• Meta plans a multibillion-dollar global underwater cable project spanning 31,000 miles.
• The project will ramp up data transmission and connect the US to India, Brazil, and South Africa.
• Meta says it aims to improve global connectivity and support innovation in artificial intelligence.

Meta has unveiled plans to spend billions of dollars as part of its multi-year ambition to build the world's longest subsea cable and accelerate AI innovation.

In a blog post on Friday, the company said its new Waterworth Project will cover over 50,000 kilometers, or about 31,000 miles, making the project's cable longer than the Earth's 24,901-mile circumference.

The Waterworth Project aims to connect five continents, linking the US to India, Brazil, South Africa, and other key regions.

A Meta spokesperson told Business Insider the company anticipates the project will be done toward the end of this decade. They said they don't have specifics to share on the cost, but the blog post said it would be a "multibillion-dollar, multi-year investment" to improve global connectivity. Last November, TechCrunch reported the company may spend over $10 billion on a nearly 25,000-mile underwater cable project led by Meta's South Africa office that the company would 100% own.

Subsea cables form an integral part of the world's internet infrastructure, shuttling data around the world at close to the speed of light thanks to their fiber optic technology. In its blog, Meta noted that cables spanning the world's oceans account for the transfer of "more than 95% of intercontinental traffic."

Meta sees the subsea cables as vital to unlocking future AI innovation as CEO Mark Zuckerberg increasingly shifts the company's focus to generative AI.

Last month, the company announced plans to boost its spending up to $65 billion this year as it seeks to build vast data centers capable of training and hosting the increasingly powerful large language models at the heart of the generative AI boom.

According to Meta's blog post, the Waterworth Project aims to ramp up data transmission capacity by using a fiber optic cable containing 24 fiber pairs instead of the typical systems that use 8 to 16 fiber pairs.

It said the project's features include a first-of-its-kind routing to optimize the cable installed in deep water at depths up to 7 kilometers, or about 4.3 miles. It also said it would use "enhanced burial techniques" in shallow, high-risk areas to protect against damage from ship anchors and potential hazards, which would maintain cable resilience.

"As AI continues to transform industries and societies around the world, it's clear that capacity, resilience, and global reach are more important than ever to support leading infrastructure," it said.

The project's announcement comes after tankers dragging their anchors have severed undersea cables in recent months in the Baltic Sea and East China Sea.

Officials in Europe have accused Russia of sabotaging undersea cables, while Taiwan has said it suspects China is behind the damage off its northern shores.

Cable resilience is key to the global financial system, which depends on a vast network of undersea cables that crisscross the sea floor, carry $10 trillion worth of transactions every day, and power Wall Street's global trading and communications.

"We've driven infrastructure innovation with various partners over the past decade, developing more than 20 subsea cables," Meta's blog post said.

"With Project Waterworth, we can help ensure that the benefits of AI and other emerging technologies are available to everyone, regardless of where they live or work."

• Apple finally has an AI partner in China: Alibaba.
• Alibaba will integrate its AI into iPhones in China, its chairman, Joe Tsai, said on Thursday.
• Apple faces a tough task to turn around sliding smartphone sales in China that may go deeper than AI.

It looks like Apple has finally found a partner to bring its full vision of generative AI to China: Jack Ma's Alibaba.

Apple will be hoping the partnership, confirmed Thursday by Alibaba's chairman Joe Tsai, drives more iPhone sales in China. But its success is far from guaranteed at a time when the company is facing tough competition from domestic rivals like Huawei.

"Apple has been very selective. They talked to a number of companies in China, and in the end, they choose to do business with us," Tsai said in an interview at the World Government Summit in Dubai. "They want to use our AI to power their phones."

Apple has been searching for a local partner to bring the full version of its Apple Intelligence suite of features to Chinese consumers. Beijing's rules prevent it from bringing its partnership with ChatGPT maker OpenAI to the country.

As a result, Apple Intelligence has not yet been introduced to Apple's most important international market. While the Alibaba partnership brings that one step closer, it still has some regulatory hurdles to pass.

Some analysts predict that Apple Intelligence will boost the company's fortunes in China at a time when its smartphone sales are continuously sliding in the face of fierce competition from local smartphone makers such as Huawei, Xiaomi, and Vivo.

Last month, Apple reported an 11% year-on-year revenue drop to $18.5 billion in Greater China, adding further woes to the health of its sales in the region after a 7.7% drop there to $72.5 billion in its previous fiscal year.

In that time, market share has shifted toward domestic firms as their devices have begun to incorporate 5G capabilities that can rival those of top-end iPhones. These devices have also provided a made-in-China alternative that appeals to growing national sentiment in the country.

In the last quarter of 2024, for instance, Huawei saw its smartphone sales jump 15.5% year-on-year, per figures from research firm Counterpoint, following the release of its high-end Mate 70 series of smartphones in November. Apple's smartphone sales fell 18.2%.

Apple's China problems loom large

The other uncertainty for Apple is whether any of its generative AI features will move the needle among consumers spending more carefully.

Counterpoint's associate director Ethan Qi noted last month that China's smartphone market saw a slowdown in the final quarter of 2024 — overall smartphone sales fell 3.2% year-on-year in the quarter — as "consumers adopted cautious spending behavior."

Though Apple is betting on generative AI to help trigger a fresh upgrade cycle, analysts have been left underwhelmed by what Apple Intelligence offers, raising concerns about the value consumers might see in the technology.

Following rumors of an Apple-Alibaba deal earlier this week, analysts at investment bank Jefferies wrote that they thought it "would be unlikely to boost sales" of the iPhone 17 in China.

While the analysts recognized that Alibaba's Qwen AI model has left a strong impression on the industry, having demonstrated "strong performance in global AI leaderboards," there are concerns that Apple will find it hard to use the model to leverage data from its own apps and deliver the "highly intelligent, personalized services" it seeks to offer.

Chinese iPhone users "rarely use any Apple apps" and choose to opt for local ones like WeChat and Baidu instead, the analysts added.

Markets have reacted positively to The Information's initial report that Alibaba and Apple had agreed to a partnership. On Wednesday, Alibaba shares in Hong Kong climbed over 8%.

Apple did not immediately respond to Business Insider's request for comment.

A partnership between Apple and Alibaba would not mark the first time the two companies have floated the possibility of working together.

In 2014, Apple CEO Tim Cook publicly addressed the prospect of a "marriage" between Apple Pay and Alibaba's payment platform, Alipay, by sharing his respect for the company's founder, Jack Ma. He said he likes to work with "those that "push us, and we like to push them."

"Those partners we work the best with, and I think Jack has a company that's exactly like that," Cook said. Some 11 years on, he'll hope that Alibaba can still be that company to Apple.

• "Vibe coding" is Silicon Valley's latest buzzword, coined by OpenAI cofounder Andrej Karpathy.
• It means AI tools like Replit Agent can do the heavy lifting in coding to build software quickly.
• Whilst it lowers the barrier to coding, experts told BI it has its pitfalls.

Silicon Valley isn't just coding anymore. It's "vibe coding."

Using AI to write code has been gaining traction for years, but now, a new buzzword coined by Andrej Karpathy, the computer scientist who cofounded OpenAI, is capturing the movement.

This month, he described what he sees as a new kind of coding in which "you fully give in to the vibes" and "forget the code even exists."

It's an approach that defies conventional wisdom in the tech industry: that developing software demands virtuosic skill from engineers.

"It's not really coding — I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works," Karpathy, who also led Tesla's AI operations for five years, wrote on X.

AI's ability to write code has come on leaps and bounds since ChatGPT's release in late 2022. Less than two months after the chatbot's release, Karpathy said, "The hottest new programming language is English" — an allusion to how smart prompting can generate good lines of code.

Software engineers have remained in hot demand since then, but the arrival of AI that can "vibe" code into existence has some industry leaders predicting big changes ahead.

OpenAI CEO Sam Altman said during a visit to India in early February that he expected software engineering to be "very different by the end of 2025." Mark Zuckerberg also said last month on the Joe Rogan Experience that AI would soon do the work of midlevel Meta engineers.

How Silicon Valley is starting to vibe code

To give a sense of what "vibe coding" looks like in action, Karpathy shared a few ways in which he's been using AI.

In one example, he said he has been using a digital workspace tool called Composer — made by OpenAI and A16z-backed startup Cursor AI — alongside the Sonnet model from AI lab Anthropic.

Cursor's Composer tool is an AI coding assistant that it says can help users "explore code, write new features, and modify existing code." When used with Anthropic's AI, a popular choice for programmers looking to AI for assistance, making an app from scratch becomes easier. That's because the AI just needs to be ushered to take steps with a user's guidance.

In another example, Karpathy said he could just "talk" to Composer by using SuperWhisper — an AI-powered voice-to-text tool.

As he noted, this means he can do things with code in Composer without having to "barely even touch the keyboard." If there's a mistake, his approach appears to be just as simple: "When I get error messages, I just copy and paste them in with no comment; usually, that fixes it."

Others are doing similar at a time when AI coding agents can be set simple instructions to do the heavy lifting that would have once required seasoned engineers to spend hours reading through reams of code — or beginners a seriously steep learning curve.

"For a total beginner who's just getting a feel for how coding works, it can be incredibly satisfying to build something that works in the space of an hour," Harry Law, an AI researcher at the University of Cambridge, told Business Insider.

Amjad Masad, CEO of Replit, a software company backed by A16z and Y Combinator, addressed Karpathy's original post, saying that "75% of Replit customers never write a single line of code."

Replit provides software used by coders to start projects, called an "online integrated development environment," but aims to offer a version that brings AI into the mix so apps can begin to be built with simple prompts. "Vibe coding is already here," Masad wrote on X.

Menlo Park Labs, a startup that builds generative AI consumer applications, is also all in on vibe coding. Its founder, Misbah Syed, is a big believer in the method.

Syed told BI he uses it for the startup's products like Brain Docs, which lets users convert a PDF to an explainer video with slides. Syed said if it makes mistakes, he feeds back the errors, and it usually fixes them. For him, the approach means that "if you have an idea, you're only a few prompts away from a product."

Vibe coding has its downsides

For all its potential benefits, experts see some risks with vibe coding.

"Ease of use is a double-edged sword," Law said. "Beginners can make fast progress, but it might prevent them from learning about system architecture or performance."

According to Law, overreliance on AI can also create technical debt, which means that it can become unmanageable when scaling or debugging code, a process engineers routinely have to go through. "Security vulnerabilities may also slip through without proper code review," he told BI.

A senior software engineer at Microsoft, who spoke with BI on the condition of anonymity as he is not authorized to speak to the media, feels the vibe coding concept is "a little overhyped."

"LLMs are great for one-off tasks but not good at maintaining or extending projects," he said, referring to large language models. "They get lost in the requirements and generate a lot of nonsense content."

A16z venture capitalist Andrew Chen said last week that while it was "brilliant" to be able to use "the latest AI codegen tools to do 'vibe coding,'" he found it enormously "frustrating" at the same time.

Karpathy acknowledged some of these limitations in his original post, noting that sometimes, an AI model "can't fix a bug." Still, he has found that he can "just work around it or ask for random changes" until the errors disappear.

With AI already pushing the limits of what was previously possible for programmers, it may soon be time for an industry vibe check.

• Aidan Gomez is impressed by DeepSeek but doesn't think its AI is enterprise-ready.
• In an interview with Business Insider, the Cohere CEO said companies are looking for custom models.
• Gomez, an ex-Google researcher, said DeepSeek has validated his view that AI can be cheaper.

Aidan Gomez, the CEO of Cohere, felt vindicated in his belief that powerful AI didn't need to be so expensive when DeepSeek released a model that would go on to blow a $1 trillion hole in the US stock market.

"I think it validated Cohere's strategy that we've been pursuing for a while now," the 28-year-old computer scientist said in an interview with Business Insider. "Spending billions of dollars a year isn't necessary to produce top-tier tech that's competitive."

But while Gomez, an ex Google researcher, considers DeepSeek's R1 "a really impressive release," he's not convinced it should be a serious option for businesses.

He said organizations are looking for customized AI models rather than something off the shelf — and are cautious about giving AI tools access to sensitive data.

"We don't see the enterprises that we sell to relying on R1 to power their systems," Gomez said. "We don't see it as a competitor on our side."

Why DeepSeek isn't enterprise-ready

As the leader of a $5.5 billion company building AI for enterprises, Gomez has a clear business reason to make this case. But, as one of the eight Google Brain researchers who co-authored the 2017 seminal "Attention is All You Need" paper that sparked the generative AI boom, his position carries weight.

Nor is he alone, with some US firms trying to adopt DeepSeek running into a host of troubles, BI previously reported.

For Gomez, DeepSeek isn't a quick win for businesses — regardless of how impressive its tech might be.

"What we're seeing from enterprises is that they don't just want to buy a model," he said. "You're going to have to build something with that model, you're going to have to deploy a lot of technical resources to see value, and it will take time."

To unlock "a new tier of value," he thinks enterprises must carefully consider how they customize core AI technology with their proprietary data.

In November, amid an industry-wide debate over whether AI performance gains had hit a wall, leaders cited private and synthetic data as key resources that organizations must tap into to maintain a competitive edge.

It's a point echoed by Gomez. And, as concerns rumble over DeepSeek being "back-ended by servers in China" — US lawmakers are seeking to ban the startup's software from government devices — the Cohere CEO said enterprises must put privacy first if models are to touch "more and more sensitive data."

"That's something that will unlock usage in enterprises because right now, they're hesitant to build systems that touch sensitive data," he said. "Our competitors treat it in a way that's less secure."

DeepSeek has not returned Business Insider's requests for comment about its data privacy policies.

All about AI agents

While Gomez thinks Deepseek's R1 is impressive, he believes the real value will come from transforming a base model into a tool that's proving to be another hot area for the industry this year: agentic AI.

Software programs that can perform tasks autonomously have been high on the agenda of business leaders this year. Agentic AI was a hot topic at Davos, while Nvidia CEO Jensen Huang said at the Consumer Electronics Show that 2025 will be the year it takes off.

Toronto-based Cohere, one of a handful of companies competing with AI rivals like OpenAI, Google, and Anthropic, is focusing on bringing AI agents to enterprises.

Last month, Cohere introduced its early-access program for North, its own agentic AI that's designed to meet specific workloads.

Gomez sees it as another way for Cohere to gain an edge over companies that just want to deploy a base model like DeepSeek's R1.

Gomez declined to say how much it can cost Cohere to adapt a platform like North to the needs of specific enterprises. He said after an "upfront investment," agents can "operate quite autonomously" once they're plugged in and then allow enterprises to "start reaping the value."

DeepSeek the disruptor

Despite his concerns about DeepSeek, Gomez views the startup spun out of a Chinese hedge fund as a positive disruptive force for the AI industry.

"The fact that they published their training efficiency numbers let people see that it doesn't need to be so capital-intensive to publish fantastic models," he said.

AI leaders continue to scrutinize DeepSeek's claims that it produced AI on par with the performance of Silicon Valley's best models at a fraction of the cost. In the meantime, investors are questioning whether large AI infrastructure spending is still justified.

Addressing the implications of Sam Altman's $500 billion Stargate project, Gomez said that "spending more and more" on infrastructure for training AI models, rather than "inference," is a mistake. Inference refers to an AI model making predictions or decisions about new data, while training is the process of building a model's capabilities.

"I think DeepSeek's a big proof point of that," he said.

The other implication of DeepSeek's emergence out of seemingly nowhere is the validation of an open-source approach.

While there is debate over whether DeepSeek's AI is truly open-source — it has secured an MIT license and made its model weights open, but hasn't disclosed the data it used to train it — Gomez sees "great technology" coming from both open-source and closed-source players.

Still, enterprises will need more than an impressive Chinese model to build powerful AI into their operations. As Gomez put it, "It's not just enough to download a model."

• Daniel Penny, the Marine veteran acquitted in a subway killing, has been hired by Andreessen Horowitz.
• A16z said Penny would take the role of deal partner to support the VC firm's American Dynamism team.
• He was found not guilty of criminally negligent homicide last year after choking a subway rider.

Daniel Penny, the Marine veteran who was cleared last year in a subway killing that drew widespread attention, has joined Andreessen Horowitz as a deal partner.

Penny will support the venture capital firm's American Dynamism team, its website said.

In 2023, Penny was arrested and charged in connection with the death of Jordan Neely, a 30-year-old homeless street performer whom Penny restrained using a chokehold on the New York City subway in May of that year. Neely, who had a history of mental illness, was said to be yelling at passengers.

A New York jury deadlocked on a manslaughter charge, which was dismissed, but found Penny not guilty of criminally negligent homicide in December.

In an internal statement seen by The Free Press, David Ulevitch, a general partner at Andreessen Horowitz, also known as A16z, said, "Daniel is a Marine Corps veteran who served his country and, in a frightening moment in a crowded New York City subway car, did a courageous thing."

Vice President JD Vance reacted to Penny joining A16z in an X post, saying, "Incredible news."

A16z did not immediately respond to a request for comment from Business Insider.

An early investor in Silicon Valley giants such as Facebook and Stripe, the firm set up its American Dynamism group in 2023 to invest in founders and companies that support "the national interest."

The venture capital giant, led by the investors Marc Andreessen and Ben Horowitz, has defined the national interest broadly, including aerospace, defense, and public safety as well as education, housing, and manufacturing.

Among the group's top investments so far is Anduril, a drone maker cofounded by the billionaire Palmer Luckey. It was most recently valued at $14 billion.

Advising the new White House administration has been a top priority for A16z's Andreessen. In a podcast episode in December, the firm's cofounder said he had spent about "half" his time at Mar-a-Lago after Election Day discussing policy issues with Donald Trump, then the president-elect.

• DeepSeek's powerful, cheap AI models have taken the tech world by storm.
• Altman said OpenAI will embrace one of DeepSeek's popular approaches.
• The CEO said OpenAI has "been on the wrong side of history" when it comes to model weights.

When rivals take a different approach and succeed, it sometimes pays to change course.

This is what Sam Altman said OpenAI will do, according to a Reddit AMA session on Friday.

The discussion touched on several AI topics, but in particular Altman was asked about DeepSeek, which has taken the tech world by storm after rolling out top-performing AI models that are relatively cheap to use.

One Reddit user asked if OpenAI could show "all of the thinking tokens." This refers to the chain of thought that new "reasoning" AI models use to break tasks into smaller steps — similar to how humans think through complex challenges.

OpenAI's o1 and o3 models use this reasoning approach, however they don't show any of the intermediate thinking steps to users, and instead just show the final answer.

DeepSeek's reasoning models, such as its R1 offering, show every step to users. When Business Insider demoed DeepSeek with the Chinese lab's DeepThink setting, it shared about 16 pages of mathematical steps before providing the correct answer to a tough question.

On Friday, Altman said OpenAI would follow DeepSeek's approach. "Yeah we are gonna show a much more helpful and detailed version of this, soon. Credit to R1 for updating us," he wrote.

Open-source and open weights

Meta chief AI scientist Yann LeCun has said the biggest takeaway from DeepSeek's success is the value of open-source AI models versus proprietary ones.

Meta's Llama models are mostly open-source, letting anyone access important details such as weights and parameters for free. Sharing the inner-workings of models like this allows other developers and many companies to customize these models for their own use.

Despite its name, OpenAI has taken a more closed approach to AI development so far. Most of its models are proprietary and the startup charges for access.

During the Reddit AMA on Friday, Altman was asked if OpenAI would consider releasing some of its model weights, and publishing some research.

"Yes, we are discussing. I personally think we have been on the wrong side of history here and need to figure out a different open source strategy; not everyone at OpenAI shares this view, and it's also not our current highest priority," Altman replied.

Do you work at Meta? Contact this reporter from a nonwork email and device at [email protected] or [email protected]. You can also reach him securely via Signal at +1408-905-9124. Your identity will be protected.

• Apple avoided a hit to its share price from DeepSeek, but it has another China-made problem.
• iPhone sales in the country have fallen again.
• An 11% slide in sales last quarter is raising concerns about Apple amid tough competition in China.

DeepSeek may not have triggered a sell-off panic for Apple, but the iPhone maker has another China headache.

While Apple avoided the pummelling handed to its Silicon Valley peers this week by Chinese AI startup DeepSeek, the company had its own China-related problems on Thursday as it reported an 11% year-on-year drop in revenue in the country to $18.5 billion.

The slide in last quarter's sales is a sign of deepening struggles for Apple in its most important international market — and is raising concerns among some investors and analysts who had forecast better results.

In its last financial year, Apple's net sales in Greater China decreased 7.7% from $72.5 billion the previous year. The year before that, net sales decreased by 2.2%. Understanding if the decline can be stopped is now key for Apple watchers.

During Thursday's earnings call, CEO Tim Cook took the 11% China decline head-on by explaining that "over half of the decline" was driven by changes in "channel inventory."

He also had another explanation. "It's the most competitive market in the world," Cook said.

That competition has been a particular threat to Apple's smartphone sales in China, where domestic companies such as Huawei and Xiaomi have been enticing consumers with 5G-enabled devices that now appear to rival the capabilities of iPhones.

Figures from research firm Counterpoint show that in the last three months of 2024, Apple's smartphone sales fell 18.2%, while Huawei's grew by 15.5% as it was boosted by the launch of its latest high-end smartphone series, the Mate 70.

In a research note after Thursday's earnings, Logan Purk, senior analyst at Edward Jones, said that "sales in China were weak and will remain a point of debate among investors," particularly as Apple contends with "fierce competition in China for 5G phones."

That said, analysts also see an opportunity for Apple to counter the competition later this year once it rolls out one of its biggest bets to China: Apple Intelligence.

The tech giant's generative AI features, unveiled in June last year, are yet to be rolled out in China and are seen as a key driver of a huge upgrade cycle in iPhones in the future.

Dan Ives, a Wedbush analyst, said in a note on Friday that his firm expects growth to "markedly rebound" in China to "double-digits" year-on-year starting in the June quarter, with the "anticipated April rollout" of the Apple Intelligence in the country.

His confidence was boosted by the fact that "markets where Apple Intelligence was available" saw clear year-on-year performance "outpace markets without this AI rollout."

Still, Apple's AI push in China remains untested, with domestic competitors in the country vying to show local consumers that their technology can match America's best.

Apple will need to show Chinese consumers that iPhones are fit for the AI era.

• The Chinese startup DeepSeek shook the tech world and markets when it released R1, its new AI model.
• The West is now trying to reproduce R1 on its own terms and cut out Chinese servers.
• Recreating R1 from scratch can help researchers build better models and validate DeepSeek's claims.

Silicon Valley doesn't want to get caught out again. It's scrambling to replicate DeepSeek's AI model, the cheaper Chinese tech that shook Wall Street and is freely available for anyone to adopt.

Companies like Microsoft and Amazon have already made versions of DeepSeek's R1 models available on their cloud platforms. This allows people to use the models, which appear to match the capabilities of models from rivals like OpenAI, while keeping data from being sent to servers in China.

But there are also attempts to replicate DeepSeek's cost-efficient AI from the ground up — and see whether all the Chinese AI lab's market-moving claims hold up.

One major effort is being led by Hugging Face, a platform on which researchers in artificial intelligence's open-source community can collaborate and share their trade-research notes and ideas for free.

Leandro von Werra, the head of research at Hugging Face, told Business Insider that the company expected to complete its replication efforts within weeks. He described the mood at Hugging Face as "kind of like, 'Avengers assemble,'" as they dissect the inner workings of R1.

DeepSeek obtained open-source licensing for its model from MIT, which means a lot of the vital components of the recipe needed to build R1 have been laid out in the company's publicly available technical paper.

However, there are some elements of R1 that remain unclear.

In a December paper on V3, DeepSeek's earlier model, the Chinese company said the training cost $5.6 million in total. The cost was calculated based on its use of H800 graphics processing units, a less powerful version of Nvidia's top chips, at a rental price of $2 per GPU hour.

Right now, no one can be quite sure what the actual development cost of R1 was, von Werra told BI.

DeepSeek's research paper also did not share what was required to bake reasoning capabilities into V3 to then produce R1.

That said, von Werra thinks it won't remain a mystery for long. "I don't know about the compute number. We can only guess at this time," he said. "I think one thing that's exciting about our reproduction is we're going to find out pretty quickly if the numbers hold up."

Learning from DeepSeek

Some corners of Silicon Valley responded swiftly to the launch of DeepSeek. This week, Meta set up "war rooms" for its researchers to analyze DeepSeek, The Information reported. Sam Altman, the CEO of OpenAI, said Tuesday his company would accelerate the release of "better models."

DeepSeek's decision to publish its findings and make its R1 model open gives researchers worldwide insight into its novel approach.

The main technique used to make R1 so capable was "pure reinforcement learning," DeepSeek's paper said. This, Hugging Face researchers said in a blog on Tuesday, can "teach a base language model how to reason without any human supervision."

The researchers also know more specific technical details about why R1 caused such a stir in Silicon Valley and wiped $1 trillion from US stocks on Monday. For instance, the reasoning model is what's known as a "mixture of experts" model — industry-speak for a model that can be "pre-trained with far less compute." It also involves subtle changes to its architecture by introducing techniques like "multitoken prediction," first introduced by Meta, that make models more efficient.

Hugging Face's von Werra said details like this from DeepSeek had helped the industry better understand how a reasoning model like OpenAI's closed-source o1 was built. "Everybody thought this is the secret that is going to take awhile to crack," he said.

DeepSeek comes to platforms

Days after its launch, DeepSeek flew to the top of Apple's top free apps chart. While everyone wanted to try the latest AI tool, DeepSeek's policy of storing user data in China prompted security concerns.

This spurred US companies to make R1 available on their own platforms so customers could use the Chinese AI model while cutting out China's servers.

Lin Qiao, the CEO of Fireworks AI and former head of the PyTorch team at Meta, told BI that one clear reason for doing so was to ensure AI developers and users continued to get access to top model innovations.

"The approach we have been taking is always to enable state-of-the-art models for developers the fastest," she said. "DeepSeek is one example."

Her company, founded in 2022, made R1 available on its platform after congratulating DeepSeek for "pushing the boundaries of what's possible in open models." It has been made available via its serverless service, as well as through on demand and for enterprise customers.

Others have followed suit. On Wednesday, Microsoft announced that it was making R1 available in its model catalog on its AI development platform Azure AI Foundry to make it "accessible on a trusted, scalable, and enterprise-ready platform."

Asha Sharma, a corporate vice president at the tech giant, wrote in a blog that R1 "offers a powerful, cost-efficient model" but one that it had done "rigorous red teaming and safety evaluations" on before introducing as a model to its library.

Amazon Web Services is making a similar move. Swami Sivasubramanian, the vice president of AI and data at AWS, said this week that the company's "commitment to AI accessibility" meant R1 was being made available on its platforms such as SageMaker and Bedrock.

Perplexity CEO Aravind Srinivas, whose AI startup made R1 available to users of its search platform this week, said Tuesday in an X post that downloading the model onto its servers could also help control the way it responds to user queries and ensure those queries don't go to servers in China.

DeepSeek's AI appears to censor sensitive information about China, such as refusing to answer questions about the 1989 Tiananmen Square protests. Srinivas said that Perplexity's version of R1 had no censorship and shared its accurate response to what happened in Tiananmen Square.

David Sacks, the White House's AI czar, offered one reason Perplexity's R1 integration was an important way to reproduce R1 in the West. "This is one of several ways that you can try DeepSeek R1 without downloading the app or sharing any data with a Chinese company," he said on X.

Many see DeepSeek as an example of China challenging American AI hegemony using a tried-and-tested playbook. OpenAI, which has a lot to lose from DeepSeek making its technology freely available, said on Wednesday it's investigating whether the Chinese firm "inappropriately" replicated its models for training.

For von Werra, it's a full-circle moment. The whole field started as open source, so seeing efforts to make a leading reasoning model available for free is welcome, he said.

"I think in the end, everybody's going to get better models and do cooler things," he said. "I feel like it's a win-win situation."

• SoftBank is in talks to invest up to $25 billion in OpenAI, which could value it around $300 billion.
• The deal could make OpenAI the joint-second most valuable tech company after SpaceX.
• The funding round could see OpenAI pump $15 billion into the Stargate AI infrastructure project.

SoftBank is preparing to lead fresh investment into OpenAI at a $300 billion valuation, Business Insider understands — a move that would value the ChatGPT maker the same as TikTok owner ByteDance.

The Japanese investment giant is in talks to invest up to $25 billion in OpenAI, the Financial Times first reported.

While SpaceX is the most valuable private company in the world, the funding round could push OpenAI up to second spot alongside Chinese tech giant ByteDance. OpenAI currently ranks as the third-most valuable private tech firm.

Talks between OpenAI and SoftBank are ongoing, which means details around the the round size and valuation are subject to change. If the AI juggernaut hits its target valuation of $300 billion, it would nearly double its current valuation of $157 billion.

SpaceX's valuation is around $350 billion after the Elon Musk-owned company agreed to buy back $1.25 billion worth of stock at $185 a share in December. ByteDance, the parent company of TikTok, similarly catapulted its valuation to $300 billion following a buyback offer, offering investors a price of about $180 per share, the Wall Street Journal reported in November.

OpenAI could become one of three companies globally with a centibillion valuation; Stripe and Shein follow suit with valuations of $70 billion and $66 billion, respectively.

Prior to SoftBank's latest investment talks, OpenAI raised nearly $20 billion from investment heavyweights, including Thrive Capital, Khosla Ventures, and Nvidia.

If SoftBank invests $15 to $20 billion into OpenAI, it would overtake Microsoft — which has poured around $13 billion into the company so far — as OpenAI's lead investor.

Such a deal would mark the most significant bet yet on the generative AI boom from SoftBank and form a key part of the wider ambitions of its billionaire founder, Masayoshi Son, to usher in an era of "artificial super intelligence."

Son, who made his original fortune from a timely bet on Alibaba in the dot-com era, has previously spoken about his nonstop use of OpenAI's ChatGPT, and predicted AI that is 10,000 times smarter than humans will arrive by the midpoint of the next decade.

A fresh investment from SoftBank would also further deepen its relationship with OpenAI after the companies announced plans last week to form Stargate, a joint venture that aims to spend up to $500 billion on AI infrastructure projects in the US over the next four years.

SoftBank's Son will serve as chairman of the project, which was unveiled at the White House by President Donald Trump. Initial equity funders include Oracle and the UAE's MGX alongside SoftBank and OpenAI. The project will begin deploying $100 billion immediately, according to OpenAI.

It is not yet clear where the companies will source the capital from, with OpenAI currently lossmaking. It has been suggested that SoftBank's proposed equity investment could allow OpenAI to invest around $15 billion in Stargate.

SoftBank declined to comment. OpenAI did not immediately respond to a Business Insider request for comment.

• President Donald Trump is threatening to impose tariffs of up to 100% on chips made in Taiwan.
• That could be bad news for Nvidia, which relies on Taiwan's TSMC for its chip supply.
• Nvidia's value declined by 17% on Monday in a market rout triggered by China's DeepSeek.

President Donald Trump's pledge to impose tariffs on semiconductors made in Taiwan could deal a fresh blow to Nvidia, whose shares dropped by 17% on Monday in a DeepSeek-induced sell-off.

Tech firms like Nvidia have long relied on Taiwan's TSMC, the world's largest contract semiconductor manufacturer, to make the specialist AI chips they design in the US.

That relationship was threatened on Monday as the president delivered a speech to Republicans in which he said tariffs on Taiwan would be aimed at returning the production of chips to the US.

"They left us and went to Taiwan," Trump said, seeming to refer to US firms that source processors from TSMC.

Trump said companies wouldn't want to pay a "25%, 50%, or even a 100% tax."

US tariffs on Taiwan's semiconductors could result in a steep increase in costs to Nvidia and other significant customers, such as Apple and AMD. Chip manufacturing efforts in the US are less developed and more expensive than those in Taiwan.

The threat of tariffs risks a double blow for Nvidia, one of TSMC's largest customers. On Monday, after the Chinese startup DeepSeek released an AI model, Nvidia lost about $589 billion in market value.

Investors reacted with panic, fearing that Nvidia's chips may face a decline in demand. DeepSeek's new model claims to have achieved performance levels similar to a frontier model created by OpenAI but with fewer and less-advanced chips.

An Nvidia spokesperson declined to comment.

It's unclear whether Trump will proceed with tariffs on Taiwan, but US leaders have increasingly exercised caution about dependence on Taiwan for chips.

China has long overshadowed the independently governed island with the prospect of invasion. A war could hugely disrupt the US economy, endangering the supply of the chips vital to swaths of the US tech sector.

The COVID-19 pandemic also exposed the vulnerability of global supply chains.

President Joe Biden, as part of his CHIPS Act, sought to encourage more semiconductor firms to set up business in the US by offering incentives such as tax breaks. TSMC has moved some of its operations to the US, opening chip manufacturing plants in Arizona as part of a $65 billion initiative.

While Trump has veered toward imposing tariffs to bolster US chip production, the US's chip manufacturing sector could take years to develop the same capacity as Taiwan's. That could mean higher prices for hardware that relies on chips from Taiwan, such as Apple's iPhones, Nvidia's GPUs, and AMD's processors.

"If the argument is that this is the way to force it to move here, TSMC is already moving here," William Reinsch, a senior advisor with the Center for Strategic and International Studies, told Yahoo Finance last year after Trump floated the idea of imposing tariffs.

"They're already building a fab plant in Arizona," he added. "That's all already underway and the tariffs aren't going to make that move any faster. If anything, they might complicate the effort."

Taiwan responded to Trump's tariff proposal by pointing to the relationship between the Taiwanese and US economies.

In a statement reported by Reuters on Tuesday, Taiwan's economy ministry said: "Taiwan and the U.S. semiconductor and other technology industries are highly complementary to each other, especially the U.S.-designed, Taiwan-foundry model, which creates a win-win business model for Taiwan and U.S. industries."

AI is among the sectors where the US and China are in an intensifying battle for global technological dominance. On January 21, Trump announced, alongside the leaders of OpenAI, SoftBank, and Oracle, a $500 billion initiative to boost AI infrastructure in the US.

TSMC did not immediately respond to a request for comment from Business Insider.

• Chinese startup DeepSeek triggered panic among investors in top AI companies like Nvidia.
• DeepSeek claims to have built AI that rivals OpenAI's o1 but with less compute.
• That could mean lower demand for AI chips, but some analysts and AI leaders don't agree.

Did the market overreact?

That's the question tech investors are asking after they got walloped on Monday when Chinese AI lab DeepSeek shocked markets with a model that runs on fewer, less advanced chips — wiping out $1 trillion in stocks.

Trillions were expected to be spent on AI infrastructure, but DeepSeek's breakthrough challenged that assumption. Still, some believe the panic was premature, and their reasoning is to do with computational power, a key factor in scaling AI.

Is the DeepSeek sell-off overblown?

Hamish Low, an analyst at research firm Enders Analysis, told Business Insider that the reaction to the chip stock sell-off seems "quite overblown" as "being able to use compute much more efficiently," a key claim of DeepSeek's R1 release, "is by no means bad for compute demand."

Several tech leaders, such as Microsoft CEO Satya Nadella, have taken to social media to make a similar point by citing the Jevons Paradox, the idea that as the cost of using a resource falls, demand will go up — not down.

As Nadella put it on X: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of."

Or, as former Intel CEO Pat Gelsinger put it in an X post on Monday, "Computing obeys the gas law."

He added, "Making it dramatically cheaper will expand the market for it. The markets are getting it wrong, this will make AI much more broadly deployed."

That suggests AI leaders want more efficiency alongside more computing power.

Ethan Mollick, a Wharton professor who studies AI, echoed this point. "Everyone in the space is compute constrained," he wrote in an X post on Monday. "More efficient models mean those with compute will still be able to use it to serve more customers and products at lower prices & power impact."

Similarly, Bernstein analysts wrote in a Monday investor note that their "initial reaction does not include panic." The analysts, also citing the Jevons paradox, said that "any new compute capacity unlocked is far more likely to get absorbed due to usage and demand increase vs impacting long-term spending outlook at this point."

Meanwhile, Dan Ives, a Wedbush analyst, used a note to remind investors of the bull case for Nvidia. He wrote that while launching a competitive model for consumers was one thing, Nvidia's "broader AI infrastructure" involving robotics, for instance, "is a whole other ballgame."

AI model developers have also been very clear about their intent to buy more AI hardware in the near future. Last week, both OpenAI and Meta announced massive plans to drastically increase their investment in AI chips and relevant infrastructure.

The ChatGPT maker announced a $500 billion initiative called Stargate to that end, while Meta CEO Mark Zuckerberg said his company was increasing its capital expenditure on AI this year to $65 billion.

Taken together, these initiatives signal a serious willingness from top AI players in Silicon Valley to continue spending on the products sold by the companies on the negative end of the market rout triggered by DeepSeek.

Bearish signals

However, for other industry watchers, there remains a sobering rationale behind Monday's market sell-off.

Javier Correonero, an equity analyst at Morningstar, told BI that investors will be conscious that if DeepSeek's claims hold true, then there is reason to question if Big Tech firms like OpenAI, Meta, and others need to spend billions of dollars on securing extra chips.

"In my view, in the short-medium term, this could be bearish because maybe now the Big Tech firms that are doing all the capex will start focusing more on optimizing all their existing AI infrastructure rather than keep on acquiring more," he said.

Enders Analysis' Low made a similar point, telling BI that "DeepSeek is maybe just acting as a trigger point here for much broader investor unease around the returns on Big Tech AI capex and Nvidia's continued rise."

In the meantime, investors will continue to reel from the fallout of Monday's market rout while Silicon Valley leaders unpack how DeepSeek achieved so much with seemingly so little.

• DeepSeek's new AI model has triggered a sell-off in top AI stocks.
• Former Intel CEO Pat Gelsinger said Monday that the markets had got it wrong.
• Gelsinger said DeepSeek's claims of creating AI cheaply would drive demand for chips, not reduce it.

Former Intel CEO Pat Gelsinger said that "the markets are getting it wrong" after investors drove a panic sell-off in top AI stocks in response to concerns that DeepSeek would weaken demand for advanced chips.

Top AI companies, including Nvidia, saw hundreds of billions of dollars wiped off their valuations on Monday after a new AI model from the Chinese startup DeepSeek claimed to have emulated the success of a leading model released by OpenAI just months ago — and at a fraction of the cost.

DeepSeek's new model, R1, released on President Donald Trump's Inauguration Day, appeared to have accomplished the feat with fewer and less powerful chips than those used by top AI labs in America. The development has raised concerns that chip stocks like Nvidia may see demand recede.

Nvidia, which has added trillions of dollars to its market capitalization since the start of the ChatGPT boom, suffered a wipeout of as much as $500 billion on Monday, triggering the biggest stock market rout in US history.

Gelsinger posted Monday on X to suggest that the market's assumptions were wrong. He said that instead of reducing demand, making computing "dramatically cheaper" and more efficient to use — as DeepSeek appears to have done — "will expand the market for it."

The former Intel boss, who retired from the company in December after struggling to capitalize on the AI boom, also suggested that the Chinese engineers at DeepSeek "had limited resources, and they had to find creative solutions" to squeeze performance out of their models.

The AI industry has insisted that models become smarter when fueled by more computing power in the form of chips loaded in data centers, where they're trained and hosted.

Because of tough export controls, Chinese companies have struggled to access the best chips from America. That said, DeepSeek has not explicitly said how much computing power is behind the R1 model it released last week.

Gelsinger's views echo those of others in the AI industry who feel the market has overreacted to DeepSeek's claims of greater computing-power efficiency.

Ethan Mollick, a Wharton professor, wrote on X that he's "not sure why people assume this will make compute less valuable," adding: "More efficient models mean that those with compute will still be able to use it to serve more customers and products at lower prices & power impact."

• OpenAI's ChatGPT once gave it a moat that protected it from competition from rivals.
• That moat is looking less secure as the likes of Chinese startup DeepSeek close the gap.
• OpenAI now seems ready to build a new moat with a $500 billion infrastructure project called Stargate.

OpenAI has a new $500 billion project called Stargate. Consider it an extraordinary attempt by the ChatGPT maker to build itself a whole new moat in the face of growing competition from rivals at home and a buzzy startup in China.

This week, OpenAI revealed plans to shift focus and spend $500 billion on AI infrastructure over the next four years. Stargate will focus on hardware powering advanced AI software, such as data centers and energy supply — all with Donald Trump's blessing.

It is a colossal amount of money, but OpenAI has two critical reasons for spearheading the initiative alongside key partners SoftBank, Microsoft, and Oracle.

First, reaching what OpenAI boss Sam Altman calls an era of "superintelligence" depends on having access to more computing power. The industry insists that AI models will get smarter when given more of it, making ownership and access a priority.

Stargate, then, offers OpenAI a chance to build itself the kind of formidable "moat" that Warren Buffett once said companies need to ensure they maintain a competitive advantage against rivals. A vast infrastructure empire belonging to OpenAI would certainly help give it that.

The second reason for betting big on infrastructure through Stargate is even more critical: OpenAI may have no choice but to build itself a new moat in response to a Chinese startup that this week caused a stir in Silicon Valley.

Why OpenAI may need a new moat

When OpenAI first introduced ChatGPT to the world, it was clear that the combination of a powerful AI model and an intuitive, consumer-friendly interface gave it a moat that was the envy of Silicon Valley.

How much of a moat ChatGPT offers OpenAI has come under increasing debate.

The latest challenge to OpenAI's moat came on Monday as a little-known Chinese startup called DeepSeek stunned America's top AI researchers after revealing a new AI model called R1. In a paper, it said R1 rivals the frontier AI model OpenAI introduced just four months ago, o1.

Like o1, it reasons, meaning it thinks carefully before responding. According to DeepSeek, which has obtained an MIT license to open-source its model, R1 achieves "performance comparable to OpenAI o1 across math, code, and reasoning tasks."

That an open-source Chinese company has been able to release a free model competitive with the best OpenAI has to offer has not gone unnoticed among AI's most prolific leaders.

Jonathan Ross, the CEO and founder of Groq, an Nvidia rival backed by BlackRock and Samsung, responded to a question in Davos this week about DeepSeek's latest release by claiming "open models will win."

"We cannot do closed models anymore and be competitive — open always wins," the former Google engineer said while noting that it will be harder to differentiate between models soon. As he put it, "The models are not going to be particularly special for long."

Gary Marcus, a cognitive scientist and AI researcher, told Business Insider that he has been adamant for a year now that "closed source players have almost no obvious moat" and that it is "certainly even more true now" following DeepSeek's release.

It's not the first time OpenAI's moat has been challenged.

In May 2023, when Google was busy playing catch up to ChatGPT, an internal memo written by an engineer inside the search giant, published by research firm SemiAnalysis, warned that "we have no moat, and neither does OpenAI."

Why? Because, in the engineer's view, the gap in capabilities between the models being worked on in Silicon Valley versus the open-source community was "closing astonishingly quickly."

We are, of course, almost two years on from the warning issued in that memo, and OpenAI has since then convinced investors that it is of market-leading value, having completed a $6.6 billion funding round at a $157 billion valuation in October.

Nathan Benaich, founder and general partner of AI-focused venture capital firm Air Street Capital, offered Business Insider one explanation for why OpenAI has retained favor, noting ChatGPT has things like ease of use and reliability that make it a top pick for customers.

"OpenAI, I think, has won the spot of being CocaCola for consumers," he said. "Who is the Pepsi? The gap is large."

Avijit Ghosh, an applied policy researcher at Hugging Face, doubled down on the point, telling Business Insider that while DeepSeek's R1 performance "challenges conventional wisdom about technical moats in frontier AI," he thinks the reality is more nuanced.

"The real differentiation isn't in raw model performance anymore — it's in how these capabilities are integrated into practical applications and systems, and tech behemoths are still leagues ahead in being able to vertically integrate and commodify their models into products," he said.

OpenAI offered a demonstration of this on Thursday following the release of Operator — an AI agent that serves as an assistant capable of booking everything from dinner reservations to travel.

This is clearly impressive, but resting competitive advantage on AI agents and an easy-to-use interface may still not be enough in the face of an entire industry making "agentic AI" its focus. The AI industry's pace of innovation is relentless.

Building the Stargate moat

Stargate offers OpenAI a chance to build a new moat that could well be tougher to replicate than a large language model or AI agent for a few reasons. For one, having more computing power is vital to scaling models to greater levels of intelligence, experts say.

Dylan Patel, founder of research firm SemiAnalysis, told Business Insider: "Stargate is extremely important for OpenAI to be competitive because the access to compute is what enables them to keep scaling."

Groq's Ross, meanwhile, told a panel in Davos that he sees access to infrastructure being so important in the future that "what countries are going to be tussling over is how much compute they have access to."

If OpenAI does manage to get Stargate built, it seems it will do so to serve its own purposes. A report from the Financial Times this week said Stargate will exclusively serve OpenAI — a sign that the project is meant to give the company an edge.

OpenAI did not respond to multiple requests for comment from Business Insider.

Clearly, then, the Stargate announcement suggests OpenAI sees a chance to build a new moat. It just can't be expected to come easy.

Futurum Group CEO Daniel Newman noted that while the project is likely to work as a VC play, where the outcome is expected to be delivered at an unknown date in the future after serious investment, "raising $500 billion is going to be a real challenge."

OpenAI remains lossmaking, and its key Stargate partners, like SoftBank and Oracle, the main equity funders alongside the UAE's MGX, have cash on hand that falls far short of the $100 billion initially being deployed. From a sheer capital position, a new moat won't be cheap.

Building the data centers and clean energy sources forming this new moat will require nationwide coordination spanning multiple initial technology partners, including Microsoft, Nvidia, Oracle, OpenAI, and Arm. Construction of a data center site in Texas is underway, but the project will need to go much further.

Hugging Face's Ghosh also thinks the "'compute is everything' narrative misses the mark," as small teams like Deepseek have shown "remarkable results through clever engineering."

Still, OpenAI seems ready to go all in on Stargate to build a new moat. Its future leadership in the industry may well depend on its success.

• An AI startup in China just showed how it's closing the gap with America's top AI labs.
• The Chinese startup DeepSeek released a new AI model last Monday that appears to rival OpenAI's o1.
• Its reasoning capabilities have stunned top American AI researchers.

Donald Trump started his new presidency by declaring America must lead the world. He just got a warning shot from an AI crack team in China that's ready to show that US technological supremacy is not a given.

Meet DeepSeek, a Chinese startup spun off from a decade-old hedge fund that calculates shrewd trades with AI and algorithms. Its latest release, which came on the day Trump was inaugurated, has left many of America's top industry researchers stunned.

In a paper released last Monday, DeepSeek unveiled a new flagship AI model called R1 that shows off a new level of "reasoning." Why it has left such a huge impression on AI experts in the US matters.

Some of Silicon Valley's best-resourced AI labs have increasingly turned to "reasoning" as a frontier of research that can evolve their technology from a student-like level of intelligence to something that eclipses human intelligence entirely.

To accomplish this, OpenAI, Google, Anthropic, and others have focused on ensuring models spend more time thinking before responding to a user query. It's an expensive, intensive process that demands a lot from the computing power buzzing underneath.

As a reminder, OpenAI fully released o1 — "models designed to spend more time thinking before they respond" — to a glowing reception in December after an initial release in September. DeepSeek's R1 shows just how quickly it can close the gap.

DeepSeek narrows the gap

What exactly does R1 do? For one, DeepSeek says R1 achieves "performance comparable to OpenAI o1 across math, code, and reasoning tasks."

Its research paper says this is possible thanks to "pure reinforcement learning," a technique that Jim Fan, a senior research manager at Nvidia, said was reminiscent of the secret behind making Google DeepMind's AlphaZero a master at games such as go and chess from scratch, "without imitating human grandmaster moves first." He wrote on X that this was "the most significant takeaway from the paper."

DeepSeek, which launched in 2023, said in its paper that it did this because its goal was to explore the potential of AI to "develop reasoning capabilities without any supervised data." This is a common technique used by AI researchers. The company also said that an earlier version of R1 called R1-Zero gave them an "aha moment" in which the AI "learns to allocate more thinking time to a problem by reevaluating its initial approach."

The end result offers what the Wharton professor Ethan Mollick described as responses from R1 that read "like a human thinking out loud."

Notably, this level of transparency into the development of AI has been hard to come by in the notes published by companies such as OpenAI when releasing models of a similar aptitude.

Nathan Lambert, a research scientist at the Allen Institute for AI, said on Substack that R1's paper "is a major transition point in the uncertainty in reasoning model research" as "until now, reasoning models have been a major area of industrial research without a clear seminal paper."

Staying true to the open spirit, DeepSeek's R1 model, critically, has been fully open-sourced, having obtained an MIT license — the industry standard for software licensing.

Together, these elements of R1 provide complications to US players caught up in an AI arms race with China — Trump's main geopolitical rival — for a few reasons.

First, it shows that China can rival some of the top AI models in the industry and keep pace with cutting-edge developments coming out of Silicon Valley.

Second, open-sourcing highly advanced AI could also challenge companies that are seeking to make huge profits by selling their technology.

OpenAI, for instance, introduced a ChatGPT Pro plan in December that costs $200 a month. Its selling point was that it included "unlimited access" to its smartest model at the time, o1. If an open-source model offers similar capabilities for free, the incentive to buy a costly paid subscription could diminish.

Nvidia's Fan described the situation like this on X: "We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive — truly open, frontier research that empowers all."

DeepSeek has shown off reasoning know-how before. In November, the company released an "R1-lite-preview" that showed its "transparent thought process in real time." In December, it released a model called V3 to serve as a new, bigger foundation for future reasoning in models.

It's a big reason American researchers see a meaningful improvement in the latest model, R1.

Theo Browne, a software developer behind a popular YouTube channel for the tech community, said that "the new DeepSeek R1 model is incredible." Tanay Jaipuria, a partner investing in AI at Silicon Valley's Wing VC, also described it as "incredible."

Awni Hannun, a machine-learning researcher at Apple, said a key advantage of R1 was that it was less intensive, showing that the industry was "getting close to open-source o1, at home, on consumer hardware," referring to OpenAI's reasoning model introduced last year.

The model can be "distilled," meaning smaller but also powerful versions can run on hardware that's far less intensive than the computing power loaded into servers in data centers many tech companies depend on to run their AI models.

Hannun demonstrated this by sharing a clip on X of a 671 billion-parameter version of R1 running on two Apple M2 Ultra chips, responding with reason to a prompt asking whether a straight or a flush is better in a game of Texas Hold'em. Hannun said its response came "faster than reading speed."

AI censorship

R1 does appear to have one key problem. The former OpenAI board member Helen Toner pointed out on X that there were demos of R1 "shutting itself down when asked about topics the CCP doesn't like."

Toner did suggest, however, that "the censorship is obviously being done by a layer on top, not the model itself." DeepSeek didn't immediately respond to a request for comment.

It is worth noting, of course, that OpenAI has introduced a new model called o3 that's meant to be a successor to the o1 model DeepSeek is rivaling. Lambert said in his blog post that OpenAI was "likely technically ahead," but he added the key caveat that the o3 model was "not generally available," nor would basic information such as its "weights" be available anytime soon.

Given DeepSeek's track record so far, don't be surprised if its next model shows parity to o3. America's tech leaders may have met their match in China.

• AI leaders are preparing to take America into what Sam Altman calls the "Intelligence Age."
• Getting there will depend on building vast amounts of new AI infrastructure on US soil.
• Whether investments in this infrastructure will ever pay off is another matter.

America is ready to reach a new age of intelligence. Getting there —and staying ahead of rival nations in the AI race — depends on a plan to transform the physical world that's becoming more formidable by the day.

Leaders driving the AI boom entered 2025 by getting louder about the radical transformation they say is needed on US soil to deliver an era of AI-led superintelligence: more data centers, more chip plants, and more power infrastructure.

By taking root in the physical world — huge data center facilities depend on complex wiring, hardware, and integration with power infrastructure across vast amounts of landmass — the hope is that AI software could one day transform society the way the Industrial Age did.

Sam Altman, the CEO of OpenAI, calls this next leap the "Intelligence Age." In a September blog post, Altman said its defining characteristic would be "massive prosperity." However, he cautioned that without enough infrastructure, "AI will be a very limited resource that wars get fought over and that becomes mostly a tool for rich people."

Last week, in one of his final executive orders, President Joe Biden signaled intent to build more at home with plans to lease acres of federal land to private sector firms with the know-how to develop complex AI infrastructure. The intent to build is likely to continue after Donald Trump's inauguration on Monday, as tech leaders rally around the incoming president and put AI among the top priorities on his agenda.

Biden's executive order followed the release of a blueprint from OpenAI a day earlier, which claimed "the economic opportunity AI presents is too compelling to forfeit" by not building the infrastructure needed.

Data centers, power plants, and chip manufacturing plants will all cost money — a lot of money. Goldman Sachs estimates that roughly $1 trillion will be spent in the next few years alone to develop the infrastructure needed to bring today's AI models closer to superintelligence.

It's why the big question investors and companies must now grapple with is whether or not they are willing to put up money for a vision of the future that is hardly guaranteed.

The case for building AI infrastructure

Altman has offered no shortage of reasons for spending so much money on achieving superintelligence.

Ensuring technological hegemony over China is one. As his company said last week, "there's an estimated $175 billion sitting in global funds awaiting investment in AI projects" that "will flow to China-backed projects" and strengthen Beijing if not directed to the US.

Another is that superintelligence could unlock unimaginable prosperity for society. Altman recently said that "if we could fast-forward a hundred years," the prosperity from superintelligence would feel just as unimaginable as today's world would to a lamplighter, a person employed to light and maintain street lights until about the 1950s.

The third reason is perhaps more surprising. In a blog published at the start of the year, Altman said his company is now confident that it knows how to build artificial general intelligence, a term often interchanged with superintelligence despite their differences.

It's a combination of factors that will, in some way, have triggered the flood of comments from those who want to play their part in developing the infrastructure needed to deliver superintelligence.

In a blog published this month titled "The Golden Opportunity for American AI," Microsoft president Brad Smith said the company planned to spend $80 billion alone this year on data centers. "Not since the invention of electricity has the United States had the opportunity it has today to harness new technology to invigorate the nation's economy," he said.

Last year, in conjunction with BlackRock and others, the tech giant unveiled a fund focused on AI infrastructure with an investment potential of up to $100 billion.

In an interview with Semafor last month, Google CEO Sundar Pichai said that he was ready to work on a "Manhattan Project" for AI once Donald Trump takes office, underscoring the scale of the development and investment needed by invoking the World War II program that eventually produced the atomic bomb.

Meanwhile, Japanese conglomerate SoftBank committed $100 billion to investing in the US over the next four years, focusing on AI and related infrastructure.

A risky investment

While there is clear intent to develop AI infrastructure, it's not clear if or when the investments will pay off — for two key reasons.

First, much of the infrastructure needed in the US faces an uphill struggle to get built.

Take chip plants. US companies like Nvidia, Google, AMD, and others that specialize in designing chips have developed a significant reliance on Taiwanese firm TSMC to manufacture those chips in the Far East, where a combination of cheap, skilled labor, economies of scale, and a long history of government support for the semiconductor sector has made the incredibly expensive business of manufacturing chips easier to pull-off.

Simply throwing capital at projects aimed at getting chips manufactured in the US won't cut it. Efforts to build chip manufacturing plants at home have been taking shape — the Biden administration's CHIPS Act has provided billions of dollars of grants to semiconductor firms in the US — but there remains a huge gap between the capabilities of manufacturers in the East versus those at home.

The AI boom has been kind to TSMC, with its value roughly doubling last year to $1.1 trillion. US chip manufacturer Intel, meanwhile, more than halved to around $85 billion.

Clean power infrastructure, increasingly focused on nuclear power, also faces challenges. Returns on investment in nuclear power projects meant to provide clean energy to intensive data centers are highly uncertain. These projects also face significant regulatory hurdles.

In December, for instance, the States of Texas, Utah, and Washington D.C.-based company Last Energy sued the Nuclear Regulatory Commission over claims that the government agency was applying the same risk analysis to small modular reactors as it was to large-scale power plants. These SMRs, as they're known, are meant to make access to nuclear power cheaper, given their compactness and greater affordability versus traditional nuclear plants. But even these face roadblocks.

The second big reason that investors may want to approach infrastructure investment with caution is that the emergence of superintelligence remains highly speculative.

Altman's claim that there is now a clear path to AGI is worth taking seriously, as new models like OpenAI's o3 released in December demonstrate increasingly sophisticated reasoning capabilities that do more than just parrot their training data.

That said, there have been rumblings across the industry recently about AI models hitting a wall in terms of performance improvements.

Without really serious advances in capabilities, then, or a clearly defined path forward to superintelligence, it is not clear how or when these colossal bets on AI infrastructure will pay off. But with China and other nations showing no sign of slowing down, it is clear that the cost of not being in the AI race could be far greater.

• TikTok is back online in the US after Donald Trump promised to delay a ban.
• TikTok is working for users who already have the app, but unavailable to download from app stores.
• The Supreme Court upheld the TikTok divest-or-ban law on Friday.

After going dark for millions of Americans this weekend, TikTok came back online following a pledge from Donald Trump to keep it going. There's just one problem: the law says it should still be offline.

TikTok's status in the US was thrown into confusion ahead of Trump's inauguration on Monday after the president-elect promised to extend the deadline on a law threatening to ban the social media app unless it was sold by its Chinese owner ByteDance.

The divest-or-ban law signed in April last year on national security grounds was upheld by the Supreme Court on Friday, meaning TikTok would officially be offline in the US on January 19 unless it had a new owner. On Saturday, TikTok went dark for American users ahead of the ban.

But following Trump's pledge over the weekend to delay the law's effect with an executive order — he asked tech companies to "not let TikTok stay dark" — the companies responsible for making TikTok available in the US appear to have taken diverging paths over the law.

"In agreement with our service providers, TikTok is in the process of restoring service," the company said in a statement on Sunday.

TikTok also thanked Trump for his assurance that its service providers — which include Apple, Google, Oracle, and cloud provider Akamai — "will face no penalties" by providing TikTok to over 170 million Americans.

Not all service providers seem to have been convinced that's the case.

Both Apple and Google have opted to comply with the law upheld by the Supreme Court by blocking new downloads of TikTok on their respective app stores; TikTok remained unavailable on their app stores ahead of Trump's inauguration on Monday.

It means people in the US can use TikTok if they have already installed it, but they cannot download it from Apple or Google's app stores.

Meanwhile, The Information reported on Sunday that key partner Oracle was turning its TikTok servers back on to support that process — despite the ban. Under the terms of a 2020 deal, US TikTok user data is stored on Oracle Cloud.

TikTok, Oracle, Apple, Google, and additional service provider Akamai Technologies did not immediately respond to Business Insider's request for comment.

How this saga will unfold once Trump takes office continues to face uncertainty.

Though Trump said on his social media platform Truth Social that "there will be no liability for any company that helped keep TikTok from going dark" before his promised executive order, it appears that Apple and Google have chosen to stand by the law that officially bans TikTok.

Not standing by it could come with risk: the law states that US service providers that make TikTok available in the country face a penalty of $5,000 for each person who uses the app. TikTok has over 170 million users in the US so that fine could top $850 billion.

The decision to stick by the law has also drawn praise from some quarters of the Republican Party. In a statement, Senators Tom Cotton and Pete Ricketts commended tech companies for following the law.

"The law, after all, risks ruinous bankruptcy for any company that violates it," they said. "Now that the law has taken effect, there's no legal basis for any kind of 'extension' of its effective date."

• Apple has temporarily disabled AI summaries of news notifications in a new software release.
• The feature has faced criticism from media outlets over concerns that AI generates factual errors.
• Apple has been betting big on AI to trigger an iPhone upgrade cycle.

Apple is pausing a feature in its artificial intelligence software designed to summarize news notifications following backlash from media companies that said the technology was making major errors.

The technology giant unveiled the change to its Apple Intelligence platform on Thursday following its latest software release to developers. The iOS 18 beta 3 update shows that generative AI-enabled notification summaries are "temporarily unavailable" for news apps.

In a statement, Apple confirmed to Business Insider that notification summaries for news and entertainment will be temporarily unavailable with its latest beta software releases across iPhones, Macs, and iPads. "We are working on improvements and will make them available in a future software update," the company said.

The decision comes just months after the company introduced its generative AI platform, Apple Intelligence, to users. Apple touted the summarized notifications feature as one that surfaces "what's most important" to users.

However, the feature has prompted criticism from some outlets, who complained that the generative AI powering Apple Intelligence notifications repeatedly made mistakes when summarizing news headlines.

The BBC raised serious complaints with Apple last month after an AI summary of a story on Luigi Mangione — the suspect behind the killing of UnitedHealthcare CEO Brian Thompson — falsely told some iPhone users that Mangione had shot himself.

It's the latest sign of the challenges Silicon Valley companies face with the technology that their industry has rallied around since the launch of ChatGPT.

Major tech companies like Google and OpenAI have also seen their AI generate major inaccuracies, raising concerns about the technology's potential to spread false information.

It's also a sign that Apple faces teething problems with a technology that CEO Tim Cook has called a "new chapter in Apple innovation."

Analysts have been examining whether Apple Intelligence can match the performance of generative AI offerings from rival firms and trigger an iPhone upgrade cycle.

Not all have been convinced it will. Apple received a "sell" downgrade this month from Craig Moffett, senior analyst at MoffettNathanson, who told Bloomberg that there are concerns consumers are "unmoved by AI functionality."

• President Joe Biden signed an executive order that aims to speed up AI infrastructure projects.
• It directs the DOD and DOE to lease land to the private sector for building AI data centers.
• The executive order requires developers to build clean energy sources to power their data centers.

President Joe Biden has signed an executive order to accelerate the scale-up of AI infrastructure in the US to meet the massive energy and supply needs of the "cutting-edge" technology.

Biden said on Tuesday that the executive order will direct the departments of Defense and Energy to lease federal sites to the private sector for building "frontier AI infrastructure at speed and scale."

According to the White House, private sector access to federal sites for building "gigawatt-scale AI data centers" will be determined by "competitive solicitations" for proposals.

Private companies leasing the land would be required to cover the cost of building and operating the sites, including the build-out of clean energy sources for powering data centers.

Biden, who is serving his final days in office, said the US cannot take its lead in AI for granted. The technology is set to have "profound implications for national security and enormous potential to improve Americans' lives if harnessed responsibly," Biden said.

"That is why today, I am signing an historic Executive Order to accelerate the speed at which we build the next generation of AI infrastructure here in America, in a way that enhances economic competitiveness, national security, AI safety, and clean energy," Biden said.

The move from Biden comes at a critical time for the AI sector.

Industry leaders say progress for leading AI companies — which have invested billions of dollars in the technology — depends on an extraordinary build-out of infrastructure.

Expensive data centers and clean energy supplies are needed to train, host, and run new models sustainably, while chip plants that supply vital computing power are being demanded closer to home after years of reliance on chipmaking facilities overseen in the East by the likes of Taiwan's TSMC.

In September, OpenAI boss Sam Altman wrote in a blog that "if we don't build enough infrastructure, AI will be a very limited resource that wars get fought over, and that becomes mostly a tool for rich people."

Biden addressed the need for the US to become self-sufficient, noting in his statement that "we will not let America be out-built when it comes to the technology that will define our future."

The executive order will likely be a boost for private sector companies that have shown increasing interest in AI infrastructure development over the past few years.

In September, Microsoft and BlackRock announced the creation of a $30 billion megafund designed to drive "significant infrastructure investment" and "enhance American competitiveness in AI." The fund's total investment potential is $100 billion, they said.

SoftBank boss Masayoshi Son, whose company is betting big on AI, also committed in December to investing $100 billion in AI infrastructure in the US over the next four years after meeting President-elect Donald Trump.

• Apple still can't sell iPhone 16s in Indonesia due to a ban in place since October.
• Indonesia's industry minister said that Apple doesn't meet rules to source some materials locally.
• That's despite Apple's plan to invest $1 billion in building an AirTag factory in the country.

Apple is still banned from selling the iPhone 16 in Indonesia after a top government official dismissed its $1 billion plan to meet local investment requirements.

Indonesia's industry minister, Agus Gumiwang Kartasasmita, said in a briefing on Wednesday that Apple's proposal, which included a plan to build an AirTag factory in the country, did not fulfill the demands of its local regulations.

Apple has been blocked from selling its latest iPhone model in the country of over 280 million people since October. It currently falls short of a requirement for tech firms to locally source at least 40% of the material in their smartphones and tablets.

Indonesia's investment minister, Rosan Roeslani, said on Tuesday that Apple had "committed" to building the AirTag factory on Batam Island, with operations beginning in early 2026.

However, Kartasasmita said on Wednesday that the factory would not be enough to reverse the ban, with AirTags considered just an accessory, according to comments reported by Bloomberg.

"As of this afternoon, the government does not have a basis for issuing the local content certificates" that Apple needs to sell its flagship device in Indonesia, Kartasamita said. "Apple needs to negotiate with us so that we can issue a certificate."

The comments came after Reuters reported that Indonesia's industry minister had met with Apple representatives on Tuesday to discuss the tech giant's plans to comply with investment expectations and get the iPhone 16 ban lifted.

However, Kartasasmita's comments on Wednesday suggest negotiations between Indonesia and the world's most valuable company had failed to reach a resolution, extending Apple's drought on iPhone 16 sales in Indonesia.

Apple's difficulties in Southeast Asia's largest economy have been deepened as smartphone rivals such as Samsung have pushed forward with their own efforts to meet Indonesia's regulatory demands.

"There's no deadline for compliance," Kartasasmita said. "If Apple wants to sell the iPhone 16, and especially if they plan to launch the iPhone 17, the decision is entirely up to them."

Apple did not immediately respond to BI's request for comment.

• Jensen Huang has unveiled a platform called Cosmos to simulate scenarios to train real-world robots.
• Huang likened it to Marvel superhero Doctor Strange simulating millions of versions of the future.
• The Nvidia boss said at CES that physical AI is the "next frontier" of artificial intelligence.

In "Avengers: Infinity War," Marvel superhero Doctor Strange looks into the future to see over 14 million different outcomes of the galactic battle against supervillain Thanos. Jensen Huang thinks it's the kind of power needed to reach "the next frontier of AI."

In a keynote address at CES in Las Vegas on Monday, the Nvidia CEO introduced Cosmos, a platform that aims to make "physical AI" a reality by simulating endless real-world scenarios for robots and autonomous vehicles to study and gain a deeper understanding of their environment.

According to Huang, the path to this next frontier — in which autonomous hardware becomes a common sight in daily life — has been limited until now because of data availability. As he put it, "Physical world data is costly to capture, curate, and label."

That's where Nvidia Cosmos comes in, for Huang at least. "You could have it generate multiple physically-based, physically plausible scenarios of the future," he told the Las Vegas audience. "Basically, do a Doctor Strange."

Nvidia's next frontier is coming

Here's how it works. Cosmos ingests text, image, or video prompts to generate videos with virtual renderings of real-world environments, lighting, and more.

Developers of robots and autonomous vehicles can then use these virtual creations to provide their technology with synthetic data for reinforcement learning — a research technique used to teach AI models — as well as test and validate the models behind the physical AI.

According to an Nvidia blog post, Cosmos can also be used along with Omniverse, the company's platform for creating 3D graphics and metaverses, to "generate every possible future outcome an AI model could take to help it select the best and most accurate path."

Cosmos itself starts with a strong, foundational understanding of real-world environments. It has been trained on 20 million hours of video focusing on everything from humans walking and "dynamic nature" to camera movements, Nvidia said.

If robots and autonomous vehicles are to become a widespread reality, as other industry leaders like Elon Musk think, they'll need a highly sophisticated understanding of these kinds of scenarios.

"It's really about teaching the AI, not about generating creative content, but teaching the AI to understand the physical world," Huang said.

There's a good reason Huang is talking up physical AI. While Nvidia has grown by roughly $3.3 trillion since the start of the generative AI boom, thanks to high demand for its chips needed to train AI models, the business isn't completely free of threats.

Some of Nvidia's Big Tech customers, such as Amazon and Google, are developing chips of their own to reduce their dependence on Nvidia. The company made 87.7% of its $35.1 billion revenue last quarter from its chip and data center business.

As Business Insider's Emma Cosgrove also notes, the semiconductor industry has historically been brutal. Companies typically experience boom and bust cycles as interest in niche chips can come in waves. There is an incentive then for Huang to diversify Nvidia's sources of income.

Time will tell if Cosmos can offer the path forward to Nvidia's next frontier. Development of robots that can navigate complex world environments has taken shape slowly, despite companies like Google, Boston Dynamics and Figure AI deploying increasing amounts of capital on developing these technologies.

Huang himself noted during his CES keynote that he expects autonomous vehicles to represent the "first multi-trillion dollar robotics industry."

With autonomous cars already on the road in certain locations from companies like Waymo and Cruise, this could be the case. During CES, Huang shared that Nvidia had struck a new partnership with Toyota to help power its autonomous vehicle ambitions.

Getting to a world where robots roam freely among humans will take considerably more effort, however. Huang will hope that Cosmos starts to provide the superpowers needed to pull off such a feat.