Apple Is Quietly Killing the AI Data Centers (And It's Working) Part 2

[00:00:00] Speaker 1: On October 15th, Meta released what it calls the most capable American OpenAI model to date. Here is the twist: Meta plans to spend up to $145 billion on AI infrastructure this year. Yet its own announcement says this $816 billion parameter model runs on just two Apple desktops, two Mac studios, with 512 gigabytes of memory each. Think about that. The biggest datacenter spenders on earth just showed developers a way to skip the datacenter. The catch? Those two machines cost over $44,000. In part two, we follow the money. If you watched part one, you know the argument: Communities across America are blocking datacenters, and Apple has quietly built the one kind of machine that can run serious AI without one. A fair objection to that argument was simple: The biggest machine was not even on sale yet. Nobody knew its price, and no major AI lab had backed the idea. Since then, three things happened in just a few weeks: On October 15th, Meta released a giant model built to run on Apple desktops. On October 28th, Apple began selling the 512-gigabyte Mac Studio. And in the same two weeks, the biggest cloud companies reported what they spent in the third quarter. Put those three events side by side, and a pattern shows up. That pattern is the subject of this video. Let's start with the model, because the details matter. MUSE Spark Open has 816 billion parameters in total, but it uses a design where only 42 billion of them are active for any single word it produces. Think of a hospital, with hundreds of specialists. You only see the few who match your problem. That design is why a model this big can respond at a usable speed. It also has a context window of 1 million tokens, which is roughly several long novels held in mind at once. Meta built it for coding and long multi-step work. Open-wait means anyone can download the model files and run them on their own machines. And here is the part that matters for this channel. Meta said the model was optimized for the Mac Studio with the M5 Ultra using Apple's own MLX software. Apple's machine learning researchers published a technical note the same week. They explained that at full precision, the model's parameters alone would need more than 1.6 terabytes of memory. Compressed to 4-bit precision, it shrinks to about 493 gigabytes. That is just under the 512-gigabyte ceiling of one machine. But the model also needs working memory while it runs. So one machine is not enough. It takes two. Apple's new chief executive, John Turnus, publicly congratulated the Meta team. Two companies with a reportedly chilly history were suddenly promoting each other. Now, the numbers. And I want to be careful here because these come from Apple's own testing and not from an independent lab. Apple ran the model on four M5 Ultra Mac Studios linked with Thunderbolt 5 cables and measured 36.73 tokens per second. With two machines using a simpler split, the speed was 15.74 tokens per second. Those tests used a short context of 8,192 tokens. So this is the best case, not a typical workday. How do four separate computers act like one? Apple added a low-latency link over Thunderbolt 5 that lets one machine read another machine's memory directly, and its MLX software splits the model across them. Each pair of machines is wired together with two cables to double the bandwidth. With two machines, the model is cut in half. One machine runs the first layers, and the other runs the rest. With four, every layer is spread across all of them, which is harder to build, and is the reason the speed more than doubles. When Apple pushed the context to 65,536 tokens, which is more like feeding the model a long report, the wait before the first word appeared was about 40 seconds with two cables per link and about 50 seconds with one. So local AI at this size is not instant. It is fast enough to be useful and slow enough to notice. Here is some arithmetic. 36.73 tokens per second, running nonstop, adds up to about 3.2 million tokens in a day from a single stream. Over a year, that is more than 1 billion tokens. Real workloads vary, so treat this as a sizing exercise and not a guarantee. If you know what your company pays per million tokens in the cloud, you can multiply it yourself and see how the math looks. And the power. Apple says a Mac Studio with the M5 Ultra peaks around 300 watts. Four of them peak around 1,200 watts, which is less than a common space heater. That is an entire cluster running an 816 billion parameter model. Why is the new chip faster? Because generating text is mostly limited by how quickly the machine can read the model from memory. The M5 Ultra moves 1.2 terabytes per second, up from 819 gigabytes per second on the M3 Ultra. Apple reports that text generation runs up to 1.6 times faster, and the wait for the first word is up to 4.6 times shorter. Now, here is a point most coverage skipped. You do not need the giant model to get value. Apple's Note also tested smaller open models. A 27 billion parameter Quen model used up to 68 gigabytes of memory, and a 128 billion parameter Mistral model used up to 158 gigabytes, even with a long context of 65,536 tokens. Both fit inside a single machine, at 19.69 and 7.70 tokens per second with shorter context. A 256 gigabyte machine costs about $10,800, and it can hold models like these. Now the price, because this is where most people stop watching. On October 28th, Apple began selling the 512 gigabyte configuration. The minimum price is $22,299. That is $11,500 more than the same machine with 256 gigabytes. Two machines cost $44,598. Four cost $89,196, before any extra storage. In part one, I guessed the 512 gigabyte machine could land near $36,000 each, which would have put a four-machine cluster near $146,000. The real number came in lower. Is $89,000 cheap? No. Look at what it is being compared with. Now follow the money. In the third quarter of 2026, Amazon, Alphabet, Microsoft, and Meta together spent roughly $200 billion on capital expenditure in just three months. Trackers differ slightly on the exact total, but they all land near that number, and it is a record. For the full year, Amazon guides to about $220 billion. Alphabet to $195 to $205 billion. Microsoft to about $190 billion. And Meta to $135 to $145 billion. Here is what changed. The cash to pay for all of this is getting tight. Meta reported $34.9 billion of capital spending in the quarter, and its free cash flow turned negative at about minus $2.1 billion. After the report, its stock fell about 9%. Across the industry, share buybacks by Alphabet, Amazon, Microsoft, Meta, and Oracle fell to $4.4 billion in the quarter, the lowest since 2017. Alphabet, Meta, Amazon, and Oracle are now spending more on data centers and equipment than their operations bring in. The days of paying for the build-out from spare cash are ending, and borrowing fills the gap. Alphabet and Amazon have both told investors that spending will rise again in 2027, so the pressure to find cheaper ways to serve AI is only growing. Now return to the Meta paradox. A company burning cash on data centers just released a model designed to run somewhere else. Why? I cannot see inside Meta, so treat what follows as analysis and not fact. There are three plausible reasons. One, an open model spreads fast, and the more people build on it, the more it sets the standard. Two, every question answered on a customer's own machine is a question Meta does not have to pay to serve. Three, very few desktop machines can hold a model this size in one memory pool. So partnering with Apple is the shortest road to getting it into people's hands. None of these reasons require Meta to believe data centers are a bad idea. They only require Meta to believe that not every workload should live in one. Meanwhile, the neighborhoods have not gone away. In part one, we covered Data Center Watch and its first half total of almost $200 billion blocked or delayed. Here is an update that makes the picture more honest. The second quarter was lower than the first, with 45 projects instead of 75. Opposition did not disappear, but it did not keep rising either. Separately, the Financial Times reported, using a tracker called RegLayer, that at least 47 large projects were delayed, withdrawn, or blocked in 2026. The third quarter report has not come out yet, and that is the number to watch. The point is simple. Local resistance raises the cost and the time of every new data center. Local hardware lowers the cost of not building one. When one side of a pair gets harder and the other side gets easier, money moves. So who actually loses? Not the company's training giant models. Training a frontier model still takes thousands of specialized chips working together. And a shelf of Macs does not replace that. And the cloud is not shrinking. Microsoft reported $627 billion of contracted future revenue earlier this year, and Google Cloud's backlog was above $460 billion in the spring. Demand for cloud AI is real, and it is growing. The pressure lands on something narrower. Think of a company with 300 employees that wants a private assistant to read its contracts, its tickets, and its internal documents. Until this year, the realistic choices were a cloud contract with a meter running on every question, or a rack of specialized hardware with its own power and cooling needs. Now there is a third choice: a shelf with four quiet boxes. Run the numbers for that company. $89,196 divided by 300 employees is about $297 per person, paid once. That is the price of a decent chair and a monitor, and it buys a private AI system that the company owns. There is also a reason that has nothing to do with price. Some data simply should not leave the building. Law firms, clinics, accounting offices, and manufacturers hold records that their own rules, or their own customers, say must stay inside. Apple's pitch for this machine leans on exactly that: running big models with complete privacy and without a token meter. For these buyers, the question is not whether local is cheaper. It is whether local is the only option they are allowed to use. And notice who buys first: not the average consumer, and not the hobbyist. It is the team with a real budget, a real compliance problem, and an engineer who can wire four machines together. That is a small market today. But every big shift in computing started with a small group that had a good reason to pay too much. Even if most companies never buy one, the existence of that third choice changes the negotiation. A cloud provider quoting a price now has to compete with what the customer could own for about $89,000. That puts a ceiling on what the cloud can charge for this kind of work. You do not need to win every customer to cap what the others will pay. Now the parts that do not fit the exciting story. First, prices keep rising. Memory is scarce across the whole industry. Apple removed the 512GB option from the older M3 Ultra machine back in March, and the 256GB option in May. It raised the Mac Studio's starting price in June. The M5 Ultra starts at $5,499, but that version has only 96GB. The same memory shortage that strains data center budgets is straining Apple, too. There is one more detail worth knowing. Apple dropped the Mac Pro from its lineup in March 2026, which makes the Mac Studio the most powerful desktop Apple sells. So this small silver box is no longer a side product. It is now the flagship, and Apple is marketing it for on-device AI. Second, the best models are not open. The strongest closed models from the largest labs cannot be downloaded at any price. Local AI means open-weight AI, and it depends on labs like Meta continuing to release strong ones, on terms that Meta sets. Third, the model at the center of this story needs at least two machines, and for the best speed, four. That is an engineering project, and not a plug-in appliance. Someone has to set up the cluster, monitor it, update it, and fix it when it fails. That unglamorous work is where most small companies will need help. Fourth, the 1.5 terabyte rumor is still a rumor. The report points to an M7 Ultra around 2028, and it depends on memory supply. Apple has confirmed nothing. Step back and look at the pattern. In part one, we described a company that was written off as late to AI. Look at what happened since. Apple did not win by announcing a bigger building. It got a public endorsement from the company that is building some of the biggest buildings. Is that vindication? Partly. The part that is proven is the hardware. A frontier-scale open model really does run on desktop machines. And Apple's tests, Meta's announcement, and independent coverage all agree on that much. The part that is not proven is the economics at scale. We do not yet have independent tests of a four-machine cluster serving a real company for months. That is the evidence that would turn this from a clever demo into a new category. Here is your simple watch list for the next 90 days. 1. The data center watch report for the third quarter. If opposition keeps falling, the community story weakens. If it rebounds, it strengthens. 2. Delivery dates for the 512 gigabyte machine. Because some European buyers were quoted late November. 3. Independent benchmarks of a four-machine cluster, not just Apple's own. 4. Nvidia's next earnings report, expected in mid-November. Because that company sells the other side of this trade. 5. Whether another major lab follows Meta and tunes an open model for desktop hardware. 6. Whether Apple says anything official about a machine with even more memory. So is Apple quietly killing the AI data center? Not yet. And frankly, anyone who tells you otherwise is selling something. What Apple is doing is quieter. It is building the second option. And in a market where the first option is getting more expensive, slower to build, and harder to get approved, a second option is how providers lose their pricing power. If this video gave you something new, subscribe. Because in part three, we go inside the machine. And look at why unified memory suits this one job and where it falls short. And tell me in the comments which side you are on. Is this the start of a real shift or just an expensive demo? And I'll see you in the next one!