The markets aren’t bracing for an AI crash — yet
Transformer Weekly: Blackwell chips and China, White House warning to pro-AI super pac, and OpenAI’s restructure
Welcome to Transformer, your weekly briefing of what matters in AI. And if you’ve been forwarded this email, click here to subscribe and receive future editions.
NEED TO KNOW
Nvidia’s Blackwell chips weren’t on the table at Trump’s talks with Xi.
The White House is annoyed at the Leading the Future super pac.
OpenAI completed its for-profit restructure.
But first…
THE BIG STORY
Amid all the talk of an AI bubble, it seemed mostly business as usual when some of the world’s biggest tech companies released results this week, at least as far as the markets were concerned.
Google parent Alphabet posted more than $100b in quarterly revenue for the first time, powered by its cloud business riding demand for enterprise AI services.
Gemini is also apparently “generating billions in quarterly revenue” and reaching 650m users, according to CEO Sundar Pichai.
Amazon, too, saw its cloud business outperform expectations, helping push its share price up more than 10%.
Microsoft’s higher than anticipated spending forecasts sent its shares down slightly after the results, despite its Azure cloud business performing unexpectedly well — again due partly to AI demand.
Meta was the outlier. Its plans to spend between $70b and $72b — mostly on compute capacity — spooked investors who see a company still lagging behind competitors on AI, sending its stock down 10%.
Microsoft, Meta and Google racked up $78b in capital expenditures last quarter, an 89% increase from a year ago, and Amazon alone spent $89.9b in the quarter — all mostly for data centers and the chips to fill them. Microsoft, Amazon and Google are all, like Meta, increasing their forecast spending, too.
Google now expects between $91b and $93b this financial year, up from $85b.
Microsoft plans to spend $120b, up from $88b.
Amazon said it now expects to spend $125b, up from $118b.
Though the response to the earnings was mixed, what it didn’t suggest was anyone worrying about a looming crash.
Meta took a big hit not because of a lack of faith in AI, but a lack of faith in Meta’s ability to profit from it.
Mark Zuckerberg insisted that it’s sensible to “frontload building capacity, so that way we’re prepared for the most optimistic cases,” and that this would mean “superintelligence arrives sooner.” Investors weren’t buying it.
There’s also a key difference between Meta and Google, Amazon or Microsoft: Meta doesn’t sell computing power to other companies. Servicing the race for AGI still looks like a decent bet to Wall Street.
The week brought other signals seemingly suggesting a crash isn’t imminent.
Microsoft’s market cap pushed above $4t after what was seen as a favorable result from OpenAI’s restructuring. Apple, hot on its heels, also hit the $4t milestone.
Nvidia’s valuation rose to make it the first $5t company, driven in part by expected Trump-enabled sales to China. If anyone’s worried about circular funding deals, the market isn’t showing it.
Of course, this is the kind of pattern you would see before a crash. The bubble has to inflate enough to burst, and the market reaction to Meta and Microsoft suggest at least some nervousness around big spending and uncertain returns is creeping in.
It’s also notable that much of the media coverage before and after the results invoked the specter of a bubble, prompted in part by warnings from the likes of JPMorgan CEO Jamie Dimon, the Bank of England and of course Sam Altman and Jeff Bezos.
Bubbles and crashes are as much about confidence as they are about fundamentals. The mood music has changed, even if the markets aren’t yet listening.
— Jasper Jackson
THIS WEEK ON TRANSFORMER
Where have the really big AI models gone? — Lynette Bye explains why scaling took a break.
What you need to know about the OpenAI restructure — Celia Ford breaks down the new board and financial setup.
Japan’s unusual approach to AI policy — Eleanor Warnock on how Japan’s punishment-free approach to AI regulation might be a model for others.
Audits, not essays: How to win trust for enterprise AI — Synthesia’s Alexandru Voica on the importance of standards.
THE DISCOURSE
Casey Newton has reservations about AI firms’ commitment to tackling mental health concerns:
“Even as it works to reduce emotional dependence on ChatGPT for some subset of distressed users, OpenAI will be working to increase users’ overall dependence on its products for work, entertainment, web browsing and more ... It would be great if AI companies continue to see emotional dependence on their products as a problem to be solved. But it seems likely that, at least to some of them, it will eventually be evidence that the product is working as intended.”
Jasmine Sun wishes tech critics would admit that lots of people like having AI friends:
“Price and stigma were the only reasons that paying for friends wasn’t already more common. Cheap chatbots in your pocket have now solved both, and millions of people have AI friendships that carry all the emotional intensity of human ones.”
Anthropomorphic AI is “a devil’s bargain,” Sun wrote:
“If companies encourage human-AI relationships at scale, they should expect user revolts, lawsuits, and responsibility for the psychological chaos that results.”
Steven Adler warned against trusting OpenAI’s claims about mitigating the risks of AI smut:
“If the company really has strong reason to believe it’s ready to bring back erotica on its platforms, it should show its work…People deserve more than just a company’s word that it has addressed safety issues. In other words: Prove it.”
Zvi Mowshowitz detailed some of the reasons he signed a statement calling for a pause in building superintelligence.
“At the same time as everything counts, the core reasons these problems are wicked are fundamental. Many are technical but the most important one is not. If you’re building sufficiently advanced AI that will become far more intelligent, capable and competitive than humans, by default this quickly ends poorly for the humans.”
POLICY
Trump’s meeting with Xi Jinping led to a one-year trade deal that will see China suspend rare earth metal export controls.
While they did talk about Nvidia’s access to the country, Trump said they didn’t discuss the most powerful Blackwell chips.
US Trade Representative Jamieson Greer told reporters Blackwells unlikely to be part of future negotiations. “We do have crown jewels in the United States, the Blackwells being part of those, so those were not really on the table.”
The suggestion that Blackwells would be on the table had earlier in the week got lots of heat from the likes of Zvi Mowshowitz, Miles Brundage, Dean Ball and Caleb Watney.
The rest of Trump’s Asia tour produced deals with Japan and South Korea to collaborate on AI, semiconductors, and biotech.
Malaysia, Cambodia and Thailand also agreed to cooperate on rare earth minerals and export controls on China.
Sens. Josh Hawley and Richard Blumenthal introduced the GUARD Act, which would ban AI chatbots for minors.
The bill would force AI firms to include disclosures from chatbots that they aren’t human every 30 minutes.
Providing chatbots that encourage “suicide, non-suicidal self-injury, or imminent physical or sexual violence,” would be a criminal offense with potential fines of up to $100,000.
US Customs and Border Protection signed a two year deal to use Altana’s AI platform for trade enforcement, forced labor and counternarcotics risk detection.
The US Department of Energy and AMD formed a $1b partnership to build two supercomputers for scientific research using AI. Los Alamos National Laboratory is building two supercomputers based on Nvidia’s Vera Rubin platform.
Saudi Arabia is reportedly pouring money into its AI buildout in a bid to become the world’s third-biggest provider of compute.
The country is building three new data centers it claims will provide compute 30% cheaper than what’s available in the US.
The head of Humain, the state-backed company set up in May to oversee AI projects, said it was working on deals with xAI, Amazon and Microsoft.
EU officials reportedly hosted a delegation from China in Brussels on Thursday in a bid to calm tensions caused by the US-China trade battles, as well as the Netherlands’ decision to seize control of chipmaker Nexperia.
INFLUENCE
The White House is reportedly not happy about the pro-AI super pac Leading the Future and its plans to fund candidates from both parties.
“Any donors or supporters of this group should think twice about getting on the wrong side of Trump world,” an anonymous White House official told NBC.
Leading the Future has reportedly launched a “dark money” offshoot called Build American AI. The non-profit plans to run millions of dollars of ads promoting its legislative agenda, including pushing for federal, not state, AI regulation.
A bunch of former national security advisors were among the signees of a letter organised by Americans for Responsible Innovation urging Congressional leaders to pass the GAIN Act.
“AI is rapidly becoming the backbone of advanced intelligence, cyber defense, and autonomous systems. Maintaining priority access to compute is as vital to our long-term security as maintaining priority access to energy, rare earths, or steel was in past decades.”
Microsoft’s general manager for US policy Gerry Petrella said the revised GAIN Act looked “really positive.”
“You would then have policy in place where we were making sure we keep as much of that advanced compute that we need in the United States, but… creating a pathway for those American companies to build and export American AI around the world to our allies.”
Nvidia held its first developers conference — the “Super Bowl of AI” — in DC.
Jensen Huang tailored his effusive speech to Trump’s “pro-America” agenda, reportedly thanking the crowd for “making America great again.”
Huang told reporters: “I come with only one purpose — only just to inform and to be in service of the president.”
He also brushed off AI bubble fears: “All of these different AI models we’re using — we’re using plenty of services and paying happily to do it.”
Industry groups including the Chamber of Commerce and the Information Technology Industry Council called on congress to end the shutdown, citing delays to permitting for new AI projects.
Tech companies have reportedly made public officials and land sellers sign NDAs on data center projects.
OpenAI said the US needs to build 100GW of energy generation capacity each year to stay ahead of China.
In a blog post it said: “Electricity… [is] a strategic asset that is critical to building the AI infrastructure that will secure our leadership on the most consequential technology since electricity itself.”
INDUSTRY
OpenAI
OpenAI completed its reorganization into a for-profit public benefit corporation, eliminating caps on returns while keeping some nonprofit governance controls.
Microsoft received a 27% stake in the company.
Softbank’s approved the $22.5b remaining of its investment in OpenAI — contingent on their restructuring.
The Midas Project criticized the move, citing “significant concerns” about “whether this restructuring adequately preserves OpenAI’s founding commitments to humanity.”
OpenAI is reportedly laying the groundwork for an IPO that could value the company at up to $1t, according to Reuters.
The WSJ reported that OpenAI’s promise to remain in California was key to securing Attorney General Rob Bonta’s approval for the restructuring.
It released research revealing that millions of ChatGPT users show signs of reliance, psychosis, or suicidal intent.
Around 0.03% of messages “indicate potentially heightened levels of emotional attachment to ChatGPT,” and 0.05% “contain explicit or implicit indicators of suicidal ideation or intent.”
The company says it has since made GPT-5 safer: Compared to GPT-4o, GPT-5 showed a 39-52% decrease in delusions, suicidal ideation, self harm, and emotional reliance.
Sam Altman shared internal goals to develop an “automated AI research intern” by 2026 and a “true AI researcher by 2028.”
The company is also reportedly moving towards an “AI cloud,” and has committed $1.4t to infrastructure.
OpenAI released gpt-oss-safeguard, an open-weight reasoning model for safety classification tasks.
Atlas, OpenAI’s new browser, raised security concerns due to its extensive data collection and vulnerability to prompt injection attacks.
PayPal will become ChatGPT’s first embedded digital wallet.
OpenAI is reportedly working on AI for music generation.
Nvidia
Nvidia announced new deals with Palantir, CrowdStrike, and Eli Lilly.
Jensen Huang also announced plans to optimize AI for factories.
Samsung said it is using 50,000 Nvidia GPUs to power a new semiconductor “AI Megafactory.”
Nvidia has backed 59 AI startups without being a VC firm, Bloomberg reported.
Blackwell chips are now being manufactured in Arizona.
Nvidia took a $1b stake in Nokia to supply AI-powered computers for wireless networks.
Microsoft
Azure and 365 experienced a major outage hours before the company’s earnings release.
Users can now vibecode apps with Copilot.
WSJ columnist Jonathan Weil criticized Microsoft for not providing a “clear number” for their OpenAI losses or its investment structure.
Australia’s Competition and Consumer Commission sued Microsoft for allegedly misleading customers about AI-bundled Microsoft 365 price increases.
xAI
Elon Musk launched Grokipedia, a bare-bones AI alternative to Wikipedia.
Musk’s lawyer vowed to continue legal action against OpenAI despite its restructuring.
Even if Musk wins, it may be hard to unwind the change, The Information reported.
NextEra Energy partnered with Google to restart a nuclear plant in Iowa to power data centers.
Google will offer free AI Pro subscriptions to millions of users in India.
Google upgraded chat in NotebookLM, expanding its context window and conversation memory.
Anthropic
Amazon opened an $11b AI data center in Indiana to train Anthropic’s models on its own chips.
Claude now works directly in Excel (via a Clippy-esque sidebar).
The company’s unglamorous focus on corporate customers may make for a better business model than their rivals, theorized WSJ writer Asa Fitch.
Others
Character.AI said it would ban under-18s from interacting with its chatbots days after the Bureau of Investigative Journalism revealed disturbing interactions with AI characters, including one modelled on Jeffrey Epstein.
Meta’s AI app downloads are soaring following the Vibes launch, but engagement still trails Sora and doesn’t hold a candle to Instagram.
Substrate raised $100m to develop a new, lower-cost semiconductor manufacturing process using particle accelerators.
SemiAnalysis explained how Substrate’s new X-ray lithography tool could disrupt ASML’s monopoly on advanced chipmaking equipment.
Uber plans to launch 100,000 Nvidia-powered autonomous vehicles by 2027.
IBM released four open-source Granite 4.0 Nano AI models with as few as 350m parameters — small enough to run on a laptop or web browser.
Cursor launched Composer, a new coding model that’s 4x faster than similar models.
Intel held talks to acquire AI chip startup SambaNova.
Qualcomm unveiled new AI200 chips aimed at challenging Nvidia and AMD.
Its stock jumped 20% after the announcement.
MiniMax launched M2, an open-source model that lags just behind top American model capabilities at 8% of Claude Sonnet’s price.
Mercor, which recruits humans to train AI chatbots, finalized a funding deal valuing it at $10b.
Amazon announced 14,000 white-collar layoffs, partly due to AI.
Kyle Vogt’s The Bot Company, which builds household robots, is eyeing a $4b valuation.
MOVES
A fifth of OpenAI’s current employees previously worked at Meta, The Information reported.
Meta moved former metaverse executive Vishal Shah to AI products.
Researcher Ariel Kwiatkowski resigned from Meta, citing “better places to build AGI” that won’t create “stepmom chatbots and infinite slop machines.”
YouTube CEO Neal Mohan announced a major reorganization focused on AI and voluntary layoffs.
Mikhail Shapiro, award-winning biomolecular engineer, joined Merge Labs, Sam Altman’s new brain-computer interface startup.
Sora engineer Will Depue returned to OpenAI to lead a team focused on a moonshot he says has “a small, but significant, chance of leading to ASI.”
Advertising chief John Nitti left X after just 10 months.
RESEARCH
Anthropic model psychiatry lead Jack Lindsey published a paper reporting evidence of “some degree of introspective awareness” in Claude models.
Lindsey cautioned against interpreting these results as a sign that any AI system is conscious.
Scale AI and the Center for AI Safety launched the Remote Labor Index to measure how well AI agents can do remote work.
Anthropic Fellows and researchers at Thinking Machines Lab reported that frontier models behave differently when forced to choose between competing values.
For example, Grok 4 stood out for being willing to respond to harmful requests, while Claude 3.5 Sonnet frequently refused benign ones.
Epoch AI found that decentralized AI training across multiple sites is feasible, potentially solving power constraints for large AI models.
Researchers from Shanghai AI Lab and elsewhere released EgoThinker, a framework that teaches AI models to infer intentions in first-person video — a key capability for wearables and robots.
New survey results from American University’s Sine Institute found that young adults are generally apprehensive about AI.
Only 21% said they “feel more excited and positive” than “concerned and anxious.” Over half believe AI will limit their career opportunities.
Beyond job loss, many respondents cited concerns that AI could compromise critical thinking, spread misinformation, and increase loneliness.
BEST OF THE REST
Wired released a big (and beautiful) package about AI.
Highlights include “An Algorithm for Consciousness” and “The Argument for Letting AI Burn It All Down”
A16z’s Speedrun program is funding “an unappetizing buffet of bad ideas,” 404 Media reports.
Among them: a “Bible-based AI buddy,” an “AI powered credit card,” and “mythological pets that turn wellness into play.”
China is using Deepseek to power military tech, Reuters reported.
A journalist attempted to live without AI for 48 hours. It wasn’t easy.
Billboards featuring satirical messages about AI have been popping up in New York and San Francisco from a seemingly fake company called Replacement.AI, reported SFist.
China is building a $226m wind-powered underwater data center that uses seawater for cooling.
Developing countries such as Indonesia are pursuing “AI decolonization” to get tech companies to locate servers locally, reports WSJ.
Understanding AI has 16 charts that explain the AI boom.
Paris Hilton has been training her own AI for years, according to The Information.
MEME OF THE WEEK
Thanks for reading. If you liked this edition, forward it to your colleagues or share it on social media. Have a great weekend.
Great roundup, but I think the most concerning part of OpenAI's transition to a for-profit is that the nonprofit isn't being fairly valued.
The nonprofit is getting ~20% of the for-profit, when a fair number should be more like 50%, which is why Zvi calls it potentially the biggest heist in history.