JUNE 6, 2026 | YOUR DAILY AI INTELLIGENCE BRIEFING
GOOGLE HANDS ELON MUSK $920 MILLION EVERY SINGLE MONTH FOR COMPUTE POWER — THE BIGGEST AI INFRASTRUCTURE DEAL IN HISTORY
So let me make sure I have got this straight. Google, one of the most powerful technology companies in the world, a company that has its own custom AI chips called TPUs, a company that has spent years building one of the most sophisticated computing infrastructures on the planet, has just signed a deal to rent computer power from Elon Musk’s rocket company.
The number is $920 million a month. Not a one-time fee. Not a small pilot. Every month. For almost three years, running from October 2026 through June 2029. The deal covers roughly 110,000 Nvidia GPUs, some CPUs, memory, and related hardware. Do the math and you are looking at approximately $30 billion over the life of the contract.
This is, by any reasonable measure, a genuinely unusual set of circumstances. Google is one of the founding members of the AI arms race. They created the Transformer architecture that made large language models possible in the first place. They have DeepMind, one of the most respected AI research labs on earth. And yet demand for their Gemini Enterprise agent platform is apparently so high that they cannot build new data centers fast enough to keep up.
Which is where SpaceX comes in. SpaceX, the company whose core business is putting rockets into orbit and getting humans to Mars, has apparently built or is in the process of building a data center full of Nvidia GPUs that Google wants access to. And not just a little access. Nearly a billion dollars a month worth of access.
The timing is also worth noting. SpaceX is preparing to list its stock on the Nasdaq, and this deal landed just one week before that is expected to happen. A $920 million monthly commitment from Google is not the worst thing to have in your back pocket when investors are deciding what your company is worth. Whether that timing is coincidence or not is a question for more cynical minds than mine.
What this really tells you is that the AI infrastructure crunch is not a theoretical problem anymore. The largest technology companies in the world are scrambling for compute in ways that were hard to imagine even two years ago. When you are paying nearly a billion dollars a month to rent GPUs from a company whose main business is launching satellites, the conventional wisdom about who builds AI and how has clearly changed.
Google’s own spokesperson confirmed the deal was made to secure bridge capacity to meet surging demand for Gemini Enterprise. Bridge capacity. Nearly a billion dollars a month is their bridge. Whatever the destination costs is apparently even more than that, which is a sentence that would have sounded absurd in any previous decade of the technology industry.
Read the full story at TechCrunch
CHATGPT NOW READS YOUR OLD CONVERSATIONS WHILE YOU SLEEP AND CALLS IT “DREAMING” — YOUR AI KNOWS YOU BETTER THAN YOUR THERAPIST
OpenAI has just flipped a switch on something they are calling Dreaming V3, and I want to be honest with you: the name is doing a lot of heavy lifting here. The idea is that ChatGPT now runs a background process after your conversations end, reads through years of your old chats, and builds up a synthesized picture of who you are, what you care about, and what you are working on.
No instruction required. No “remember this.” Just the AI quietly going through your conversational history like someone reading your diary while you are in the other room.
Now before you get too comfortable or too alarmed, here is how it actually works. Dreaming V3 replaces the old saved memories system where you could tell ChatGPT to remember specific things. Instead, it synthesizes memory automatically from many conversations, building a running portrait of your preferences, ongoing projects, recurring concerns, and time-sensitive context. OpenAI says factual recall has jumped from 41.5 percent in 2024 to 82.8 percent now. That is a meaningful leap and not a number you can dismiss.
There are user controls. You get a readable memory summary showing what ChatGPT has figured out about you. You can add or update things. You can tell it which topics to avoid raising. And there is a privacy toggle if you want to turn the whole thing off. So it is not quite as dystopian as the framing might suggest.
But let us be real for a second. The appeal of this feature and its slight eeriness come from exactly the same source. The system is better because it quietly reads your old conversations without being asked. Most people do not sit around explicitly managing what their AI remembers about them. They just talk, and then they come back later, and the AI is either helpful or frustrating depending on how well it knows them. Dreaming V3 is OpenAI’s attempt to solve that friction permanently.
The efficiency angle is also interesting. OpenAI says they cut the compute cost of serving this memory system by about 5x, which is why they can now offer it to free users for the first time. Previously, the advanced memory features were only available on paid tiers. That changes now. So the version of ChatGPT that is free to use is now quietly learning who you are in the background, which is either a great product improvement or something your privacy-minded friends will post about for the next two weeks, depending on your circle.
The rollout started June 4 for Plus and Pro subscribers in the United States. Free tier and international users are next. If you use ChatGPT regularly and have not looked at what it remembers about you lately, now would be a good time to check.
Read the full announcement at OpenAI
BIPARTISAN BOMBSHELL: CONGRESS DROPS A 269-PAGE BILL THAT WOULD MAKE EVERY STATE AI LAW IN AMERICA UNENFORCEABLE FOR THREE YEARS
Last Thursday, a bipartisan group of House members dropped a 269-page discussion draft called the Great American Artificial Intelligence Act. That name tells you everything you need to know about the current political moment, but the substance inside is genuinely significant and worth understanding properly before the cable news cycle reduces it to a five-second chyron.
The bill was put together by Representatives Jay Obernolte and Lori Trahan, along with four other lawmakers from both parties. The broad goal is to create a federal framework for AI governance before the patchwork of state laws gets any messier. There are currently something like 300 AI-related bills active across various state legislatures, and the argument being made here is that you cannot have a coherent national AI policy if every state is writing its own rules for the same technology.
Here is the part that has made consumer advocates and labor groups very unhappy. The bill includes a three-year preemption of state laws related to AI development. Meaning that for three years after this passes, if it passes, states cannot enforce their own AI laws that overlap with what the federal bill covers. California’s AB 2013, which requires AI model developers to publish summaries of their training data, would be gone. Parts of California’s content watermarking law would also be preempted.
The argument in favor is straightforward. AI companies need to build and deploy products across all 50 states without running into 50 different compliance frameworks. Innovation moves faster with clear national rules than with a legal maze. And the federal framework in the bill does include real requirements, namely that large frontier AI developers have to publish plans for managing catastrophic risks before releasing new models, report serious safety incidents to regulators, and submit to third-party audits.
The argument against is also straightforward. States have historically been where consumer protections actually get built and tested, often faster than Congress moves. Stripping states of their authority for three years, during what is probably the most consequential stretch of AI deployment in history, is not a small thing. Public Citizen called it a bill that strips states of their authority to protect consumers, workers, and children. Brad Carson of Americans for Responsible Innovation described the preemption clause as a generational mistake.
This is a discussion draft, not a final bill, so the scope of preemption could still change before anything gets voted on. But the fact that it landed with genuine bipartisan sponsorship means it has real legs. The tech industry has been quietly pushing for federal preemption of state AI laws for years, arguing that a single national standard is the only workable path forward. This bill is the most serious attempt yet to make that happen.
Read the full breakdown at Axios
$30 BILLION TO BUILD AI DATA CENTERS IN INDIA: THE INFRASTRUCTURE ARMS RACE PLANTS ITS FLAG IN SOUTH ASIA
AirTrunk, the data center company backed by Blackstone, has just announced it will commit $30 billion to building data center capacity in India by 2030. The target is 5 gigawatts of total capacity, which would make it one of the largest single infrastructure investment commitments to India’s digital sector ever announced by any company in the country’s history.
Let me put 5 gigawatts in perspective for you. That is roughly the power consumption of a medium-sized city. That is how much electricity this data center build-out is expected to require when fully operational. If you want one number that captures how much raw power the AI boom is going to eat up over the next few years, this is a good one to save.
AirTrunk has been making big moves in rapid succession. Earlier this year it committed large sums to Australia and then to Japan, and now India is getting the biggest commitment of all. The pattern is clear: Blackstone is betting that demand for AI compute is going to be geographically distributed in ways that Western-centric infrastructure simply cannot serve efficiently. India has 1.4 billion people, a rapidly growing technology economy, and a regulatory environment that has become increasingly welcoming to large foreign infrastructure investment under the current government.
The AI industry’s hunger for data center capacity has created something that looks almost like a construction boom running simultaneously on multiple continents. Microsoft, Google, Amazon, and Meta have all announced multi-billion dollar data center investments over the last 18 months. But most of those have been concentrated in the United States, parts of Europe, and East Asia. AirTrunk is explicitly betting that South Asia is the next major frontier, and a $30 billion public commitment is about as explicit as it gets.
This matters beyond the headline number. India has been actively trying to position itself as an alternative to China in the global technology supply chain, and a commitment of this scale strengthens that argument considerably. It also means that AI infrastructure, the physical hardware and real estate that makes large language models actually run, is becoming a geopolitical asset in ways that the mainstream coverage of AI does not yet fully appreciate.
The 2030 timeline is ambitious. Building 5 gigawatts of capacity requires not just capital but land acquisition, power grid integration, cooling infrastructure, and regulatory approvals at a scale that India has not always processed quickly. But the fact that a Blackstone-backed company is willing to announce this publicly and commit to the number suggests they have done serious homework on what is actually achievable.