has been pretty efficient at maintaining their attention, adapting to how they are thinking in that moment, and repeating something for the sixth time without making it feel like punishment. It's surprisingly useful at absorbing some of that friction.

Since last year, their progress reports went from being a little "below satisfactory" to being "above satisfactory." Both of their grades have improved, but the more valuable result may be that studying has become more interactive and less adversarial. I am still involved, but I can spend more time observing where they are struggling instead of simply reading questions from a worksheet.

I also use AI to digitize their school handouts into a calendar

Domanski uses ChatGPT to scan paper handouts and transfer the info to his Outlook calendar. Chris Domanski/ChatGPT

AI also helps me manage the constant stream of information that comes with having two grade-school children.

Every Monday, I open their take-home folders and take pictures of school flyers, calendars, permission slips, and classroom notices, then use ChatGPT to extract dates, deadlines, supplies, and action items. It turns paperwork that would otherwise require another round of mental sorting into something much easier to track.

The prompt I use is: "You'll be given images of take-home folders. Please review this for key dates and things that require action or follow-up on behalf of anyone. Add relevant tasks to my calendar." ChatGPT then exports the scanned events as .ics files that I then import into a school calendar in Outlook.

At this point, ChatGPT hasn't really dropped the ball. It is really nice not to discover at 8 p.m. that there is a project due the next morning — and then spend the rest of the evening scrambling to figure out what we need.

AI frees me up to have more fun with my kids

Domanski also uses AI to generate custom images for his daughters to draw. Chris Domanski/ChatGPT

We also do quite a bit of generative art, asking ChatGPT to provide an image for my daughters to draw on their own.

Originally, we'd ask for something like Disney's Ariel and Elsa dancing underwater in a sandcastle. ChatGPT declined the request, citing copyright restrictions. I asked it to provide an open-source alternative, such as a generic red-headed mermaid and blonde ice princess. Everyone was happy.

Now, a few times a week, my daughters can draw or paint whatever character crossover they want.

AI does not so much create new free time as help us get much more out of the time we already have. We spend less time looking for things, figuring out instructions, or trying to understand a new concept from scratch, and more time actually doing projects, making art, playing games, or just having fun together.

I made my kids a voice-mode AI tutor to help them with spelling and math. It saves me time to have more fun with them.

AI systems developed using data from the Ukrainian battlefield are set to play a crucial role in safeguarding UK defense facilities, rail networks, and energy infrastructure from potential threats posed by protesters and foreign adversaries. This initiative comes from an agreement between the UK and Ukraine.

The scheme allows private firms to tap into the extensive data sourced from Ukraine’s Avengers AI lab, marking a pioneering collaboration in the UK's approach to defense technology. A trial will commence at a UK defense location, focusing on creating a model capable of identifying both protests and attacks from hostile entities. This involves utilizing AI-enhanced sensors embedded in fiber-optic cables to monitor various movements.

If the technology proves effective, the government anticipates its future application in airports, prisons, railways, and energy sectors.

The recent decision to proscribe the activist collective Palestine Action came in light of their July 2022 breach at RAF Brize Norton, which was a worrying lapse in security. The group managed to vandalize two military aircraft and leave without apprehension. Palestine Action has defended its actions by claiming it aimed to disrupt Britain’s involvement in actions they view as genocidal and criminal in the Middle East.

Private companies looking to engage with this data will need approval from the Ministry of Defence (MoD) and will utilize a secure platform for access, which has raised concerns among privacy advocates. Initial pilot programs have received backing from three British AI firms: Sintela, Mind Foundry, and Skyral. Sintela, based in Bristol, has previously entered a $35 million contract with the Trump administration focusing on surveillance at the US-Mexico border.

This collaboration not only allows for the incorporation of operational drone datasets and mission footage into British security frameworks but also includes insights into Russian sabotage tactics against vital Ukrainian systems. The comprehensive military data accumulated over the past four years offers a more nuanced view than previous models, which largely relied on publicly available information.

The UK government envisions employing AI models to swiftly assess threats to critical infrastructure, including railways and military bases, by using advanced fiber-optic cable systems capable of detecting distinct patterns of movement around sensitive areas. Sources from the MoD clarified that this monitoring includes identifying potential protests, not solely actions from foreign states.

Prime Minister Andy Burnham emphasized the opportunity for British firms to meld their AI innovations with battlefield insights from Ukraine, thereby enhancing national security while ensuring the protection of vital infrastructures across the nation.

Future initiatives could explore developing new AI chips for drones, robotics, and automated weapon systems aimed at bolstering defense capabilities.

Steven Murdoch, a professor of computer science at University College London, noted that while the underlying technology for movement detection via fiber-optic sensors is established, it's uncertain how the Ukrainian data will enhance its effectiveness. He acknowledged, however, the potential for discovering valuable insights related to emerging threats that may soon challenge Western Europe.

"What's valuable now could be significant for what other nations will confront down the line," he remarked, citing drones as a relevant example.

Defense Secretary Wes Streeting highlighted Ukraine's role in advancing defense innovation in its fight against Russian aggression. He asserted that leveraging the UK's scientific and technological prowess alongside Ukraine’s battlefield experience would bolster defense and security for both countries.

UK plans to utilize battlefield data from Ukraine to enhance AI training for safeguarding sensitive locations.

Dimly lit command center with flickering screens displaying Ukraine battlefield footage and live drone shots, streams of glowing code mapping a fortified site’s blueprint on a central terminal, a Union Jack draped on the wall, military gear casting long shadows in a tense, protective atmosphere

UK to use Ukraine battlefield data to train AI to protect sensitive sites - The Guardian

- UK and Ukraine will use AI models trained on battlefield data to enhance security systems.  
- Private companies can access Ukraine's AI lab data to develop new surveillance technologies.  
- Fiber-optic sensors will detect movements around sensitive sites to prevent potential attacks.  

AI models trained on Ukrainian battlefield data will be used to stop protesters and foreign states targeting UK defence sites, railways and energy plants under a deal struck between London and Kyiv.

Private companies will also be given access to the vast trove of data from Ukraine’s Avengers AI lab to help build new systems, the first agreement of its kind in the UK.

The scheme will be piloted at a UK defence site, building a model to identify protesters or a hostile-state attack. It uses AI-optimised sensors in buried fibre-optic cables to detect different kinds of movement.

The government has suggested the technology, if proven workable, could be used at airports, prisons, railways and energy infrastructure.

The controversial proscription of the activist group Palestine Action was sparked after the group breached RAF Brize Norton on scooters last June, a highly embarrassing security lapse as a result of which the group was able to deface two military aircraft with spray paint before escaping the site without being detained.

Palestine Action said at the time that it had targeted RAF Voyager aircraft used for transport and refuelling, and that “activists have interrupted Britain’s direct participation in the commission of genocide and war crimes across the Middle East”.

Private companies will be able to access the data once approved by the MoD on a secure platform, though the data sharing is likely to raise alarm among privacy campaigners.

Pilot projects have already been approved with three British AI companies – Sintela, Mind Foundry and Skyral. Sintela, a Bristol-based company that specialises in fibre-optic sensors, recently struck a $35m (£26m) deal with the Trump administration to provide surveillance on the US-Mexico border.

The partnership with the Ukrainian AI lab means new UK security systems can be trained on operational drone datasets, strike mission footage and flight profiles, likely to include Russian sabotage attempts against critical Ukrainian infrastructure.

The cache of four years’ worth of military data contains a far deeper picture than models outside Ukraine have previously been trained on, which had been mostly restricted to open-source data.

The government eventually hopes to use the AI models to predict quickly when infrastructure such as railways, the power grid or military bases might be coming under threat or stress from hostile actors.

That would primarily come from the laying of fibre-optic cables that are able to identify unique movements of people or vehicles around sensitive sites. Ministry of Defence sources said that would specifically include when a site might be targeted by protesters – not just hostile states.

The prime minister, Andy Burnham, said it would be an opportunity for British companies to combine their AI expertise with Ukraine’s battlefield data, which would “accelerate the development of capabilities that not only strengthen our national security but better protect our critical infrastructure and growth in every postcode here at home”.

Other potential projects could include new AI chips to power drones, robotics and automated weapons systems.

skip past newsletter promotion Free newsletter | Every weekday Sign up to First Edition Our morning email breaks down the key stories of the day, telling you what’s happening and why it matters Enter your email Sign up after newsletter promotion

Steven Murdoch, a professor of computer science at University College London, said technologies to detect movement from sensors in fibre-optic cables had been around for decades and that it was unclear if data from Ukraine would refine their capabilities.

“I haven’t heard anything that is very convincing, but there is a plausible scenario where they could see something that could be predicted easily and isn’t in any publicly known dataset,” he said.

Ukraine, however, has a lot of other data that may be valuable, especially related to new threats that countries in western Europe might face in the short to medium term.

“What’s valuable is something we’re seeing today which other countries will be worrying about in the future,” Murdoch said. “Anything to do with drones, for example.”

The defence secretary, Wes Streeting, said Ukraine was “pushing the frontier of defence innovation as they bravely resist Putin’s illegal invasion”.

“By combining the UK’s world-leading scientists, engineers and tech companies with Ukraine’s unique battlefield experience and data, we will strengthen both of our nations’ defence and security,” he said.

UK to use Ukraine battlefield data to train AI to protect sensitive sites

Recent findings from OpenRouter indicate that agentic AI workloads demand a staggering 15 times more tokens compared to standard chat requests. What accounts for this significant difference? 

Imagine an AI agent tasked with investigating a company for a potential investment. The agent must access financial databases, analyze relevant news and filings, utilize a sub-agent for peer comparisons, and model valuations—all while synthesizing this information into a cohesive recommendation. During this process, these agents and their sub-agents engage in continuous reasoning until the objective is fulfilled, creating a growing need for tokens. As each step builds on the last, effective handling of extensive context becomes crucial for the performance of agentic AI.

This escalation in token consumption is consistent across various applications of agentic AI, whether in software development, customer support, or in-depth research.

As the adoption of agentic AI expands across different sectors, it is vital for the underlying infrastructure to efficiently cater to this rising token demand.

New performance metrics reveal that NVIDIA’s Vera Rubin NVL72 systems provide up to 30 times greater throughput per megawatt compared to the NVIDIA GB300 NVL72 when handling agentic workloads. This inference throughput was assessed using the SemiAnalysis AgentX workload, which involved recorded real-world sessions of agentic coding, maintaining actual context growth, tool interactions, and sub-agent activations. For energy-limited AI factories, this improvement translates directly to a 30-fold increase in agentic tasks executed for the same energy consumption.

These preliminary findings regarding the Vera Rubin NVL72 highlight NVIDIA’s rapid advancements in innovation. Ongoing software enhancements promise to boost performance further across both the Vera Rubin NVL72 and GB300 NVL72 systems.

When discussing agentic workloads, the nature of the work is distinctly different from that of chat or document summarization, where the input-output sequences typically range from 1,000 to 8,000 tokens. In contrast, agentic sessions can see context accumulate through multiple steps, often resulting in hundreds of thousands of input tokens, with variations in both input and output lengths. Thus, performance measurement strategies need to evolve to reflect the entirety of the agent workflow, rather than focusing on a single inference request.

The following performance results are reflective of real-world coding scenarios with agentic AI.

The NVIDIA Blackwell platform showcases superior performance across various agentic models, including Kimi K3, MiniMax M3, GLM5.3, Qwen3.5, and DeepSeek V4 Pro. For instance, the GB300 NVL72 exhibits up to 15 times superior throughput per megawatt compared to the NVIDIA Hopper architecture when utilizing the DeepSeek V4 Pro model. This enhancement stems from the benefits of a larger GPU domain and code-signed software that significantly boosts inference efficiency.

Vera Rubin enhances this advantage, providing as much as 30 times higher throughput per megawatt compared to GB300 NVL72 on the same DeepSeek V4 Pro model. These initial results, derived from the SemiAnalysis AgentX workload and currently pending review from SemiAnalysis, do not yet account for Vera CPU performance in tool interactions.

Through technologies such as NVIDIA's DSX MaxLPS, power management across GPU, rack, and workload levels enables the provisioning of up to 40% more GPUs under the same power budget, pushing throughput per megawatt to new heights in large-scale AI environments.

Furthermore, the relationship between throughput per megawatt and the cost of token production is significant. The Vera Rubin NVL72 achieves costs that are up to 35 times lower per million tokens compared to GB300 NVL72, allowing agents to operate continuously at scale throughout the entirety of their customers' workloads. 

In power-limited AI factories, the throughput rate directly influences revenue, while the cost per million tokens dictates profit margins.

To achieve optimal performance for agentic AI, advanced inference optimizations are crucial. NVIDIA’s Vera Rubin NVL72 facilitates a wide array of such optimizations, thanks to a highly integrated design across all layers of the platform.

Disaggregated serving separates context processing from response generation, allowing each function to scale independently, while rate matching coordinates the token production speed between prefill and decode GPUs to maximize overall efficiency. Large-scale expert parallelism distributes expert sub-networks across the GPU domain, and distributed KV-caching expands memory access, ensuring previously processed context is readily available without needing to be recomputed. 

KV-aware routing directs incoming requests to GPUs that have relevant cached context, minimizing redundant calculations during extensive sessions. Additionally, fused CUDA kernels like MegaMoE streamline multiple computation and communication tasks into single processing passes, ensuring GPUs remain active rather than idle.

The enhanced fifth-generation Tensor Cores and third-generation Transformer Engine in NVIDIA Rubin GPUs accelerate the inference process significantly. Through NVFP4 quantization, model weights are compressed to 4-bit precision, optimizing both memory use and throughput while maintaining output quality.

The NVL72 scale-up domain, integral to the architecture of both Vera Rubin and Grace Blackwell, facilitates high-bandwidth, low-latency communication between GPUs—essential for methods like large-scale expert parallelism and distributed KV-caching. The purpose-built NVIDIA NVLink technology and sixth-generation NVLink Switches provide tenfold higher packet rates with threefold lower latency compared to standard Ethernet options.

NVIDIA's software ecosystem—including optimized CUDA kernels, TensorRT LLM inference runtimes, and NVIDIA Dynamo serving frameworks—is meticulously designed alongside the hardware to enhance inference operations.

While the current results illustrate the capabilities of the Vera Rubin NVL72, the comprehensive platform encompasses a seven-chip architecture that includes elements such as the NVIDIA Vera CPU, Groq 3 LPU, NVLink 6 Switch, BlueField-4 DPU, Spectrum-6 SPX, and ConnectX-9 SuperNIC, all specifically engineered for deploying agentic workflows at scale.

NVIDIA continues to prioritize extreme codesign principles in collaboration with its partners, ensuring that the Vera Rubin system is fully operational and scaling within the tech ecosystem.

NVIDIA Vera Rubin NVL72 Establishes a New Efficiency Benchmark for AI Agents with Up to 30x More Work Per Watt

A bright modern lab bathed in warm morning sunlight from floor-to-ceiling windows, spotlighting a gleaming NVIDIA Vera Rubin NVL72 chip mounted on a polished pedestal; golden sunbeams dance across its intricate circuits as faint wisps of energy-efficiency graphs swirl around it, with a large NVIDIA logo etched prominently on the pedestal.

Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents - NVIDIA Blog

- Agentic AI workloads demand 15x more tokens than simple chat requests due to complexity.
- NVIDIA Vera Rubin NVL72 achieves 30x higher throughput per megawatt than GB300 NVL72 systems.
- The architecture enables significant performance optimizations for efficient agentic AI workload processing.

According to OpenRouter data, agentic AI workloads consume 15x more tokens than a simple chat request. Why?

Consider what happens when an AI agent researches a company for an investment decision. The agent queries financial databases, searches news and filings, invokes a sub-agent to run peer comparisons and model valuations, then synthesizes everything into a recommendation. Agents and sub-agents keep reasoning until the task is done, driving increased token demand. With every step, the accumulated tokens become the input to the next, making long-context handling central to agentic AI performance.

The same pattern plays out across every agentic use case, from software development to customer service to deep research.

As agentic AI moves into production across industries, the infrastructure running it needs to meet that token demand efficiently.

New measured performance data shows NVIDIA Vera Rubin NVL72 systems deliver up to 30x higher throughput per megawatt than NVIDIA GB300 NVL72 on agentic workloads. NVIDIA measured this inference throughput data using the SemiAnalysis AgentX workload, consisting of recorded real-world agentic coding sessions, with actual context growth, tool calls and sub-agent spawning preserved. For power-constrained AI factories, that translates directly into 30x more agentic work for the same energy footprint.

These early results for Vera Rubin NVL72 demonstrate NVIDIA’s accelerated pace of innovation. With continuous software optimizations, performance across both Vera Rubin NVL72 and GB300 NVL72 will continue to improve.

Vera Rubin NVL72: 30x Higher Throughput per Megawatt and 35x Lower Token Cost

Agentic workloads look fundamentally different from chat or document summarization, where input and output sequences typically range from 1K to 8K tokens. In agentic sessions, context accumulates across steps and can reach hundreds of thousands of input tokens, with wide variability in both input and output lengths across requests. Performance measurement must evolve to capture the full agent workflow rather than a single inference request.

The results below reflect performance measured on real-world agentic coding trajectories.

In SemiAnalysis AgentX, the NVIDIA Blackwell platform delivers leading performance across multiple agentic models including Kimi K3, MiniMax M3, GLM5.3, Qwen3.5 and DeepSeek V4 Pro.

For example, GB300 NVL72 delivers up to 15x better throughput per megawatt than the NVIDIA Hopper architecture on the DeepSeek V4 Pro model, giving customers a high-performance foundation to run agentic workloads. This leap reflects the advantage of a larger scale-up GPU domain and codesigned software in delivering significantly better inference efficiency.

Vera Rubin extends that advantage, lifting the performance across the entire Pareto curve, to deliver as much as 30x higher throughput per megawatt than GB300 NVL72 on the DeepSeek V4 Pro model. These early results, measured using the SemiAnalysis AgentX workload and currently pending SemiAnalysis review, don’t yet reflect Vera CPU performance for tool calling.

NVIDIA DSX MaxLPS technologies manages power across the GPU, rack and workload levels to provision up to 40% more GPUs within the same megawatt budget, pushing throughput per megawatt further at AI factory scale.

Throughput per megawatt also directly impacts the cost of every token produced. At up to 35x lower cost per million tokens than GB300 NVL72, Vera Rubin NVL72 can run agents continuously, at scale, across the full breadth of customers’ workloads.

For power-constrained AI factories, throughput per megawatt determines AI factory revenue and cost per million tokens determines the profit margin on that revenue.

Extreme Codesign for Agentic Scale

Modern inference optimization spans a range of techniques that are especially critical for agentic AI. NVIDIA Vera Rubin NVL72 enables all of these and more through extreme codesign across every layer of the platform to deliver multifold performance gains.

Disaggregated serving separates context processing (prefill) from response generation (decode) so each scales independently.

Rate matching synchronizes the speeds at which prefill GPUs and decode GPUs produce tokens to maximize efficiency.

Large-scale expert parallelism distributes expert sub-networks in mixture-of-experts models across the scale-up GPU domain.

Distributed KV-caching extends memory across the scale-up GPU domain, while KV-cache offloading tiers less-active context to host and storage, keeping previously processed context accessible without recomputation.

KV-aware routing directs incoming requests to the GPUs that already hold the relevant cached context, reducing redundant computation across long sessions.

Fused CUDA kernels like MegaMoE combine many computation and inter-GPU communication operations into a single execution pass, keeping GPUs active rather than waiting for data.

NVIDIA Rubin GPUs’ enhanced fifth-generation Tensor Cores and the third-generation Transformer Engine accelerate both prefill and decode stages of inference. NVFP4 quantization compresses model weights to 4-bit precision, reducing memory footprint and increasing throughput without sacrificing output quality.

The NVL72 scale-up domain, a defining architecture across Vera Rubin and Grace Blackwell, enables the high-bandwidth and low-latency inter-GPU communication essential for techniques such as large-scale expert parallelism and distributed KV-caching. Purpose-built to power this scale-up domain, NVIDIA NVLink interconnect technology and NVLink Switches, now in their sixth generation, deliver 10x higher packet rates and 3x lower latency than off-the-shelf Ethernet alternatives.

Spanning optimized CUDA kernels, inference runtimes like NVIDIA TensorRT LLM and serving frameworks like NVIDIA Dynamo, NVIDIA’s software stack is codesigned with the hardware to enable inference optimizations.

While the results above reflect current Vera Rubin NVL72 performance, the full platform is a seven-chip architecture that also includes the NVIDIA Vera CPU, Groq 3 LPU, NVLink 6 Switch, BlueField-4 DPU, Spectrum-6 SPX and ConnectX-9 SuperNIC, all purpose-built for AI factories deploying agents at scale.

Extreme codesign also extends to NVIDIA’s co-engineering with its partners. Vera Rubin is in full production and is scaling across the ecosystem.

Learn more about the NVIDIA Vera Rubin platform.

Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents

Amidst the bustling environment of the New York Stock Exchange, a Goldman Sachs partner has expressed concern about the implications of artificial intelligence (AI) proliferating on Wall Street. Chris Churchman, who oversees the bank's prominent AI initiative known as Marquee, cautioned that over-reliance on these advanced systems may diminish critical thinking skills among emerging financial professionals.

In a recent episode of Goldman's "Exchanges" podcast, shared exclusively with CNBC, Churchman highlighted the potential pitfalls of AI integration. He likened the situation to how technological advancements have impacted personal navigation and memory, warning that essential analytical skills could be lost if algorithms take over cognitive tasks.

"Reasoning remains crucial," Churchman emphasized. "The ability to conceptualize and articulate arguments is vital, yet we are increasingly delegating this reasoning."

The ongoing drive to incorporate AI in trading and banking practices presents a double-edged sword for the finance sector. While it enhances current profitability, there is a looming risk of undermining the very talent needed for future growth. The automation of routine tasks, which traditionally enabled new bankers and traders to develop their decision-making skills, threatens to erode the foundational workplace culture that shapes junior employees into experienced professionals.

The trend may even lead to a reduction in the demand for junior bankers. In a report from last year, CNBC noted that Wall Street firms were exploring AI solutions to decrease the number of junior staff in relation to their senior counterparts.

Churchman stressed the necessity for banks to strike a balance between leveraging AI technologies and maintaining the apprenticeship model that has historically defined the industry. Drawing from his experience running currency trading at UBS before joining Goldman in 2021, he stated, "Learning comes from hands-on experience, much of which is based on tacit knowledge that has never been documented."

It is critical for Goldman to safeguard this innate knowledge possessed by its most skilled professionals and ensure that it is passed down to future generations, he added.

For example, junior traders typically gain invaluable insights by responding to client pricing inquiries under the guidance of seasoned risk managers. Churchman acknowledged that while this process could be automated, it raises concerns about the competence of senior traders who may not fully grasp the complexities involved.

He advocated for systems that empower employees to make significant decisions in unpredictable, high-stakes scenarios rather than relegating them to passive roles. Despite being one of the leading investment banks globally, Goldman Sachs has yet to finalize how it will navigate the ongoing transition brought about by AI advancements, Churchman remarked, noting his role as co-chair of the firm's Global Banking and Markets AI working group.

Goldman Sachs partner cautions against the significant risk of allowing AI to take over bankers' critical thinking abilities.

A moody, shadow-filled trading room with a single warm spotlight on a suited banker slumped at his desk, head in hands. Behind him looms a giant metallic AI cog with crawling wires that seem to reach for his brain. On the dark glass wall above towers a large, embossed Goldman Sachs logo in muted gold, casting ominous shadows across the floor.

Goldman Sachs partner warns of 'huge danger' in letting AI replace bankers' reasoning skills - CNBC

- Goldman Sachs partner warns AI could hinder young financiers' reasoning abilities on Wall Street.  
- AI's adoption in finance might reduce junior banker roles and erode essential skills.  
- Firms must balance AI use with preserving the apprenticeship culture for future talent.  

A screen displays the the company logo for Goldman Sachs on the floor at the New York Stock Exchange (NYSE) in New York City, U.S., May 7, 2025.

A Goldman Sachs partner leading one of the bank's flagship artificial intelligence projects warned that AI's spread across Wall Street risks hobbling the thinking capabilities of the next generation of financiers.

"There's a huge danger here that in the era of AI, we outsource our reasoning to these models, and we have cognitive atrophy that stops us being able to reason from first principles ourselves," said Chris Churchman, who leads Goldman's digital platform for institutional clients called Marquee.

The comments came during the latest episode of the firm's "Exchanges" podcast, according to a transcript provided exclusively to CNBC.

Just as people lost navigation and memorization skills with modern inventions, bankers risk losing analytical abilities if algorithms handle all the heavy lifting, Churchman said.

"Reasoning is still important," he said. "You still need to reason about [problems] and structure it into an argument, and now we're delegating reasoning."

Wall Street's push to enmesh AI into all of its trading and banking processes could be a kind of devil's bargain: It will make the industry more profitable today while potentially eroding the talent it needs for tomorrow. With AI taking over more of the routine work that has traditionally taught young bankers and traders how to think and make decisions, firms risk sacrificing the culture that turns junior employees into seasoned Wall Street talent.

It could even reduce the need for junior bankers in the first place. Last year, CNBC reported that Wall Street firms were examining ways of using AI to lower the ratio of junior bankers to senior employees.

Banks need to find a balance between using AI and preserving Wall Street's apprenticeship culture, said Churchman, who ran currency trading at UBS before joining Goldman in 2021.

"You learn by doing, and a lot of knowledge is tacit, it was never written down," he said.

Goldman needs "to make sure we don't lose that tacit and intuitive knowledge that some of our best people have today [and] to ensure the next generation have it too," Churchman said.

For instance, junior traders learn by fielding client pricing requests under supervision of experienced risk takers, Churchman said.

"We can absolutely automate that," he said, "but then do we get the senior traders that fully understand?"

Systems must be designed so that employees still call the shots in high-stakes, high-uncertainty decisions rather than becoming passive operators, Churchman said.

Even Goldman, one of the world's top investment banks, hasn't yet "figured out" how it will manage the transition the company has begun, said Churchman, who is also co-chair of the firm's Global Banking and Markets AI working group.

Rising memory expenses and decreasing funds: The cost of tech's AI expansion is escalating.

More from AI

Superintendent: Orange Schools taking initial steps into the realm of AI

Porsche signs $1.5 billion agreement for AI implementation with India's top IT services company Tata Consultancy.

Rep Chip Roy emphasizes the need for an important discussion in Washington regarding AI.

CNBC's The China Connection newsletter: AI firms are hurrying to assist robots in acquiring human skills.