In a recent memo, Kasriel acknowledged the concerns raised by users regarding the privacy measures implemented during the initial launch, which underwent extensive risk assessments. He emphasized the team's confidence in these protections but also recognized feedback about the handling of personal data on work devices, issues related to battery longevity, and the desire for greater control over the timing of data capture.

Meta reduces its plans to monitor employees' clicks and keystrokes for AI training.

A bright, sunlit open-plan office with a large Meta logo mounted on the back wall. In the foreground, a desktop monitor displays real-time click and keystroke counters frozen at zero with a red “PAUSED” stamp across them. Warm natural light streams through floor-to-ceiling windows as a handful of adult employees stand up, stretch, and walk away in relief from their desks.

Meta scales back plan to track workers' clicks and keystrokes to train AI - BBC

- Company maintains confidence in initial privacy protections after extensive risk reviews.  
- Users express concerns about personal data on work devices and battery life.  
- Demand for greater control over capture functionality was highlighted by users.  

"While we remain confident in the privacy protections we put in place at launch, which went through several layers of risk review, we have heard your concerns about personal data on work devices, battery life, and wanting more control over when capturing happens," Kasriel said in the memo.

Meta scales back plan to track workers' clicks and keystrokes to train AI

When it comes to your privacy preferences, you have several options to tailor your experience. By selecting 'Accept all', you permit us and our partners to utilize technical identifiers along with browsing and search data for various purposes, including analytics, personalized advertising, content delivery, and audience research. This involves the collection of information via cookies and the use of precise geolocation data.

If you prefer not to have us or our partners utilize cookies and your personal information for these specific uses, you can choose 'Reject all'.

For those interested in specifically fine-tuning their choices, the option to 'Manage privacy settings' is available.

Remember, you can revoke your consent or modify your preferences at any point. Simply click the links labeled 'Privacy and Cookie settings' or 'Privacy dashboard' found on our websites and applications. For further details regarding the handling of your personal data, please refer to our Privacy Policy and Cookie Policy.

Your options for privacy

An over-the-shoulder view of an adult at a wooden desk bathed in warm morning light, reaching to slide a glowing “Privacy” toggle on a sleek translucent interface panel hovering above a laptop. Floating icons of a padlock, an eye with a slash, and a shield softly illuminate the scene, suggesting thoughtful choices. A steaming coffee cup and scattered notes hint at careful decision-making.

I run a workplace management platform. Here's how AI is reshaping both the culture and physical form of offices. - Yahoo

- Users can accept or reject the use of cookies and personal data.  
- Clicking 'Manage privacy settings' allows customization of privacy choices.  
- Consent can be withdrawn anytime through 'Privacy and Cookie settings' links.  

Your privacy choices

Accept all', we and technical identifiers and browsing and search data, for analytics, personalised advertising and content, advertising and content measurement, and audience research and services development. If you click '', we and our partners , including 250 who are part of the IAB Transparency & Consent Framework, will also store and / or access information on a device (in other words, use cookies) and use precise geolocation data and other personal data such asand browsing and search data, for analytics, personalised advertising and content, advertising and content measurement, and audience research and services development.

If you do not want us and our partners to use cookies and personal data for these additional purposes, click 'Reject all'.

If you would like to customise your choices, click 'Manage privacy settings'.

You can withdraw your consent or change your choices at any time by clicking on the 'Privacy and Cookie settings' or 'Privacy dashboard' links on our sites and apps. Find out more about how we use your personal data in our Privacy Policy and Cookie Policy.

Your privacy choices

In a recent discussion, Paul Triolo from DGA Albright Stonebridge Group shed light on China’s significant steps to enhance its capital markets, reflecting a sophisticated approach to its artificial intelligence (AI) industrial strategy. Triolo emphasized the implications of the public offerings of Chinese firms involved in AI hardware and advanced language models. He also explored how U.S. export restrictions and the inclusion of certain Chinese companies on various entity lists might impact the growth and development of the AI sector in China. This evolving landscape raises important questions about the future dynamics of global competition in AI technology.

Paul Triolo: China's opening of capital markets reflects the advancement of its AI industrial policy.

A warm-lit scene of a golden microchip sprouting green shoots that morph seamlessly into upward-trending stock market graphs; a single Chinese flag waves gently in the background, symbolizing the maturity of China’s AI industrial policy and its unleashed capital markets.

China's 'unleashing' of capital markets shows a maturity of its AI industrial policy: Paul Triolo - CNBC

- China demonstrates maturity in AI industrial policy through capital market developments.  
- Public listings for AI hardware and large language models are gaining momentum.  
- U.S. export controls may significantly impact China's AI industry and companies.  

Share Share Article via Facebook Share Article via Twitter Share Article via LinkedIn Share Article via Email

China's 'unleashing' of capital markets shows a maturity of its AI industrial policy: Paul Triolo

Paul Triolo of DGA Albright Stonebridge Group discusses the public listing of Chinese AI-related hardware and large language models, as well as how the U.S.' export controls and placing of Chinese companies on various entity lists could affect the sector.

China's 'unleashing' of capital markets shows a maturity of its AI industrial policy: Paul Triolo

Shares of prominent Chinese technology firms Alibaba and Baidu experienced a significant uptick on Thursday following their collaboration with Apple to implement their artificial intelligence tools. Alibaba's shares, listed in Hong Kong, surged by 5% after the company announced that its Qwen AI model would be incorporated into Apple’s services within China. In the U.S., Alibaba's shares also closed slightly higher after a spokesperson confirmed to CNBC that "Qwen will be integrated into Apple Intelligence experiences across iOS, iPadOS, macOS, and visionOS for users in China."

Baidu's shares in Hong Kong climbed by 4% as the company revealed its ongoing partnership with Apple to develop AI features for iPhones available in China. This collaboration comes on the heels of reports from late June indicating that Baidu's AI chip division, Kunlunxin, is eyeing an initial public offering that could value the unit at around $50 billion. In a recent announcement, the Cyberspace Administration of China designated Apple Intelligence, along with several other AI services from companies like Huawei Technologies, as approved service providers for smartphone-based applications.

The partnership between Apple and Alibaba is expected to enhance user experience by allowing seamless access to Qwen's capabilities, such as text and image comprehension and generation, without requiring users to switch between different applications, according to an Alibaba representative. Apple has yet to provide a response to requests for additional comments from CNBC.

Alibaba and Baidu stocks soar in Hong Kong following Apple AI collaboration

A sunlit Hong Kong trading floor with warm morning light streaming through floor-to-ceiling windows. In the center, a rising green arrow on an electronic ticker carries oversized 3D logos of Alibaba, Baidu, and Apple climbing skyward. The Hong Kong flag flutters subtly in the background as suited traders exchange confident glances at soaring share prices on nearby screens.

Alibaba and Baidu shares jump in Hong Kong on Apple AI partnership - CNBC

- Alibaba and Baidu's shares rose following their partnership with Apple for AI tools.  
- Alibaba confirmed its Qwen AI model will integrate with Apple services in China.  
- Baidu is collaborating with Apple on AI features for iPhones in China.

In this article 9888-HK Follow your favorite stocks CREATE FREE ACCOUNT

Chinese tech company Baidu, best known for its search engine, also operates cloud, mapping and other internet-based services. Bloomberg | Bloomberg | Getty Images

Shares of Chinese tech giants Alibaba and Baidu rose Thursday on their partnership with Apple for deploying their AI tools. Hong-Kong listed shares of Alibaba rose 5% after the company confirmed that its Qwen AI model would be integrated into Apple services in China. U.S.-listed shares of Alibaba had closed slightly higher overnight after an Alibaba spokesperson told CNBC that "Qwen will be integrated into Apple Intelligence experiences within iOS, iPadOS, macOS, and vision OS for users in China."

Stock Chart Icon Stock chart icon Alibaba HK shares

Baidu 's Hong Kong-listed shares gained 4% as the company confirmed that it was working with Apple on Apple Intelligence features for iPhones in China. This comes amid reports in late June that its artificial intelligence chip unit Kunlunxin is targeting an initial public offering in the city, which could value its affiliate at $50 billion. The Cyberspace Administration of China in a notice on Wednesday included Apple Intelligence, along with six other smartphone-based AI services including Huawei Technologies, in a list of approved service providers. The Apple-Qwen combination will allow users to access the model's capabilities, "like text and image understanding and generation, without needing to jump between tools," the Alibaba spokesperson added. Apple did not immediately respond to CNBC's request for comments.

Stock Chart Icon Stock chart icon Baidu hk shares

Alibaba and Baidu shares jump in Hong Kong on Apple AI partnership

Mira Murati, previously the Chief Technology Officer at OpenAI and now the CEO of Thinking Machines, recently spoke at the Wall Street Journal's WSJ Tech Live Conference held in Laguna Beach, California, on October 17, 2023. This marks a notable moment in her career following the founding of Thinking Machines Lab in February 2025, alongside several senior researchers from OpenAI. The company made headlines by raising an unprecedented two billion dollars in its seed round, achieving a remarkable twelve billion dollar valuation without releasing any products. Major investors include tech giants like Nvidia, AMD, Cisco, Andreessen Horowitz, and Jane Street, setting high expectations for the emerging firm.

This week, Thinking Machines unveiled its inaugural model, Inkling. It employs a mixture-of-experts transformer architecture featuring an impressive total of 975 billion parameters, with 41 billion being active per token. Inkling was trained on an extensive dataset of 45 trillion tokens, which encompasses text, images, audio, and video. It is designed to accommodate a context window of one million tokens. Despite these formidable numbers, the model's performance has not set it apart from its competitors. Thinking Machines readily acknowledges this, stating that Inkling is not currently the strongest model available, whether open or closed. Benchmark tests such as Humanity's Last Exam, Terminal Bench, and SWE-Bench Verified reveal that it lags behind leading Chinese open models, such as Zhipu's GLM 5.2 and Moonshot's Kimi K2.6.

What stands out is the fact that this twelve billion dollar American lab, staffed by the creators of ChatGPT, has introduced a model that does not outperform its free Chinese counterparts. A deeper investigation into this phenomenon reveals some intriguing insights.

For the past eighteen months, U.S. authorities have accused Chinese labs of basing their models on pilfered American technology. Following DeepSeek's impressive debut in early 2025, the rapid response from OpenAI and government officials was that DeepSeek had utilized distillation techniques on OpenAI models. Distillation is the process of training one's model using the outputs of a more advanced one. Subsequent Congressional inquiries have asserted that Chinese AI companies were engaging in coordinated distillation efforts aimed at U.S. frontier models, leading to accusations of theft.

With this context, let's delve into Inkling. Its architecture closely resembles that of DeepSeek-V3, a fact that both technical analysts and the Thinking Machines team itself have noted shortly after its release. The company admits that Inkling's post-training involved supervised fine-tuning with synthetic data derived from open-weight models, including the Chinese Kimi K2.5 model developed by Moonshot AI.

This raises an important point: an American lab has essentially embraced a Chinese model architecture and utilized a Chinese dataset for training its flagship product. While nobody is labeling this as theft—and rightly so, as the Chinese models in question are open-source and licensed for use—there exists a noticeable disparity in the narrative. When Chinese labs draw from American models, it’s considered criminal; however, when American labs learn from Chinese models, it is approached as engineering innovation. Such selective rhetoric poses challenges for coherent technology policy-making.

This brings us to a crucial question: why launch a model like Inkling, which isn't even the best in the market? The strategy appears calculated, benefiting from a protective market environment. Inkling is licensed under the open-weight Apache 2.0, predominantly marketed as a chatbot but also as a customizable platform through Tinker, the company’s fine-tuning tool. Thinking Machines seems to be banking on enterprises prioritizing adaptability over sheer capability.

However, could American companies not adapt superior Chinese models instead? They attempted to do so but found themselves gravitating towards Chinese open models, which are both high-quality and cost-effective. These Chinese models now represent about 45 percent of the enterprise tokens processed through OpenRouter. Companies like Coinbase have significantly cut down their AI expenditures by leveraging models like GLM and Kimi, while Cursor has developed its Composer model based on Kimi.

Despite this trend, the U.S. government is actively seeking to restrict access to these models. The State Department recently issued warnings to American businesses regarding the potential risks associated with Chinese models. Additionally, investigations have been initiated into firms like Airbnb and Cursor due to their use of Chinese AI tools. Proposed procurement bans are in the works, making it increasingly risky for any company involved with government contracts—or facing the prospect of congressional scrutiny—to engage with Chinese open models. 

Inkling is positioned to fulfill the gap left by these restrictions. It serves as the compliant alternative, not the supreme one, but rather a sanctioned low-risk option.

Finally, it is essential to reflect on the background of Thinking Machines’ founder. Mira Murati’s tenure as OpenAI's CTO means she possesses unparalleled insight into the inner workings of OpenAI’s technology. Given her expertise and the significant resources available to her, her decision to adopt an architecture inspired by DeepSeek and a training strategy influenced by Kimi carries weight. This choice suggests a shifting landscape in the realm of open research, particularly with OpenAI experiencing a noticeable slowdown. The rollout of GPT-5.6 has been limited, while Kimi K2.6 has surpassed GPT-5.5 in certain benchmarks, marking a significant moment where an open-weight model has bested a leading proprietary model. The competitive landscape between American closed labs and their Chinese counterparts is tightening, with the distinctions between them narrowing on multiple fronts. When a former leader from OpenAI opts for a design derived from Chinese models instead of her previous employer’s methodologies, it compels the market to reevaluate assumptions about the strength of OpenAI’s position.

Murati is aware of OpenAI's secrets; her AI indicates a preference for China.

A moody, shadow-filled high-tech control room: at center, a lone figure in silhouette (Mira Murati) stands poised. To her left, a large monitor glows with the word “OpenAI” in stark white text overlaid on schematics and circuit diagrams. To her right, another screen displays a vivid red Chinese flag superimposed on golden neural-network pathways. Dramatic overhead spotlight casts deep shadows, dark muted palette, cinematic tension.

Murati Knows OpenAI’s Secrets. Her New AI Signals She Prefers China’s. - Forbes

- Mira Murati's Thinking Machines Lab launched its first model, Inkling, with notable specifications.  
- Inkling, while impressive, underperforms compared to free Chinese models like Kimi K2.6.  
- The model serves as a low-risk alternative amid increasing scrutiny of Chinese AI usage.

Mira Murati, former Chief Technology Officer of OpenAI, now CEO of Thinking Machines, speaks during The Wall Street Journal's WSJ Tech Live Conference in Laguna Beach, California on October 17, 2023. (Photo by Patrick T. Fallon / AFP) (Photo by PATRICK T. FALLON/AFP via Getty Images) AFP via Getty Images

Thinking Machines Lab released its first model this week. The company was founded in February 2025 by Mira Murati, the former Chief Technology Officer of OpenAI, along with a cohort of senior OpenAI researchers. It raised a record two billion dollar seed round at a twelve billion dollar valuation before shipping a single product. Nvidia, AMD, Cisco, Andreessen Horowitz and Jane Street backed it. Expectations are enormous.



The newly released model is called Inkling. It is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active per token. It was pretrained on 45 trillion tokens spanning text, images, audio and video. It supports a one million token context window. These are very impressive numbers, yet the performance in comparison with other models is less impressive. To their credit, Thinking Machines concede this themselves. In the company’s own words, Inkling is not the strongest model available today, open or closed. On benchmarks like Humanity's Last Exam, Terminal Bench and SWE-Bench Verified, it trails leading Chinese open models including Zhipu's GLM 5.2 and Moonshot's Kimi K2.6.



It is quite remarkable that a twelve billion dollar American lab staffed by the people who built ChatGPT have released a model that underperforms free Chinese alternatives. It is worth diving into, and when we do, the details end up being quite revealing.



Distillation Wars

For eighteen months Washington has accused Chinese labs of building their models on stolen American intellect. When DeepSeek stunned the AI community in early 2025, the immediate response from OpenAI and from US officials was that DeepSeek had distilled OpenAI’s models. Distillation means using a stronger model’s outputs to train your own. Congressional investigators have since claimed that Chinese AI firms ran coordinated distillation campaigns against US frontier models. The word used in Washington is theft.



With that background in mind, let’s take a look at Inkling. Its mixture-of-experts design largely follows DeepSeek-V3. That is not my characterization. It is plainly visible in the published architecture and has been noted by technical analysts within hours of release. And by Thinking Machines’ own admission, Inkling's post-training was bootstrapped with supervised fine-tuning on synthetic data generated by open-weight models including Kimi K2.5, a Chinese model from Moonshot AI.



In other words, an American lab adopted a Chinese architecture and distilled a Chinese model to train its flagship release. Nobody is calling this theft. And frankly, nobody should. The Chinese models in question are open weight and permissively licensed. Learning from published work is how science advances. But the asymmetry in rhetoric is now impossible to ignore. When Chinese labs learn from American models it is larceny. When American labs learn from Chinese models it is engineering. You cannot run a serious technology policy on that kind of selective outrage.



So Why Launch A Model That Isn’t The Best?

The obvious question is why release Inkling at all. The answer is found in a shrewd strategy that benefits from protectionism. Inkling is open weight under Apache 2.0. It is being pushed primarily as a chatbot but as a base for customization through Tinker, the company’s fine-tuning platform. Thinking Machines is betting that enterprises care less about the smartest general model than about a model they can make their own.



But can American companies not do that with better Chinese models? They tried to. American companies flocked to Chinese open models because they are excellent and nearly free. Chinese models came to account for roughly 45 percent of enterprise tokens routed through OpenRouter. Coinbase cut its AI bill nearly in half by moving its agents to GLM and Kimi. Cursor built its Composer model on Kimi.



But Washington is now moving to shut that door. The State Department warned American companies this month about the risks of Chinese models. House committees launched probes into Airbnb and Cursor over their use of Chinese AI. Procurement bans are being drafted. For any company that touches government work, and increasingly for any company that fears congressional subpoenas, Chinese open models are becoming untouchable regardless of what the final rules say. The risk overhang is massive, and that is all that’s needed to push companies in other directions.



Inkling exists to fill that vacuum. It is the compliant base model. Not the best base model. Just the permitted, low-risk one.



What Inkling’s Architecture Tells Us About OpenAI

Let’s revisit for a moment who the founder of Thinking Machines is. Mira Murati was OpenAI’s CTO. She knows exactly what is inside OpenAI's stack. She served as interim CEO during the November 2023 board crisis. Few people on earth have better visibility into what OpenAI has built and what it can build. Given a blank slate, two billion dollars and Nvidia's newest GB300 systems, she chose an architecture that follows DeepSeek and a training bootstrap that leans on Kimi.



That is a not insignificant data point about where the technical frontier of open research actually sits. It arrives alongside a measurable deceleration at OpenAI itself. GPT-5.6’s rollout was limited. Kimi K2.6 edged past GPT-5.5 on SWE-Bench Pro, the first time an open-weight model surpassed a leading proprietary model on that benchmark. The gap between American closed labs and Chinese open labs has compressed to months on many tasks and has inverted on some. When the person who ran OpenAI’s technology organization builds from Chinese blueprints rather than from anything resembling her former employer's approach, the market should - like any good Bayesian thinker - update its beliefs about how large OpenAI's remaining moat really is.

Murati Knows OpenAI’s Secrets. Her AI Suggests She Prefers China’s.

Recent developments in the scientific community are paving the way for a groundbreaking method of energy production known as fusion, bringing it ever closer to being a reality for humanity. 

Significant strides have been made in the realm of plasma containment—an essential aspect of fusion, where the challenge lies in maintaining superheated plasma long enough for hydrogen nuclei to combine and produce significant energy outputs. The concept behind fusion involves merging two light atomic nuclei to create a heavier nucleus, which releases a substantial amount of energy in the process.

Kevin Zondervan of Promethean Action highlights the progress being made, particularly with a startup named Avalanche, which is focused on the development of small fusion reactors for various vehicles. He notes that we are witnessing remarkable achievements, such as breakeven plasma confinement times reaching as high as 22 minutes and contracts emerging for the purchase of fusion power. While most current initiatives are aimed at electrical power plants, there is a vision for extending fusion technology to vehicles, planes, ships, and even smaller, localized reactors for remote applications.

Before realizing this potential, however, significant foundational work in fusion technology must take place. Many experts believe that artificial intelligence could emerge as the initial consumer of this type of energy. 

Dennis Whyte, a key figure in fusion development, shared insights during a TED Talk event in June. With both entrepreneurial and academic experience—including involvement with MIT—Whyte underscored that the field is still in its early stages. "To turn fusion into a practical energy solution, we need to operate at scale and cost that make it viable," he explained. He described a current project in Devens, Massachusetts, which aims to demonstrate an energy gain where the output from fusion plasma exceeds the initial input, all while focusing on commercial viability.

Looking ahead, Whyte suggests that we could see a prototype of a fusion power plant by 2030, as commitments for construction and financial backing are starting to materialize. He also pointed out the potential relationship between AI and fusion, noting that AI may be ideally suited to benefit from the local, high-power density sources offered by fusion, particularly since it avoids certain supply chain vulnerabilities. Innovations in fusion may be further accelerated by incorporating new ideas, as technological advancements from various fields often enhance established scientific principles.

AI Accelerates Energy Development in Fusion Research

A spacious fusion research facility flooded with warm natural light, a ring-shaped tokamak glowing faintly with plasma in the center, engineers in lab coats studying AI-driven data on sleek monitors, curves of neural-network graphs and real-time simulations subtly overlayed on the screens, the fusion core humming softly behind a protective viewport—capturing AI’s role in accelerating clean energy breakthroughs.

Working On Fusion: AI Drives Energy Acceleration Processes - Forbes

- Scientists are advancing in fusion energy, improving plasma containment for energy generation.  
- Startup Avalanche is designing small fusion reactors for vehicles and commercial applications.  
- MIT's Dennis Whyte projects fusion power plants could be operational by 2030.

A vibrant illustration of a glowing atomic structure with electron orbits on a dark background, symbolizing concepts in quantum physics, energy, and scientific research. getty

We’re making headway, in the scientific world, on a powerful form of energy generation. It’s called fusion, and it seems, more than ever, that sort of thing is almost within humanity’s grasp.

For one thing, scientists have been making great advances in plasma containment. That’s important, because one of the abiding challenges in the fusion puzzle is keeping the superheated plasma in place long enough to allow hydrogen nuclei to fuse, and get net energy outputs.

That’s why it’s called “fusion” – the general idea is that two light atomic nuclei combine to form a heavier nucleus. That releases energy – a whole lot of it.

“The closer we get to realizing commercial fusion power for electricity generation, the more we see records and barriers broken around the world,” writes Kevin Zondervan at Promethean Action, chronicling the efforts of a startup called Avalanche in designing small fusion reactors for vehicles. “Breakeven plasma confinement times of up to 22 minutes, contracts to buy fusion power, plans for siting commercial fusion plants, etc. All of those plans are for electrical power plants. What about cars, trucks, planes, ships, and spaceships? What about your little reactor for your farm in the middle of nowhere?”

Before that happens, though, we’ll need the less sexy process of building basic fusion technology and proving the theory to a high level, and there’s a general consensus that we wouldn’t be doing all of this unless AI was “the first customer” for this type of energy generation.

Thoughts from a Leader in Fusion I was talking recently to Dennis Whyte about all of this at a TED Talk event in June. Whyte has both the corporate and the academic credentials to speak with authority: he has started multiple companies, and is also a member of MIT faculty on energy. Whyte stressed that all of this is, essentially, in its infancy. “You actually have to do it at a scale and a cost that would eventually make fusion practical,” he said. With that in mind, he described a project that is currently underway in Devens, Mass. (See more about this type of thing here .) “Its purpose is to not only demonstrate energy gain, namely that you get more energy out of fusion plasma than you started with, but it's also doing that at an obviously commercial scale,” Whyte said.

A Timeline “It should be developed according to plan,” Whyte said of related efforts, “that by 2030 we'll see a demonstration of something that starts to look like what a fusion power plant would be, and we're actually starting to see the commitments, both financially and locally, about actually constructing these so that they would start to put power on the grid.” What does AI have to do with it? Whyte explained: “AI is a particularly interesting and very recent advancement in this,” he said, “because AI will almost certainly be the first customer for fusion. (see my reference above) Because it's almost exactly tuned up for what AI needs, namely, local high power density sources that also don't have vulnerability to (certain) supply chains.” Work on fusion, Whyte suggested compounds as new ideas come along. He described that process this way: “That often happens in technology: innovations that come in from the outside actually start amplifying what is established science elsewhere.”

Working On Fusion: AI Drives Energy Acceleration Processes

In recent years, stocks related to artificial intelligence (AI) infrastructure have been at the forefront of market growth. However, they've recently faced some challenges, prompting concerns about a possible decline in data center expansion. Nonetheless, this atmosphere appears more akin to a typical market correction following a robust period of gains.

Three semiconductor stocks that stand out as excellent buying opportunities amid the current downturn are Nvidia, Advanced Micro Devices (AMD), and Broadcom. Each of these companies boasts significant growth potential, and investments in AI data centers are expected to remain sturdy for the foreseeable future.

Starting with Nvidia, the company has experienced a decline in stock price that has brought its valuation down to a forward price-to-earnings (P/E) ratio of 16 based on analysts' expectations for its fiscal year ending January 2028. This valuation positions Nvidia as one of the most attractive options in the semiconductor space. The company's CUDA software platform has created substantial competitive advantages in the AI model training market, as a large portion of foundational AI code runs on CUDA and is tailored for its graphics processing units (GPUs).

Nvidia is also adeptly navigating the shift towards more inference and agentic AI workloads. It has evolved from merely a GPU manufacturer to a comprehensive AI infrastructure provider, creating servers specifically designed for various AI applications. Following its acquisition of Groq, Nvidia has incorporated chips optimized for inference into its CUDA ecosystem, while its networking segment is rapidly growing.

With a promising trajectory ahead, Nvidia remains a strong buy, especially given its current valuation dip.

Next up is Advanced Micro Devices. The company is well-poised to capitalize on two prominent AI trends: inference and agentic AI. AMD’s chip portfolio allows it to pursue a greater share of the AI inference market, as evidenced by its significant GPU contracts with major players like OpenAI and Meta Platforms.

Inference relies heavily on swift memory access, and AMD is focusing its innovations in this area. Its chiplet architecture enables GPUs to accommodate more memory, and through its recent acquisition of the memory optimization platform MEXT, it can virtually enhance memory capacity without compromising performance, thereby minimizing costs for clients.

At the same time, AMD is set to benefit from the surge in agentic AI, which demands greater computing power from central processing units (CPUs). As a long-standing leader in data center CPUs, AMD is strategically positioned to meet the rising demand, with projections suggesting that the CPU market in data centers could double to $120 billion by 2030. The company is already developing CPUs specifically for agentic AI applications.

Given its substantial revenue growth potential, AMD is another compelling stock to consider post-sell-off.

Lastly, Broadcom has emerged as a key player in the shift to custom AI accelerators favored by hyperscalers looking to reduce costs. The company partnered with Alphabet to create Tensor Processing Units (TPUs), and with Alphabet poised to invest up to $190 billion in AI infrastructure this year, Broadcom is on track for rapid expansion. Additionally, Alphabet's agreement to provide Anthropic with $21 billion in TPUs further enhances Broadcom’s prospects.

As interest in TPUs continues to grow, other large tech companies are seeking Broadcom’s expertise in developing tailored AI chips. The firm anticipates that this market segment could surpass $100 billion by its fiscal year ending in 2027, with Citigroup estimating that Broadcom's AI revenue might reach $180 billion by fiscal 2028. The company’s data center networking division is also expanding quickly, and after sealing a $30 billion deal with Apple, its non-AI chip sector is prepared for a rebound.

Broadcom's stock currently trades at a forward P/E of 20 based on fiscal 2027 estimates. With its significant growth potential, this valuation appears quite low, making Broadcom an attractive investment opportunity.

Three AI Stocks I Favor to Purchase Amid Ongoing Chip Price Decline

A bright morning sunbeam streams through a sleek office window, illuminating a bronze bull sculpted from circuit-board traces. Three metallic logos—Nvidia, Microsoft, Google—are etched into its sides, glowing with optimism. Behind it, glass shelves stacked with semiconductor chips spill over the edge, red arrows streaking down to show the chip sell-off. The warm natural light highlights the bullish AI stocks charging ahead.

My 3 Favorite AI Stocks to Buy on the Continued Chip Sell-Off - The Motley Fool

- AI infrastructure stocks, including Nvidia, AMD, and Broadcom, face recent market pressures.  
- Nvidia's valuation has decreased, making it a strong buy with significant growth potential.  
- AMD and Broadcom are well-positioned in AI, targeting rapid growth in their markets.  

While artificial intelligence (AI) infrastructure stocks have helped lead the market higher over the past few years, more recently, these stocks have come under pressure. There is some fear of an eventual slowdown in the data center build-out, but this does look more like a typical market breather after a nice run.

Three of my favorite semiconductor stocks to buy on this sell-off are Nvidia (NVDA +0.29%), Advanced Micro Devices (AMD 3.40%), and Broadcom (AVGO +1.28%). All three still have huge growth opportunities in front of them, and spending on AI data centers should remain strong for many years.

Nvidia

The pullback in Nvidia's stock has taken its valuation down to a forward price-to-earnings ratio (P/E) of 16 times analysts' estimates for its fiscal 2028 (which ends in January 2028). That makes it one of the best bargains in the chip space. Given the moat its CUDA software platform has established, the company is set to continue dominating the market for AI model training, as most foundational AI code was written on CUDA and optimized for its graphics processing units (GPUs).

Expand NASDAQ : NVDA Nvidia Today's Change ( 0.29 %) $ 0.61 Current Price $ 212.41 Key Data Points Market Cap $5.1T Market cap calculated using publicly traded shares outstanding only. Does not include unlisted, private, or dual-class non-traded shares. Implied market cap may vary. Day's Range $ 206.05 - $ 213.81 52wk Range $ 164.07 - $ 236.54 Volume 4.5M Avg Vol 158.1M Gross Margin 74.15% Dividend Yield 0.13%

And while the use of data center processing power is shifting toward more inference and agentic AI workloads, the company is also well positioned here. Nvidia has transformed itself from a simple GPU maker into a complete AI infrastructure player, offering end-to-end servers designed for specific AI tasks. Its acquisition of Groq gave it chips designed specifically for inference, which it has incorporated into its CUDA ecosystem. Meanwhile, its networking portfolio has become the fastest-growing part of its business.

With strong growth still ahead, Nvidia remains a top stock to own, and a good buy at its discounted valuation.

AMD

Advanced Micro Devices is currently riding two of the hottest trends in AI: inference and agentic AI. The company's chip offerings make it much better positioned to take a larger slice of the AI inference pie, and it already has large GPU deals in place with OpenAI and Meta Platforms.

Expand NASDAQ : AMD Advanced Micro Devices Today's Change ( -3.40 %) $ -18.61 Current Price $ 529.52 Key Data Points Market Cap $894B Market cap calculated using publicly traded shares outstanding only. Does not include unlisted, private, or dual-class non-traded shares. Implied market cap may vary. Day's Range $ 509.63 - $ 558.89 52wk Range $ 149.22 - $ 584.73 Volume 957.6K Avg Vol 36.4M Gross Margin 47.09%

Inference is much more about fast memory access than raw compute power, and this is where AMD has focused its efforts. Its chiplet design allows its GPUs to be packaged with more memory, while its recent acquisition of memory optimization platform MEXT will allow it to virtually expand memory capacity without sacrificing performance, helping customers reduce costs.

At the same time, the company is set to ride a powerful wave in agentic AI. While other types of AI workloads have largely needed GPUs to provide their processing power, agentic AI workflows require more participation from central processing units (CPUs).

AMD has long been a leader in data center CPUs, and as AI agents proliferate, the need for CPUs is expected to grow rapidly. In fact, the GPU-to-CPU ratio in new data centers is expected to shrink from 8 to 1 for training to 1 to 1 for agentic AI. AMD has projected that the data center CPU market will double in size to $120 billion by 2030. It is already developing CPUs specifically for agentic AI.

With huge revenue growth ahead of it, AMD is a top stock to buy after its sell-off.

Broadcom

Broadcom has been one of the biggest beneficiaries of the trend among hyperscalers to deploy custom AI accelerators to help save costs. It helped Alphabet develop its Tensor Processing Units (TPUs), and with the search giant set to spend up to $190 billion on AI infrastructure this year, Broadcom is set to see rapid growth. Adding to that, Alphabet has agreed to sell Anthropic $21 billion worth of TPUs.

Expand NASDAQ : AVGO Broadcom Today's Change ( 1.28 %) $ 4.99 Current Price $ 394.10 Key Data Points Market Cap $1.9T Market cap calculated using publicly traded shares outstanding only. Does not include unlisted, private, or dual-class non-traded shares. Implied market cap may vary. Day's Range $ 386.79 - $ 398.86 52wk Range $ 273.00 - $ 495.00 Volume 690K Avg Vol 26.3M Gross Margin 65.66% Dividend Yield 0.65%

The success of TPUs led other hyperscalers to turn to Broadcom for help in developing custom AI chips. It expects this to grow into a more than $100 billion business in its fiscal 2027, while Citigroup has projected that Broadcom's AI revenue could rise to $180 billion in its fiscal 2028. The company also has a fast-growing data center networking business, and after signing a $30 billion deal with Apple (AAPL +3.95%), its non-AI chip business also looks set for a turnaround.

The stock is trading at a forward P/E of just 20 times fiscal 2027 estimates. Given its potential explosive growth, that's too cheap, and makes it an attractive buy.

My 3 Favorite AI Stocks to Buy on the Continued Chip Sell-Off

On Wednesday, Jim Cramer from CNBC expressed that it is crucial for companies to demonstrate the financial benefits of their investments in artificial intelligence. He stated, "I need cold hard return facts," highlighting a growing skepticism about the returns on these substantial tech expenditures.

The ongoing surge in AI has led to significant investment by tech firms, with projections indicating that total capital outlays for AI could exceed $1 trillion by 2027. While Cramer maintains a hopeful outlook on the long-term prospects of AI, he emphasized that the market is in need of concrete evidence that these investments are yielding significant financial benefits for their clients.

Cramer voiced concern that, so far in the earnings season, companies utilizing AI have not provided compelling data to show revenue increases or cost efficiencies linked to the technology. "We’re still early in the earnings season, but already we are not hearing anything material about the use of AI," he remarked.

He expressed particular disappointment with the banking sector, citing its potential to leverage AI for process automation and efficiency. However, financial institutions have yet to demonstrate that AI is significantly impacting their results. "It's valuable, but nothing that can raise numbers," he said, adding that it has not improved efficiency ratios or reduced hiring.

Despite the boom benefiting AI infrastructure firms, Cramer noted that many companies employing the technology have yet to see similar gains. He pointed to firms like Anthropic and Micron, which have thrived, but questioned why the end clients are not reporting substantial savings from AI.

Cramer highlighted that only a few companies, such as fintech leader Block and cybersecurity firm Cloudflare, have directly linked job reductions to their AI initiatives. Block announced its AI-related layoffs in February, while Cloudflare did so in May. Critics also caution that some businesses may use AI as a trendy justification for downsizing, coining the term "AI washing."

In conclusion, Cramer warned that without more companies showcasing tangible returns from AI investments, skepticism will only grow, impacting the major tech players among the big spenders in the industry. "The longer we wait to hear how actual clients are profiting, the more we will scrutinize the tech giants that seem to be thriving," he concluded.

Jim Cramer expresses the need for undeniable evidence that AI is yielding results.

A shadowy trading desk in a darkened room, warm tungsten light spilling across cracked ice shards on the table. On the screens behind, bold “AI” overlays flickering stock charts. A sharply dressed analyst, face half-lit by the monitors, grips a chunky ice block etched with the words “COLD HARD PROOF.” A small hammer lies discarded at his side, shards glistening in the tense, dark atmosphere.

Jim Cramer says he needs 'cold hard' proof that AI is paying off - CNBC

- Jim Cramer demands companies show measurable financial returns from AI investments to maintain optimism.  
- Analysts predict AI spending could exceed $1 trillion by 2027, yet results remain unclear.  
- Banks have disappointed Cramer with few signs of AI improving efficiency or profitability.  

CNBC's Jim Cramer said Wednesday that it's time for companies to prove artificial intelligence is paying off.

"I need cold hard return facts," the "Mad Money" host said. "Or, I, too, will grow more skeptical than I am now."

The AI boom has fueled massive spending by technology companies, and there's no end in sight. Analysts estimate total AI capital expenditures could climb above $1 trillion in 2027. While Cramer said he remains optimistic about the long-term opportunity, he argued the market needs more evidence that those investments are translating into measurable financial returns for customers.

Cramer said one of his biggest concerns this earnings season is that companies adopting AI have largely failed to point to meaningful revenue gains or cost savings from the technology.

"We're still early in the earnings season but already we are not hearing anything material about the use of AI," he said.

Banks, in particular, have disappointed him on the AI front. Cramer said financial institutions seemed like natural beneficiaries of AI because of the potential to automate processes and improve efficiency, yet management teams have offered little evidence that the technology is materially improving results.

It hasn't been complete crickets, Cramer said. "It's valuable, but nothing that can raise numbers. It's not helping the efficiency ratio that we can tell and it's not allowing them to cut back on hiring. Does that mean AI is a bust? No. But I don't see it making much difference."

While AI infrastructure companies continue to benefit from the spending boom, Cramer said the same cannot yet be said for many of the businesses buying the technology.

"Sure Anthropic is getting a return ... The component companies are doing well," he said, alluding to companies like memory-chip maker Micron , whose profits have soared. "But shouldn't the ultimate clients ... be able to cite at least a couple of million in savings?"

Cramer said only a handful of companies, most notably fintech firm Block and web-security provider Cloudflare , have clearly attributed recent layoffs to AI adoption. Block did so in February, while Cloudflare's job cuts were disclosed in May. Plus, critics argue some companies may also cite AI as a buzzy excuse for cuts, leading to the creation of the term "AI washing."

Ultimately, Cramer said that if more businesses do not begin reporting tangible returns, the AI skeptics will grow louder, with ramifications for the tech industry's big spenders.

"The longer we go without hearing how actual clients make money, the longer we'll take days like today, when it seems that the hyperscalers are making money," with a grain of salt, he said.

Jim Cramer says he needs 'cold hard' proof that AI is paying off

President Donald Trump took to Truth Social on Wednesday to criticize New York Governor Kathy Hochul for her recent moratorium on data center construction. He urged the state to revise this policy "IMMEDIATELY," as New York has become the first in the nation to implement such a ban.

In his post, Trump emphasized the importance of data centers in the job market, stating, "One of the biggest Driving Forces in the Future for Jobs, are Data Centers." He described these facilities as powerful assets that generate substantial revenue for the states where they are established. He asserted that Governor Hochul had halted both ongoing and future data center projects due to political motivations.

The governor's executive order, which she signed on Tuesday, imposes a temporary ban on the development of large data centers consuming 50 megawatts of electricity or more, lasting for a duration of up to one year. Trump expressed strong disapproval, stating, "New York State has made a terrible decision" in this regard.

Trump criticizes New York's AI data center moratorium, urges the state to alter its policy 'right away'

Donald Trump stands at a sunlit outdoor podium, gesturing emphatically toward a sleek AI data center in the distance. The facility is ringed by red “MORATORIUM” barrier tape on a chain-link fence. A single New York state flag flutters behind him. Warm, natural morning light highlights the urgent, confrontational mood.

Trump blasts New York AI data center moratorium, says state should change policy 'immediately' - CNBC

- Trump criticized Gov. Hochul's data center moratorium as detrimental to job growth.  
- The moratorium prohibits large-scale data centers utilizing 50+ megawatts of power.  
- Trump urged immediate change to the policy, calling it a "terrible decision."  

President Donald Trump slammed New York Gov. Kathy Hochul's new data center moratorium on Wednesday, saying in a Truth Social post that the state should change its policy "IMMEDIATELY" after it became the first in the U.S. to impose a ban of its kind.

"One of the biggest Driving Forces in the Future for Jobs, are Data Centers," he wrote. "They are big, strong, bold, and Money Machines for the State in which they are built. Governor Kathy Hochul, for political reasons, has terminated all Data Centers being built, or to be built, in New York State."

The new state executive order, signed by the governor Tuesday, bars for up to a year the construction of large-scale data centers that use 50 or more megawatts of power.

"New York State has made a terrible decision," Trump wrote Wednesday.

Trump blasts New York AI data center moratorium, says state should change policy 'immediately'

Imagine the intelligence of an AI system powered by thousands of advanced computer chips — impressive, right? Now, consider the cognitive abilities of a one-year-old child. While infants may lack the capacity to write software, tackle complex equations, or engage in philosophical discussions, they possess an incredible ability to learn and understand their surroundings with remarkable efficiency.

Unlike modern AI models, which need massive datasets and consume energy equivalent to that of a small nation, babies quickly recognize new objects after just a few glimpses and acquire knowledge through brief observations and hands-on experiences. This innate learning capability offers profound insights that could pave the way for advancements in artificial intelligence.

Research teams from Meta, Stanford University, the University of Tokyo, and France's École Normale Supérieure are pioneering efforts to tap into this potential. They have initiated a unique challenge aimed at mimicking the learning prowess of infants, which can result in more efficient AI with reduced costs and energy consumption.

Dubbed the EgoBabyVLM Challenge, this initiative evaluates how effectively vision-language models (VLMs) — systems that learn from both text and images — can interpret the world through the eyes of a child. The challenge involves a model analyzing around a thousand hours of video footage recorded from the perspectives of babies and toddlers. This innovative approach seeks to inspire AI developers to create algorithms that replicate the remarkable learning techniques employed by the youngest among us.

AI Hasn't Outpaced an Infant's Intelligence—At Least Not Yet

A cozy nursery corner bathed in warm morning light: a wooden mobile of tiny circuit boards and robot figurines hangs above a glowing laptop showing simple neural-net diagrams; a plush teddy bear sits beside colorful ABC blocks, hinting that AI still plays in the baby’s sandbox, soft pastel tones, natural light.

AI Isn’t Smarter Than a Baby—Yet - WIRED

- Babies learn efficiently through minimal observation and interaction, unlike current AI models.  
- Researchers aim to create baby-like AI to reduce costs and energy consumption.  
- The EgoBabyVLM Challenge tests AI models on a baby's perspective using infant video data.  

If you think an artificial intelligence model running on thousands of cutting-edge computer chips is smart, allow me to introduce you to the concept of a 1-year-old.

OK, so babies might not be able to write computer programs, solve advanced math problems, or debate philosophical ideas. But unlike today’s AI models, which consume an ocean’s worth of training data and as much energy as a small country, babies learn to make sense of the world with amazing efficiency. They identify new objects after seeing them once or twice, and they learn through fleeting observation and physical interaction.

When it comes to improving AI, babies—and the architecture of their brains—might hold crucial insights. Building a more baby-like version of AI could make frontier models less costly and less energy intensive, and it might also be valuable if AI-powered robots are to learn about their environments in a more natural way.

To explore this bold new frontier, researchers at Meta, Stanford University, the University of Tokyo, and France’s École Normale Supérieure developed a new test that highlights the learning skills of babies and pushes AI researchers to design algorithms that match them.

The EgoBabyVLM Challenge judges how well vision language models, or VLMs, which learn from both text and imagery, can make sense of the world as a baby sees it. It requires a model to describe the world after ingesting about a thousand hours of video collected from cameras strapped to the heads of infants and toddlers. (Yes, really.)

AI Isn’t Smarter Than a Baby—Yet

Apple has generally been conservative with its acquisition spending, but this approach may evolve as the tech giant aims to enhance its AI processing capabilities. Reports indicate that Apple has been engaging with semiconductor manufacturers and financial institutions regarding potential purchases to strengthen its server operations.

As highlighted by The Information, which first broke this story, Apple has encountered challenges with the performance of servers utilizing its M2 Ultra chips. While these chips do support some AI functions, the more complex tasks, like those executed by the Gemini model backing Siri, appear to rely on NVIDIA chips running within Google Cloud infrastructure. Apple has attempted to use its servers for these demanding applications, but its existing setup seems to fall short.

Bloomberg recently noted that a server chip based on the anticipated M7 Ultra will not be available until 2029, but Apple is set to enhance its current infrastructure with M5 Ultra chips shortly. Although there were plans for the introduction of a new server chip, codenamed "Baltra," this year, those timelines may have been pushed back. In a significant move, Apple secured a partnership with Broadcom, committing to a purchase of $30 billion worth of chips produced in the United States.

Apple's proficiency in chip design has largely been focused on consumer products, making it reasonable for the company to seek additional assistance on the server side. Its foray into chip manufacturing began with the acquisition of PA Semi for $278 million in 2008, but historically, Apple has not been one to make extravagant acquisitions. Earlier this year, it purchased the AI startup Q.ai for nearly $2 billion, constituting its second-largest acquisition after the $3 billion spent on Beats more than a decade ago.

Considering the crucial role chips play in the AI sector, any future acquisitions in this area may come at a premium. However, Apple possesses ample financial leverage if it opts to pursue this strategy, with $45.6 billion in cash and cash equivalents available as of the end of March.

Apple is allegedly seeking to acquire AI chip firms.

A sunlit high-tech expo hall with skylights casting warm natural light over rows of futuristic AI chip prototypes displayed on sleek pedestals. In the foreground, a sharply dressed buyer studies a metallic chip through a magnifying glass, a large Apple logo prominently hanging on an overhead banner. Startup booths and discreet tech signage line the background, hinting at multiple chip companies awaiting selection.

Apple Is Reportedly Shopping For AI Chip Companies - Engadget

- Apple is exploring acquisitions to enhance its AI processing capability and server performance.  
- Performance issues with M2 Ultra chips led Apple to upgrade to M5 Ultra chips.  
- Apple has significant cash reserves, enabling potential high-value acquisitions in the semiconductor sector.  

Apple doesn't tend to spend a ton of money on buying companies, but that might change as it looks to shore up its AI processing power. The company is said to have been in talks with semiconductor makers and bankers about possible acquisitions in order to bolster its servers.

According to The Information, which first reported the news, Apple has faced performance issues with servers that run on its M2 Ultra chips. Those are used for some AI tasks, though the heavy lifting (such as the Gemini model that's behind Siri AI) is seemingly handled by NVIDIA chips on Google Cloud. Apple is said to have tried using its own servers for that purpose, but its infrastructure is evidently insufficient.

Bloomberg reported this week that a server chip based on the M7 Ultra won't be ready until 2029, but noted that Apple will soon upgrade its infrastructure with M5 Ultra chips. Apple reportedly planned to debut a next-gen server chip (codenamed "Baltra") this year, but that timeline appears to have slipped. Last week, Apple struck a deal with Broadcom to buy $30 billion worth of chips that the latter makes in the US.

Apple's chip design expertise is primarily in the realm of consumer devices, so it makes sense that the company would look to bring in more support on the server side. It got into making its own chips in the first place after buying PA Semi for $278 million in 2008, but Apple doesn't typically splash much cash on acquisitions. It bought AI startup Q.ai for almost $2 billion this year. That was its second largest acquisition after the $3 billion it paid for Beats over a decade ago.

Given the importance of chips to AI companies, Apple might have to pay a premium for any acquisitions in that domain in the near future. It has plenty of flexibility if it decides to go down that route, though. As of the end of March, it had $45.6 billion ​in cash and cash equivalents.

Meta reduces its plans to monitor employees' clicks and keystrokes for AI training.

More from AI

Your options for privacy

Paul Triolo: China's opening of capital markets reflects the advancement of its AI industrial policy.

Alibaba and Baidu stocks soar in Hong Kong following Apple AI collaboration

Murati is aware of OpenAI's secrets; her AI indicates a preference for China.