About this transcript: This is a full AI-generated transcript of Apple Just Sent Big Tech A MASSIVE Warning About AI Data Centers from Williamson, published November 17, 2026. The transcript contains 3,921 words with timestamps and was generated using Whisper AI.
"Across America, towns are saying no to AI data centers, and almost nobody has noticed what Apple is doing at the same moment. In the first six months of 2026 alone, local communities blocked or delayed almost $200 billion worth of these giant buildings. Now a quiet silver box from Apple, small..."
[00:00:00] Speaker 1: Across America, towns are saying no to AI data centers, and almost nobody has noticed what Apple is doing at the same moment. In the first six months of 2026 alone, local communities blocked or delayed almost $200 billion worth of these giant buildings. Now a quiet silver box from Apple, small enough to sit on a desk, is going after the reason to build them at all. Today we count down the four Apple moves behind that push, and the early signs say it is working. One of these machines ran a model with 671 billion parameters on less than 200 watts. A rumored future chip could hold 1.5 terabytes of memory. The four biggest cloud companies are spending about $725 billion this year, and all of it leans on one shaky idea. More Americans would accept a nuclear plant nearby than an AI data center. That is not a joke. It comes straight from a national survey. For years, critics called Apple's chip design a nice trick for laptops, and most of the tech world wrote the company off in AI. Now that same design has a real price tag, starting at $5,499, and it is not a drawing on a whiteboard. Somewhere under all of this sits one number that changes the whole story, and it is not the one in the rumors. That is before the strangest part, and if that part already surprised you, what comes next is the reason this story exists. Tap like so more people see it, and subscribe so you are here when we get there. To see how a company that was told it had lost got this far, we have to start with the moment people first noticed. A single desktop computer loaded an AI model with 671 billion parameters, and the whole machine drew less than 200 watts. That is about what three old light bulbs use. The computer was the Mac Studio with the M3 Ultra chip, which Apple released in March 2025. The model was the full version of DeepSeq R1, and even squeezed into a smaller format it filled around 404 gigabytes of memory. Parameters are the tiny settings a model learns as it trains, like billions of dials inside its brain. Every dial needs a place to sit in memory, so a bigger model needs a bigger room. That room is why the test spread so fast. Models of that size usually run across several graphics cards working together, with heavy power supplies and cooling to match. The Mac Studio did the job alone, and produced roughly 17 to 18 tokens per second. A token is a small chunk of a word, so the answer came out faster than most people can read. The shock was not the speed though, the shock was that it ran it all on a machine that fits on a desk. The top memory option of 512 GB bytes pushed the price to about $9,499, which is a lot for a desktop, and very little for a server room. Memory is the most boring line on any spec sheet. Nobody lines up outside a store for it the way they do for a new camera, but for AI, memory works like a gate. If a model does not fit, it cannot run at all, no matter how fast the chip is. A strong gaming graphics card holds around 24 to 32 GB, so a model the size of DeepSeq R1 would need more than a dozen of them. One spec most people skip decides who gets to use serious AI and where that AI has to live. The road to that test started in 2020, when Apple stopped buying processors from Intel and began designing its own chips for the Mac. Most of the industry shrugged. But Apple also made an odd choice inside those chips. A normal computer gives the main processor and the graphics chip separate pools of memory, so data gets copied back and forth between them. Apple built one shared pool instead and called it unified memory. The main processor, the graphics processor, and the neural engine all read from the same place. It works like a kitchen where every cook shares one big fridge, instead of each cook owning a tiny one and passing food across the room. Critics waved that off as a clever laptop feature and nothing more. The serious AI world was going a very different way, with thousands of graphics cards packed into giant buildings. Each card had its own small pool of very fast memory. That design is brilliant for training the largest models, and it is also the reason those buildings need so much power. So while rivals announced spending plans in the hundreds of billions, Apple looked badly behind. Its smarter Siri was late, and most headlines said the company had missed the biggest shift since the smartphone. Apple never answered with a giant data center of its own. It just kept adding memory to its machines. The rest of the world, meanwhile, was renting. For the last three years, a business that wanted strong AI paid someone else for it. A cloud provider owned the chips, the building, the cooling, and the power bill. The business sent its data away, waited for an answer, and paid for every token. It works like a taxi, with the meter always running. Every question from every worker adds a little more to the bill, and that meter made a small group of companies very powerful. It also set off a building spree, with almost no match in computing history. Alphabet, Amazon, Microsoft, and Meta have guided to roughly $725 billion of spending in 2026. That is close to $2 billion every day, weekends included. It is also about 77% more than the $410 billion they spent in 2025. Analysts at Bank of America and Evercore expect the yearly total to pass $1 trillion in 2027. Microsoft alone is planning around $190 billion this year, and it still told investors it will stay short of capacity through 2026. A company spending that much is saying it still does not have enough. All of that money leans on one idea, which is that serious AI can only live inside huge buildings owned by a few companies. The M3 Ultra test was the first public crack in that idea, because a giant model had just run on a shelf with no cooling tower in sight. One desktop does not threaten budgets like those, of course. It cost as much as a used car, and it was only one box. But it proved the idea had a weak spot. A crack only matters if someone keeps pushing on it, and that is where the story picks up speed. Number 3. The M5 Ultra Mac Studio The older Mac had a weak spot, and the community found it fast. When testers fed the M3 Ultra a very long prompt of around 13,000 tokens, its answers slowed to about 6 tokens per second. A prompt that long is like handing someone a thick report and asking for a summary. That is also exactly what a real office wants to do all day. A machine that shines on short questions but drags on long documents makes a great demo and a poor co-worker. If Apple wanted real business work, it had to fix that first. On August 25, 2026, only 17 months after the last model, Apple announced a new Mac Studio with the M5 Ultra chip. The fix sits at the center of it. The chip has up to 36 processor cores and up to 80 graphics cores. For the first time on an Ultra chip, neural accelerators are built right into those graphics cores. These are small parts made for one job, which is the kind of math AI runs on. Think of a busy kitchen that suddenly gets a station built only for slicing. The menu stays the same, but dinner comes out much faster. The most important number on this machine is one that did not grow. Memory still tops out at 512 GB bytes, the same ceiling as before. What changed is how fast that memory moves, at 1.2 TB per second, which Apple says is 50% higher than the last model. That may sound like a small detail, but it is the secret of the whole machine. When an AI model answers you, the holdup is often not the math. The holdup is how quickly the machine can read those billions of dials out of memory. A genius stuck in a library with one narrow door will still take all day. Apple built a much wider door, and that is why a desktop can now do work that once needed a full rack of servers. Apple's own claims are bold. It says the M5 Ultra has up to 4.3 times the peak AI compute of the M3 Ultra. It also claims up to 4 times faster prompt processing in LM Studio, an app people use to run models on their own computers. Those results came from Apple's own tests in July 2026, so treat them as the best case. Independent testers are still catching up. If the real numbers land even close, though, that crawl on long documents stops being the story. The new Mac Studio starts at $5,499, and the 256 GB version, with 16 TB of storage, tops out at $18,299. Apple set late October for the 512 GB version, and at last check its price had not been announced. None of this is cheap for a home office, and the real target looks like actual offices anyway. Apple's pitch is simple. Run large models on your own machine, in private, without counting tokens or worrying about rising cloud costs. That is a sales line, and it is also a direct shot at the taxi meter. Consider a company with a few hundred workers. On rented cloud power, the bill grows with every question. So the more the staff loves the tool, the more it costs. Success gets punished. On a machine the company owns, the thousandth question of the day costs the same as the first, which is a little electricity. That one change flips how a boss thinks about AI. It stops being a bill to watch, and starts being a tool to use as much as possible. This is where the M5 Ultra starts to touch the bigger claim in this story. Every job a company moves onto its own machine is a job the cloud no longer gets paid for. To make that move, the company runs open weight models, the kind anyone can download and keep. Deep Seek R1 is one of them. They are like cookbooks you buy once, instead of meals you have to order every night. The catch is real, and it deserves a straight answer. The strongest closed models from the big labs cannot be downloaded at all, so nobody can run them on their own desk. A local setup only stays useful if open weight models keep pace. Training is a separate problem too. Building a new frontier model from scratch still takes enormous clusters of special chips, and Apple is not replacing that or trying to. Training is like writing a whole encyclopedia, and running a model is like looking things up in it. Apple is only going after the looking up, but that is the part most offices need every single day. So the M5 Ultra changed the question from whether a desktop could run big AI, to whether a business should. It turned a strange test into something a buyer could put on a purchase order, and that is a much bigger threat to anyone selling rented power. Yet one number on that spec sheet stayed stuck, and the reason reaches far beyond Apple. That is where this story gets complicated. Number 2. The M7 Ultra Apple has not said one word about this chip, and it may be the most important Mac that does not exist yet. Bloomberg's Mark Gurman, the reporter Apple watchers follow most closely, reported that Apple is engineering an M7 Ultra. It could support up to 1.5 terabytes of unified memory in a Mac studio expected around 2028. One terabyte is about 1,000 gigabytes, so that is nearly three times the memory in today's top machine. DeepSeq R1 would fit inside more than three times over, with room left for other work. A desktop with that much memory would have sounded silly only a few years ago. Gurman himself said the public may never get that version. It depends on the memory industry, and the memory industry is in the middle of a shortage. Apple has confirmed nothing. There is also a recent warning from history. An earlier report floated 768 GB buys for the M5 Ultra, and what actually shipped was 512 gigabits. A rumor can show you which way a company is heading, but it cannot tell you what will land in stores. That shortage is already showing up on Apple's own price tags. In June 2026, Apple raised the starting price of the Mac Studio by $500, from $1,999 to $2,499. The new M5 Ultra costs $1,500 more to start than the M3 Ultra did at its launch. Apple has trimmed memory options over the past year, and in March 2026, it dropped the Mac Pro from its lineup altogether. The same squeeze that makes data centers more expensive is also hitting the company trying to replace part of what they do. Everyone in this story is bidding for the same scarce chips. So why does a chip that may never reach the public rank this high? The date on it is the reason. A data center is a long bet, built to pay for itself over many years. The trillion-dollar year that analysts expect comes in 2027, and a 1.5-terabyte desktop could land right behind it. Anyone signing building plans today has to ask what a desk will be able to do by the time the concrete dries. Even as a rumor, the M7 Ultra puts a clock on the idea that big AI needs a big building. And those buildings now have a second problem that money cannot easily solve. That second problem is permission. Data Center Watch, a research project run by a firm called 10A Labs, tracks local fights over these sites. In the first quarter of 2026, it counted at least 75 projects worth about $130 billion, blocked or delayed by local opposition. That is close to one project a day. It was the largest quarter on record, and it roughly matched the total for all of 2025. The second quarter added another 45 projects worth about $68 billion, which is how the six-month total climbed so high. One caution matters here, because those figures mix projects stopped for good, with projects only slowed down. So this is not a list of permanent cancellations, but the direction is easy to read. The people behind those numbers are organizing quickly. At the end of 2025, Data Center Watch counted 396 opposition groups. By March 2026, there were 833 of them, spread across 49 states. That is more than double in about three months, and it leaves just one state out. More than 300 state bills about data centers were filed in the first six weeks of the year. Proposals to pause new building across a whole state appeared in 14 states. These are not a few loud voices at one meeting anymore. The nuclear comparison came from Gallup, in a survey taken in March 2026. It found that 71% of Americans oppose an AI data center in their area, and 48% oppose one strongly. Only 27% were in favor. When Gallup asked about a nuclear power plant nearby, 53% said no. So a plant that splits atoms is an easier cell than a building full of computers. People named electricity use, water use, noise, and rising utility bills. Supporters mostly named one benefit, which was jobs. The fight also does not split neatly along party lines, so no single party can simply wave it away. The logic from a neighbor's side is simple: the answers from a data center go out to customers far away, while the noise and the power bill stay in town. A family in that county may never use the service at all, but it still lives beside the building. That is a hard trade to sell at a public meeting, no matter how big the budget behind it is. Look at what all of this leaves the industry with. It has cash, chips, and customers lined up, but it cannot simply buy a town's approval. Apple needs no town's approval to put a box on a desk. That contrast is why a rumored chip matters, even if it slips a year or never arrives in that exact form. One side is climbing a memory curve, while the other side fights sight by sight. But the strangest thing about waiting until 2028 is that you might not have to. Number 1: The 4 Mac Studio Cluster The number hiding under all of this is 2 terabytes, and Apple can reach it without any rumor coming true. Take 4 Mac Studios with 512 gigabytes each, connect them, and you have 2 terabytes of memory sitting on a shelf. That beats the 1.5 terabytes everyone is waiting on for 2028. It takes 4 boxes instead of 1, but nobody has to wait on an unconfirmed chip. That much room is enough for some of the largest open models in the world, with no server farm in sight. The feature that makes this work has a long name, which is Thunderbolt 5, with remote direct memory access. In plain words, one Mac can reach straight into the memory of the Mac next to it, without a slow chain of steps in between. Several machines then act like one shared space. Apple says a group of 4 delivers up to 3 times faster AI inference than a single machine, and inference simply means a model answering questions. Our kitchen with one big fridge has now knocked through the walls into 3 neighboring kitchens. A real company can run the math on this in an afternoon. Apple has not given us every price, so let's guess on the high side. Suppose the 512 GB version costs twice what the $18,299 machine does. That is about $36,000 each, and 4 of them come to around $146,000, paid once. That is real money, but a mid-sized company can approve it in one meeting. Microsoft's budget for this year alone could buy about 1.3 million of those clusters. For that one payment, a business gets a system it owns, data that never leaves the building, and no meter running in the background. The jobs most at risk are not the biggest ones in AI, they are the ones in the middle. Think of searching a company's own files, helpers for staff, number crunching on private records, and coding agents. Think of summarizing contracts and medical notes that nobody wants to hand to an outside company. A law firm has every reason to keep client papers in the building, and a clinic feels the same way about patient records. That middle layer is where a huge share of everyday business AI is expected to live. It is also exactly the layer that a shelf of four Macs can serve. Meta's planned Hyperion campus in Louisiana has been described as reaching 5 gigawatts of capacity. Divide 5 billion watts by the 200 watts from that older Mac test, and you get 25 million Mac studios running flat out. That is not a fair comparison, and it should not be treated as one. A data center trains frontier models and serves hundreds of millions of people, which a Mac studio cannot do. But it shows a difference in kind. One approach packs a huge power demand into one county and asks the people there to live with it. The other spreads small loads across law firms, clinics, factories, and accounting offices that already have electricity. That means no new transmission lines, and no fight over a zoning map. The cluster still has to prove itself, and a careful buyer should say so. Independent tests of four machines running the largest open models are what matter now, not Apple's own charts. The price could also climb if the memory shortage gets worse. So is Apple really killing data centers? No, not in the way a headline suggests. Spending is still climbing, and nothing in the numbers says these buildings are going away. But Apple is wearing down is the business case, and that is a different thing. A new data center only pays off if a steady stream of customers has no better choice than renting. Each job that moves from a rented rack to a machine in the office is a customer whose bill stops growing. No single day will make the the news. The damage shows up as demand thinning out in the middle of the market, just as towns make new sites harder and more expensive to win. Towns squeeze from one side, and hardware squeezes from the other. That is what working looks like here, slow and almost invisible. Look at who ended up with the leverage. Residents who went to those hearings were told they were standing in the way of the future. Apple spent years hearing that it was behind. Today one of those groups is holding up projects worth billions, and the other is selling the most memory ever put in a personal computer. Neither had the biggest budget in the room. Both kept pressing on the one thing a giant building cannot easily fix, which is where the power comes from and who has to live beside it. The next three years will settle this. The real price and delivery of the 512 gigabyte machine come first, along with any sign that the memory shortage is easing. Then 2028 will show whether the M7 Ultra report holds up. The signal to watch most closely is an earnings call. If any one of the four big cloud companies cuts its spending plans, that will say more than any outside estimate could. And if even a slice of business AI leaves the cloud for a desk, the people pouring all that concrete face a new problem. For years, the only question was how fast they could build. The new question is, who will fill those buildings once they are done? In some town that already said no, part of the answer may be a quiet silver box on a shelf at the local clinic, and nobody had to hold a hearing for it. So that is the countdown. One desktop test, a faster chip, a rumor, and a shelf of four machines, all pushing on the old belief that powerful AI has to live in giant warehouses. The buildings are not going away, but the reason to build every one of them is getting harder to defend. If this changed how you see the AI boom, give the video a like, subscribe to the channel, and turn on notifications for the next story. How long have you been watching this channel? Tell me in the comments.