The prevailing anxiety surrounding artificial intelligence and employment has often suggested a scenario where machines are set to take over all jobs, leaving only the most imaginative and inherently human roles intact. However, a new study by a prominent economist in the field of automation challenges this perspective and offers a viewpoint that is both comforting and somewhat disconcerting.

Pascual Restrepo, an associate professor of economics at Yale University and a leading researcher in automation's impact on labor markets, posits in a working paper published by the National Bureau of Economic Research that widespread automation of human jobs may not occur in a world dominated by artificial general intelligence (AGI). His argument isn't centered on AI's inability to automate tasks but rather on the assessment that many current jobs may not be significant enough to warrant replacement.

In his paper, titled "We Won’t Be Missed: Work and Growth in the AGI World," Restrepo suggests that instead of automating a broad spectrum of careers, AI might focus its computational power on "bottleneck" jobs that are crucial for future advancements, such as addressing existential threats, preventing asteroid impacts, or developing fusion energy. Consequently, many roles within the labor market could remain unchanged.

Restrepo clarifies that AGI doesn't render human abilities obsolete but rather shifts their value. As computational resources become scarce, human skills will be re-evaluated on how much it costs to replicate them with technology. In a scenario where compute power and human skills are the only scarce resources, average wages may rise, but the role of human labor would diminish relative to economic growth.

His research differentiates between two categories of work in the AI-driven economy. "Bottleneck" work includes essential tasks necessary for progress—energy production, infrastructure maintenance, scientific advancement, and national security. In contrast, "supplementary" work comprises roles that the economy could forgo yet still grow, such as arts, customer service, design, and labor in hospitality. While bottleneck tasks could be automated, supplementary roles might not attract sufficient computational resources for full automation, leading to their preservation.

This distinction provides some solace for those in creative and interpersonal professions, who may find their jobs relatively safe—not due to any unique human qualities, but simply because the cost of automating these areas would be prohibitively high compared to more pressing challenges that AI needs to address.

However, Restrepo adds a more sobering note: surviving automation doesn’t guarantee benefiting from economic growth. In an AGI-dominant future, he predicts a decoupling of wages from GDP growth. Currently, when the economy expands, workers usually see a corresponding increase in wages. This link may disappear when AI manages the essential tasks needed for economic advancement, leading to the replacement of human contributions with computable costs.

One of Restrepo's most striking conclusions is that labor’s proportion of GDP could potentially approach zero. The total computational capacity in the economy could vastly exceed that of human brains, which means that human labor would become increasingly marginal in comparison to computational gains. Consequently, wealth generation would likely concentrate with those who own computational resources rather than workers themselves.

This perspective presents significant implications regarding the ownership of AI technology. Concerns have already surfaced regarding wealth inequalities exacerbated by AI, with figures like BlackRock CEO Larry Fink noting that AI could replicate patterns of wealth concentration even more pronounced than those seen today, where a tiny fraction of Americans holds a vastly disproportionate share of wealth.

Restrepo suggests that addressing this impending disparity may involve strategies like universal basic income or treating computational resources as public assets that can be equitably distributed.

The discussion around automation transitions further complicates the picture. Restrepo outlines two distinct paths: a "compute-binding" scenario, where AI integration is gradual and provides time for workers to adapt, and an "algorithm-binding" phase, characterized by rapid advancements that could lead to abrupt wage variations and heightened inequality within the workforce.

As evidenced by current trends in skilled trades in 2026, certain professions are witnessing significant wage premiums as demand surges, notably in data center constructions, highlighting how these dynamics are unfolding already. 

Restrepo reassures that the collective economic position of workers will not necessarily decline due to AGI, given that the economy's output can surpass current levels, but warns that gains may not be distributed equally. The reality is that while total labor income could rise, if the benefits are concentrated among a wealthy few, many may remain economically disadvantaged.

The paper's title encapsulates the existential stakes of an AGI-dominated society: "We Won’t Be Missed." It implies a future where the intrinsic value of work—once tied to societal contribution and acknowledgment—is diminished, potentially severing the connection between effort and societal regard.

Restrepo’s insights underscore that the crucial issue is not merely the automation of jobs but the relevance of those jobs in the face of increasingly advanced AI systems. It may not be a question of whether AI will replace your job, but rather whether that job ever held substantial significance in the broader economy. 

In summary, while AI might catapult productivity to new heights, the resulting economic landscape will require consideration of ownership and equitable distribution, as the growth from AI advances may not trickle down to the workforce as it once did.

A Yale economist believes AGI won't take over most jobs due to their lack of value.

A sunlit academic study with a Yale pennant on the wall, a thoughtful economist in tweed standing before a chalkboard covered in dozens of job icons—nurse, teacher, plumber—with red X’s over each, while a half-drawn robotic arm trails off unfinished. Natural light streams through tall windows, underscoring the idea that most jobs aren’t worth automating for AGI.

A Yale economist says AGI won’t automate most jobs—because they’re not worth the trouble - Fortune

- Most human jobs won't be automated; many are deemed too unimportant for replacement.  
- Economic growth may occur without benefiting workers, as wages decouple from GDP.  
- Jobs in supplementary sectors could survive due to high costs of AI replication.  

The conventional fear about artificial intelligence and jobs runs something like this: the robots are coming for everything, and only the most creative, deeply human work will survive. A new paper by one of the world’s leading economists of automation turns that assumption on its head—and in doing so, arrives at a conclusion that is simultaneously more reassuring and more unsettling than the standard nightmare scenario.

Pascual Restrepo, an associate professor of economics at Yale University and one of the field’s foremost researchers on automation and labor markets, argues in a working paper published by the National Bureau of Economic Research that most human work won’t be automated in an era of artificial general intelligence. The reason isn’t that AI lacks the capability. It’s that most of what people do for a living simply isn’t important enough to bother replacing.

“The model opens up the intriguing possibility that much of today’s work may not be essential for future growth and may never be automated,” Restrepo writes in the paper, titled We Won’t Be Missed: Work and Growth in the AGI World. “Instead, compute may be directed toward bottleneck work critical for future progress—such as reducing existential risks, defending against asteroids, or mastering fusion energy—leaving large parts of the labor market unchanged.”

Not obsolete—just irrelevant

The main point, he argues, is that fundamentally, “AGI does not render human skills obsolete; it revalues them.” The new scarcity in the economy isn’t skilled labor or intelligence; it’s compute. This means that skills are valued at the opportunity cost of compute required to replicate them.

“In fact, if compute and human skill are the only scarce resources, average wages are higher in a post-AGI world. On the other hand, labor’s relative role shrinks.”

His analysis extends this logic to assume that compute will go to the areas that are most valuable for economic growth, leaving jobs that are less important to be filled by humans.

Two kinds of work in the AI economy

The paper draws a sharp distinction between two types of work. “Bottleneck” work consists of tasks that are essential for economic growth—things like producing energy, maintaining infrastructure, advancing science, and national security.

“Supplementary” work, by contrast, is everything the economy can do without and still expand: arts and crafts, customer support, hospitality, design, academic research, even the work of professional economists. In Restrepo’s framework, the economy will eventually automate every bottleneck task using compute—the raw computational resources of AI systems. But supplementary work? AI may simply ignore it.

That sounds like good news for the baristas and the novelists. Jobs in hospitality, live performance, and socially intensive work could survive largely intact, Restrepo argues, not because of any special human magic, but because the massive computing resources needed to fully replicate them would never justify the expense when AI has bigger problems to solve.

Crucial bottleneck work, in Restrepo’s telling, is very science-fiction sounding: “reducing existential risks, defending against asteroids, or mastering fusion energy.” Socially intensive work, on the other hand will include hospitality, live performances and entertainment: non-essential for future growth, costly to replicate with compute, and thus likely to remain human. “These domains could continue to offer familiar and meaningful work.”

Surviving automation is not the same as sharing in growth

But here is where the paper delivers its more sobering message. Surviving automation and prospering from economic growth are two very different things.

In an AGI world, Restrepo shows, wages would become decoupled from GDP. Today, as the economy grows, workers tend to share in that growth as wages rise and living standards improve. In the post-AGI economy he models, that link breaks. Once AI systems handle all the tasks essential for growth, economic expansion is driven entirely by adding computational resources.

Human work, whether essential or supplementary, is valued not by its contribution to growth, but by what it would cost to replace it with compute. That ceiling is, in the long run, a low one.

Labor’s share of GDP goes to zero

The paper’s starkest finding is that labor’s share of GDP converges to zero. Total computational resources in the economy could eventually reach 10⁵⁴ floating-point operations per second. The computing power of all human brains combined amounts to roughly 10¹⁸ flops.

In an economy where wages are anchored to what compute would cost to replicate human work, human labor becomes economically marginal—not worthless, but negligibly small relative to the overall pie. “Most income will accrue to owners of computing resources,” the paper concludes.

That means the distribution question of who owns the compute becomes the defining political and economic challenge of the AGI era. Already, that question is becoming urgent. BlackRock CEO Larry Fink warned in his closely watched annual letter that AI “threatens to repeat that pattern at an even larger scale—concentrating wealth among the companies and investors positioned to capture it,” noting that the top 1% of U.S. households now hold more wealth than the bottom 90% and that AI is likely to exacerbate this gap.

Restrepo notes that in such an economy, “one approach is to redistribute these gains through universal income. Another is to treat compute as a public resource—akin to land or natural capital—and distribute its returns broadly.”

Two modes of automation

The paper also makes important distinctions about the path to that future, and not all of them are comforting for workers navigating the transition today. Restrepo identifies two modes of automation. In a “compute-binding” transition, AI adoption is constrained by available hardware; adjustment is gradual, wages follow continuous paths, and workers have time to reallocate.

In an “algorithm-binding” transition—the one that looks more like the current moment, with AI capabilities advancing in sudden leaps—the picture is jagged and destabilizing. “Inequality may rise sharply: workers whose tasks cannot yet be automated enjoy large temporary wage premiums, while others face sudden wage declines as theirs are,” he writes.

This bears a strong resemblance to what’s happening in the trades as of 2026, with electricians, plumbers and HVAC technicians commanding strong premiums, especially on data-center construction. Construction workers on data center projects currently earn an average of about $81,800 annually—roughly 32% more than those on non-data center builds—according to data from Skillit, an AI-powered hiring platform.

Some electricians are pulling in $260,000 a year, with electrical work accounting for an estimated 45% to 70% of total data center construction costs. The U.S. will need roughly 300,000 new electricians over the next decade, in addition to replacing the 200,000 expected to retire.

We won’t be poorer—but we may not be richer either

Restrepo does offer one piece of meaningful reassurance: workers as a group are not made worse off by the transition. Because AGI expands what the economy can produce, total labor income in the post-AGI world—across all workers—is higher than in the pre-AGI baseline.

The arrival of AI cannot make us collectively poorer, the paper argues, because we could always retreat to a no-AI zone and produce exactly as we did before. The fact that we don’t means the new arrangement is better in aggregate. “The arrival of AGI cannot make us collectively worse off,” Restrepo writes.

But that collective gain is cold comfort if it is concentrated at the top of the income distribution—among the companies, investors, and nations that own the data centers.

Indeed, 40% of Americans currently lack meaningful exposure to capital markets, according to Fink. And without structural intervention—he suggests tools like tokenization and expanded retirement investment options—the AI-driven boom will leave them further behind.

‘We Won’t Be Missed’

The paper’s title, borrowed from its closing argument, captures the existential wager of the AGI economy. “Historically, work provided not only income but also recognition that one’s efforts improved society’s well-being,” Restrepo writes. “Work gave people the sense that they would be missed. In an AGI world, that connection is severed.”

Today, he notes, if half the workforce stopped showing up, the economy would collapse. In the AGI world, we would not be missed.

For Restrepo—whose work with Nobel laureate Daron Acemoglu has shaped the economics profession’s understanding of automation for more than a decade—the message is not one of despair, but of clear-eyed reckoning. The question is not whether AI will take your job. It may be that your job was never important enough for the question to matter.

For this story, Fortune journalists used generative AI as a research tool. An editor verified the accuracy of the information before publishing.

A Yale economist says AGI won’t automate most jobs—because they’re not worth the trouble

OpenAI appears to be undergoing significant internal changes, and it's not solely due to the usual pre-IPO butterflies. 

As reported by Bloomberg, the organization responsible for ChatGPT is simultaneously placing two high-ranking executives on prolonged medical leave. Kate Rouch, the chief marketing officer, is stepping back to focus on her recovery from cancer. While she intends to maintain limited engagement with the office if her health permits, her primary focus will be on her treatment.

Additionally, Fidji Simo, who serves as the CEO of artificial general intelligence development — a pivotal role within OpenAI — is also taking medical leave. Simo aims to return to her responsibilities after several weeks, during which she will be seeking treatment for a neuroimmune condition.

Compounding these departures, the company’s chief operating officer, Brad Lightcap, is transitioning into a new position as the special projects lead. This shift will require the chief revenue officer to absorb some of Lightcap's former responsibilities as COO.

With these moves unfolding, it's understandable to feel overwhelmed by the changes at OpenAI, especially as the firm has recently scaled back on major projects like Sora, the AI video generation initiative developed in collaboration with Disney. They are also grappling with significant challenges related to their data center expansion.

Amidst this upheaval, OpenAI is pursuing an aggressive strategy to go public. However, it remains uncertain how the newly restructured executive team will handle the intricacies and pressures of this transition.

In a statement provided to Bloomberg, the company expressed confidence in its leadership: “We have a robust team focused on our top priorities: pushing forward frontier research, expanding our global user base of nearly one billion users, and enhancing enterprise applications. We are in a strong position to maintain continuity and drive momentum.”

Two OpenAI Executives, Including the AGI CEO, Taking Medical Leave

Warm morning light floods a sleek, minimalist OpenAI lobby. Two sharply dressed figures—one wearing a subtle red medical armband—walk away through glass doors. Behind them, “OpenAI” appears in simple black text on a white wall. On the reception desk nearby, a hospital ID badge and an open medical file lie abandoned under a potted plant, hinting at their sudden medical leave.

Two OpenAI Execs, Including CEO of AGI, Going on Medical Leave - Futurism

- OpenAI's CMO Kate Rouch is on medical leave due to cancer treatment.  
- CEO of AI development Fidji Simo is taking time off for health reasons.  
- COO Brad Lightcap is transitioning to special projects, reshuffling leadership roles.  

Sign up to see the future, today Email address Sign Up Thank you!

Something’s going on at OpenAI — and it isn’t just the pre-IPO jitters.

According to Bloomberg, the company behind ChatGPT is sending not just one, but two top executives away on extended medical leave.

The company’s chief marketing officer Kate Rouch is reportedly stepping down to recover from cancer, the publication reports, though she’ll have a contingency in place to make limited returns to the office if her health allows.

And Fidji Simo, OpenAI’s CEO of artificial general intelligence development — and arguably one of the AI company’s most important cogs — is taking medical leave. Simo ultimately plans to return to her position, but will be gone for several weeks to find treatment for a neuroimmune condition.

On top of all that, the company’s chief operating officer Brad Lightcap is also being shuffled into a new role. According to Bloomberg, Lightcap will be the new special projects lead, forcing the chief revenue officer to take up some of his previous COO work.

If your head is spinning, just imagine what it’s like inside OpenAI’s hermetically sealed offices. The company has recently jettisoned some of its banner projects like Sora — the AI video generation model which included a massive partnership with Disney — while it juggles major bottlenecks with its data center buildout.

Against that chaotic backdrop, the company has begun an all-out push to go public, though it remains to be seen how its now thoroughly reshuffled executive branch handles the pressure.

In a statement to Bloomberg, the company insisted it has a “strong leadership team focused on our biggest priorities: advancing frontier research, growing our global user base of nearly one billion users, and powering enterprise use cases. We’re well-positioned to keep executing with continuity and momentum.”

More on OpenAI: Sam Altman Opens Up About Telling CEO of Disney That It Had All Been Smoke and Mirrors

Two OpenAI Execs, Including CEO of AGI, Going on Medical Leave

OpenAI is currently experiencing a significant shift in its leadership, as Fidji Simo, the CEO responsible for artificial general intelligence (AGI) deployment, has announced she will take a medical leave for several weeks. This development comes amid broader executive restructuring within the organization, stirring concern about stability during a crucial phase as OpenAI strives to achieve its ambitious goals for AGI.

The news, confirmed by OpenAI after speculation circulated among industry insiders, reveals that Simo's absence will be felt at a pivotal moment for the AI leader. Simo is in charge of deploying AGI technology, which aims to create systems that can perform tasks at or beyond human capability. While the specific details of the ongoing restructuring have not been publicly shared by OpenAI, sources indicate that it is significant.

Joining OpenAI in 2024, Simo previously held the position of CEO at Instacart, overseeing its transition to a publicly traded company. Her prior experience spans a decade at Meta, where she ascended to lead the Facebook app. Simo's appointment marked a strategic shift for OpenAI, highlighting its commitment to infusing experienced executives into its operations as the organization evolved from a research-focused entity to a major player in tech innovation.

However, maintaining stable leadership has proven challenging for OpenAI. Over the last two years, the company has seen numerous high-profile exits, including the departure of co-founder and chief scientist Ilya Sutskever in 2024, along with a tumultuous episode involving CEO Sam Altman that saw him briefly removed and then reinstated in late 2023. These fluctuations have raised concerns about the internal cohesion of the company as it navigates the delicate balance between its foundational nonprofit goals and the commercial demands posed by significant investors, particularly Microsoft, which has committed over $13 billion to the AI firm.

OpenAI's AGI Leader Fidji Simo Takes Medical Leave During Reorganization

A minimalist, sunlit OpenAI office suite: a frosted-glass door etched with “AGI Chief – Fidji Simo” and the word “OpenAI” (text only, no logo). Inside, an empty executive chair pushed back from a sleek desk; a folded blazer draped over its back, a stethoscope and medical bracelet left atop scattered papers, a half-drunk coffee mug tipped on its side. Soft morning light streams through floor-to-ceiling blinds, casting long shadows across the polished floor and hinting at sudden absence and underlying turmoil.

OpenAI's AGI Chief Fidji Simo Takes Medical Leave Amid Shake-Up - The Tech Buzz

- Fidji Simo, CEO of AGI deployment, is taking a medical leave for several weeks.  
- OpenAI is undergoing significant executive restructuring, raising concerns about leadership stability.  
- The company's leadership transitions may impact its ambitious goal of developing artificial general intelligence.  

OpenAI is navigating a major leadership transition as Fidji Simo, the company's CEO of AGI deployment, takes medical leave for several weeks. The move comes as the world's most valuable AI startup undergoes what sources describe as significant executive restructuring, raising questions about stability at the top just as the company races toward artificial general intelligence. The timing adds uncertainty to OpenAI's ambitious AGI roadmap at a critical moment in the AI race.

OpenAI just confirmed what insiders have been whispering about for days - the company's leadership structure is shifting, and it's happening while a key executive steps away. Fidji Simo, who holds the critical role of CEO of AGI deployment, is taking medical leave for what the company describes as "several weeks," according to reporting from Wired.

The timing couldn't be more sensitive. Simo oversees one of the most consequential functions at OpenAI - the deployment strategy for artificial general intelligence, the long-sought goal of creating AI systems that can match or exceed human capabilities across virtually any task. Her absence comes as the company undergoes what sources familiar with the matter describe as significant executive restructuring, though OpenAI hasn't publicly detailed the scope of these changes.

Simo arrived at OpenAI in 2024 after a high-profile tenure as CEO of Instacart, where she led the grocery delivery platform through its public debut. Before that, she spent a decade at Meta, rising to lead the Facebook app. Her hire signaled OpenAI's intent to bring in seasoned operational leadership as it transitioned from research lab to product powerhouse.

But leadership continuity has become a recurring challenge for OpenAI. The company has weathered multiple executive departures over the past two years, including co-founder and chief scientist Ilya Sutskever's exit in 2024, and the brief but dramatic ouster and reinstatement of CEO Sam Altman in late 2023. Each transition has raised questions about internal alignment as the company balances its original nonprofit mission with commercial pressures from partners like , which has invested over $13 billion.

OpenAI's AGI Chief Fidji Simo Takes Medical Leave Amid Shake-Up

For many years, video games have served as a critical platform for testing the capabilities of artificial intelligence. From the early days of checkers programs to advanced systems that have triumphed in chess and Go, it seemed that each achievement brought us closer to machines exhibiting human-like intelligence. However, a recent paper by Julian Togelius and his co-authors challenges this commonly held view. They assert that, despite some significant wins, today's AI still encounters a fundamental obstacle: the ability to play a game it has never encountered before.

Notable advancements in game AI typically revolve around systems specifically optimized for a single game. While these systems can outperform humans under specific conditions, their ability to adapt dwindles with even minor changes in the rules, visuals, or environment.

This inflexibility uncovers a significant drawback. Human intelligence encompasses more than mastering tasks; it includes the capacity to adapt to new situations. Video games, with their wide variety of mechanics and goals, present an excellent environment for testing this adaptability. The authors highlight that games challenge an extensive range of cognitive abilities, such as spatial reasoning, long-term planning, social understanding, and experiential learning. Yet, current AI systems struggle with such comprehensive challenges.

Reinforcement learning, a prevalent method driving many current successes, involves systems that learn through trial and error, refining their skills over millions or even billions of simulations. However, these systems often become overly specialized, excelling only in the precise scenarios they were trained on while failing to generalize effectively. Even slight changes, like color shifts or spatial adjustments, can drastically hinder a trained agent's performance.

In contrast, planning-based systems, which are often utilized in chess or Go, offer greater adaptability. These systems analyze potential moves and outcomes rather than relying solely on previous training. However, they rely on speedy and accurate simulations, a capability that most modern video games—and real-world scenarios—fail to deliver on a large scale.

Although large language models—crucial to today’s most visible AI applications—could appear to be a promising fit, their performance in unfamiliar games is surprisingly inadequate. Even in instances where they perform acceptably in well-known games, this is typically due to extensive, game-specific support structures. These systems require additional tools to interpret game states, manage memory, and carry out actions, and without this specialized infrastructure, their effectiveness diminishes significantly.

This discrepancy likely stems from the nature of their training data. Language models are built upon vast quantities of text, lacking experience with game states and actions. Consequently, they miss the embodied understanding and interactive engagement that come with gaming.

According to the authors, achieving genuine general gameplay abilities would necessitate an AI capable of learning a new game from the ground up in a time frame comparable to that of a skilled human—potentially several tens of hours—without leaning on prior knowledge or extensive simulations.

This threshold is far beyond the capabilities of current technology. Contemporary reinforcement learning models require significant amounts of data, while language models fail to possess the means to acquire and refine knowledge through prolonged interactions. Addressing this disparity will likely require innovative architectures and new learning paradigms.

The consequences of this research extend beyond gaming. The capacity to adapt to new contexts is a vital aspect of what is envisioned with artificial general intelligence (AGI). If an AI struggles to handle a novel video game—a relatively controlled and simplified environment—it is unlikely to manage the complexities of the real world.

Interestingly, the paper highlights one area where AI has demonstrated effectiveness: computer programming. The authors suggest that coding can be treated as a game with defined rules, specific objectives, and immediate feedback available through debugging and testing. Modern AI has become adept at this task due to extensive training on its structure and associated data.

However, outside of these structured situations, AI capabilities remain restrained.

In conclusion, the researchers advocate for maintaining games as a focal point for evaluating AI—not as isolated challenges but as part of a dynamic, evolving array of tests for adaptability and innovation. A truly intelligent system wouldn't just learn to efficiently play new games; it could potentially create captivating games of its own.

Reasons AI has yet to conquer a new video game

Dimly lit gaming den with dark wood walls and deep shadows; a sleek robotic arm clutches a game controller, its LEDs flickering out as the big screen blazes “Game Over.” Empty strategy guides and crushed energy drink cans litter the desk. In warm, low-key accent light, a lone adult figure’s silhouette leans forward triumphantly. Rich textures, moody shadows, and a cinematic feel convey the AI’s struggle to conquer the new video game.

Why AI still can’t beat a new video game - Brooklyn Eagle

- Current AI excels in specific games but struggles with unfamiliar ones due to brittleness.  
- Reinforcement learning systems overfit to trained scenarios, failing to generalize to new experiences.  
- True adaptability requires new AI architectures capable of learning games without prior exposure.  

METROTECH — For decades, video games have served as a proving ground for artificial intelligence. From early checkers programs to systems that conquered chess and Go, each milestone has seemed to bring machines closer to human-like intelligence. But a new paper by Julian Togelius and colleagues argues that this narrative is misleading. Despite impressive victories, today’s AI still struggles with a deceptively simple challenge: playing a game it has never seen before.

Most headline-grabbing successes in game AI rely on systems that are finely tuned to a single game. These systems can achieve superhuman performance but only within narrow boundaries. Change the rules, visuals or environment even slightly, and their competence can collapse.

This brittleness reveals a deeper limitation. Intelligence, as humans experience it, is not just about mastering one task but adapting to new ones. Video games, with their enormous variety of mechanics and goals, offer an unusually rich testbed for that kind of flexibility. As the authors note, games collectively probe a wide range of cognitive skills, from spatial reasoning and long-term planning to social intuition and learning through trial and error. Yet modern AI systems fall short on this broader challenge.

One major approach, reinforcement learning, has powered many recent breakthroughs. These systems learn by trial and error, improving through millions — or billions — of simulated plays. But they tend to overfit, becoming experts at the exact scenarios they were trained on while failing to generalize. Even minor changes, such as shifting colors or positions on a screen, can render a trained agent ineffective.

Planning-based systems, such as those used in chess or Go, offer more generality. They simulate possible moves and outcomes rather than relying on prior training. But they depend on fast, accurate simulations — something that most modern video games, and certainly the real world, cannot provide at scale.

Large language models, the technology behind today’s most visible AI tools, might seem like a promising alternative. After all, they can write essays, generate code and solve complex reasoning tasks. But when it comes to playing unfamiliar games, they perform surprisingly poorly.

Even in cases where language models appear to succeed — such as playing well-known games — the results often rely on elaborate, game-specific scaffolding. Systems are augmented with tools to interpret game states, manage memory and execute actions. Strip away this custom infrastructure, and performance drops sharply.

The gap likely exists due to the nature of training data. Language models are trained on vast amounts of text, not on sequences of game states and actions. As a result, they lack the embodied understanding and interactive experience that games demand.

The authors suggest that truly general game-playing ability would require something very different: an AI that can learn a new game from scratch in roughly the same time it takes a skilled human — perhaps tens of hours — without relying on prior exposure or massive simulation.

That benchmark is far beyond current capabilities. Today’s reinforcement learning systems require far more data, while language models lack the mechanisms to accumulate and refine knowledge over extended interaction. Bridging this gap would likely demand entirely new architectures and learning paradigms.

The implications extend well beyond gaming. The ability to adapt to unfamiliar situations is central to the idea of artificial general intelligence (AGI). If an AI cannot handle a novel video game — a controlled, simplified environment — it is unlikely to cope with the unpredictability of the real world.

The paper offers a different perspective on one area where AI does excel: computer programming. Coding, the authors argue, can be viewed as a kind of “game” with clear rules, well-defined goals and immediate feedback through debugging and testing. Modern AI systems have effectively mastered this particular game because they have been trained extensively on its structure and data.

But outside such well-structured domains, their abilities remain limited.

Ultimately, the researchers propose that games should remain central to AI evaluation — not as isolated challenges but as a vast, evolving ecosystem of tests for adaptability and creativity. A truly intelligent system would not only learn to play new games efficiently but might even invent compelling ones of its own.

Why AI still can’t beat a new video game

During a recent episode of the Big Technology Podcast, Greg Brockman, co-founder of OpenAI, made a significant declaration regarding the capabilities of large language models (LLMs) and their potential to achieve artificial general intelligence (AGI). He stated that the question of whether LLMs can reach AGI has been definitively answered, asserting that the GPT architecture will indeed pave the way toward it. 

Brockman expressed confidence that we now have a clear trajectory toward AGI, indicating a transformative shift in how we understand the capabilities of reasoning models based on the GPT framework. However, this assertion raises important discussions within the AI research community regarding the depth of understanding that text-based models can achieve compared to more complex, multimodal models like Sora. 

OpenAI recently made the decision to discontinue the Sora app and model, choosing to focus their resources on smaller-scale world model research primarily for robotics, rather than on consumer-facing technologies. Brockman acknowledged Sora as an "incredible model" but indicated that it represents a different avenue in AI development compared to the GPT reasoning series. He emphasized that, given their finite computational resources, OpenAI needs to prioritize one route over the other, suggesting that timing and sequencing are crucial to achieving the ambitious applications they envision.

When prompted by podcast host Alex Kantrowitz about the potential oversight in avoiding Sora and similar models, Brockman conceded the possibility of missing a critical opportunity, stating that making choices is an inherent part of AI development. 

The broader AI research community remains divided on the path to achieving general intelligence. Yann LeCun, a prominent AI researcher, has long contended that LLMs alone are insufficient for attaining human-like intelligence. He argues that these models struggle with logical reasoning, lack a true understanding of the physical world, and do not possess characteristics such as permanent memory or hierarchical planning. LeCun supports the notion that world models are necessary for developing a deeper comprehension of environments.

Similarly, Deepmind's founder, Demis Hassabis, shares the belief that merely scaling LLMs does not equate to advancements in intelligence; further innovations will be vital in this pursuit. In this context, Francois Chollet, another notable voice in AI, suggests that intelligence should be defined by a system's ability to efficiently acquire new skills and create its own abstractions.

In response to these discussions, recent research from Deepmind researchers Richard Sutton and David Silver advocates for a shift in approach—from training AI on human-curated knowledge to enabling it to learn from its own experiences. This paradigm aims to foster a more autonomous form of intelligence. Silver has even launched a new startup focused on simulative learning methods.

On the other hand, some researchers, like Adam Brown from Deepmind, have defended the existing LLM architecture, likening its token prediction capabilities to biological evolution. He believes that, with extensive scaling, these simple processes could give rise to complex behaviors that might be interpreted as understanding or even consciousness.

The debate around the capabilities and limitations of LLMs continues unabated, highlighting the evolving landscape of AI research and the varied perspectives that contribute to this critical discourse.

OpenAI's Greg Brockman states that GPT reasoning models are close to achieving AGI.

A sunlit rooftop observatory at dawn, a sleek brass telescope etched with “GPT Reasoning Models” aimed toward a radiant star on the horizon labeled “AGI.” Warm golden light floods white stucco walls, where “OpenAI” text with no logo is stenciled above an open archway. Soft morning light and long shadows convey optimism and discovery.

GPT reasoning models have "line of sight" to AGI, says OpenAI's Greg Brockman - the-decoder.com

- Greg Brockman claims GPT architecture will lead to achieving artificial general intelligence (AGI).  
- Debate continues in AI research about text models versus multimodal models for understanding.  
- Many researchers remain skeptical about LLMs achieving human-like intelligence or continuous learning.

Ask about this article… Search

OpenAI co-founder Greg Brockman says the debate about whether large language models can achieve general intelligence (AGI) is settled. The GPT architecture will lead to AGI.

"I think that we have definitively answered that question—it is going to go to AGI. Like we see line of sight," OpenAI President Greg Brockman says about the GPT reasoning models in the Big Technology Podcast.

It's a bold claim. Brockman is essentially declaring one of the central open questions in AI research settled: Can models primarily trained on text develop a real understanding of the world? Or does that require multimodal world models like Sora? Among AI researchers, the technical approach remains hotly debated.

Ad

OpenAI recently shut down the Sora app and model. World model research will continue for robotics, but on a smaller scale and without consumer-facing products.

Ad DEC_D_Incontent-1

Brockman calls Sora an "incredible model," but says it sits on "a different branch of the tech tree" than the GPT reasoning series. With limited computing power, pursuing both at the same time isn't feasible for OpenAI. For Brockman, it's less about the relative importance of the two approaches and more about "sequencing and timing." The applications "we've always dreamed of are starting to come into reach," and the way to get there is through the GPT architecture.

When host Alex Kantrowitz asks whether OpenAI could be missing something crucial by skipping Sora-style world models, pointing out that Deepmind's Demis Hassabis had said Google's "Nano Banana" image model felt particularly close to AGI, Brockman acknowledges the risk: "In this field you do have to make choices. Right? You have to make a bet."

Ad

Researchers remain divided on whether LLMs can reach general intelligence

Whether purely text-based models can achieve general intelligence is far from settled in the broader AI research community. Renowned AI researcher Yann LeCun has argued for years that LLMs won't lead to human-like intelligence. In his view, LLMs have a very limited understanding of logic, don't understand the physical world, have no permanent memory, cannot think rationally, and cannot plan hierarchically. Instead, he's betting on so-called world models to develop a comprehensive understanding of the environment. Deepmind founder Demis Hassabis holds a similar view: LLM scaling alone isn't enough, and further breakthroughs are needed.

AI researcher Francois Chollet defines intelligence as the ability to learn new skills efficiently. What matters is how well a system can independently form abstractions. While current language models can be placed on an intelligence scale, they rank very low. Outside their training domain, they have to relearn everything from scratch. Continuous learning could help address this gap.

Ad DEC_D_Incontent-2

This view aligns with a broader line of research. In a recent paper, Deepmind researcher Richard Sutton and former Deepmind researcher David Silver called for a paradigm shift. Instead of training on human knowledge, systems should learn from their own experience. Silver has since founded his own startup focused on simulation learning.

Ad

Ex-OpenAI researcher Jerry Tworek, one of the key minds behind OpenAI's reasoning model breakthroughs, also describes his research field of deep learning as "done." The next step, he says, is building simulations of human work where AI systems can learn skills. His new startup, Core Automation, is dedicated to this approach.

Not everyone shares this skepticism, though. Deepmind researcher Adam Brown recently defended the potential of the current LLM architecture. He compares the token prediction mechanism to biological evolution: a simple rule that, through massive scaling, creates emergent complexity that people perceive as understanding. Brown argues this complexity could even lead to consciousness.

GPT reasoning models have "line of sight" to AGI, says OpenAI's Greg Brockman

Recently, Nvidia CEO Jensen Huang stirred up quite a buzz during a conversation with podcaster Lex Fridman, where he asserted that artificial general intelligence (AGI) has already been achieved. This claim touches on a long-term goal of researchers working in artificial intelligence, even though the exact definition of AGI is still debated among experts. Generally, AGI refers to a type of AI that possesses human-level intelligence, but there is significant contention over how to accurately characterize and quantify intelligence itself.

Fridman presented a unique criterion for AGI, proposing that if AI could initiate and expand a technology business to a valuation of $1 billion, it could be classified as AGI. Huang responded that he believed this level of achievement was already upon us, although he qualified his statement by suggesting that the $1 billion threshold didn’t have to be permanent. “You said a billion,” he noted, indicating that the value could fluctuate.

Most AI researchers appear to disagree with this particular definition, arguing that it is too specific. While it focuses on business success, AGI is typically understood as encompassing a broader array of cognitive skills similar to those of humans, many of which may not be relevant for running a business. However, within the research community, there remains no consensus on an ideal definition, leaving the term somewhat elusive. Notably, several prominent AI organizations, boasting a combined market value exceeding $1 trillion, claim they are striving toward AGI. Some researchers even avoid using the term due to its ambiguous nature, while others suggest that companies exploit the ill-defined concept to generate hype around their advancements.

The intrigue surrounding Huang’s comments underscores ongoing challenges within the AI landscape regarding the measurement of AGI. Just prior to Fridman's podcast, researchers at Google DeepMind—including co-founder Shane Legg, who played a key role in popularizing AGI in the early 2000s—released a research paper aimed at establishing more scientific criteria to assess whether AI has achieved general intelligence. Titled “Measuring Progress Toward AGI: A Cognitive Framework,” the study introduces a “Cognitive Taxonomy” that identifies ten essential cognitive faculties, such as reasoning and social cognition, as indicators of general intelligence.

The authors suggest that AI systems should be evaluated across these faculties and their performance compared to a standard sample of human adults with at least a secondary education background. They highlight that contemporary AI models have a fragmented cognitive profile: while excelling in specific areas like mathematics and factual recall, they lag in others, such as experiential learning or social understanding. To qualify as AGI, an AI must reach median human performance across all ten dimensions, according to the researchers.

As part of their efforts, the team also announced a $200,000 contest on Kaggle to solicit external assistance in developing evaluations for five cognitive faculties that lack robust benchmark tests. This paper is among several recent initiatives to inject scientific rigor into the measurement of intelligence.

Earlier last year, a group led by Dan Hendrycks at the Center for AI Safety published their own framework and metrics for AGI, dividing it into ten cognitive domains based on a validated human intelligence model. Their study generated "AGI Scores" for existing AI models, with OpenAI's GPT-5 scoring just 57%, indicating a significant gap in matching a well-educated adult's cognitive capabilities.

Adding to the conversation is François Chollet, a machine learning researcher, who proposed the ARC-AGI benchmark as a means to evaluate how well current AI systems learn new skills rather than what they already know. His benchmark comprises visual puzzle challenges that require nuanced reasoning—tasks where AI struggles compared to humans.

Chollet's team recently unveiled ARC-AGI-3, a more sophisticated version of the benchmark featuring interactive tasks that require AI agents to navigate new situations and learn continuously, reflecting abilities that humans excel at but AI is still striving to master. These new benchmarks showcase the community's push toward creating empirical measures of AGI, yet defining intelligence remains a longstanding challenge.

Historically, in 1950, British mathematician Alan Turing grappled with creating a definition for intelligence prior to the establishment of the term "artificial intelligence." Instead of a rigid definition, Turing formulated the "Imitation Game," later known as the Turing Test, which posited that an intelligent machine should be indistinguishable from a human in conversation. However, the test itself faced scrutiny when early chatbots like Eliza demonstrated the ability to deceive users without possessing true understanding.

The Turing Test's limitations prompted the evolution of the term “artificial general intelligence,” first used in a 1997 paper by Mark Gubrud and later popularized by Legg in the early 2000s. Gubrud described AGI as a system that could rival human cognitive capabilities across various tasks. Yet the term still lacks clarity, with various definitions circulating in both academic and corporate environments, often causing confusion around its implications.

For instance, DeepMind has embraced AGI as part of its corporate mission since its founding in 2010, followed by OpenAI explicitly stating it would work toward AGI in its foundational principles. Notably, OpenAI's financial goal for AGI has been set at a profit benchmark of $100 billion, yet the company has not approached this threshold, raising further questions about what constitutes AGI.

Even as OpenAI's CEO, Sam Altman, claims proximity to achieving AGI through other metrics, he sometimes acknowledges the term's vagueness. Meanwhile, Microsoft's portrayal of AI projects like GPT-4 has faced critique for exaggerating their capabilities in relation to AGI.

All these factors illustrate the complexity of defining and measuring AGI amid evolving technology and corporate narratives. Huang, a pivotal figure in Nvidia's ascent to a market value exceeding $4 trillion, recognizes these nuances. His acknowledgment of the gap between human cognition and AI capabilities suggests that while the conversation around AGI is vibrant, we might need a new standard—perhaps “artificial Jensen intelligence”—to truly capture and celebrate the singular achievements of innovative minds within the AI landscape.

Nvidia's Jensen Huang claims 'we've reached AGI,' yet there's disagreement on its definition.

A grand tech conference stage bathed in warm golden sunlight, Jensen Huang standing at a podium with his back to the viewer, gesturing toward a massive NVIDIA logo printed center-stage. Behind him, a sweeping curved screen glows with “AGI?” in bold letters encircled by drifting question marks. In the dimly lit audience below, shadowy silhouettes of executives and reporters lean forward, some whispering, some raising their hands in puzzled debate—an atmosphere charged with excitement and uncertainty.

Nvidia’s Jensen Huang says ‘we’ve achieved AGI.’ But no one can agree on what that means - Fortune

- Nvidia's CEO Jensen Huang claimed AGI has already been achieved during a podcast.
- Google DeepMind proposed a new scientific framework for measuring artificial general intelligence.
- Current AI models show uneven cognitive abilities, lacking in key areas like learning and reasoning.

Last week, Nvidia CEO Jensen Huang made headlines when he told podcaster Lex Fridman that AGI—artificial general intelligence—had already been achieved.

AGI has long been the ultimate goal of many artificial intelligence researchers. That’s been the case even though there is no universally accepted definition of the term. It generally means AI that is as intelligent as humans, but there is a fierce debate over exactly how to define and measure “intelligence.”

In this case, Fridman had offered Huang a very unusual metric for AGI: Could AI start and grow a technology business to the point where it was worth $1 billion? Fridman asked if Huang thought AGI by this definition could be achieved within the next five to 20 years. Huang said he didn’t think that amount of time was necessary. “I think it’s now. I think we’ve achieved AGI,” he said. He then hedged, noting the company didn’t necessarily have to remain that valuable. “You said a billion,” Huang told Fridman, “and you didn’t say forever.”

Few AI researchers agree with the definition of AGI that Fridman offered Huang, which was both more specific (a company worth $1 billion), but also more narrow than most AGI definitions (which tend to refer to matching a vast range of human cognitive skills, not all of which might be needed to build a successful business.) But AI researchers also disagree with one another over what a better definition should be. The term remains stubbornly amorphous despite the fact that several leading AI companies, with collective market valuations of more than $1 trillion, say that AGI is what they are racing towards. Some computer scientists avoid using the term at all precisely because they say it is perpetually undefined and unmeasurable. Others say tech companies like using the term for completely cynical reasons—precisely because it is ill-defined, it’s easy for companies to build hype by claiming big strides towards achieving the fabled milestone.

The buzz over Huang’s AGI remarks only serves to highlight this quandary at the heart of the AI boom.

Trying to measure AGI

In fact, just days before Fridman dropped his podcast, researchers at Google DeepMind—including DeepMind cofounder Shane Legg, who first helped popularize the term AGI in the early 2000s—published a new research paper that proposed a more scientific way to define and assess whether AI models had achieved general intelligence. The paper, “Measuring Progress Toward AGI: A Cognitive Framework,” draws on decades of research in psychology, neuroscience, and cognitive science to construct what its authors call a “Cognitive Taxonomy.”

The taxonomy identifies 10 key cognitive faculties—including perception, reasoning, memory, learning, attention, and social cognition—that the researchers argue are essential for general intelligence. The framework then proposes evaluating AI systems across all 10 faculties and comparing their performance to a representative sample of human adults with at least the equivalent of a secondary education.

The paper’s key insight is that today’s AI models have a “jagged” cognitive profile: They may exceed most humans in some areas, like mathematics or factual recall, while dramatically trailing even average people in others, like learning from experience, maintaining long-term memories, or understanding social situations. An AI model would need to at least match median human performance across all 10 areas to be considered AGI, the Google DeepMind researchers suggest.



The researchers also announced a contest with a $200,000 prize pool on the popular machine learning competition site Kaggle for outside researchers to help build evaluations for the five cognitive faculties where existing benchmark tests are weakest.

The DeepMind paper is only the latest in a string of recent attempts to put the measurement of intelligence on more rigorous footing.



Last year, a team led by Dan Hendrycks at the Center for AI Safety, and that included deep learning pioneer Yoshua Bengio, published their own AGI framework and metrics. That paper also divided general intelligence into 10 separate cognitive domains, drawing on a framework for human intelligence developed by three psychologists—Raymond Cattell, John Horn, and John Carroll—that is the most empirically validated model of human cognition. It produced “AGI Scores” for existing AI models; the most capable system tested, OpenAI’s GPT-5, which was released in August 2025, scored just 57%, falling far short of matching a well-educated adult across all the cognitive dimensions.

One of the most ambitious practical attempts to highlight what today’s AI systems still cannot do is the ARC-AGI benchmark, created by well-known machine learning researcher François Chollet. Chollet’s core argument is that intelligence should be measured not by what a system already knows, but by how efficiently it can learn new skills.

The ARC-AGI benchmark consists of visual puzzle tasks involving grids of colored cells. Each task shows a few examples of an input grid being transformed into an output grid according to a hidden rule, and the test-taker must figure out the rule and apply it to a new input. For a human, grasping the pattern typically takes seconds. For frontier AI models, these puzzles remain surprisingly difficult, because they require the kind of flexible, abstract reasoning—spotting symmetries, understanding spatial relationships, inferring rules from a handful of examples—that current systems struggle with.

This month, Chollet and his collaborators launched ARC-AGI-3, the latest and most demanding version of the benchmark. Unlike earlier editions, which presented static puzzles, ARC-AGI-3 is interactive: AI agents must explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously over multiple steps—abilities that come naturally to humans but that remain at the frontier of AI research.

Taken together, these new benchmarks represent a growing effort within the AI research community to replace vague definitions about AGI with something closer to scientific measurement. But as these researchers are the first to admit, the difficulty of defining intelligence is as old as the study of thinking itself—and has plagued artificial intelligence as a field from its very earliest days.

Defining intelligence

In 1950, before the term “artificial intelligence” had even been coined and when mathematicians and electrical engineers were just starting to build the first modern computers, the famed British mathematician and computer pioneer Alan Turing wrestled with the fact that it was extremely difficult to formulate a definition of intelligence.



Rather than attempting one, Turing proposed an assessment he called “the Imitation Game,” which later became better known as the Turing Test. It stipulated that a machine should be considered intelligent when it can hold a general conversation with a person, via text, and a second human judge, reading the exchange, cannot reliably determine which participant is the machine and which the human. It was, in essence, an “I’ll know it when I see it” approach to intelligence.

But the Turing Test soon proved problematic too. Eliza, a chatbot developed at MIT in the mid-1960s, was designed to mimic a psychotherapist. Most of its responses followed hard-coded logical rules; Eliza often answered users with questions such as “Why do you think that is?” or “Tell me more” to cover up its weak language understanding. And yet Eliza fooled some people into believing it understood them. Eliza came close to passing the Turing Test even though on almost every other measure it came nowhere close to human cognitive abilities. And, in fact, a more sophisticated chatbot called “Eugene Goostman” officially passed a live Turing Test competition in 2014, again without touching most human cognitive skills.



Today’s large language models converse far more fluently than Eliza ever could, they still cannot match humans across the full spectrum of cognitive abilities—they hallucinate facts, struggle with long-horizon planning, and cannot learn from experience the way a person does.

Compared to the Turing Test, the term “artificial general intelligence” is a relatively recent one. It was first coined in 1997 by Mark Gubrud, then a graduate student at the University of Maryland, who used the neologism in a 1997 paper he presented at a conference on nanotechnology. He used the phrase “advanced artificial general intelligence” to describe AI systems that could “rival or surpass the human brain in complexity and speed, that can acquire, manipulate, and reason with general knowledge, and that are usable in essentially any phase of operations where a human intelligence would otherwise be needed.” But the paper quickly vanished in obscurity.

Then, in the early 2000s, Legg—who would go on to cofound DeepMind—independently coined the same term. He was collaborating with computer scientists Ben Goertzel, Cassio Pennachin, and others on a book about potential ways to create machine learning systems that would be able to address a wide range of problems and tasks. They wanted a term that would distinguish the ambition of these systems from the narrow machine learning algorithms then in vogue, which, once trained, could only tackle a single, narrow task. Goertzel considered calling this more general AI “real AI” or “strong AI,” but Legg suggested “artificial general intelligence” instead, unaware of Gubrud’s earlier usage. He also suggested the term be abbreviated as AGI. This time, AGI took off.



In Goertzel’s book he defined AGI as “AI systems that possess a reasonable degree of self-understanding and autonomous self-control, and have the ability to solve a variety of complex problems in a variety of contexts, and to learn to solve new problems that they didn’t know about at their time of creation.”



The definition was useful for separating work on general AI systems from narrow machine learning ones, but it too contained a fair an unhelpful amount of ambiguity: What did “reasonable degree” mean? Which complex problems in which contexts counted towards the standard?



Legg would later compound this ambiguity by offering a more casual definition of AGI that was in some ways narrower (it didn’t talk about self-understanding, for instance) but equally vague. For instance, he told The Atlantic’s Nick Thompson last year, “I define an AGI to be an artificial agent that can do the kinds of cognitive things that people can typically do. I see this as the natural minimum bar.” But which things? And which people?



Questions like this have continued to swirl around AGI. Does the term mean software that matches the cognitive abilities of an average human? Or the abilities of the humans with the highest IQs? Or the best expert in each individual domain of knowledge? The Hendrycks and Bengio research paper, for instance, defines AGI as matching or exceeding “the cognitive versatility and proficiency of a well-educated adult.” The DeepMind paper proposes measuring against a representative sample of adults. Others have used less precise formulations.

Adding to the confusion, AGI is often conflated in public discussion with a concept AI researchers call “artificial superintelligence,” or ASI—an AI that would be smarter than all humans combined. Most AI researchers consider AGI and ASI to be separate milestones, and very different in degree of sophistication, but in the popular imagination the two frequently blur together.

AGI becomes a corporate goal—and a marketing slogan

If the academic debate over defining AGI has been long and nuanced, the corporate world has introduced definitions that are, to put it charitably, idiosyncratic. DeepMind became the first company to make the pursuit of “artificial general intelligence” a business goal. Legg put the phrase on the front page of the company’s first business plan when he, Demis Hassabis, and Mustafa Suleyman cofounded the company in 2010.

Five years later, OpenAI also made building AGI its explicit mission. Its original 2015 founding principles said that the new lab—at the time a non-profit—was dedicated to ensuring “that artificial general intelligence benefits all of humanity.” Three years later, when the lab first set up a for-profit arm, it published a charter that defined AGI “as highly autonomous systems that outperform humans at most economically valuable work.” Now, for the first time, AGI was being measured by financial metrics, not mere cognitive ones.

And, as it turned out, OpenAI would soon secretly set a highly specific financial threshold for AGI. When Microsoft first invested $1 billion into OpenAI’s for-profit arm in 2019, the tech giant’s agreement with the AI startup made it OpenAI’s preferred commercialization partner for any AI model the lab developed up to, but crucially not including, AGI. At the time, it was reported that the decision of when AGI had been achieved would be at the discretion of OpenAI’s non-profit board.



But, crucially, according to reporting by tech publication The Information in 2024, when Microsoft agreed to invest a further $10 billion into OpenAI in 2023, its contract with OpenAI contained a clause that defined AGI as a technology that could generate at least $100 billion in profits.

OpenAI is nowhere near that mark. The company has reportedly told investors it made $13 billion in revenues last year, but still managed to burn through $8 billion in cash. It does not expect to break even until 2030.

Despite being far short of the financial threshold for AGI in its contract with Microsoft, OpenAI CEO Sam Altman has often made statements that suggest OpenAI is close to achieving the AI milestone as measured by other benchmarks. In a post to his personal blog in January 2025 titled “Reflections,” Altman wrote that OpenAI was “now confident we know how to build AGI as we have traditionally understood it” and that the company was beginning to turn its aim towards superintelligence. In a subsequent essay titled “Three Observations,” he wrote that systems pointing toward AGI were “coming into view.” Yet, at other times, Altman has seemed to acknowledge AGI’s weakness as a concept. Around the same time as his “Reflections” blog post, Altman told a Bloomberg News interviewer that AGI “has become a very sloppy term.”



Microsoft has also chosen to ignore the financial definition of AGI it struck with OpenAI when it suited the company’s marketing purposes. In March 2023, a team of Microsoft researchers published a 154-page paper about GPT-4 provocatively titled “Sparks of Artificial General Intelligence,” arguing the model could “reasonably be viewed as an early (yet still incomplete) version” of AGI.



The paper was widely criticized for hyping the abilities of GPT-4 for commercial purposes. Even Altman distanced himself, calling GPT-4 “still flawed, still limited.”The new research and benchmarks from Google DeepMind and the Hendrycks-Bengio team makes some progress towards establishing a yardstick for AGI, one rooted in decades of study of human intelligence. And what’s clear is that today’s best AI models still don’t measure up to breadth and depth of human cognitive abilities.

Huang, the Nvidia CEO, knows this, just as he was no doubt fully aware of the social media frenzy and headlines he would generate by saying AGI had been achieved. We know Huang knows this because later in the same podcast in which he said “AGI is achieved” he also said that the popular OpenClaw AI agents, which can be powered by any of the top AI models from companies such as Anthropic and OpenAI, could never replicate Nvidia. “Now, the odds of 100,000 of those agents building Nvidia is zero percent,” he said.

Huang is not just Nvidia’s CEO. He is also the company’s founder and the person who has run the company for 33 years, piloting it past near-bankruptcy at one point, to see it now worth more than $4 trillion, making it one of the most valuable companies on the planet. In many ways, Huang is a singular genius. But he’s also a very human one. So maybe we need a new standard, not AGI but AJI—artificial Jensen intelligence. When AI reaches that level, the AI boosters on social media who breathlessly amplified Huang’s AGI claim will really have something to get excited about.

Nvidia’s Jensen Huang says ‘we’ve achieved AGI.’ But no one can agree on what that means

As an investor, I strategically allocate my capital between low-cost funds and select individual stocks that I believe offer favorable risk-reward scenarios. My investment strategy consists of a balanced mix, with roughly half of my portfolio focused on each approach. While I enjoy delving into macroeconomic factors and fundamental analysis, I am acutely aware that market momentum and investor sentiment ultimately drive market performance. Consequently, I refrain from trying to time the market and only purchase stocks with the intention of holding them for a minimum of ten years.

In terms of fundamental analysis, it's well-known that markets tend to be forward-looking, yet few grasp the true implications of this notion. I avoid simplistic evaluations based solely on P/E ratios or dismissing a stock as “cheap.” Instead, I perceive markets as the intersection of supply and demand, highlighting the importance of forecasting human behavior.

I strive to look beyond the numerical data to comprehend what the market perceives in a stock, which often entails analyzing sectors, industries, and long-term growth trajectories. This method demands creativity, curiosity, a certain level of innocence, and the willingness to sometimes swim against the tide.

I reside in Geneva, Switzerland, which is reflected in my online username. I hold a Master’s Degree in Business and contribute to the investment conversation alongside fellow contributor "Rex Investing" on Seeking Alpha. All insights and analyses shared here are solely my own. For more updates, you can follow me on Twitter @x.com/GenevaInvestor or connect with me on medium.com/@genevainvestor.

Iran Is Distraction, Artificial General Intelligence Is The Key, Nvidia Is The Investment (NVDA)

A sunlit modern trading desk bathed in warm morning light. Centered on the desk is a softly glowing glass sculpture of a neural-network brain, symbolizing AGI as the clear signal. To its left sits a slightly blurred Iranian flag on a small stand, hinting at background noise. On the right, a sleek monitor dominates the scene, its screen emblazoned with a large Nvidia logo and rising green candlestick charts. Soft shadows and natural light emphasize focus on the brain and the bullish Nvidia trade.

Iran Is Noise, Artificial General Intelligence Is The Signal, Nvidia Is The Trade (NVDA) - Seeking Alpha

- The author prefers low-cost funds and asymmetric stock bets in their investment strategy.  
- They focus on understanding market behavior rather than solely relying on P/E ratios.  
- Long-term market trends and human behavior play significant roles in their investment decisions.  

Decoding markets beyond P/E. As an investor, I either put my money into low cost funds or in single stocks that (I think) are asymmetric bets. My portfolio is roughly 50/50 between the two. I like to write about Macro and Fundamentals, with the (painful) awareness that Momentum and Sentiment are what really matters. Thatâs why I never try to time the market and I only buy stocks if I am willing to hold them for at least 10 years.When it comes to fundamentals, everybody knows the market is forward looking, but few understand what that means. I donât look at a P/E number and decide to buy if a stock is âcheapâ. I see markets as literally just the meeting point between demand and supply. Predicting human behavior is key.I always try to understand what the market is seeing in a stock beyond the numbers, which often implies trying to understand sectors, industries and long term growth trends. My approach requires ingenuity, curiosity and a good dose of naivete, as well as being comfortable in (sometimes) going against the current.I am based in Geneva, Switzerland (hence my SA name) and I have a Masterâs Degree in Business. Friend "Rex Investing" is also a contributor to Seeking Alpha. All opinions and analysis are exclusively my own.You can follow me on Twitter @ x.com/GenevaInvestor. I am also on medium.com/@genevainvestor.

Iran Is Noise, Artificial General Intelligence Is The Signal, Nvidia Is The Trade (NVDA)

The current market landscape suggests that significant investments, totaling trillions, are being directed towards the belief that artificial general intelligence (AGI) is on the horizon. Major companies are tangled in a competitive race, pouring vast resources into the development of AI technologies.

I am David B. McMillan, an investor focused on fundamental valuations. My investment strategy centers around identifying undervalued securities based on their expected future cash flows. Additionally, I implement a tactical allocation approach, taking more aggressive positions when equity prices are lower and scaling back during higher price levels. My academic background includes a Bachelor of Science in Physics and a Bachelor of Arts in Philosophy from UCSB, and I am currently a CFA Level 2 candidate. My primary interest lies in the aerospace and defense sectors, though I also follow retail and technology companies closely. Over my 12 years of investing, I have witnessed and documented impressive returns in AI, tech, and crypto sectors prior to their mainstream recognition—investing in NVDA in 2017 yielded an 8000% gain, PLTR at its IPO saw a 1870% rise, AMD also in 2017 provided a 3700% return, and TSLA in 2016 resulted in a 3400% gain. By 2018, I had already included all of the "Magnificent Seven" stocks in my portfolio before they garnered that nickname. My demo portfolio, which commenced in April 2025 with roughly $8,000 of my own capital, currently boasts a Sharpe ratio of 3.49, outpacing IVV's 2.42 over the same duration. My daily average return stands at 0.30%, compared to IVV’s 0.14%.

As for disclosures, I hold a beneficial long position in BRK.B shares through stock ownership, options, or derivatives. I authored this article myself and share my personal opinions, and I am not receiving any compensation for my content apart from what I earn from Seeking Alpha. There are no existing business relationships with any companies mentioned herein.

Seeking Alpha offers a disclaimer noting that past performance is not indicative of future results. This article does not constitute any investment recommendations tailored to any specific investor. The opinions expressed herein may not necessarily reflect the views of Seeking Alpha. Furthermore, Seeking Alpha is not a licensed securities dealer or investment adviser, and its analysts come from a mix of individual and professional investing backgrounds, not all of whom hold licenses or certifications from regulatory entities.

AGI Setback, AI Surge: The Unexpected Winners and Losers in AI

A split-scene composition: on the left, deep shadows swallow a shattered humanoid robot bust and a lone, disheveled executive in a dark suit slumped against cracked concrete; on the right, warm natural sunlight illuminates a small team of modern entrepreneurs high-fiving beside sleek robotic arms assembling golden microchips; between them, a seesaw-like steel beam pivots, tilting toward the sunny side—no artificial blue tones.

AGI Bust, AI Boom: Why AI’s Winners And Losers Aren’t Who You Think - Seeking Alpha

- Trillions are invested in AI, with expectations of imminent artificial general intelligence advancements.  
- Major companies compete fiercely in an expensive race to develop AI technologies.  
- David B McMillan focuses on fundamental valuation and tactical allocation in investing strategies.

What AI future is the market currently pricing? Right now, trillions of dollars are being invested as though artificial general intelligence ( AGI ) is inevitable. Big companies are engaged in an expensive, frenzied arms race

My name is David B McMillan and I am an investor interested in fundamental valuation. My philosophy is fundamental investing - I seek to identify underpriced securities relative to their potential future cash flows. I also use tactical allocation, investing more aggressively when equity prices are lower, and more conservatively when they are higher. I have a BS in Physics and BA in Philosophy from UCSB, and am currently a CFA Level 2 candidate. I am mostly interested in covering stocks in the aerospace and defense sector, but I am also interested in retail and tech companies. I have a 12 year investing track record, with documented investments in AI, tech, and crypto themes before they were widely understood - NVDA in 2017, 8000 percent gain; PLTR at IPO, 1870 percent gain; AMD in 2017, 3700 percent gain; TSLA in 2016, 3400 percent gain. Had all of Mag 7 in my portfolio by 2018, before those stocks were called the Mag 7. My current demo portfolio, started in April 2025 with about $8k of my my own capital, is so far achieving a Sharpe ratio of 3.49 compared to IVV of 2.42 in the same time period. My average time-weighted return is 0.30 percent per day vs IVV at 0.14 percent per day.

Analystâs Disclosure: I/we have a beneficial long position in the shares of BRK.B either through stock ownership, options, or other derivatives. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.

Seeking Alpha's Disclosure: Past performance is no guarantee of future results. No recommendation or advice is being given as to whether any investment is suitable for a particular investor. Any views or opinions expressed above may not reflect those of Seeking Alpha as a whole. Seeking Alpha is not a licensed securities dealer, broker or US investment adviser or investment bank. Our analysts are third party authors that include both professional investors and individual investors who may not be licensed or certified by any institute or regulatory body.

AGI Bust, AI Boom: Why AI's Winners And Losers Aren't Who You Think

Arm has unveiled the AGI CPU, a cutting-edge family of data center processors featuring up to 136 cores, marking a significant shift in the company’s approach. This innovation is built on TSMC's advanced 3nm technology and integrates Neoverse V3 cores, developed in collaboration with Meta. For the first time in its 35-year history, Arm is delivering its own production processors rather than licensing its intellectual property to other partners.

Designed for what Arm describes as "agentic AI infrastructure," this CPU is optimized for the critical task of orchestrating accelerators and efficiently managing data movement in expansive AI systems.

With a robust architecture, the AGI CPU can host up to 136 Neoverse V3 cores, each capable of reaching speeds of up to 3.2 GHz under full load and peaking at 3.7 GHz. This performance is achieved within a thermal design power (TDP) of 300 watts. The processor supports a substantial 12 channels of DDR5 memory, operating at speeds up to 8800 MT/s, resulting in over 800 GB/s of total memory bandwidth—equivalent to 6GB/s per core—while attempting to maintain latency below 100 nanoseconds. Additionally, it features 96 PCIe Gen6 lanes and integrates support for CXL 3.0, facilitating memory expansion and pooling.

Arm's reference design includes a 10U dual-node server conforming to the Open Compute Project’s DC-MHS specifications, where two AGI CPUs can be housed within a blade. Each standard air-cooled 36kW rack can accommodate 30 blades, yielding a staggering total of 8,160 cores. Moreover, Arm has teamed up with Supermicro to offer a high-performance liquid-cooled configuration that can support 336 chips and exceed 45,000 cores.

According to Arm, the AGI CPU surpasses the latest x86 platforms in performance per rack by more than double. However, it’s worth noting that this claim relies on Arm's internal assessments, and independent benchmarks are yet to validate these figures.

While GPUs have traditionally dominated discussions around AI hardware, there is an increasing need for more powerful general-purpose computing solutions as systems like OpenClaw gain traction. Arm aims to address this growing demand, potentially catering to a segment often overlooked by major players like Nvidia and Micron, who have heavily focused on AI-centric applications.

Among the early adopters of the AGI CPU is Meta, which plans to integrate the processor with its proprietary MTIA accelerators. Santosh Janardhan, Meta's head of infrastructure, emphasized their collaborative effort and commitment to a long-term development roadmap.

Arm has also secured commercial interest from companies such as Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. Sachin Katti, leading industrial compute at OpenAI, remarked that the AGI CPU is set to enhance OpenAI's infrastructure, specifically within the orchestration layer essential for managing extensive AI tasks.

Historically, Arm has functioned as an IP licensing enterprise, allowing partners like Apple, Nvidia, and AWS to craft customized chips using Arm's architectural designs. With the introduction of the AGI CPU, Arm expands its offerings to include production-ready silicon that clients can deploy directly, in addition to its IP licensing and Compute Subsystems programs.

Arm indicated that the AGI CPU lineup will coexist with the existing Arm Neoverse CSS product offerings, and they have already committed to future products. The company is keen to clarify that this move complements rather than competes with its current licensees. However, it remains to be seen how Arm will navigate this landscape while providing chips to the same data centers that house competing products from Nvidia, AWS, Google, and Microsoft.

Arm expands from IP into AGI CPU silicon with a 136-core data center chip aimed at AI infrastructure, featuring Meta as the primary partner.

Detailed macro shot of a gleaming 136-core silicon chip engraved with the Arm logo, resting on open circuit blueprints in a bright, sunlit R&D lab. Beyond a glass wall, a row of sleek server racks bathed in warm natural light displays a large Meta logo, highlighting their lead partnership in AI infrastructure.

Arm moves beyond IP with AGI CPU silicon — 136-core data center chip targets AI infrastructure with Meta as lead partner - Tom's Hardware

- Arm announced the AGI CPU, a 136-core data center processor co-developed with Meta.  
- The CPU is designed for agentic AI infrastructure and offers over 800 GB/s memory bandwidth.  
- Early customers include Meta and OpenAI, indicating strong demand for powerful general-purpose compute.  

Arm today announced the AGI CPU, an up-to 136-core data center processor family that the company designed and will sell as finished silicon. The chip, built on TSMC's 3nm process with Neoverse V3 cores, was co-developed with Meta and represents the first time in Arm's 35-year history that the company has shipped its own production processor rather than licensing IP to partners.

The AGI CPU has been designed for what Arm calls "agentic AI infrastructure," the CPU-side orchestration work required to coordinate accelerators and manage data movement in large-scale AI deployments.

136 Neoverse V3 cores at 300 watts

The chip packs up to 136 Neoverse V3 cores running at up to 3.2 GHz all-core and 3.7 GHz boost across two dies, all within a 300-watt TDP. It supports 12 channels of DDR5 memory at up to 8800 MT/s, delivering more than 800 GB/s of aggregate memory bandwidth or 6GB/s per core with a target of sub-100ns latency. I/O includes 96 PCIe Gen6 lanes and native CXL 3.0 support for memory expansion and pooling.

Article continues below

Arm's reference platform is a 10U dual-node server compliant with the Open Compute Project's DC-MHS standard. Two AGI CPUs fit per blade, and a standard air-cooled 36kW rack holds 30 blades for 8,160 cores total. Arm has also partnered with Supermicro on a liquid-cooled 200kW configuration that houses 336 chips and more than 45,000 cores.

Arm claims the AGI CPU delivers more than two times the performance per rack compared to the latest x86 platforms. That figure is, of course, based on the company’s own internal estimates at this stage, not independent benchmarks.

GPUs have made most of the headlines surrounding AI hardware to date, but there’s a demand for more powerful general-purpose compute as agentic systems like OpenClaw explode in popularity. Arm is clearly hoping it can meet and cash in on this demand — and hopefully that won’t be to the detriment of non-AI customers, who seem to have long since been forgotten by the likes of Nvidia and Micron.

OpenAI among early customers

Meta served as the lead partner on the project and plans to deploy the AGI CPU alongside its custom MTIA accelerators. Santosh Janardhan, head of infrastructure at Meta, said the two companies worked together on the chip and are committed to a multi-generation roadmap.

Stay On the Cutting Edge: Get the Tom's Hardware Newsletter Get Tom's Hardware's best news and in-depth reviews, straight to your inbox. Contact me with news and offers from other Future brands Receive email from us on behalf of our trusted partners or sponsors

Beyond Meta, Arm confirmed commercial commitments from Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom. Sachin Katti, head of industrial compute at OpenAI, said the AGI CPU will play a role in OpenAI's infrastructure by strengthening the orchestration layer that coordinates large-scale AI workloads.

Arm has historically operated as an IP licensing company. Its partners, from Apple to Nvidia to AWS, design their own chips using Arm's instruction set architecture and core designs. The AGI CPU adds a third option alongside IP licensing and Arm's Compute Subsystems (CSS) program: Arm-designed, production-ready silicon that customers can deploy directly.

Arm said the AGI CPU product line will continue in parallel with the Arm Neoverse CSS product roadmap, and that follow-on products are already committed. The company seems keen to point out that this is an additive move rather than a pivot that competes with existing licensees, though how Arm manages that as it sells chips into the same data centers as Nvidia Grace, AWS Graviton, Google Axion, and Microsoft Cobalt remains to be seen.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Arm moves beyond IP with AGI CPU silicon — 136-core data center chip targets AI infrastructure with Meta as lead partner

During an event in San Francisco on Tuesday, Arm introduced its inaugural chip, stepping away from its traditional blueprint approach. The new 136-core CPU, named the AGI CPU, is set to be deployed on a large scale later this year by its prominent customer, Meta.

This British chip designer’s latest offering marks a pivotal shift in its business strategy. Historically, Arm has licensed core IP or instruction set architectures for datacenter products. However, with the AGI CPU, it is venturing into producing its own Arm-branded processors specifically focused on facilitating agentic AI capabilities.

While the AGI CPU features Arm Neoverse V3 cores, it will not execute AI models directly – that function typically falls to GPUs or advanced AI ASICs. Instead, Arm envisions this CPU as the backbone powering AI agents, positioning it in direct competition with Nvidia’s standalone Vera CPUs showcased at GTC last week. 

According to Mohamed Awad, Arm's Executive Vice President of cloud AI, the CPU is pivotal for achieving artificial general intelligence (AGI).

Despite the spotlight on GPUs in recent years, the emergence of agentic systems, such as OpenClaw, has rekindled the need for general-purpose computing. These systems require CPU cores and memory for coding, task automation, and reinforcement learning vital for training future models. Arm predicts a dramatic rise in CPU demand due to this trend, hence its strategic development of the AGI CPU.

The AGI CPU operates at 300 watts and boasts 136 Neoverse V3 cores, with clock speeds reaching 3.7 GHz (3.2 GHz base). This processor is manufactured using TSMC’s advanced 3 nm technology and is designed with 2 MB of L2 cache per core, providing a total of 128 MB of shared system-level cache.

Awad emphasized the focus on creating an efficient design, intentionally omitting non-essential accelerators to maximize die space for targeted performance. "Legacy CPUs often included unnecessary support for outdated applications. We wanted our design to be solely focused on what this device needs,” he remarked.

In contrast to Nvidia’s Vera architecture, which employs simultaneous multithreading, Arm’s approach provides one thread per core, guiding more predictable performance scaling. The CPU includes 12 channels of DDR5 memory, likely distributed as six channels per die, supporting speeds up to 8800 MT/s. With an aggregate bandwidth of 825 GB/s, this translates to 6 GB/s available for each core.

In a bid to reduce latency, Arm integrates both memory and I/O operations within the same die as the compute cores. Each socket will be recognized by the operating system as two separate NUMA domains. For connectivity, the processor is outfitted with 96 lanes of PCIe 6.0, along with support for CXL 3.0.

Meta, already utilizing Nvidia’s Arm-based Grace CPUs and planning to adopt Vera chips, is among Arm's first significant CPU customers. The company has also validated two different Open Compute Project (OCP) rack designs. The first is a 36 kW air-cooled rack that accommodates 30 compute blades, summing up to 8,160 cores. Additionally, Arm has confirmed an even denser 200 kW liquid-cooled rack featuring 42 eight-node servers, resulting in an impressive 45,696 cores—more than double Nvidia's Vera ETL256 CPUs at 22,528 cores.

Arm’s customer list extends beyond Meta, with notable early adopters including OpenAI, SAP, Cerebras, Cloudflare, F5, SK Telecom, and Rebellions. The chip is envisioned not just for AI agents but also as a head node for tailored accelerators, or as a conventional CPU for networking and storage solutions. OEM partners, like Lenovo, are actively developing 19-inch systems that incorporate this new chip.

Historically, enterprise clients have faced limited options for Arm datacenter silicon, with Ampere Computing being the sole significant non-cloud competitor. Set to launch later this year, Arm's AGI CPU promises a transformative potential in the tech landscape, though whether it will usher in the era of The Singularity remains to be seen.

Arm develops its own 136-core AGI CPU to capitalize on AI trends.

A sunlit high-tech showroom where a gigantic, futuristic CPU has been sculpted into a sleek train locomotive bursting forward on glowing circuit-board rails. Its metallic shell is etched “136-core AGI,” and the Arm logo is prominently emblazoned on the engine’s side. Soft natural light pours in through floor-to-ceiling windows as professional engineers in silhouette watch in awe.

Arm rolls its own 136-core AGI CPU to chase AI hype train - theregister.com

- Arm introduced its AGI CPU, a 136-core processor aimed at artificial general intelligence.  
- Meta and other major companies will be among the first to deploy the CPU.  
- The design focuses on deterministic performance and integrates memory with compute functions.  

Arm unveiled its first homegrown silicon — yes, an actual chip, not another shake-n-bake blueprint — during an event in San Francisco on Tuesday, and said that flagship customer Meta is set to deploy the 136-core CPU at scale later this year.

Dubbed the AGI CPU, the British chip designer’s first Arm-branded datacenter processor is designed with agentic AI in mind. You heard it here first folks, artificial general intelligence (AGI) is here and it’s a ... it’s a CPU.

The new hardware represents a sea change in the British chip designer's business model. While Arm is no stranger to datacenter silicon, its involvement in those products up until now has been to license the core IP or instruction set architecture necessary to build them.

Despite the hypemaxxed branding, the chip’s Arm Neoverse V3 cores won’t be running AI models themselves. That’s a job for GPUs or one of the growing number of high-end AI ASICs. Instead, Arm sees its first datacenter CPU powering AI agents. In this respect, the chip will compete directly with Nvidia’s standalone Vera CPUs and rack systems detailed at GTC last week.

“We think that the CPU is going to be fundamental to ultimately achieving AGI,” Mohamed Awad, Arm’s EVP of cloud AI, told El Reg.

While GPUs have gotten the lion’s share of attention in recent years, the rise of agentic systems like OpenClaw have brought the need for general-purpose compute back into view. These frameworks need CPU cores and memory to write and execute code, automate tasks, and facilitate the reinforcement learning used to train next gen models.

Arm is betting on the proliferation of these agents to drive a four-fold increase in CPU demand, and it’s positioning its latest chip to capitalize on this trend.

A CPU built for AI

Arm’s AGI CPU is a 300-watt part with 136 of its Neoverse V3 cores clocked at up to 3.7 GHz (3.2 GHz base), spread across two dies fabbed on TSMC’s 3 nm process. The processor features 2 MB of L2 cache per core along with 128 MB of shared system-level cache (SLC).

According to Awad, a concerted effort has been made to avoid including accelerators or functions which eat up die area and ultimately don’t benefit the target workload.

“The way that legacy CPUs had been built worried about things like support for legacy applications,” he said. “We specifically didn't want to add things that weren't going to...be 100 percent utilized in the mission of this device.”

He added, “This is a clean sheet design meant to address all that.”

Unlike Nvidia’s Vera, Arm has opted to forego simultaneous multithreading for its agent-optimized processors, with Awad arguing that one thread per core allows for more deterministic performance scaling.

The CPU is fed by 12 channels of DDR5 — presumably 6 channels per die — with support for memory speeds up to 8800 MT/s. At 825 GB/s of aggregate bandwidth, that works out to 6 GB/s per core.

Unlike many modern CPUs, the chip’s memory and I/O functions are integrated into the same die as the compute in an effort to minimize latency. Because of this, each socket will be exposed to the operating system as two distinct NUMA domains.

Finally, for I/O, the processor is equipped with 96 lanes of PCIe 6.0 connectivity and support for CXL 3.0.

Maximum density

Meta, which is already deploying large numbers of Nvidia’s Arm-based Grace CPUs and plans to use the company’s Vera chips, will also be among Arm’s first major CPU customers.

As part of these efforts, Arm says that its validated two different OCP rack designs. There's a 36 kW air-cooled rack with 30 compute blades totalling 8160 cores per rack.

The company has also validated an even denser 200 kW liquid-cooled rack with 42 eight-node servers, which works out to 45,696 cores. For reference, that’s more than twice the core count of Nvidia’s Vera ETL256 CPU racks at 22,528.

For more than just agents

Meta isn’t the only customer lining up to gobble Arm’s new processors. OpenAI, SAP, Cerebras, Cloudflare, F5, SK Telecom, and Rebellions are also listed as early customers.

In addition to AI agents, Arm sees applications for the chip as a head node for custom accelerators, or even as a general purpose CPU for networking or storage. In fact, we’re told that OEM partners, including Lenovo, are already working on 19-inch systems using the chip.

Up until now, enterprise customers have had limited choices with regard to Arm datacenter silicon, with Ampere computing being the only non-cloud-based player in town.

Arm’s AGI CPU is set to arrive later this year. Whether it’ll be what actually brings about The Singularity is another matter entirely. ®

Correction: In a briefing before the event, Arm gave us incorrect numbers for the first air-cooled rack: It is in fact 36kW with 30 blades, for a total of 8,160 cores.

Artificial General Intelligence News