Chinese AI makers have learned to build powerful AI models that perform just short of the U.S.'s most advanced competition while using far less money, chips and power.

Why it matters: American policies restricting the flow of top-end AI semiconductors and know-how to China may have helped maintain a short U.S. lead at the outer reaches of the AI performance curve — but they've also accelerated Chinese progress in building high-end AI more efficiently.

Catch up quick: In late December, Hangzhou-based DeepSeek released V3, an open-source large language model whose performance on various benchmark tests puts it in the same league as OpenAI's 4o and Anthropic's Claude 3.5 Sonnet.

• Those are the most advanced AI models these companies currently offer to the broad public, though both OpenAI and Anthropic have next-generation models in their pipeline.

Stunning stat: Training V3 cost DeepSeek roughly $5.6 million, according to the company.

• OpenAI, Google and Anthropic have reportedly spent hundreds of millions of dollars to build and train their current models, and expect to spend billions in the future.
• AI pioneer Andrej Karpathy called DeepSeek's investment "a joke of a budget" and described the result as "a highly impressive display of research and engineering under resource constraints."

Between the lines: In an interview last year, DeepSeek CEO Liang Wenfeng said, "Money has never been the problem for us; bans on shipments of advanced chips are the problem."

• The V3 model was trained on Nvidia H800 chips, a less-powerful version of a chip the U.S. banned for export to China in 2022. Export of the H800 was then prohibited when the U.S. tightened controls again the following year.

The big picture: Some U.S. officials have argued for restricting China's access to advanced AI chips even further in hopes of slowing the country's development of the technology.

• On Monday, the Biden administration announced another big round of export controls aimed at choking the supply of chips to China via third-party countries.

What's next: Advances like V3 and OpenAI's powerful new "reasoning" model, o3, have lent weight to recent claims by OpenAI CEO Sam Altman and other industry leaders who predict the industry is closing in fast on artificial general intelligence (AGI). (Plenty of other observers remain skeptical.)

• AGI — or AI that can solve problems and perform tasks at a human or beyond-human level — is a holy grail for AI researchers, and many in the industry and U.S. government believe the technology's first developer will win a massive economic, scientific and security edge.
• Biden's latest export controls have led some observers to conclude the government shares the growing sense that AGI is close.
• "This is a 'break in case of emergency' policy, and the Biden administration identified that the emergency is that AGI is just a few years away," Gregory Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies, tells Axios.

Yes, but: AGI is also not well-defined, and both optimists and pessimists have complained that it's become a moving goalpost.

They say it's hard to turn a battleship around, but Mark Zuckerberg just about-faced his globe-spanning, $1.5 trillion-value, 3 billion-user company — transforming Meta from a bastion of Silicon Valley's socially progressive neoliberalism into a full-on MAGA hive.

Why it matters: After Zuckerberg's embrace of Trump and Trumpism, Silicon Valley is holding its breath to see whether a whole row of tech dominoes is about to fall in the same direction.

Some early signs of wobble:

• Google, Apple, Microsoft and Amazon have all lined up to give $1 million each to the Trump inauguration. (Apple's donation came personally from CEO Tim Cook.)
• Amazon put $40 million behind a Melania Trump documentary.

State of play: So far, while Meta's competitors have ritualistically expressed their willingness to work with the new administration, none of them has gone as far as Zuckerberg in donning the corporate equivalent of a MAGA hat.

• This week he announced Meta was ending its fact-checking program; loosening restrictions on hate speech; and shutting down the company's diversity efforts.

Publicly traded companies with billions of customers generally try not to alienate any large bloc of the public. Becoming closely aligned with either side of the U.S.'s red/blue divide risks limiting a business's market reach.

• Until now, Elon Musk has been the striking exception to this rule. Zuckerberg makes two — and his moves open the door wider for anyone else who wants to emulate them.

Yes, but: Zuckerberg, unlike his rival CEOs, has absolute voting control of his company.

• As he said in a three-hour conversation with Joe Rogan Friday, "Because I control our company, I have the benefit of not having to convince the board not to fire me."
• None of the other members of tech's trillion-dollar club can move with the same speed or independence, even if they wanted to.

Case in point: Apple has always aimed, and often managed, to transcend mere politics and inhabit a separate dimension making "great products that people love."

• Cook worked with Trump last time around, winning "Tim Apple" as a sobriquet but also occasionally speaking out about the importance of protecting immigrant workers and LGBTQ rights.
• Google and Microsoft — both with immigrant, non-white CEOs — also offered some pushback on those issues during Trump's first administration as well.
• Tech leaders united to oppose Trump's ban on transgender troops and, in a variety of combinations, took on some of his immigration policies.

This time around, these firms are quietly signaling they want to cooperate with the new Trump team on issues — like competition with China — where they see common ground.

• All these companies were the target of a federal antitrust full-court press in the Biden era, and each is wondering whether they can get a reprieve from the new team in D.C.
• Every tech giant also wants to expand the use of skilled-worker visas, and that has already led to friction with the immigrant-hostile MAGA movement.

During the first Trump term, an activist young tech work force occasionally took to the barricades to protest government policies and pressure reforms from their employers.

• But multiple rounds of layoffs and a messy pandemic recovery have left workers feeling more insecure.
• While Zuckerberg's new policies have sparked internal debate and dissent at Meta, the CEO may welcome an exodus of dissenters as a sort of voluntary lay-off.
• In a Threads post, he described users who might quit Meta's platforms in protest as "virtue-signaling."

What we're watching: With each fresh controversy the new administration touches off, tech CEOs will have to navigate a maze involving Trump's demands for loyalty, employees' emotions and wishes, and their own strategies.

• If Zuckerberg's experiment of committing a whole megacompany to taking overt sides in America's vast culture war pays off, others could follow.

The bottom line: Trump used to say that Zuckerberg would "spend the rest of his life in prison." But the incoming president's relationships with business leaders are strictly transactional, and Meta's CEO is probably resting a lot easier now.

Distinguishing truth from falsehood is frustrating, endless, thankless work — and now Mark Zuckerberg is walking away from it.

The big picture: Facebook's latest content-moderation pivot looks like part of a plan to win over Donald Trump as he takes power again. But the field Zuckerberg is abandoning is one he never wanted to play on in the first place.

State of play: The founders of social media giants like Facebook, Instagram, Twitter and TikTok didn't expect to end up in what the industry came to call the "content moderation" business — and what many critics, and now Zuckerberg himself, denounce as "censorship."

• Policing online speech costs a fortune to do right. It's impossible to make everyone happy. You're bound to make mistakes. And users' wishes keep changing.
• The whole effort is a distraction from what's always been Facebook/Meta's top priority — boosting engagement to sell more ads.
• Meta faces huge challenges this year, particularly an April trial in the Federal Trade Commission's suit to unwind its decade-old acquisition of Instagram and WhatsApp. The new Trump-friendly approach to content moderation is one of many efforts to win over the new administration, which is open about rewarding friends and punishing enemies.

Zuckerberg staked out a free-speech position in a 2019 speech at Georgetown. A few months later, he said that social media networks shouldn't try to be "arbiters of truth" — but at the same time Facebook was ramping up its truth arbitration.

• After taking blame for spreading misinformation during the 2016 election and violating users' privacy during the Cambridge Analytica scandal, Facebook was under enormous pressure to clean up its act, and the company made big investments in expanding its moderation efforts.
• It was also in 2019 that Facebook started a program using third-party fact-checking organizations from a variety of political perspectives to help it identify and limit the spread of potentially dangerous misinformation.

The fact-checking program has drawn fire throughout its existence.

• The kinds of topics it confronted — controversies over climate science, COVID-19 and vaccines, charges of election fraud — are often both matters of fact or science and also flashpoints for partisan rage.
• Believers in fact-checking insist that there's value to society in telling the public what is — and isn't — authoritative information, grounded in vetted research and verifiable records, in fields like medicine and public affairs.
• Critics say there's always another point of view that deserves to be heard, and blocking any perspective is a form of censorship.

Between the lines: Facebook tried to solve some of its content moderation headaches by setting up the independent Oversight Board.

• The company handed the Oversight Board hundreds of millions of dollars beginning in 2019 to build a kind of Supreme Court for user complaints.
• It's been particularly effective in sorting out complex issues beyond U.S. borders.
• But it hasn't insulated Zuckerberg and Meta from criticism by American conservatives and Congressional committees.
• Notably, Meta's announcements Tuesday failed to mention the Oversight Board at all.

Zoom out: Zuckerberg calls Meta's new approach a "back-to-our-roots" embrace of free expression. But there's never been any medium where absolute free speech reigned.

• Platform owners have legal obligations to governments of countries they operate in to obey the law.
• In the U.S. that means dealing with laws governing what Zuckerberg describes as "legitimately bad stuff" like "drugs, terrorism, child exploitation."

A second category where platform owners have generally felt an obligation to intervene is speech that could cause imminent harm.

• That might include death threats, violent conspiracies, or even plans to attack the U.S. Capitol.

Then there's the category of hate speech.

• It bedevils social-media owners, because it's constantly shifting and varies across cultures. But in any given time and place there are some slurs that violate public norms, and a global public platform can't just ignore them.
• Along with the company's other new content policies, Meta has now updated its Community Guidelines on "hateful conduct" specifically to allow "allegations of mental illness or abnormality when based on gender or sexual orientation." A variety of other changes in its rules appear to significantly loosen the knot around what many users will consider hate speech.

A final category of content moderation — most relevant to fact-checking — is misinformation (widely shared but inaccurate info) and disinformation (misinformation with deliberate bad intent).

• Zuckerberg and other social media owners hate playing content cop in this realm and would rather let users sort such issues out for themselves.
• His plan is for Meta to copy X's Community Notes approach, which lets users flag other users' posts for inaccuracies.

What we're watching: When Elon Musk rewrote Twitter's old content rules for X, the platform's never-decorous conversations deteriorated further. Today you don't have to look far on X to find posts espousing racism and antisemitism or deriding LGBTQ people.

• Musk is proud of what he's done with X, but it hasn't helped turn around his business.
• We don't yet know how Zuckerberg's version of "more free speech" will play out, but if Meta's platforms get nastier and uglier, too, advertisers could be spooked — and users who aren't on the MAGA side of the fence could flee.

Our thought bubble: Decades of human experience online shows that running any kind of community platform is like gardening — if you let the weeds go wild, the flowers will choke.

• This isn't Mill's free market of ideas; it's a world where bad speech drives out good speech.
• And if you think there's no such thing as bad speech, you haven't been on the internet for the last 25 years.

Wednesday's extraordinary display of Elon Musk's social media power saw the world's richest man funnel the anger of his hundreds of millions of online followers against a bipartisan compromise funding bill.

Why it matters: Musk's arresting demonstration of government-by-tweet left Washington reeling — but the magic trick will get harder to repeat.

The big picture: Shooting down a spending bill is a lot easier than passing one. Stopping government in its tracks is much easier than making it work.

• Beginning Jan. 20, Republicans will essentially control all three branches of government, and making it work will be their job.

How it works: Musk's X has become a partisan megaphone fine-tuned to promote its owner's own messages to a user base of MAGA supporters, libertarians, crypto enthusiasts, and fans of Musk himself.

• Its influence is strongest with Republican House members because its base overlaps with theirs.
• The other half of America has been voting with its feet, with many Democrats and progressives abandoning X in favor of a variety of online alternatives.

Between the lines: Musk's Wednesday call to arms took place at his point of maximum power.

• The legislators he wanted to pressure answer to the votes of the people who listen to him.

Yes, but: Republicans' control of Congress beginning in January will be precarious.

• The GOP has a margin of only a couple of votes in the House, and the new Senate majority will often hang on a handful of Republican moderates' votes.
• When it comes time to build any kind of coalition to pass a new law or a spending measure, Musk's megaphone is likely to prove a clumsy tool.
• If the Musk-led "Department of Government Efficiency" recommends cuts to Social Security or health care, mustering the same kind of support could be a lot harder — and, historically, opponents of such measures have found effective ways to exert their own "vox populi."

Right now, Trump and Musk seem to be on the same page. That could change any time.

• Trump-watchers suggest the president-elect's jealousy reflex could already have been triggered by Musk's high-profile power flex.

The "bully pulpit" got its name over a century ago, and American leaders have always urged the people to rise up and write, call or email their representatives.

• Musk's pulpit moves strikingly faster — getting results within hours on Wednesday.

What to watch: If Musk can repeat the trick, he will have ushered us into a new era.

• Then we'll have to figure out whether what's happening is really direct democracy, or rule by unelected billionaire.

Go deeper: Musk's unprecedented power

OpenAI is putting its powerful video creation tool, Sora, in the hands of basically every one of the millions of people who pay for ChatGPT, the company said yesterday.

Why it matters: We're about to witness an at-scale experiment in what it means for people to create and consume large quantities of video content that is photorealistic but fake.

The big picture: OpenAI released Sora precisely because it wants the results of that experiment.

• "We're introducing our video generation technology now to give society time to explore its possibilities and co-develop norms and safeguards that ensure it's used responsibly as the field advances," the company wrote in a post introducing the new product.

Catch up quick: When OpenAI gave the public a sneak peek at Sora last February, the clips the company showed caused jaws to drop — but also triggered an allergic reaction in Hollywood.

• Many artists view the advent of AI-generated video as a fraud and an insult. Just as with text-based AI, many creators also believe OpenAI trained its tool using copyrighted works.

What they're saying: "We don't want the world to just be text," OpenAI CEO Sam Altman said in a live-streamed announcement yesterday. Video is "important to our culture."

The company said in a statement that the latest Turbo version of Sora, which will be offered as a standalone product to ChatGPT Plus and Pro customers, is "significantly faster" than the version the firm previewed. It lets users generate videos up to 20 seconds long.

An early review by Marques Brownlee, who got to play with Sora before the release, details some of what OpenAI admits are the "many limitations" of the tool.

• In particular, Sora's videos regularly seem to have problems handling basic physics. Objects appear and disappear, particularly when something else moves in front of them.

OpenAI isn't trying to pull the wool over anyone's eyes here. In Monday's live-streamed launch, Sora team leaders said the tool was "not about generating feature-length movies" but instead provides a "co-creative dynamic" so users can explore new ideas.

• The company imagines Sora not as a filmmaker but as a filmmaker's sounding board.
• The new version of the tool lets you drop your own images in as prompts. There's also a timeline editor that lets you add new prompts at specific moments in a video.
• In a demo at Monday's Fortune Brainstorm AI conference in San Francisco, OpenAI also showed off a feature — still in private testing — that allows users to narrate videos using AI-generated versions of their own voices, in multiple languages.

The company said it understands that putting Sora into so many hands could cause problems.

• Altman said OpenAI wants to prevent illegal use of the tool, "but we also want to balance that with creative expression."
• Users' ability to upload images of people will be "limited at launch," OpenAI's blog post said.
• Moderation of user content in Sora will be "starting a little conservative," but if it "doesn't quite get it right, just give us that feedback," Altman said.

Friction point: Sora will be available globally, except in Europe and the U.K. for now — presumably because of stringent EU privacy laws.

OpenAI's competitors are also moving fast with AI video-making offerings in the meantime.

• Google last week rolled out its Veo video generator to business partners, and Meta is developing Movie Gen — but neither is yet available to the general public.

Zoom out: Today's online world isn't exactly experiencing a shortage of brief videos.

• Logged-in Sora users can view a bottomless feed of "featured" Sora-made videos, alongside the prompts that created them.
• OpenAI suggests aspiring video producers will want to use this feed as a source of inspiration, and doubtless they will — but this viewing mode also feels a lot like TikTok or YouTube.

Some observers Monday applauded Sora's capabilities, while others predicted it would deluge us with AI "slop."

What we're watching: OpenAI aims to prevent outright illegal uses of Sora to create child sexual abuse material, impersonation and other problematic material.

• But even if the company succeeds, internet users have a way of pushing the boundaries of new tools like this until every possible awful output has been demonstrated.

The bottom line: Sora's evolution will give us all an early glimpse of how our social and political systems handle broad exposure to AI-made video.

Go deeper: Video-making AI tools are moving into general use

Competition in AI is less a single race than a triathlon: There's a face-off to develop the most advanced generative AI foundation models; a battle to win customers by making AI useful; and a struggle to build costly infrastructure that makes the first two goals possible.

Why it matters: Picking a winner in AI depends on which of these games you're watching most closely. And the competition's multi-faceted nature means there's more than one way to win.

State of play: Below, we offer an overview of where the key players in generative AI stand in each of this triathlon's events.

Between the lines: The benchmarks used in the industry to compare performance among AI models are widely criticized as unreliable or irrelevant, which makes head-to-head comparisons difficult.

• Often, developers' and users' preference for one model over another are very subjective.

OpenAI

Two years after the launch of OpenAI's ChatGPT kicked off the AI wave, the startup remains AI's flagship.

• It has raised about $22 billion and is in the process of retooling itself from a safety-oriented nonprofit to a globe-spanning for-profit tech giant.

Yes, but: OpenAI's last major foundation model release, GPT-4, is now nearly two years old. A long-awaited successor had its release pushed back into 2025 amid a swirl of reports that its advances may not be game-changing.

• Meanwhile, OpenAI has pushed the field's edge with innovations like its "reasoning" model, o1, and impressive voice capabilities.

Models: OpenAI still has a lead, but it's shrinking.

Customers: OpenAI has direct access to a vast pool of over 200 million weekly active ChatGPT users and indirect access to the huge installed base of Microsoft users, thanks to its close alliance with that giant.

Infrastructure: OpenAI is highly dependent on Microsoft for the cloud services that train and run its AI models, though it has recently begun an effort to expand its partnerships.

Anthropic

Like Avis to OpenAI's Hertz, Anthropic seems to be trying a little harder. It has also faced fewer distractions from high-profile departures and boardroom showdowns than its competitor.

Anthropic was founded by ex-OpenAI employees aiming to double down on OpenAI's commitment to caution and responsibility in deploying AI.

• But it has now raised roughly $14 billion and begun to embrace OpenAI's philosophy of putting AI in the public's hands to pressure test its dangers.

Models: Anthropic's Claude 3.5 Sonnet is widely viewed as a worthy competitor to GPT-4 that, in some cases, even surpasses it.

Customers: Claude's usage numbers are much smaller than ChatGPT's, but the company is partnering with big and medium-sized firms looking for a counterweight to Microsoft.

Infrastructure: Amazon, which recently invested $4 billion into Anthropic, is putting its massive cloud resources behind the company, and Google's parent Alphabet has also provided some investment and support.

Google

Google's long-term investments in AI research made generative AI's breakthrough possible.

• But ChatGPT's overnight success caught the search giant flat-footed.
• Google has spent much of the last two years in catch-up mode — uniting its DeepMind research team and Brain unit and injecting its Gemini AI across its product line.

Models: Google's Gemini is very much in the same league as OpenAI's and Anthropic's models, though some reports suggest that it hasn't found as much pickup among AI developers.

Customers: By pushing its own AI summaries to the top of search results and integrating its AI with its Android mobile operating system, Google has ensured that its own AI would get in front of a global user base.

Infrastructure: Google has the know-how and the resources to scale up as much AI power as it needs, but the field's competitive frenzy has left it off-balance.

Meta

Meta has embraced and promoted open source AI via its Llama models.

• The strategy is a way to avoid becoming dependent on a competitor for AI services — the way it found itself reliant on Apple and Google in the smartphone era.
• Meanwhile, Meta has been deploying its own custom version of the technology, dubbed Meta AI. The chatbot lives inside Messenger and WhatsApp, has taken over search in Instagram and powers the assistant on Meta Ray-Ban glasses. The future has Meta AI even spitting out its own posts.

Models: Meta's models haven't directly taken on OpenAI and its competitors. Instead, they've offered better performance at smaller scales and the cost savings and freedom that the open source approach allows.

Customers: Over 3 billion social media users provide Meta with an enormous pool of consumers, while some business customers will be won over by Llama's low price and adaptability.

Infrastructure: Meta doesn't run its own B2B cloud, but has plenty of experience scaling up data centers.

Other players

Microsoft has tied its AI fate to OpenAI, but it has also begun to build out its own in-house strategy.

• It's developing its own models in a project led by DeepMind co-founder and former Inflection CEO Mustafa Suleyman.

Amazon has concentrated on meeting the AI boom's enormous demand for cloud services, but it's also investing in its own series of models.

xAI, Elon Musk's venture, raised about $6 billion in the spring and another $5 billion this month — along the way building what it calls the world's largest AI data center at impressive speed.

• But it's not yet clear how xAI intends to compete beyond making vague promises around freedom of speech.

Apple has played catch-up as it works to weave Apple Intelligence into its mobile and desktop operating systems and upgrade its Siri assistant.

Other races

AI's leading competitors are also engaged in an increasingly tough scramble to find more training data for their models.

• Tech giants that have been stockpiling data for years have a natural advantage.
• But data access faces challenges from copyright law, distrust of the technology by creators and the public, and privacy commitments some companies have made.

Zoom out: Most of these companies have also explicitly committed themselves to the quest for artificial general intelligence, or AGI — a level of autonomous intelligence that matches or exceeds human capabilities.

• But since everyone has a different definition of AGI, it may be tough for anyone to lay a claim to achieving it first.

Elon Musk is a wild card in the tech industry's frantic effort to game out where a Trump-dominated Washington will come down on AI regulation.

State of play: It's a reasonably safe bet that President-elect Trump will trash President Biden's modest moves to set limits on AI development and give companies a free hand to do what they want — and beat China.

Yes, but: Musk, who has been at Trump's right hand since his election victory, has two very different personas when it comes to AI regulation, and no one knows which of them will be whispering in Trump's ear.

Musk has been obsessed with AI doomsday scenarios for at least a decade.

• He co-founded OpenAI in 2015 and provided the initial cash for the nonprofit, all in the name of protecting the world from runaway super-intelligence.
• In 2014, he told MIT students that AI needs regulatory controls: "With artificial intelligence we are summoning the demon. In all those stories where there's the guy with the pentagram and the holy water, it's like — yeah, he's sure he can control the demon. Doesn't work out," said Musk.
• In August, Musk came out in support of California legislation that would impose new obligations on developers of advanced AI. The bill passed the state legislature over industry opposition, but got vetoed by Gov. Gavin Newsom.

On the other hand, Musk has also launched a crusade against what he calls "woke" AI.

• He built his new AI startup, xAI, around a commitment to abandoning guardrails against hate speech and misinformation in the name of freedom of speech.
• Musk's feud with OpenAI CEO Sam Altman, rooted in a fight for control over the nonprofit's early development, now centers on Musk's complaint that OpenAI's models have a built-in liberal bias and stifle conservative views.
• Last month, in an interview with Tucker Carlson, Musk said, "I don't trust OpenAI and Sam Altman. I don't think we want to have the most powerful AI in the world controlled by someone who is not trustworthy." Over the weekend, Musk reposted that quotation and video.

Friction point: Advocates of AI caution like Max Tegmark, the AI researcher who leads the Future of Life Institute, are counting on Musk to champion a tight leash on AI in Trump's councils.

• But MAGA's "anti-woke" warriors consider Musk as an ally in their fight to keep AI free of what they see as burdensome guardrails.

Musk's new role as a key Trump adviser means that when Trump takes office on Jan. 20, this contradiction will become a practical dilemma overnight.

Case in point: As Axios reported last week, the Department of Energy is working with Anthropic to test the company's Claude models, making sure they don't give users "help" in building homemade nuclear weapons.

• This sort of research has proceeded under the umbrella of President Biden's AI executive order, which Trump is expected to quickly revoke.
• Nuclear bomb-making — like bioterrorism, financial meltdowns and network sabotage — is exactly the kind of existential AI risk Musk has long warned about.
• But the measures companies are most likely to take today to limit the risk of such outcomes look a lot like what Musk and others on the right denounce as "censorship."

Between the lines: Since Musk is now a participant in the race to build bigger, faster, better AI models as founder and investor in xAI, his stance on AI regulation may become less a matter of philosophy than of self-interest.

• It also means his involvement in AI decision-making would be deeply compromised by conflicts of interest.

Our thought bubble: "AI safety" is in the eye of the beholder.

• "Aligning" AI with human welfare has long been the preoccupation of those, like Musk, who fear super-powered AI could run amok.
• But for AI makers busy deploying error-prone AI tools on a mass scale today, "safety" looks more like blocking an AI image maker from generating revenge porn, preventing a large language model from espousing antisemitism or stopping an AI agent from making racist decisions.

What we're watching: Musk's sway with Trump is high right now. But the president-elect's favors are fickle, and there are already signs that Musk may be overplaying his hand.