What We Got Right (and Wrong) in Our Predictions for 2025
We predicted a bunch of big tech deals and other events for last year. Some happened. Most didn’t.
What’s that smell? It’s the whiff of humble pie being eaten by The Information’s prognosticators.
A year ago, we made a series of predictions about tech deals that would get done, executives who would leave their companies and other events that we thought might happen in 2025. Many of those predictions did not come to pass. In our defense, when you have a policy of making bold and specific predictions, as we do, strikeouts are a lot more common than home runs. Thankfully, there were a couple of home runs among our predictions too.
Here’s how we graded last year’s predictions. Tell us in the comments if you think we were too easy or harsh in our self-assessments.
What we predicted: Amazon Will Buy Lyft
What happened: A deal between the two companies seemed to make sense: Lyft is a distant No. 2 in traditional ride-hailing to Uber, which makes a deep-pocketed partner like Amazon appealing. Despite its also-ran status, Lyft still has enough scale to help Amazon bring its autonomous vehicle taxi service, Zoox, to the masses. Oh, and Lyft’s CEO David Risher is a former longtime Amazon executive. But for now, Lyft and Amazon are on their own separate paths.—Nick Wingfield
Grade: F
What we predicted: Elon Musk Will Try to Buy TikTok
What happened: There was also a logic to this prediction. President Donald Trump was eager to find American investors to buy TikTok from its Chinese parent ByteDance to avert a shutdown of the app in the U.S. due to a law passed by Congress. Musk at the time was a close ally of Trump’s and already had a stake in social media through his ownership of X. But Musk in February said publicly that he wasn’t interested in buying TikTok. And he wasn’t on the broader list of investors in a new U.S.-based TikTok entity that the app’s management shared with its employees last month.—N.W.
Grade: F
What we predicted: Nvidia Will Make Several Significant Acquisitions
What happened: The AI chip giant announced one bona fide blockbuster deal in 2025, although it was not a traditional acquisition: its $20 billion agreement in December to hire most of the staff of the chip startup Groq and pay for a nonexclusive license to Groq’s technology. It struck a similar deal with the networking startup Enfabrica a few months earlier, but that agreement was much smaller at around $900 million. Nvidia may have a more sizable deal in the works: Earlier this week, Calcalist, an Israeli financial publication, reported that Nvidia is in advanced talks to buy an Israeli AI startup, AI21 Labs, for as much as $3 billion. If that deal happens, it will be in 2026.—N.W.
Grade: B+
What we predicted: Microsoft Will Do a Deal With Anthropic
What happened: We nailed this one. We posited that Microsoft and Anthropic would cozy up as the former company sought to lessen its dependence on OpenAI, still its most important AI partner. Sure enough, in November, the companies announced a strategic partnership through which Anthropic committed to spending $30 billion to buy computing capacity from Microsoft’s Azure cloud service. Microsoft also pledged to invest up to $5 billion in Anthropic, and Anthropic made its Claude AI models more broadly available to Microsoft customers as part of the agreement.—N.W.
Grade: A
What we predicted: Amazon Will Launch Its Own Version of Ozempic
What happened: We were wrong. Amazon didn’t launch its own version of a GLP-1, the category of weight loss drugs that includes Ozempic and Wegovy, in 2025. It did, however, expand its pharmacy business in other ways, including by launching same-day medication delivery in more U.S. cities and adding medication kiosks to One Medical offices. The wait for an “Amazon Basics” version of GLP-1 drugs continues.—Theo Wayt
Grade: F
What we predicted: Telegram Will Launch a Stablecoin
What happened: Didn’t happen. The Dubai-based messaging app is popular among crypto traders and users, and we predicted that the allure of launching its own stablecoin would be irresistible for Telegram. Instead, the first wave of noncrypto firms issuing stablecoins turned out to be payment and fintech related firms, such as Western Union, Klarna and SoFi. Social media and communications apps could join that crowd, but they haven’t yet.—Yueqi Yang
Grade: F
What we predicted: An AI Agent Will Cause Chaos for a Blue-Chip Company
What happened: We were pretty close. In early December, The Information reported that a chatbot launched by the retailer Gap Inc. to field product-related questions went off the rails, discussing a bunch of questions outside its purview including about sex toys and Nazi Germany. The chatbot was powered by enterprise AI startup Sierra, cofounded by former Salesforce co-CEO Bret Taylor.
Sierra said the episode was caused by a “bad actor” trying to “jailbreak over a dozen of our customers’ AI agents.” The company said its abuse detection systems caught all but the Gap one, although the problem was fixed. As we noted at the time, it wasn’t the only such embarrassment. In July, Elon Musk’s xAI had to apologize after its Grok chatbot praised Adolf Hitler in a series of X posts.—Martin Peers
Grade: A
What we predicted: Marc Lore’s Wonder Will Acquire FreshDirect or Gopuff
What happened: Wonder didn’t acquire FreshDirect, Gopuff or another grocery business in 2025. It made two other acquisitions though: Tastemade, a media business, and Spyce, the robotics company that manufactures Sweetgreen’s automated salad-making machines.—Ann Gehan
Grade: F
What we predicted: Tim Cook Will Extend His Apple Tenure for Another Half-Decade
What happened: It’s hard to say definitively whether Apple’s CEO has committed to remain at Apple for another five years, but it seems unlikely. Apple’s most senior executives are at-will employees, without employment contracts, the company says in its filings with regulators. If Cook had made an explicit promise to the company’s board to stay at the company through a particular date, there’s a good chance Apple would have disclosed such a detail.
There’s also been considerable buzz lately about Apple stepping up its CEO succession planning, as we and others have reported. Nothing appears to be imminent on that front, but five years is a long time in a tech industry moving very quickly.—N.W.
Grade: F
What we predicted: College Athletes Will Become University Employees
What happened: We were premature. Alabama football players, Duke basketball stars and other college athletes didn’t start earning wages from their schools in 2025. It may be a question of when, not if, that happens. Last month, a bill backed by the NCAA and the U.S. Olympic and Paralympic Committee hit a wall in Congress; it would prevent college athletes from being classified as employees.
Another part of our prediction did come to pass though—namely, that the flood of name, image and likeness dollars into colleges athletes’ pockets would prompt some of them to choose college over the pros. Just last week, on Christmas Eve, Baylor University announced it signed seven-foot Nigerian center James Nnaji, who was a second-round draft pick for the NBA’s Detroit Pistons in 2023. Colorado College this week added a former minor league professional hockey player to its roster. Both developments are a sign that, for some athletes, playing college sports and landing blockbuster sponsorships is starting to look like a more appealing financial bet than grunting it out in minor leagues or sitting on pro benches.—Sara Germano
Grade: C
What we predicted: A Buyout Wave Will Hit Consumer Companies
What happened: None of the direct-to-consumer brands that we named in this prediction as likely buyout candidates—such as ThredUp and Allbirds—ended up being taken over by private equity firms in 2025. There were, however, some notable deals that weren’t that far off from the ones we predicted. For example, Rent The Runway’s longtime lender and other investors took control of 86% of the company’s shares in August and restructured its debt obligations. But Rent the Runway remains a publicly-traded company. Meanwhile, Grove Collaborative conducted a strategic review in 2025 after one of their shareholders pushed for it, but a buyer hasn’t yet materialized for the business.—Ann Gehan
Grade: C
What we predicted: AI Chatbot Companies Will Face Hundreds of Lawsuits
What happened: We were way off. Chatbot companies have faced lawsuits, to be sure, but most likely in the dozens, not the hundreds. Plaintiffs filed seven lawsuits against OpenAI in November alone alleging that the company’s ChatGPT pushed their family members into delusional conditions that drove some people to kill themselves. Just a month ago, the New York Times sued Perplexity for copyright infringement, adding to a suit the Times filed against OpenAI a couple of years ago. Amazon and the Chicago Tribune also sued Perplexity in recent months, while Reddit sued the firm in October. Other similar lawsuits in 2025 included Advanced Local Media suing Cohere in June. The lawsuits are coming, just not that many.—Martin Peers
Grade: C
What we predicted: Cisco CEO Chuck Robbins Will Leave
What happened: We whiffed badly on this one. Not only is Robbins still running Cisco, but the networking giant’s business performed well in 2025 thanks to the AI boom. The company’s stock hit an all-time high in December and closed the year up nearly 31%.—N.W.
Grade: F
What we predicted: Warner Bros. Discovery Will Rebrand Max
What happened: In July, Warner Bros. Discovery did indeed change Max’s name to HBO Max, which is what an earlier version of the service was called. The rechristening was just one episode within a much bigger drama involving the beleaguered company’s future, which reached new heights in December when Netflix and David Ellison’s Paramount made competing acquisition offers.—Abram Brown
Grade: A
What we predicted: OpenAI’s Chief Operating Officer, Brad Lightcap, Will Depart
What happened: The bottom line is that Lightcap, as of this writing, is still at OpenAI. However, leadership changes continued to chip away at his influence at the company. (Earlier such changes are what prompted our prediction.) In May, OpenAI hired then-Instacart CEO Fidji Simo to become the CEO of Applications. That role could have been a natural fit for Lightcap, who was running OpenAI’s day-to-day business operations at the time. Now, Lightcap reports to Simo, placing him a step further away from CEO Sam Altman.—Stephanie Palazzolo
Grade: F
What we predicted: Conservative-Favorite Industries Will Be Funding Hot Spots
What happened: We were right on the money in predicting defense tech and crypto would both get big investment bumps in the Trump era. Global venture capital investments in defense tech startups surged to $59 billion for the 12 months ended Sept. 30, 2025, compared to $25.2 billion for the comparable period in 2024, PitchBook estimates. Investments in stablecoins jumped too, buoyed by the Trump administration’s crypto-friendly policies, including his signing of the GENIUS Act in July, which provided the first regulatory framework for the industry. Circle Internet Group, the stablecoin issuer, pulled off a spectacular IPO in 2025 as well, though some crypto stocks and token prices cooled considerably in the latter half of the year.
There were fewer obvious signs of an investment bonanza in other startup categories on our list of potential beneficiaries. Funding in ed tech startups, for example, was flat through the end of November compared to the full year of 2024, according to CrunchBase. And there is sparse data on a funding boom for longevity startups. One exception was NewLimit, which was cofounded by Coinbase CEO Brian Armstrong and raised a $130 million Series B round from Kleiner Perkins and others in May.—N.W.
Grade: A-
What we predicted: Microsoft Will Stave Off the FTC’s Current Investigation
What happened: It’s too soon to say whether this prediction will age well or not down the road, but it didn’t come true in 2025. Andrew Ferguson, the Trump-appointed successor to Lina Khan as chair of the Federal Trade Commission, hasn’t moved to drop the agency’s far-reaching probe into Microsoft—instead, at least early on, Ferguson’s FTC continued its line of questioning by asking other companies about alleged anticompetitive behavior by Microsoft as of March 2025. That doesn’t necessarily mean the FTC will actually file a full-fledged antitrust lawsuit against Microsoft. Meanwhile, Trump himself has shown an appetite to go after Microsoft, including demanding that it fire its president of global affairs, Lisa Monaco, because she previously worked for the Obama and Biden White Houses. Microsoft has so far declined to do so.—Aaron Holmes
Grade: D
What we predicted: Trump Will Sign Child Online Safety Legislation Into Law
What happened: The president didn’t sign a comprehensive online safety law aimed at protecting children, such as the Kids Online Safety and Privacy Act, which passed the U.S. Senate in 2024 before stalling in the House. Such an event could be getting closer though. In mid-December, a House subcommittee advanced a package of 18 child online safety bills, including a new version of KOSA, for the full House Energy and Commerce Committee to consider. But compromises in that legislation have upset some lawmakers and it’s likely to face aggressive lobbying by the tech industry, making it hard to predict whether it will pass and in what form.—N.W.
Grade: D
What we predicted: Alibaba Will Acquire One of China’s Top AI Startups
What happened: Alibaba Group didn’t acquire Moonshot AI, as we predicted. The Chinese tech giant, however, did effectively acqui-hire some researchers from another major Chinese AI startup, 01.AI, which shifted its focus away from developing cutting-edge models.
Instead, the biggest event in China’s AI sector in 2025 was DeepSeek’s sudden rise to global stardom due to the release of R1, its highly capable open-source AI model. It prompted Moonshot and other Chinese AI startups to renew their focus on research and develop their own open-source models. In July, Moonshot launched a new model called Kimi K2, whose strong coding capabilities impressed app developers in the U.S. and elsewhere. On the back of Kimi K2’s success, Moonshot has been working on a new funding round at a valuation of $4 billion.—Juro Osawa
Grade: D
What we predicted: Josh Kushner Will Buy New York Magazine
What happened: In 2024, Kushner, the founder of investment firm Thrive Capital, made a handful of flashy media deals, including buying the rights to Life magazine and investing in the movie studio A24. So our speculation that he might snap up New York magazine wasn’t such a stretch. But we were wrong: Vox Media still owns the venerable magazine.—A.B.
Grade: F
Sutskever’s Fate, OpenAI’s Next Deal, A Hit Robot—and 13 Other Predictions for 2026
Our predictions for tech breakthroughs, deals and drama in the year ahead.
No one in Silicon Valley gets wealthy or powerful by dwelling on the past. Rather, the brightest minds within tech often outshine the competition by obsessing about the future.
We like to get in on the game too. In keeping with annual tradition, we’re publishing below our best educated guesses on what’ll happen in 2026: what Microsoft buys, what Amazon discards, what robot goes mainstream, and 12 other pieces of prognostication. (Over the last few days, we also thought aloud about Apple’s AI woes, Oracle’s debt strategy and xAI’s fate.) These predictions come from reasoned analysis of the companies and people in this world—and the studious reporting we do the other 364 days of the year, when we’re not moonlighting as clairvoyants.
Amazon Ditches Fresh Stores
Amazon will throw in the towel on its Fresh supermarkets, shuttering its roughly 60 locations across the U.S.
In 2025, Amazon executives embraced grocery as a critical growth area for the company’s retail business, and I expect its organic-focused Whole Foods stores won’t be in any danger. But Amazon’s vision for Fresh has grown increasingly muddled since the stores opened in 2020—initially they were meant to be a high-tech alternative to normal supermarkets, but Amazon wound up abandoning that approach in 2024.
There’s also less of an argument these days that Fresh stores can help Amazon’s online grocery business by doubling as delivery hubs. Amazon has been putting more groceries, including perishables like produce and meat, inside its normal e-commerce fulfillment centers and recently started testing 30-minute deliveries from mini-stores.—Ann Gehan
OpenAI Buys Pinterest
OpenAI will make its biggest acquisition yet when it buys Pinterest in a bold move to beef up its nascent online shopping and ads businesses.
OpenAI’s ChatGPT has gobbled up user time and attention, often at the expense of other consumer apps like Pinterest. While Pinterest has tried to go on the defense by adding AI features like conversational search, OpenAI would be most interested in the bones of Pinterest’s business, like its trove of image data and its existing advertising business, as well as relationships with merchants. Pinterest’s digital scrapbooking features could also complement OpenAI’s image- and video-generation tools and help it fend off competition from other AI heavyweights like Google, which launched a Pinterest-like feature in 2025.
Cashing out now would make sense. The company’s stock is trading around $25, roughly the same as in early 2023, giving it a market cap of roughly $17.5 billion. Pinterest cofounders Ben Silbermann and Paul Sciarra control roughly two-thirds of the company’s voting shares, and Silbermann, who left the CEO role in mid-2022, is no longer involved in Pinterest’s day-to-day management. —Ann Gehan
Google Keeps Its Adtech Monopoly
The Justice Department wants Google to sell part of its advertising technology business, but that won’t happen in 2026.
Sure, the judge in Google’s advertising antitrust case ruled in April that the company’s web ads business is indeed a monopoly. But that monopoly is in a small and still shrinking part of the overall digital ad market, since it doesn’t include apps or streaming video.
And during the September remedy trial, the judge expressed concern that a divestiture would take too long as the market for web ads is declining. Google has argued that if it were forced to exit the market, Google would have less incentive to direct advertisers’ budget to the web. Instead of divestiture, the judge could recommend a series of behavioral remedies that would make it easier to use Google’s technology with the technology of other firms and from the open source community. She is expected to deliver a verdict sometime in 2026.
Google still has to contend with a European Commission ruling that fined the company more than $3 billion in September, having earlier proposed a sale of its ad-tech business; that 2025 recommendation is awaiting a final ruling. But President Donald Trump responded to that fine by threatening further tariffs for Europe. It would be easy for the Europeans to fall in line and avoid further headaches.—Catherine Perloff
‘Continual Learning’ Becomes a Reality
OpenAI or Anthropic will release a model with “continual learning” capabilities, marking a major departure from how AI models currently work.
That will give AI models the ability to learn on the fly, like humans do, potentially with less need for data and computing power. Right now, AI models draw connections between concepts in gobs of data they’re trained on, and then use those connections to predict the next word in a sentence.
A number of AI labs are already conducting research into continual learning, and a breakthrough would have huge implications for the broader industry. For labs, the cost to train models could drop. That’s bad news for chipmakers and cloud providers, since it would tank the demand for compute.—Stephanie Palazzolo
Sutskever Sells SSI
Safe Superintelligence, the secretive AI lab cofounded by former OpenAI chief scientist Ilya Sutskever, will sell itself instead of continuing solo with its quest to reach superintelligence, or AI that surpasses human intelligence.
Sutskever vowed in 2024 that the startup wouldn’t release any products or models until it’s reached superintelligence. But the company has given little update on its progress, and you can imagine investors are getting antsy. Recently, reality seems to have caught up to Sutskever—in a November podcast, he said that SSI may actually release products before achieving superintelligence.
Given the mounting pressure on SSI to deliver, it’s not far-fetched to picture a sale instead. Sutskever has said that Meta Platforms attempted to buy the startup in 2025, which led cofounder Daniel Gross to depart and join Meta. Whether it’s Meta or another big tech acquirer, plenty of companies would be happy to have Sutskever’s research expertise.—Stephanie Palazzolo
Activist Investors Target Salesforce
Salesforce CEO Marc Benioff will face a familiar adversary: activist investors. The writing is on the wall, with Salesforce’s 20% stock price drop in 2025 reopening the door to activists snapping up shares.
Three years ago, Benioff withstood a months-long campaign from Starboard and other activist investors by cutting staff and holding off on major acquisitions. That helped profitability, but the company’s revenue growth has slowed dramatically.
This time around, activists may push Benioff to make room for new blood—perhaps by appointing a co-CEO. That didn’t work the last two times Salesforce tried it, but hey, maybe the third time’s the charm.—Kevin McLaughlin
Microsoft Stops Charging Extra for Copilot
Microsoft has had a tough time convincing companies to pay extra for its Copiliot AI assistant. In 2026, it will give up that effort entirely.
Instead, it will bundle Copilot into its 365 Enterprise suite and use the move as an excuse to raise prices a few dollars a seat for all corporate 365 customers. Microsoft already stopped charging for the consumer version of Copilot in early 2025, including it instead in the base cost of 365.—Aaron Holmes
Bret Taylor’s Sierra Gets Acquired
We’ll see at least one acquisition of an AI customer support startup by a software giant, thinning out a crowded field.
Right now, venture-backed startups Sierra, Distyl, Decagon and Pylon plus giants including Amazon and Salesforce are all vying for corporate customers. Potential acquirers may include Oracle, Microsoft and Google, which could look to upgrade existing products.
The biggest splash would be an acquisition of Sierra, valued at $10 billion in September and cofounded by former Salesforce co-CEO Bret Taylor. A deal involving one of the other startups could range between $1 billion and $3 billion. —Kevin McLaughlin
A Stock Exchange Buys a Predictions Market
A major stock exchange will buy a prediction market, a response to a fast-growing source of competition.
Polymarket, Kalshi and other prediction markets gained in popularity in 2025, letting people bet on everything from elections to Federal Reserve decisions. They threaten the likes of the New York Stock Exchange, Nasdaq and the Chicago Mercantile Exchange because prediction market users can make pure-play bets on events like election outcomes, instead of using traditional stock market bets as a proxy. Predictions on things like Fed rates also resemble derivatives that already trade on exchanges.
The interest is already there: Intercontinental Exchange Inc., the owner of NYSE, led an investment of up to $2 billion in Polymarket in October, while CME recently partnered with sports platform FanDuel to launch a prediction market.—Yueqi Yang
The Hit Robot Is a One-Trick Pony
There will be a hit robot, but it won’t be one of those much-hyped humanoids. Instead, the most popular robot will be designed to do specific household tasks like folding laundry well.
Expectations (and valuations) for companies developing humanoid robots are sky-high, thanks in part to promises that these robots will relieve people of their household chores. In fact, 1x Technologies said it intends to start delivering its svelte humanoid to customers’ homes in 2026. But the home is a notoriously tricky environment for any robot, let alone ones that have to balance on two legs and master five-fingered hands. Early buyers of humanoids will likely be disappointed when they can’t perform every chore right out of the box.
Already, souped-up Roombas that can mop as well as vacuum are gaining traction, and roboticists are toying with autonomous lamps and coffee tables. In 2026, Syncere AI will start delivering its spindly bedside robots that do two things: emit light and fold laundry. These purpose-built robots will likely gain a foothold before do-it-all home humanoids, similar to how MP3 players, pagers and cameras cluttered consumers’ pockets before smart phones supplanted them.—Rocket Drew
Chatbots Get Friendlier, Not Smarter
Chatbot makers have a problem: making their chatbots smarter isn’t attracting new users like it used to. So they’ll focus more on making the chatbots warmer and friendlier in personality. Don’t be surprised if these companies staff up on psychologists and turn out new benchmarks with names like “FrontierFriendly” or “CompassionBench.”
Those changes will help chatbots gain and retain users, but they could also bring about more instances of “AI psychosis.” That phenomenon emerged in 2025, with chatbots contributing to users’ delusional beliefs and mental health issues. Models with warmer personalities could be more inclined to agree with users than push back on their delusions.—Rocket Drew
ByteDance Makes Its Own AI Chips
ByteDance, the Chinese tech giant behind TikTok, will begin construction of its own chip-production facility, a major pivot that reflects how U.S. chip export controls have reshaped China’s tech landscape.
Until recently, it was unthinkable that a software-focused company like ByteDance would consider making its own chips, a highly capital-intensive endeavor. But ByteDance is in a squeeze.
It burns through advanced chips to power AI systems that keep billions of users scrolling, and it can no longer buy the most cutting-edge processors from Nvidia or use advanced chipmakers’ facilities in Taiwan and South Korea due to U.S. restrictions from 2022. Meanwhile, Beijing has pressured major tech firms to replace U.S. chips with homegrown alternatives, despite a significant performance gap.
ByteDance has never shied away from ambitious bets: It has pushed into cloud computing, launched in-app shopping, and experimented with education apps and gaming. Controlling its chip supply could unlock performance gains that competitors relying on off-the-shelf chips can’t match.
But opening the facility would happen over years, not months. Chinese foundries cannot procure the most advanced chipmaking equipment due to stricter export controls by America and its allies, and building even a modest facility requires billions of dollars. ByteDance’s facility would likely start small, functioning more as a research laboratory than a serious production operation. Even a partially successful effort will give ByteDance leverage with existing AI chip suppliers such as Nvidia and show Beijing it’s playing its part in China achieving technology self-sufficiency.—Qianer Liu
U.S. Reinstates Bans on Nvidia Selling H200 Chips in China
Nvidia seems poised to resume selling its H200 chips to China in 2026. That won’t last—Trump will reverse his recent approval of Nvidia H200 chip exports to China and return to the ban established by the Biden Administration.
The H200s, used in data centers to develop AI products, deliver approximately six times the performance of the H20, a weakened chip that Nvidia created to meet U.S. export restrictions for China. Allowing that level of computing performance into Chinese AI labs undermines America’s strategy of maintaining a technological edge over Beijing.
The policy reversal will likely come by June, as intelligence assessments and pushback from national security officials make the issue impossible to ignore. Defense and intelligence officials will present evidence showing how quickly Chinese labs are using H200 chips to close the AI gap, making the case that short-term U.S. trade benefits cannot justify long-term strategic risks.—Qianer Liu
States Prevail in AI Regulation
In 2025, the Trump administration, Congressional Republicans and a collection of Silicon Valley figures tried to block states from regulating AI, arguing that a single federal standard is better. Trump and his allies will likewise be thwarted in 2026, paving the way for more state AI bills to pass.
In July, it appeared that supporters of a state ban would succeed in adding a provision on the issue to the “One Big Beautiful Bill Act,” Trump’s sweeping tax and spending measure, which was enacted July 4. The provision would have imposed a 10-year moratorium on state AI regulation, but the Senate struck it down in a 99-1 vote. A separate effort to attach it to a defense bill in December also failed. More recently, Trump signed an executive order seeking to block state AI laws.
Blocking states from regulating AI can’t happen through an executive order; Congress would need to pass a law, and that seems unlikely to happen. The Democrats won’t go for it, and they are favored to win back the House in the 2026 midterm elections. Meanwhile, the Republican Party is divided on AI issues, and several Republican state governors have come out against the administration’s plan for the states. Plus, the public has a lot of concerns about the risks of AI, which, as I reported recently, could become a big issue in the midterms and influence how people vote.—Sylvia Varnham O’Regan
Tesla’s Cybercab Comes With a Steering Wheel
Tesla’s two-seat Cybercab, which the company says it will start producing in April, will come with a steering wheel and rearview mirrors. That will undercut the idea that these vehicles will be used for fully autonomous ride-hailing—without any need for humans to take over—anytime soon.
Tesla says the self-driving software powering the Cybercab will be 100% reliable 100% of the time. And so far, the company has posted concept models of the vehicle on social media that don’t have mirrors or steering wheels, though early test versions of the Cybercab spotted in Palo Alto and Austin have included them.
Although the company’s driving software has no doubt been improving, there are just too many situations where a human might need to take over, such as extreme weather or if the software temporarily goes offline. And having features like a steering wheel will help with regulators, too.—Theo Wayt
These Were The Largest Funding Rounds Of 2025
In startup circles, 2025 will be remembered as a busy year for big AI rounds.
A total of 15 companies secured venture funding rounds of $2 billion or more last year, per Crunchbase data. Among them, they amassed more than $100 billion from those financings.
The majority were generative AI companies, and they accrued most of the cash. The single largest round went to OpenAI, with its record-setting $40 billion SoftBank-backed financing in March. Four others raised rounds of more than $5 billion.
With the year now in hindsight, we thought it would be timely to take a look at who secured the largest financings. Below are the top 15, in descending order.
1. OpenAI, $40B, artificial intelligence: On March 31, OpenAI announced that it secured a $40 billion investment led by SoftBank. The deal for the San Francisco-based company is the biggest venture investment ever. Per details of the deal, SoftBank will build a syndicate of co-investors to provide $10 billion of the total, while it expects to fund the other $30 billion, with $10 billion of that amount through debt.
2. Scale AI, $14.3B, generative AI: San Francisco-based Scale AI, a provider of training data and model evaluation for AI applications, raised a reported $14.3 billion from Meta in June at a valuation of $29 billion. Under the agreement, Scale’s founder, Alexandr Wang, and some other employees joined Meta to work on its AI efforts.
3. Anthropic, $13B, generative AI: San Francisco-based generative AI unicorn Anthropic raised a $13 billion Series F round at a $183 billion valuation in September. Iconiq Capital led the round, with Fidelity and Lightspeed Venture Partners co-leading.
4. Project Prometheus, $6.2B, artificial intelligence: Project Prometheus, a startup focused on applying AI technology to physical tasks, launched with $6.2 billion in initial funding, per a November report. Jeff Bezos will reportedly serve as co-CEO, alongside Vik Bajaj, a physicist and chemist who served as CEO and co-founder of biotech startup Foresite Labs and is also known for his work at Google’s X.
5. xAI, $5.3B, generative AI: Elon Musk’s generative AI startup xAI pulled in $5.3 billion in fresh equity funding this summer, per a securities filing. Since its inception just two-and-half years ago, the Palo Alto, California-based company has raised more than $22 billion in equity and debt financing, per Crunchbase data.
6. Databricks, $4B, data and AI: Databricks announced in December that it is raising over $4 billion in a Series L financing at a $134 billion valuation, led by Insight Partners, Fidelity and J.P. Morgan Asset Management. The 12-year-old, San Francisco-headquartered company also said it crossed the $4.8 billion revenue run-rate in its third quarter, growing more than 55% year over year.
7. Anthropic, $3.5B, generative AI: Anthropic raised a $3.5 billion funding round in March led by Lightspeed Venture Partners, valuing the San Francisco-based unicorn at $61.5 billion.
8. Anduril Industries, $2.5B, defense tech: Costa Mesa, California-based defense tech startup Anduril Industries has raised $2.5 billion in a June Series G round led by Founders Fund, more than doubling its valuation to $30.5 billion post-money.
9. Anysphere, $2.3B, AI coding: Coding automation platform Cursor and parent company Anysphere raised $2.3 billion in a November Series D financing backed by Accel, Thrive Capital, Andreessen Horowitz, DST Global, Coatue, Nvidia and Google. The round set a $29.3 billion post-money valuation for the San Francisco-headquartered company.
10. (tied) Polymarket, $2B, prediction market: Intercontinental Exchange, the operator of clearing houses and exchanges including the New York Stock Exchange, announced in October that it will invest up to $2 billion into the prediction market platform Polymarket. The deal sets an $8 billion pre-money valuation for New York-based Polymarket, which lets users wager on event probabilities across markets, politics, sports and other areas.
10. (tied) Reflection AI, $2B, artificial intelligence: Reflection AI, a developer of LLM training models based on open standards, raised $2 billion in an October funding round backed by Nvidia and a long list of venture investors. The financing sets an $8 billion valuation for the New York-based company.
10. (tied) Safe Superintelligence, $2B, artificial intelligence: AI research lab Safe Superintelligence, the Palo Alto, California-based startup co-founded by OpenAI’s former chief scientist Ilya Sutskever, reportedly raised a $2 billion April round at a $32 billion valuation led by Greenoaks Capital Partners.
10. (tied) Thinking Machines Lab, $2B, artificial intelligence: San Francisco-based Thinking Machines Lab, the artificial intelligence startup launched and led by former OpenAI CTO Mira Murati, reportedly secured a $2 billion seed round at a $10 billion valuation with Andreessen Horowitz as lead investor. The financing ranks as the largest U.S. seed round of all time, per Crunchbase data.
10. (tied) Binance, $2B, cryptocurrency: Malta-based cryptocurrency exchange Binance received a $2 billion investment in March from Abu Dhabi-based investment firm MGX.
10. (tied) Mistral AI, $2B, generative AI: France-based frontier model company Mistral was valued at $13.2 billion in a $2 billion September funding led by Netherlands-based chipmaker ASML.
Inside Elon Musk’s Optimus Robot Project
The billionaire has bet Tesla’s future on humanoids, which for now rely on human helpers
Elon Musk envisions Tesla’s Optimus robots generating “infinite” revenue and becoming the “biggest product of all time.”
Optimus is still under development, with engineers working on dexterity and navigation, and some questioning its factory utility.
Tesla’s stock had rebounded in recent months amid optimism for robotaxis and humanoid robots.
The future of Tesla is an army of humanoid robots that Elon Musk says could eliminate poverty and the need for work. He has told investors the robots could generate “infinite” revenue for Tesla and have potential to be “the biggest product of all time.”
Musk has bet the company and his personal fortune on this vision of the world in which Optimus, as it is known, works in factories, handles domestic chores, performs surgeries and travels to Mars to help humans colonize the planet. Though today each robot is made by hand, Musk has proposed manufacturing millions of robots a year.
Optimus still has a lot to learn about the world before it is capable of replacing its human creators in the type of full-scale societal shift that Musk has in mind. In public appearances, the robot is often remotely operated by human engineers. On the engineering side, it has proven difficult for Tesla to create a hand for the bot with both the sensitivity and dexterity of a human. Inside Musk’s companies, some employees have questioned the usefulness of the bots for routine business operations like manufacturing.
Musk is motivated to prove the skeptics wrong. His new compensation package gives him 10 years to make Tesla a $8.5 trillion company and sell at least one million bots to customers, among other product and financial goals. Success could mean earning Musk a $1 trillion pay package, and expanding Tesla far beyond the electric vehicle industry where it made its mark.
“The car is to Tesla what the book was to Amazon,” Adam Jonas, an analyst with Morgan Stanley, said this summer. “Tesla used cars as a laboratory to get good at other things.”
On Friday, Tesla reported its vehicle sales fell 16% in the fourth quarter and dropped 9% for all of 2025, leaving it behind China’s BYD for the year. Tesla’s share price, which tumbled in early 2025 as the company’s EV sales slumped, had rebounded in recent months amid optimism in Musk’s pivot to robotaxis and humanoid robots.
Optimus is still under development, but the bot has become a familiar sight at company events and for many employees. Inside Tesla’s Palo Alto, Calif., engineering headquarters, robots routinely circle the inside perimeter of offices gathering information on how to navigate a room alongside humans.
In Tesla’s labs, the nearly 6-foot-tall machine practices rote tasks like sorting Legos by color, folding laundry and using a drill to screw a fastener, former employees said. In October, the bot made its red carpet debut at the “Tron: Ares” premiere in Hollywood, performing a choreographed fight sequence with actor Jared Leto.
Tesla is one of several companies pushing the frontiers of robotics in the hopes of cornering a nascent market. A crowd of Silicon Valley startups like 1X and Figure, other manufacturers like Hyundai’s Boston Dynamics, and Chinese robotics companies are eager to sell their robots that can fold laundry or manufacture vehicles.
Today, there are limits on how much robots can do. Many factories, including Tesla’s, rely on robotic arms to do heavy lifting or dangerous tasks like moving hot metal. Those robots are largely stationary and programmed to do specific tasks. That leaves humans with jobs that require flexibility and precision, like installing cables or seats into cars moving across an assembly line.
Humanoids have an obvious appeal: bipedal, with flexible joints, robots like Optimus are designed to function more easily in spaces meant for humans.
Roboticists, however, have struggled to design robots with enough dexterity, sensitivity and adaptability to move around freely, said Ken Goldberg, a roboticist at the University of California, Berkeley.
“I’ve heard Elon Musk say hands are the hard part. It’s true, but it’s not only the hand—it’s the control, the ability to see the environment, to perceive it and then compensate for all this uncertainty. That’s the research frontier,” Goldberg said. “Getting these robots to do something useful is the problem.”
Some Tesla analysts have struggled to price Tesla’s opportunity with humanoids given how new the industry is and exclude it from their financial models. Even Tesla bull ARK Invest, which expects Tesla’s share price to climb to $2,600 from around $400 today, left Optimus out of its model for 2029 because it doesn’t expect the product to be commercially successful until later on.
“We believe initial versions of the robot will likely have a limited set of performable tasks,” Tasha Keeney, a director at ARK Invest, said in an email. “Given Tesla’s competitive advantages in embodied AI and manufacturing scale, we expect the company to be a formidable competitor in the space.”
Morgan Stanley’s Jonas, who now covers the robotics industry, predicts that by 2050, humanoids will bring in $7.5 trillion in annual revenue across the industry globally. Capturing even a fraction of that market could supersize Tesla’s revenues, which came in at $98 billion in 2024.
At first it seemed like a joke. Musk unveiled Tesla’s bot concept at an event in 2021 with a human dancing on stage dressed in a robot costume.
“It’s intended to be friendly, of course, and navigate through a world built for humans and eliminate dangerous repetitive and boring tasks,” he said at the time. When Musk returned to the stage a year later, he demoed a prototype called Bumblebee, with visible wires and actuators.
Behind the scenes, Tesla engineers were working out of a kitchenette on campus. Soon after, the expanding group of Optimus engineers moved to a basement, then a large parking lot in another building, one former employee said. Tesla struggled to find the right parts to build its robots, and had to make certain components like its actuators, which power the bot’s movements, from scratch.
The big idea was to take Tesla’s learnings from its self-driving technology, which uses software and cameras to autonomously drive automobiles. Musk told colleagues that its cars were just robots on wheels.
The humanoids would need to learn how to move around indoor spaces and avoid safety risks like tripping and falling on top of a nearby human or pet. To solve this problem, Tesla hired human data collectors to wear cameras and backpacks, and walk around collecting training data. Tesla had people collecting data in several shifts, running 24/7.
Another solution was to collect data using Optimus itself. The company set up bots to circle the inside perimeter of its offices learning how to navigate indoors. Sometimes the bot would fall over, after which an engineer would wheel over a robot hoist and pick the bot back up.
In October 2024, at a Warner Bros. sound stage in Burbank, Calif., Musk demonstrated his vision for cities replete with autonomous vehicles and robots.
In a disco-ball-decorated rotunda, five Optimus bots performed a dance routine to Haddaway’s “What Is Love.” Elsewhere on the lot, the bots served drinks while outfitted in cowboy hats and bow ties.
Behind the scenes, Tesla engineers worked overtime to troubleshoot technical issues, according to people on the ground. While robots in the rotunda were programmed to dance, other robots at the event were teleoperated by engineers who wore body suits and virtual-reality headsets. They guided the bots’ interactions with guests, including while serving drinks behind the bar.
Each robot on the ground required constant monitoring from several engineers: one in a suit teleoperating its movements, one with a laptop, and others standing nearby to keep track of the bot’s physical performance.
Inside Tesla’s lab, Optimus proved pretty good at learning simple tasks. In May, the company shared a video that appeared to show Optimus performing various jobs in response to verbal orders from an engineer, such as putting trash in the bin, cleaning up crumbs, vacuuming, and moving a Model X part from a box. All of the activities were “learned direction from human videos,” according to the company.
Despite this progress, inside the company, some manufacturing engineers said they questioned whether Optimus would actually be useful in factories. While the bot proved capable at monotonous tasks like sorting objects, the former engineers said they thought most factory jobs are better off being done by robots with shapes designed for the specific task.
The big question, according to Goldberg, the Berkeley roboticist, is how to give robots the dexterity of humans and the ability to understand their environment well enough to complete useful but sensitive tasks, such as clearing a dinner table. “Even a child could clear a dinner table,” said Goldberg, who is also chief scientist at Ambi Robotics and Jacobi Robotics.
Some of Tesla’s competitors have concluded that legs are the problem. Evan Beard is chief executive of Standard Bots, which sells manufacturing robots on wheels. Beard said that wheels make the bots more stable, and therefore safer to work around, and easier to power down if something goes wrong.
“With a humanoid, if you cut the power, it’s inherently unstable so it can fall on someone,” Beard said. “For a factory, a warehouse or agriculture,” he said, “legs are often inferior to wheels.”
Tesla has backed away from its initial Optimus timeline of putting a commercial version to work into its own factories by the end of the year. The company is currently working on its third generation of the robot.
In Tesla marketing materials, Optimus has a role as a domestic worker watering plants, unpacking groceries and handling other household tasks, giving its owners time to hang out with their families.
“Who wouldn’t want their own personal C-3PO/R2-D2?” Musk said in November, referencing the droid characters in “Star Wars” movies. “This is why I say humanoid robots will be the biggest product ever. Because everyone is gonna want one, or more than one.”
What a Soros theory can tell us about the AI boom
So much of bubble activity is driven by feedback loops, dubbed reflexivity by the well-known investor
It is a mug’s game trying to predict the end of a boom with any precision. They last much longer than anyone might reasonably expect. That is true of bull markets, as well as economic advances. The reason is that markets and economies find ways to support themselves. George Soros, the well-known investor and philanthropist, has a term for it: reflexivity.
In a Financial Times article back in October 2009, Soros defined the concept, in terms of its impact on markets, quite succinctly. “The participants’ views influence the course of events, and the course of events influences the participants’ views,” he wrote.
It is a positive feedback loop. The same idea was at the heart of what John Maynard Keynes, the great economist, described as “animal spirits”; if businesses are confident, they will invest money and hire more workers, and this investment will boost economic growth.
In terms of asset markets, the most obvious example of reflexivity comes from the link between banking and property prices. Initially, for whatever reason, banks start lending more money to people who are buying property. The availability of additional finance pushes up demand for property — whether it is office blocks or homes — and property prices rise. This makes the bankers more confident about lending money in the property sector, as their collateral is rising in value. And it makes investors and or speculators more willing to borrow money to buy property, since it looks like a very good bet.
Debt does not have to be involved. For much of the life of cryptocurrencies, the price of digital assets such as bitcoin and ethereum has been sustained by the belief, among some investors, that they represent the wave of the future. Any weakness is thus a buying opportunity. And a rising price is a wonderful way of proselytising the crypto religion; more people are tempted to adopt the faith.
Another way in which booms can sustain themselves, in both economic and asset-market terms, is through spending on goods and services. That is clearly the case at the moment with the rush to invest in artificial intelligence.
This spending has done a lot to prop up US economic growth, at a time when job creation has stalled and consumer confidence has declined. In the first half of the year, JPMorgan estimated that AI spending contributed 1.1 percentage points to US GDP growth. In market terms, it plays a crucial role in convincing investors of the solidity of the AI boom, not least in the demand it creates for the chips made by Nvidia, the world’s most valuable company.
The buzz surrounding this spending also creates a kind of Fomo (fear of missing out) among other executives. If AI is the wave of the future, then any company that doesn’t embrace it risks being left behind. And, true to the principle of reflexivity, the race to invest makes the AI boom seem all the more substantial to investors. The obvious parallel is the late 1990s when spending on fibreoptic cable, routers and telecoms equipment soared, spurring the dotcom bubble.
The intoxicating nature of bullish sentiment indicates how these booms may eventually sow the seeds of their own destruction. In the late 1990s, it seemed that every twenty-something was either launching their own website or joining a start-up internet company with the hope of cashing in their share options. The appeal of the technology was so obvious that too many businesses were founded; only a fraction of them would ever be profitable. When it became clear, in the spring of 2000, that some businesses were running out of cash, sentiment changed.
The AI boom is different as it is focused on a few big players with strong existing business models, rather than on a host of start-ups. This means that the financial pressures are unlikely to bite as quickly.
On the other hand, AI might not be as immediately useful as many executives hope; a McKinsey study found that 80 per cent of companies that had started to use AI had yet to experience any boost to their profits. Plenty of consumers — particularly students — are enthusiastic users of AI to summarise reports and generate business proposals or essay plans. Useful stuff, but hardly the basis of a productivity miracle.
Of course, in the past, the impact of innovations such as electrification has taken decades to show up in the productivity numbers. By that stage, however, history suggests that a market boom, even if powered by reflexivity, will be long over. At some point, the growth rate in AI spending — and in Nvidia’s revenues — will slow; and then the rating that investors are willing to apply to corporate earnings will decline, along with share prices. The bandwagon will develop a wonky wheel.
Arguing that a boom must come to an end is not the same as saying the underlying technology is rubbish. AI will be useful, just as the internet is useful and the railways were very useful. That didn’t stop the other two booms from experiencing crashes. A reflex action may prolong a boom but it can also deliver a painful kick.
Meet the ‘Mad Max’-Loving CEO Challenging Nvidia With a Renegade Chip
June Paik spurned a takeover offer from Meta Platforms last year. Now his South Korean company, FuriosaAI, has an AI chip entering mass production.
SEOUL—The startup that is now one of a handful of chip makers nipping at the heels of Nvidia began in a hospital bed in Seoul a decade ago.
June Paik, a memory-chip engineer at Samsung Electronics, had torn his Achilles tendon while playing soccer at a company outing. Bedridden for months, he let his short hair grow out and passed the time watching online courses offered by Stanford University about the rising field of artificial intelligence.
He walked away with a healed ankle and a conviction: AI was set to become not just a novel technology but a new paradigm. Soon after, he left Samsung to start an AI company of his own.
“I left with absolute certainty that I had to get into the AI space,” said Paik, who kept the long hair.
He needed a core product and a team. At a computing conference in Seoul where AI was a main theme, Paik ran into old colleagues and asked a question: “Should we be getting into AI chips?” The response was an enthusiastic “yes.” In the following months, he convinced a former Samsung colleague and an old friend who was an algorithm expert to come aboard as co-founders. And in 2017, FuriosaAI was born.
Furiosa was named after the protagonist of the 2015 post-apocalyptic film, “Mad Max: Fury Road.” Paik saw parallels between the warrior’s against-the-odds journey back home and his goal of achieving startup success. The name came up once, off the cuff, and stuck. Furiosa’s AI chip is dubbed “RNGD”—short for renegade—and slated to start mass production this month.
Valued at nearly $700 million based on its most recent fundraising, Furiosa has attracted interest from big tech firms. Last year, Meta Platforms attempted to acquire it, though the startup declined the offer. OpenAI used a Furiosa chip for a recent demonstration in Seoul. LG’s AI research unit is testing the chip and said it offered “excellent real-world performance.” Furiosa said it is engaged in talks with potential customers.
Nvidia’s graphic processing units, or GPUs, dominated the initial push to train AI models. But companies like Furiosa are betting that for the next stage—referred to as “inference,” or using AI models after they’re trained—their specialty chips can be competitive.
Furiosa makes chips called neural processing units, or NPUs, which are a rising class of chips designed specifically to handle the type of computing calculations underpinning AI and use less energy than GPUs.
Paik said Furiosa’s chips can provide similar performance as Nvidia’s advanced GPUs with less electricity usage. That would drive down the total costs of deploying AI. The tech world, Paik says, shouldn’t be so reliant on one chip maker for AI computing.
“A market dominated by a single player—that’s not a healthy ecosystem, is it?” Paik said.
Paik grew up in Daegu, a city in the southeastern region of South Korea. In 1996, he entered South Korea’s prestigious Seoul National University to study electrical engineering—a topic that applied his favorite subjects, math and physics, to real-world settings.
At that time, his parents had moved to California so his father, a pastor, could do advanced studies in theology. They encouraged Paik to study in the U.S. too. After his initial hesitation, a summer spent taking computing architecture courses at University of California, Berkeley convinced him. Paik transferred to Georgia Tech, where he got a bachelor’s and a master’s degree in electrical engineering.
He started working at U.S. chip maker AMD, gaining experience in GPU design. In 2013, he moved to Samsung in South Korea to lead a small team to develop new memory-chip products.
Hanjoon Kim, who worked with Paik at Samsung and left to co-found Furiosa with him, described Paik as someone who can envision and execute product ideas on a huge scale.
“I found his approach quite striking,” said Kim, who is now Furiosa’s chief technology officer.
Paik, 48, travels everywhere with a keyboard-sized card equipped with a RNGD chip so he can show off Furiosa’s key product. A workout enthusiast, he runs and swims. On a Furiosa company outing to a coastal town, Paik and some colleagues held a race to see who could swim to a nearby island and return the fastest.
He met his wife, who also works in the AI business, through colleagues who heard she liked long-haired guys. The couple have a 3-year-old daughter.
South Korea, which boasts semiconductor know-how from homegrown firms like Samsung and SK Hynix, plus software expertise, is making a big push in AI. The government is pursuing AI development as a policy priority, hoping to become another leader in the technology alongside the U.S. and China. In recent months, OpenAI opened a new office in Seoul and Nvidia signed a major GPU-supply deal led by the South Korean government.
In Furiosa’s early years, Paik often referenced the Silicon Valley bible, “Blitzscaling: The Lightning-Fast Path to Building Massively Valuable Companies,” to emphasize the need for fast decision-making and risk-taking to achieve market dominance as a first mover.
“He’s always backed the idea that we must prioritize our long-term mission over immediate security,” Kim said.
When Furiosa’s 2017 seed investment round of just under $1 million ran out quickly, Paik took out loans. In 2019, Furiosa didn’t pay the salaries of its senior executives for several months as the firm sought to avoid lowering its valuation while it worked to close out its next round of funding.
Paik recruited globally for talent. He flew from Seoul to Princeton, N.J., to meet and convince an engineer to join Furiosa in its early days, said Jae W. Lee, director of Seoul National University’s AI Institute and a mentor to Paik. Lee recalls running into Paik at virtually every major hardware and software conference to recruit talent. Furiosa now has about 200 employees.
“He had incredible energy. I just knew he was going to make waves one day,” said Lee of his first encounter with Paik at an academic conference in 2015.
In 2024, at Stanford’s prestigious Hot Chips conference, Paik debuted Furiosa’s RNGD chip as a solution for what he called “sustainable AI computing” in a keynote speech. Paik presented data showing how the chip could run the then-latest version of Meta’s Llama large language model with more than twice the power efficiency of Nvidia’s high-end chips.
Furiosa’s booth was swarmed with engineers from big tech firms, including Google, Meta and Amazon.com, wanting to see a live demo of the chip.
“It was a moment where we felt we could really move forward with our chip with confidence,” Paik said.
Looking back, Paik sees his Achilles injury as a turning point. Even the grueling rehab, where doctors once told him he might not fully recover, helped him weather the tough times at Furiosa.
“I think it could have been a blessing in disguise,” he said.
Venezuela Hit by Explosions It Says Were Act of Aggression by U.S.
Residents reported power outages in several areas of Caracas.
Venezuela’s capital was rocked by a series of explosions as planes and helicopters flew over the city, with the government saying they were an act of aggression by the U.S.
Videos recorded by the residents showed at least six enormous blasts with plumes of black smoke rising over the explosions around Caracas, shortly before 2 a.m. local time. Power outages were reported.
U.S. Southern Command and the Department of Defense referred requests for comment to the White House, which didn't immediately respond.
- Venezuela’s government said that, in addition to Caracas, the coastal states of Miranda, Aragua and La Guaira were hit.
- Venezuela ordered a general mobilization of all the country’s social and political forces to counter what it said was an imperialist attack.
- The explosions hit the city after President Trump repeatedly threatened land strikes in Venezuela, following months of an unprecedented campaign of U.S. airstrikes on alleged drug-trafficking vessels in the Caribbean.
2026 AI Strategy: Reflexivity, Hyperscaler Debt Limits, and the Negative Alpha Basket
Drawing on the attached FT piece, I’ve mapped the 2026 AI landscape through the lens of reflexivity to identify exactly where the current feedback loop is most likely to break.
While the "virtuous cycle" of AI spending is still propping up GDP, my analysis pinpoints the emerging "wonky wheel"—specifically the disconnect between record hyperscaler CapEx and the stagnant 20% enterprise ROI. I’ve outlined a targeted monitoring checklist and a "Negative Reflexivity" short basket of high-leverage plays most vulnerable to a sentiment reversal.
### Key 2026 Monitoring Checklist
- Credit Stress: Track the spike in CDS spreads for MSFT and Google as they tap debt markets to fund $500B+ in annual build-outs.
- The Utilization Gap: Monitor if SaaS AI Upsells (NDR) stall in H1 2026; if enterprises don't see productivity gains, the hardware orders will be the first to be cut.
- GPU Resale Value: A secondary market dip in Blackwell chip prices below MSRP will be our primary signal that the supply-demand loop has flipped.
### Targeted "Negative Reflexivity" Short Basket
| Ticker | Thesis | Risk Trigger |
| ORCL | Negative FCF; $38B+ CapEx; Credit rating on "Negative Outlook." | Revenue miss from Tier-1 AI labs (e.g., OpenAI). |
| CRWV | High-leverage specialist; 2% operating margins; GPU-collateralized debt. | 20% drop in secondary GPU market prices. |
| SMCI | Cyclical hardware play with thin margins. | "Inventory Overhang" leading to order cancellations. |
| CRM/ADBE | Trading at AI premiums with slowing Net Dollar Retention. | Customer churn on $20-$30/month AI add-ons. |
Full analyse below :
## Executive Summary
The article argues that market booms are self-sustaining feedback loops. Using George Soros’s theory of reflexivity, it explains that investor confidence leads to increased spending and lending, which in turn justifies that confidence, creating a "positive feedback loop." While this can propel markets to extreme heights (as seen in the dotcom bubble or the current AI boom), these cycles eventually sow the seeds of their own destruction when the underlying reality (productivity and profits) fails to keep pace with the hype.
### The Main Idea: Reflexivity
The core concept is that markets are not passive observers of reality; they help create it. * The Loop: Investors believe in a "wave of the future" --> They pour capital into that sector --> This spending boosts GDP and corporate revenues (e.g., Nvidia’s sales)--> The rising prices "prove" the investors were right, attracting more capital.
- The AI Context: AI is currently propping up US GDP growth (contributing ~1.1 percentage points) and creating a FOMO-driven race among executives, even though 80% of companies have yet to see a profit boost from the technology.
## Strategic Conclusions
If you are managing a portfolio in this environment, the article suggests a strategy of calculated participation combined with rigorous exit-monitoring:
1. Respect the "Mug’s Game" of Timing
- Action: Avoid shorting the boom prematurely based on "valuation logic." Reflexivity means markets can stay irrational longer than you can stay solvent. The feedback loop creates its own temporary fundamental support (like the boost to GDP).
2. Monitor the "CapEx to Profit" Disconnect
- Action: Track the delta between AI infrastructure spending (Nvidia, cloud providers) and AI end-user profitability (McKinsey’s 80% stat). The "wonky wheel" appears when the buyers of the technology stop seeing a return on investment and scale back their spending.
3. Distinguish Between Technology and the Asset Bubble
- Action: Maintain a long-term bullish view on the technology’s utility (like the internet or railways) but a short-to-medium-term skepticism regarding the stock prices. The crash usually happens long before the actual "productivity miracle" (like electrification) shows up in the data.
4. Watch the "Big Players" as Sentiment Indicators
- Action: Since this boom is concentrated in high-cash-flow giants rather than fragile start-ups, look for cracks in the earnings guidance of the "Enablers" (chip makers/hyperscalers) as the primary signal for the end of the reflexivity cycle.
Key Takeaway: Ride the momentum of the feedback loop, but treat the current GDP and revenue growth as a byproduct of sentiment, not a permanent structural shift.As we start 2026, there is no better "case study" for the reflexivity theory than Nvidia (NVDA). Based on their most recent financial results (Q3 FY2026) and the current market setup, here is an analysis of how the "virtuous cycle" is playing out and where a hedge fund manager should look for the "wonky wheel."## Nvidia Analysis: The Reflexivity Loop in Action### 1. The "Positive Feedback" Phase (Current State)Nvidia’s Q3 FY2026 results (reported late 2025) perfectly illustrate Soros’s concept.
Strategy Actionable Intelligence Ride the Momentum Don't short yet. 2026 CapEx is already "locked in" by hyperscaler budgets. The reflexivity loop is still providing structural support to GDP (contributing over 1% to US growth). Watch the "End-User" ROI The article notes that 80% of companies haven't seen profit boosts. Monitor the Q1/Q2 2026 earnings of software companies (Salesforce, Adobe, MSFT Office). If they can't monetize AI, their hardware orders will eventually dry up. Monitor Credit Spreads The "kick" in a reflex action often comes from the credit market. If Microsoft or Meta’s borrowing costs continue to rise, they will be forced to prioritize buybacks/dividends over "speculative" AI CapEx. Exit Trigger The "wonky wheel" will likely be a downward revision in margin. Nvidia is guiding for 75% gross margins. If they drop to 70% due to "inventory overbuilding" (a classic end-of-boom sign), the reflexive belief in "infinite growth" will shatter.
Final Insight: In a reflexive market, the "fundamentals" (Nvidia's revenue) are a lagging indicator of "sentiment" (the belief in AI). By the time Nvidia's revenue actually misses, the stock will likely have already corrected 20-30%. Your job is to watch the buyers' balance sheets, not the seller's earnings.As of early January 2026, the AI boom has transitioned from a software "hype" cycle into a massive industrial utility build-out. The "Big Three" buyers (Microsoft, Meta, and Google) are now spending at a scale that exceeds the GDP of many mid-sized nations.For a hedge fund manager, the critical shift in 2026 is that balance sheet strength has replaced "GPU count" as the most important metric. Here is the breakdown of their 2026 CapEx plans and the potential debt "tipping points."## 1. Microsoft: The "Compute Sovereign"Microsoft has effectively transformed from a software company into a global compute utility.
Risk Factor 2026 Warning Signal Credit Rating Pressure S&P Global has already put Oracle on a "negative outlook." If Microsoft or Google see a rating downgrade, the entire AI sector will de-rate. The "Inventory" Overhang Hyperscalers are currently "stockpiling" Blackwell chips. If utilization rates (the actual usage of these chips) drop below 60%, a massive write-down cycle begins in late 2026. Circular Financing Much of the revenue growth is "incestuous" (Big Tech $\rightarrow$ AI Startups $\rightarrow$ Big Tech Cloud). A credit event at a major startup like OpenAI or Anthropic could trigger a systemic "reverse-reflexivity" loop.
### Summary for Your PortfolioThe boom is currently supported by the fact that these companies can afford the debt. However, we are moving from the "Expansion" phase to the "Execution" phase.Watchlist of Leading IndicatorsReflexivity Failure Radar: a watchlist of leading indicators designed for a hedge fund manager to monitor the transition from "Virtuous Cycle" to "Painful Kick."## 1. The Physical Bottleneck: "Power-to-Chip" RatioIn a reflexive boom, money is infinite, but physics is not. If hyperscalers (Google, MSFT) buy chips they cannot plug in, their ROI collapses.
Indicator Healthy (Keep Long) Warning (Trim Position) Failure (Short/Exit) GPU Resale Value > MSRP At MSRP < 80% of MSRP Power Grid Wait < 24 Months 24 - 48 Months > 48 Months SaaS AI Upsell Growing > 20% Growth Stall High Churn / Cancellations Big Tech Debt Debt for Buybacks Debt for CapEx Rating Downgrade (e.g. AA to A)
The following "Negative Reflexivity" Short Basket targets the companies most vulnerable to a credit squeeze, secondary market GPU gluts, and the "Utilization Gap."## The "Negative Reflexivity" Short Basket (2026 Edition)### Category 1: The Leveraged Infrastructure (The "Debt-Trap" Plays)These companies have borrowed heavily to build data centers, betting that AI demand will be permanent and price-insensitive.
Phase The Vicious Cycle Mechanism 1. The Spark A major AI lab misses a milestone or a "Big Three" buyer (like Meta) cuts 2027 CapEx guidance. 2. The Glut Cancelled orders lead to an "overhang" of chips. Secondary market prices for GPUs drop below MSRP. 3. The Credit Squeeze Lenders (Private Credit) realize their "collateral" (GPUs) is worth 30% less. They hike interest rates on AI firms. 4. The GDP Drag The 1.1% GDP boost from AI spending vanishes, leading to a broader US economic slowdown. ### Portfolio Manager's "Exit Trigger"Keep a close eye on Nvidia’s Inventory levels. Historically, when Nvidia's "Days of Inventory" starts to rise while their revenue growth slows, the reflexive cycle has peaked. In late 2025, this metric began to creep up for the first time in three years.
European football clubs lose out amid soaring US valuations in deal frenzy
US sports team valuations are dwarfing those of European football clubs as investors are drawn to closed league model
European football clubs have been left on the sidelines of a deals boom that has highlighted soaring valuations for US sports franchises and underlined the challenges facing Europe’s team owners.
Investors argue that a failure to get a grip on costs, as well as the constant threat of relegation, has kept a lid on European interest even as a flurry of deals in the US has underscored rising valuations in several sports.
Private equity group Sixth Street and US financier Dean Metropoulos bought an 8 per cent stake in the New England Patriots in September. The deal put a value of more than $9bn on the National Football League team, equivalent to more than 11 times revenues.
A few weeks later, Guggenheim Partners chief executive Mark Walter completed his purchase of a majority stake in the Los Angeles Lakers, valuing the basketball team at $10bn, equivalent to about 18 times revenue.
According to estimates from Sportico, the average enterprise value of a team in the National Basketball Association has risen to 14.1 times revenue, up from 11.8 times in 2023.
Average valuations of NFL teams, which have been boosted by the league’s decision in 2024 to allow private equity firms to buy minority stakes, have increased to 10.2 times over the same period, according to Sportico.
Of the 10 most valuable sports teams in the world, six compete in the NFL, three in the NBA and one — the New York Yankees — plays Major League Baseball. Sportico estimates that Real Madrid is the world’s most valuable football club, yet it only ranks in 22nd place overall.
“You have this very predictable, durable, sustainable growth [in the US],” said Doc O’Connor, co-founder of sports-focused private equity firm Arctos Partners, at a recent industry event in London. “The environment in European football is very, very different from North American leagues.”
Valuations of the top men’s football teams, which are concentrated in Europe, have stagnated at just 4.2 times revenue. M&A activity in European football has dropped sharply since a spate of record-breaking takeovers in 2022, according to figures from governing body Uefa.
Apollo Global Management agreed to buy a controlling stake in Atlético Madrid, Spain’s third-biggest football club, at a valuation of between €2bn and €2.5bn in 2025. The lower end of that range implies a valuation of 4.9 times its 2024 revenue.
“The level of regulation at a league level [in football] is far reduced,” O’Connor said. “You have issues like promotion and relegation, you have no effective real debt limits and no effective limits on spending [on player wages and transfer fees]. That accounts for the difference in valuations.”
While football’s financial rules are being tightened both at European and national league level, several investors and team owners have called for US-style salary caps to halt spiralling spending on players by clubs.
According to the most recent figures from Uefa, more than half of Europe’s top-flight football teams reported operating losses in 2024, with an aggregate loss of €300mn. Transfer costs pushed overall pre-tax losses up to €1.2bn, while combined debt rose 10 per cent to €28.1bn.
Other experts also cite the divergence in the media rights markets in the US and Europe as an important reason for the difference in desirability to investors.
US sports leagues have successfully negotiated steep increases in their media rights. The NBA’s latest set of deals, covering 11 years, resulted in income rising from $2.6bn a year to $6.9bn, driven by strong competition from traditional cable networks and streamers.
Meanwhile media rights for major European football leagues are either falling or showing signs of stagnation.
Even the English Premier League, the most commercially successful national competition, only managed to secure a 4 per cent increase in its UK media rights, compared with eight years ago, when the rights were last renegotiated.
As well as the closed format of the competitions US sports teams share most central revenues equally, including income from TV deals, increasing their attractiveness to investors.
“The league organisation, structure and revenue-sharing opportunities in the US vs anywhere else in the world are just very different,” said John Lambros, head of digital media and entertainment at Houlihan Lokey.
“There’s lower risk and volatility in most US sports . . . no relegation risk and predictable media rights make revenues stable and foreseeable,” he added.
Football fans in Europe, however, are resistant to the closed-league model. The widespread anger at plans for a breakaway European Super League in 2021 was partly because the competition was designed so that founding member clubs could not be relegated.