DELL PRINTS $16 BILLION IN AI SERVERS IN A SINGLE QUARTER AND MAKES IT LOOK EASY
Dell came out last week with the kind of earnings report that makes people in the investment world quietly double-check the numbers and then triple-check them. The company posted record revenue of $43.8 billion for the quarter, up 88% from a year ago. That is already a story. But the AI server number is something else entirely.
Dell sold $16.1 billion worth of AI servers in a single quarter. A year ago, that same number was a fraction of that. We are talking 757% growth year over year. To give you a sense of scale, that is more AI server revenue in one quarter than most tech companies generate in total revenue in a full year. The company also booked $24.4 billion in AI server orders during the quarter, suggesting the next few quarters are going to look just as wild. Dell then went ahead and raised its full-year AI server revenue guidance to $60 billion, which would make it one of the biggest AI beneficiaries on the planet even though Dell does not actually build the chips inside these things.
The stock had its best single day in company history, up 32%. Earnings per share came in at $5.24, against analyst expectations of $2.94. In other words, Dell beat earnings so badly it almost seems like someone forgot to write down the actual number when they made the estimate.
Here is the thing that makes this interesting beyond the raw numbers. Dell is not an AI company in the way OpenAI or Anthropic are AI companies. Dell is a hardware assembler. It takes Nvidia chips, sticks them into racks, runs some cables, and sells the whole package to hyperscalers and enterprises building out AI infrastructure. The fact that Dell is printing money this fast tells you something important: the AI infrastructure buildout is not slowing down. Not even a little. Every major company on earth is still racing to buy the compute capacity they think they will need over the next five years, and many of them are running out of patience to do it before prices go even higher. Dell is the beneficiary of that institutional panic, and right now it is laughing all the way to the bank. The quarter also proves something broader: you do not need to build the best AI model in the world to make a fortune from AI. You just need to be in the right place when everyone else needs the hardware.
MASAYOSHI SON BETS $87 BILLION ON FRANCE AS SOFTBANK TURNS EUROPE INTO ITS NEXT AI FRONTIER
Masayoshi Son has a new address for his next enormous bet, and it is in the north of France. SoftBank Group announced a plan to invest as much as 75 billion euros, roughly $87 billion at current exchange rates, in AI data center infrastructure across France. The first phase alone is 45 billion euros to deliver 3.1 gigawatts of AI compute capacity in the Hauts-de-France region by 2031, with the full buildout eventually reaching 5 gigawatts.
To put 5 gigawatts into perspective: that is roughly the total electricity output of five large nuclear power plants, poured entirely into AI compute. It is a number that sounds almost absurd when you say it out loud, but Son has been making absurd sound reasonable since the late 1990s. He lost billions on WeWork, nearly took down his entire empire, and then came back by being very right about the AI infrastructure trade at a time when most institutional investors were still unsure whether to take it seriously.
The choice of France is not random. President Macron has made it a personal crusade to position France as Europe’s AI hub, and he has been more aggressive than any other European leader about offering favorable conditions for tech investment. France has also historically been less hostile to large American tech companies than Germany or Brussels-adjacent regulators, which lowers the political risk for a deal this size. SoftBank gets below-market land costs and regulatory support. France gets tens of thousands of construction and technology jobs, substantial tax revenue, and the ability to say it is home to the largest AI data center buildout on the continent.
The timing matters because Europe has been watching the US and China race ahead on AI infrastructure with a combination of anxiety and envy. Most European nations have fallen badly behind, constrained by energy costs, regulation, and a general reluctance to build at the scale American hyperscalers have normalized. An $87 billion injection changes that picture materially. If this goes through on the stated timeline, France will have more AI compute capacity by 2031 than most European countries combined do today. Whether the rest of Europe follows or watches from the sidelines will be one of the defining technology-policy stories of the next few years.
Read the full story at Bloomberg
MICROSOFT’S AI CHIEF TELLS WHITE COLLAR WORKERS THEY HAVE 18 MONTHS BEFORE THE MACHINES TAKE OVER
Mustafa Suleyman, CEO of Microsoft AI, said something recently that the financial press mostly treated as a tech story but is really more of a social alarm bell. In a conversation with the Financial Times, Suleyman predicted that within 18 months, AI will reach human-level performance on most professional tasks. Not some tasks. Most tasks. The specific jobs he named include accounting, legal work, marketing, and project management. If you are reading this at a desk doing any of those things, he is more or less talking about your career.
What makes Suleyman’s comments worth engaging with, even if you find the timeline aggressive, is that he is not a random optimist running a startup on vision-fund money. He is the person running AI at Microsoft, which has put more money into OpenAI than any other investor on earth and is betting its entire enterprise software future on AI winning. When he says “human-level performance on most professional tasks” in 18 months, he is not guessing. He is describing the trajectory he sees from the inside of one of the best-resourced AI programs in existence.
The honest counterpoint is that his own predictions have not always landed cleanly. There is a reasonably good study from METR, a nonprofit, finding that AI actually made software developers’ tasks take 20% longer when applied to complex real-world coding work. And profit margin data from Apollo Global Management shows that while Big Tech profits soared over 20% last quarter, the broader economy saw almost no change. AI has not, yet, produced the sweeping productivity gains that would signal the jobs argument is winning in reality rather than just in presentations.
So here is where things stand. The technology is genuinely accelerating. The business case for replacing professional labor with AI is real and getting more compelling each quarter. But there is a gap between what AI can do in a controlled demo and what it reliably does at scale in messy real-world workflows. Suleyman’s 18-month window either reflects genuine insider knowledge about how fast things are about to move, or it is an aggressive prediction from someone whose professional incentives are to make AI sound more transformative than it currently is. The answer probably lies somewhere in the middle, which is still a reasonably uncomfortable place for accountants and marketing managers to be sitting right now.
Read the full story at Fortune
OPENAI LAUNCHES BIODEFENSE PROGRAM AND NAMES IT AFTER THE SCIENTIST WHOSE DISCOVERY WAS STOLEN
OpenAI just announced something that does not fit into its usual product calendar of model releases and API updates. The company unveiled the Rosalind Biodefense Program, named after Rosalind Franklin, the crystallographer whose X-ray diffraction work was essential to discovering DNA’s double helix structure and who was famously denied credit for it during her lifetime. The program will make OpenAI’s GPT-Rosalind model available to “trusted developers” building tools for biodefense, pandemic preparedness, and public health infrastructure.
The specifics include support for epidemiological modeling, early detection of biological threats, screening tools, non-pharmaceutical intervention planning, and development of medical countermeasures. OpenAI said it has already briefed the White House and several federal agencies, and it is actively working toward giving US government and allied partners access for public health and defense missions.
The timing of this announcement is not accidental. OpenAI and other major AI labs have faced growing and serious questions about the dual-use risk of powerful AI models in life sciences. The same capabilities that help researchers model how a pathogen spreads could theoretically help a bad actor figure out how to make one spread more effectively. Congress has held hearings about this. The national security community thinks about it constantly. By launching an explicitly defensive biodefense program, OpenAI is making a public statement: we know this risk exists, we are on the right side of it, and here is the proof.
It also fits neatly into OpenAI’s broader government and institutional strategy. The company launched a $4 billion consulting arm last week focused on enterprise and government clients. GPT-Rosalind appears to be part of the same push into institutional markets where security clearances, federal contracts, and long-term procurement relationships create the kind of sticky, recurring revenue that makes investors happy. AI for national security is a massive addressable market, and every major lab is currently competing hard to own a piece of it. OpenAI is clearly not planning to leave that territory to Anthropic or Google alone. As for the naming choice, honoring a woman whose scientific contribution was overlooked for decades while two men got the Nobel Prize for it, someone at OpenAI either has a sense of history or a very good communications consultant. Possibly both.
GROQ TAKES NVIDIA’S $20 BILLION, KEEPS THE COMPANY, AND NOW RAISES $650 MILLION TO PROVE IT WAS THE RIGHT CALL
Here is a story that reads like a Silicon Valley redemption arc playing out in real time. Groq, the AI chip startup best known for making inference unusually fast on its custom hardware, is raising $650 million from its existing investors. This comes a few months after what might be the strangest transaction in recent tech history: Nvidia paid around $20 billion for what was officially called a “not-an-acquisition,” taking Groq’s hardware technology and absorbing several senior employees into the chip giant, while leaving the rest of the company still standing as an independent entity. Groq’s investors got paid out in cash. Everyone walked away, technically happy.
Except Groq did not actually disappear. What remained pivoted toward building an inference cloud business, offering developers and enterprises a way to run their AI models at the high speeds Groq’s custom chips deliver. Inference, which is the processing that happens every time you send a prompt to an AI model and get an answer back, has become one of the most competitive and expensive parts of the AI stack. Training a model once is a capital-intensive one-time cost. Serving it to millions of users every day is where the real ongoing operational expense lives, and that expense compounds fast at scale.
The company is currently being led by interim CEO Adam Winter and CFO Matt Eng following the leadership transition that came packaged with the Nvidia deal. The $650 million round is structured as an internal raise from existing investors, and it is effectively guaranteed: backers Disruptive and Infinitium have agreed to cover any portion that other investors choose not to take, which means this round is less about whether Groq can raise money and more about giving a clear signal that the inference cloud direction has backing behind it.
The real question is whether Groq can carve out a durable position in an inference market that now includes Nvidia itself, Amazon Web Services, Google Cloud, Microsoft Azure, and a dozen well-capitalized startups. The competition is genuinely severe. But Groq’s chips do have a real and measurable speed advantage for specific inference workloads, particularly tasks where latency matters more than raw throughput. If the company can identify the right customers for that advantage before the $650 million runs out, there is a legitimate business here. The deal gives them enough runway to find out, which at this point in the AI race is about as good as anyone can ask for.