TheChek News Flow

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 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

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

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

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

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 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

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 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

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

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, 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

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

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 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