NVIDIA BORROWS $25 BILLION — AND WALL STREET TRIED TO HAND THEM THREE TIMES THAT AMOUNT
Source: Bloomberg | Also: CNBC
So Nvidia, a company that is currently printing money so fast the machines run hot, decided it needed even more money. On June 15th, Nvidia walked into the bond market asking for around $20 billion and walked out having raised $25 billion — after investors tried to hand them $85 billion. Eighty-five billion dollars in orders. For bonds. From a company that already has more cash than most country treasuries.
Here is what happened. Nvidia has not issued bonds since 2021, back when the AI boom was still being dismissed as hype in polite company. Since then, Jensen Huang became the most important person in global technology, the H100 became the most valuable physical object in corporate America, and demand for Nvidia chips outpaced what any government or company could actually secure. So why borrow money now? General corporate purposes, the filing says. Repay some existing debt. Keep the war chest full. Stay flexible. That is the official answer. The real story is what happened when they showed up to ask.
The deal was initially sized at around $20 billion. Demand came in so hard that they upsized it to $25 billion. They had more than three times as much investor interest as product available. The notes run from 2028 all the way out to 2056, which means some investor out there is betting Nvidia will still dominate computing thirty years from now. That is not a quarterly trade. That is a thesis about the entire future of computing infrastructure.
Think about what this says about where we are in the AI cycle. Nvidia does not need to borrow money in any conventional sense. They are one of the most profitable companies on Earth. Revenue is extraordinary, margins are spectacular, the stock has been on a multi-year run that made early investors look like prophets. And yet the bond market is so desperate for exposure to this company that investors offered three times the available supply.
That is not bubble mentality. That is war economy mentality. Investors are looking at the AI infrastructure buildout — data centers, chips, power, training clusters — and deciding they want to be attached to the company that controls the hardware for all of it for as long as possible. Maturity dates out to 2056 say it plainly: some of these buyers are making a generational bet. What does Nvidia do with $25 billion in fresh capital? Probably nothing dramatic immediately. You raise it because the terms are favorable, the demand is real, and options are better than no options. What the market has communicated, very loudly, is that the AI infrastructure investment cycle is nowhere near its end and Nvidia sits at its center. You do not generate $85 billion in orders for a company you think is peaking.
GOOGLE DESIGNED AN AI YOU CAN LITERALLY UNPLUG — AND THE GOVERNMENT GETS TO HOLD THE CORD
Here is something that would have sounded like a bad pitch for a spy movie eighteen months ago. Google’s most advanced AI model, Gemini, can now be shipped to you in a single physical server, plugged into your building, and run completely cut off from the internet. No Google cloud. No remote monitoring. No connection to anything outside the room it is sitting in. Just a box, your data, and AI humming away in complete isolation.
This product exists through a partnership with a company called Cirrascale Cloud Services and Google’s Distributed Cloud program. The hardware is a Dell-built rack server with eight Nvidia GPUs inside, wrapped in confidential computing protections. You arrange delivery. You plug it in. You use Gemini. Nobody at Google knows anything about what you asked or what it told you. That is entirely the point.
The update mechanism is almost cinematic. When Google releases a new version of Gemini, your server reconnects briefly to a private Google channel, downloads the updated model, and goes dark again. For customers who operate at classification levels where even that brief connection is impossible — intelligence agencies, defense contractors whose facilities have never touched the public internet — Cirrascale will physically swap the machine. Someone shows up at your facility. They unplug the server. They purge it completely. They leave with the old hardware. A new server arrives with the updated model. No data survives the transition. This is a feature they are proud of, not a limitation they are apologizing for.
The customer this was built for is obvious: government agencies and defense contractors working with classified data. The kind of organizations that have strict rules about what can and cannot touch an external network. Intelligence agencies. Defense laboratories. Contractors on programs where even the question being asked is potentially sensitive. These buyers have wanted AI for years and have been locked out because every AI product required sending data through someone else’s infrastructure. Google just unlocked that market.
The broader enterprise significance is real too. Medical institutions with strict patient data requirements. Law firms with attorney-client privilege concerns. Financial firms where regulators would have strong opinions about sensitive strategies running through a third-party cloud. Banks with data residency requirements. For all of these buyers, cloud-first AI has not been a preference but a legal wall. Air-gapped Gemini goes over that wall entirely.
What Google is saying to every nervous compliance officer is simple: the AI runs on hardware you control, the data never leaves your facility, and you can wipe it on demand. That removes the most common objection slowing enterprise AI adoption for two years. Enterprise AI spending has been slower than analysts predicted — this product is a direct response to that exact problem. If it delivers what it promises, Google just opened a market that has been effectively closed for the entire commercial history of AI.
OPENAI WANTS TO CERTIFY 10 MILLION AMERICANS — TO USE THE TOOL THAT MAY TAKE THEIR JOBS
You have to admire the sheer audacity of this one. OpenAI — the same company whose CEO has gone on record warning that AI will eliminate half of all entry-level white-collar jobs within five years — just launched a free workforce training program to help people learn how to use AI at work. The stated goal is to certify 10 million Americans by 2030. Launch partners include Walmart, BCG, and Accenture.
Walmart. The largest private employer in the United States, with roughly 1.6 million Americans on payroll doing jobs that are increasingly automatable. OpenAI is partnering with Walmart to train Walmart workers in AI skills. Whether that is a genuinely noble workforce development initiative or the most politely packaged layoff preparation program we have ever seen is a question worth asking out loud.
The courses are called AI Foundations, Applied AI Foundations, and Agents and Workflows. They take someone from zero AI experience to building and directing automated workflows. The underlying philosophy, which OpenAI states directly, is that people learn AI best by practicing on work that actually matters to them. Start with your real job. Apply the tools to it. See what happens. The courses are free and live now on OpenAI Academy.
The partner list tells you something. BCG is a management consulting firm whose business is advising companies on how to operate more efficiently, which often means leaner. Accenture helps organizations outsource and automate operations. BBVA is a Spanish bank among the most aggressive AI adopters in European financial services. None of these are organizations known for resisting automation. They are all, in various ways, in the business of helping companies do more with fewer people.
Both cynical and optimistic readings of this announcement are available simultaneously, and both are probably correct.
The cynical read: OpenAI knows the political conversation about AI and jobs is getting louder. Offering to train workers is a way to stay on the right side of that conversation. It costs relatively little, generates goodwill, and lets the company say it is doing something constructive about the problem its products are contributing to. It is PR. It is also fairly effective PR.
The optimistic read: AI skills are real, learnable, and valuable. People who know how to use these tools effectively are more productive and more employable. If OpenAI actually delivers on certifying 10 million Americans in four years using real distribution through real corporate partners, that is meaningful. It does not require believing OpenAI’s motives are purely altruistic to recognize that the courses themselves might genuinely help people adapt to what is coming. The 10 million number in context: the U.S. labor force is around 160 million people. Reaching 6 percent of working Americans with meaningful AI training is not transformational at the macro level. What it is, if executed seriously, is a beginning. Whether that beginning turns into something real depends on whether OpenAI treats this as a genuine program or a press release with a website behind it. The answer will be clearer within 12 months.
THE AI JOB MASSACRE IS ALREADY IN PROGRESS — 11,000 POSITIONS DISAPPEAR EVERY MONTH AND THE OFFICIAL NUMBERS DO NOT SHOW IT
Bloomberg published a piece this week with a headline claiming no one is talking about the AI jobs crisis. Given that the AI jobs crisis has been the dominant economic anxiety story for two solid years, the headline oversells the premise. But the actual point the piece makes is sharper and worth paying attention to regardless of the framing.
The crisis they are describing is not the loud, obvious version of AI job destruction. Not a factory closing. Not a company announcing mass layoffs citing automation in the press release. The thing that is actually happening is quieter, structurally harder to reverse, and nearly invisible in the official statistics.
What is happening is this: entry-level white-collar jobs are not being refilled when they open up.
A company loses a junior analyst. Historically they post the job, hire someone, that person learns the work, and in a few years becomes a senior analyst. The pipeline keeps running. What is happening now at a growing number of organizations is different. The position opens, someone runs the work through an AI tool for a few weeks, it handles enough of it well enough, and the job simply does not get posted. No termination. No severance. No announcement. The position disappears quietly, and nobody has to report a layoff.
This shows up very badly in conventional unemployment statistics because no one got fired. The person who left found something else. The headline unemployment number looks stable. What is not visible is the steady shrinkage in available entry-level positions — precisely the positions young workers, recent graduates, and career changers depend on to build experience and get started.
Goldman Sachs economists put the current direct displacement figure at around 11,000 net U.S. jobs per month being eliminated by AI. That is happening now. Not in five years. In a country that typically adds around 150,000 jobs per month in normal conditions, 11,000 monthly AI displacement is already material, and that number is almost certainly growing as the tools improve and adoption spreads.
Anthropic CEO Dario Amodei has been the most direct voice on the long-term trajectory. He has said publicly that AI could eliminate more than half of all entry-level white-collar jobs and push unemployment into the 10 to 20 percent range within five years. He has also said the AI industry needs to stop softening the message. The fact that he runs a company whose products are doing some of this displacing either makes his candor remarkable or makes it easier to deliver than if it came from a politician who would need to follow the statement with actual policy proposals.
The structural problem Bloomberg is pointing to goes beyond individual job losses. Entry-level jobs do not exist only to get low-cost work done. They exist to develop people. The junior analyst becomes the director because they spent years doing foundational work and developing judgment by doing it badly and then better. If those foundational roles disappear, who develops the domain knowledge and instincts to manage the AI doing the work? Nobody has a credible answer yet.
The optimistic counter-argument — that new jobs will emerge as they always have when technology disrupted existing ones — is probably true over some long enough time horizon. The uncomfortable question is whether the transition period will be long enough and painful enough to cause serious damage to people who are not abstractions in an economic model. The data says the displacement has already started. The policy response remains somewhere between inadequate and nonexistent. That gap tends to close eventually, one way or another, and usually not quietly.