Good morning. In honor of the poetic release of a major new AI model—read on for more—a haiku about the album Is This What We Want?, a new recording protesting U.K. government plans to let AI companies train on copyright-protected work without permission:
One thousand writers
Twelve tracks of silent protest
Can’t just “Let It Be”
Today’s news below. —Andrew Nusca
Want to send thoughts or suggestions to Data Sheet? Drop a line here.
Nvidia prepares to face off with investors demanding perfection
For years, Nvidia has accustomed investors to a world in which it outperforms their expectations every quarter. Anything less and the stock takes a beating.
Nvidia reports its fourth quarter financial results on Wednesday. That means Nvidia can’t just meet the $37 billion in revenue it forecasted, it has to exceed it. As of now, that seems to be on tap, with Wall Street expecting roughly $38 billion.
Nvidia has to do its “normal exceed” or else it will get “crapped on,” Baird managing director Ted Mortonson told Fortune.
Investors will be focused on what Nvidia executives say about how the Chinese AI upstart DeepSeek might impact spending in the U.S. tech sector. Amazon, Meta, Microsoft, and Alphabet still expect to allocate a projected $325 billion to capital expenditures this year, most of which will go toward AI research and development.
Among tech bulls, the consensus is that DeepSeek’s success only proves the pace of innovation, and therefore development, is accelerating, and that demand for Nvidia chips will only increase.
Investors will also look for any guidance on the margins for Nvidia’s new Blackwell chip.
Blackwell is projected to lower Nvidia’s margins, which the company told Wall Street to expect this quarter. However, if Nvidia’s margins fall further than anticipated, it could be a rare piece of bad news from a company that over-delivers.
Said Baird’s Mortonson: “There’s concern on gross margin and mix—that’s the bottom line.” —Paolo Confino and Greg McKenna
Microsoft ditches some data centers
Microsoft has cancelled some of its data center leases in the U.S. and pulled back on signing certain new leases, according to analysts at TD Cowen who say the cancellations amount to “a couple of hundred” megawatts worth of capacity, which is quite a lot.
Is this a sign of the long-awaited deflation of the AI infrastructure bubble? Maybe—we have Nvidia’s earnings report to look forward to, which might shed more light on that matter.
There’s certainly a lot of context into which we can place this latest nugget of Microsoft data center news.
A lot of Microsoft’s new capacity is designed to support OpenAI, which is the company’s big AI partner. But OpenAI is clearly looking elsewhere for its capacity these days. The big Stargate project will use Oracle’s data centers (albeit running Microsoft’s Azure platform) and The Information reported Friday that OpenAI expects to get three-quarters of its capacity from Stargate by 2030.
Still, Microsoft said at the start of the year that it would spend $80 billion on building out AI data centers this year. On Monday it said this plan remains on track—but it refused to comment on the TD Cowen note.
So let’s see what shakes out on Wednesday. This could be an industrywide thing, or just a reflection of Microsoft’s changing relationship with OpenAI. —David Meyer
Anthropic releases Claude 3.7 Sonnet
Anthropic has taken the wraps off its latest model, dubbed Claude 3.7 Sonnet.
The San Francisco AI startup led by siblings Dario and Daniela Amodei calls it a “hybrid AI reasoning model” because it can spit out answers in real time as well as think through them, so to speak, if given a longer duration.
Anthropic calls it “extended thinking” and says it’s best for complex queries.
“Extended thinking mode isn’t an option that switches to a different model with a separate strategy,” the company wrote in a blog post. “Instead, it’s allowing the very same model to give itself more time, and expend more effort, in coming to an answer.”
“Reasoning” is AI-speak for a deduction-like process that reduces complex queries into smaller, simpler steps. It is therefore not quite the same as human reasoning, a.k.a. critical thinking (which involves questioning or challenging the information itself).
The latest Claude model is Anthropic’s attempt to simplify and combine what are typically two or more products from its competitors. As such, it’s available to all users and developers—but free users get the non-reasoning version, and premium users get the full experience.
Claude 3.7 Sonnet’s reasoning abilities compete with DeepSeek’s R1, Google’s Gemini 2.0 Flash Thinking, xAI’s Grok 3, and OpenAI’s o3-mini.
Anthropic isn’t alone in pursuing a simpler product strategy. OpenAI CEO Sam Altman said earlier this month that his company would scuttle its o3 model release in favor of a “simplified” model called GPT-5 that integrates capabilities and allows paying subscribers to run the model at ever-higher levels of intelligence. —AN
More data
—Apple bets big on U.S., facing Trump tariffs: $500 billion in its home country over next four years, including AI server production in Texas.
—DOGE will use AI to evaluate federal workers. Summarize the week’s accomplishments and the machine will determine if it’s critical or not.
—DoorDash pays $16.75m in NY settlement. The delivery service was accused of pocketing customers' tips for delivery drivers between May 2017 and September 2019.
—Epic Games v. Apple continues. App Store boss Phil Schiller initially opposed fees for in-app transactions using outside payment systems, fearing antitrust violations.
—SpaceX sends Starlink terminals to FAA. Verizon won a contract to upgrade infrastructure in 2023; apparent conflict of interest for Elon Musk.
—SEC closes Robinhood Crypto investigation. No enforcement action under Trump administration.
—Zoom gives slightly soft sales outlook. Projected Q1 revenue is between $1.162 billion and $1.167 billion; Wall Street is looking for $1.18 billion.
—Chegg sues Google, claiming AI Overviews hurt its revenue. Shares plunged 26%.
—Stripe founders push back on RTO. Remote work solves the “two-body problem” faced by working couples, argue the brothers Collison.
Endstop triggered
