Buy a New Computer for Local AI Before It's Too Late
A 31-minute investigation into why a $95 SSD now costs $288, tracing the memory shortage from AI data centers back to the scaling-law math that makes them keep buying.
Posted
2 days ago
Duration
Format
Essay
educational
Views
6.5K
62 likes
57 · 43
Big Idea
The argument in one line.
AI data centers and your next computer draw memory from the same factories, and because every step up the scaling ladder costs ten times the last, that competition for RAM is likely to keep pushing consumer prices higher.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're planning to buy a Mac Studio, a PC or an upgrade in the next year and want to know whether to buy now or wait.
You noticed RAM, SSD or Mac prices jumping in 2026 and want the supply-chain explanation behind it.
You want to run local AI models someday and are deciding how much unified memory to lock into a machine you can't upgrade.
You hear 'scaling laws' and 'computing power' constantly and want a plain-English walk through what those words cost in chips, memory and electricity.
SKIP IF…
You want benchmarks or a spec-by-spec buying guide. This is a market and economics argument, not a hardware review.
You already track DRAM contract pricing and the AI capex cycle closely. Most of the numbers will be familiar.
TL;DR
The full version, fast.
Memory prices are climbing because AI data centers are buying enormous amounts of both RAM and flash storage, and high bandwidth AI memory uses roughly three times the wafer capacity per bit of ordinary PC memory. DRAM contract prices rose 93-98% in one quarter, Kioxia's 2026 flash output is already sold, and Micron shut down Crucial to serve AI demand. The labs keep buying because scaling laws say every 10x in compute cuts error by only about 11%, so each step needs ten times more chips, memory and power. Since Mac memory can't be upgraded and subscription AI is subsidized for now, the argument is to buy a machine with enough memory for the next few years at today's prices, financed, used or traded in if needed.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Apple Upgrade lets buyers pay $39-59 a month for a MacBook Pro they never own, setting up the question of why Apple is pushing leases.
00:43 – 02:11
02 · Apple blames AI for price hikes
Apple's statement says AI data center expansion created an extraordinary surge in memory and storage demand, forcing Mac and iPad price increases.
02:11 – 03:38
03 · The SSD that tripled
A Samsung T9 bought for about $109 is now $242 on Amazon and $288 from Samsung, with nothing added to it.
03:38 – 05:05
04 · Two kinds of memory
Flash storage keeps files, RAM holds working data. AI data centers buy both, and HBM takes about three times the wafer per bit.
05:05 – 07:02
05 · Supply already spoken for
Kioxia's 2026 flash output is sold, DRAM contract prices nearly doubled in a quarter, and HP and Microsoft report the cost hit.
07:02 – 09:14
06 · OpenAI, Crucial and Ram Altman
OpenAI's 900,000-wafer deals, Micron killing Crucial, and Sam Altman's comment that RAM sitting idle in personal computers is 'quite terrible'.
09:14 – 11:19
07 · What the labs are building
Chatbots guess the next word, and getting better at guessing forces deeper understanding. That chase is the road to AGI.
11:19 – 13:15
08 · What computing power really means
Compute is three physical purchases: chips cut from wafers, high bandwidth memory beside each chip, and electricity to run and cool them.
13:15 – 17:18
09 · The scaling ladder
The 2020 scaling paper: each 10x in compute cuts error about 11%, and a millionfold increase only halves it. Error never reaches zero.
17:18 – 21:29
10 · Where the frontier sits now
Labs stopped publishing compute after GPT-4. Outside estimates put GPT-4 past step five and GPT-6 Astra around step seven, at roughly a billion dollars.
21:29 – 23:48
11 · Counting the bills
Colossus holds nearly a million AI chips, the biggest data centers draw power like 2 million homes, and OpenAI loses money faster than it raises it.
23:48 – 25:40
12 · Why they keep climbing
The labs have counted the cost and decided the top of the ladder is worth it. If AGI never arrives, they are left with warehouses of memory.
25:40 – 28:06
13 · The computer on your desk
The base Mac Studio is $500 more than last generation, Mac memory can't be upgraded, and PC RAM kits jumped from about $200 to $1,250.
28:06 – 31:36
14 · Own it while you can
Subscriptions are subsidized for now. Finance, buy used or trade in, but get a machine with enough memory for the next few years at today's prices.
Atomic Insights
Lines worth screenshotting.
A one-terabyte Samsung T9 SSD that sold for about $95 on Black Friday 2025 was listed at $242 on Amazon a year later and $288 direct from Samsung.
High bandwidth memory for AI chips takes roughly three times the silicon wafer capacity per bit as the RAM in a normal PC.
Every gigabyte of AI memory a chip factory makes is roughly three gigabytes of ordinary PC memory it didn't make.
TrendForce reported conventional DRAM contract prices rising 93% to 98% in the first quarter of 2026, nearly doubling in three months.
HP says memory is now about 35% of the cost of building one of its PCs.
OpenAI's early deals with Samsung and SK Hynix covered up to 900,000 memory wafers a month, close to 40% of global DRAM output for one company.
Micron shut down Crucial, the consumer RAM and SSD brand, to keep up with AI data center demand.
A 5-trillion-parameter model needs about 10 terabytes of fast memory just to hold its weights, before answering a single question.
In the 2020 scaling-law paper, ten times more compute cut model error by only about 11%, so a millionfold increase only halved the error.
The scaling paper itself predicted error would level off above zero, because language has unpredictability built in.
OpenAI stopped publishing model size and training compute with GPT-4, so outsiders can no longer judge how close the labs are to their goal.
OpenAI brought in $6.7 billion and lost $12.3 billion in a single quarter while committing to roughly $1.4 trillion in compute.
A 64 GB RAM kit that cost about $200 a year ago now runs around $1,250.
Mac unified memory is fixed at purchase, so on a Mac the memory you can afford today is the memory you live with for the life of the machine.
Memory has crashed before: in 2022 and 2023 prices fell so hard Samsung cut production by about half.
Takeaway
Your next computer is competing with AI labs.
WHAT TO LEARN
The same factories make AI memory and consumer RAM, and scaling-law math means the labs need ten times more of everything at each step, so memory prices track their ambitions.
01Apple starts leasing Macs
Apple Upgrade turns a $2,000 MacBook Pro into a $39 to $59 monthly lease you never own, a sign that prices have outrun what buyers will pay outright.
02Apple blames AI for price hikes
Apple publicly blamed its Mac and iPad price increases on AI data centers creating an unprecedented surge in memory and storage demand.
03The SSD that tripled
Track a part you already own: a Samsung T9 SSD went from about $95 to $288 at Samsung in a year with no change to the product.
04Two kinds of memory
Flash storage and RAM are different products, but AI data centers buy enormous amounts of both, so both get more expensive together.
High bandwidth memory uses about three times the wafer per bit, so every AI memory order shrinks the supply left for consumer RAM.
05Supply already spoken for
When a supplier like Kioxia has a full year of output already sold, new buyers wait in line regardless of what they're willing to pay.
DRAM contract prices nearly doubled in one quarter, and memory now makes up about 35% of the cost of building an HP PC.
06OpenAI, Crucial and Ram Altman
One company's deals for up to 900,000 wafers a month could absorb around 40% of the world's DRAM, enough to reshape the whole market.
Micron exiting the Crucial brand shows memory makers will drop consumer lines to chase larger AI contracts.
07What the labs are building
Large language models work by guessing the next word, and getting better at that guess forces the model to actually follow math and code.
The labs are explicitly chasing AGI and beyond, which sets a target with no clear compute ceiling.
08What computing power really means
Computing power is a shopping list of chips, memory and electricity, not a dial, and all three scale up together.
A model's parameters each take about two bytes, so a 5-trillion-parameter model needs roughly 10 terabytes of fast memory just to load.
09The scaling ladder
Under the 2020 scaling laws, each tenfold increase in compute cuts error by only about 11 percent, so returns shrink while bills multiply.
Error never reaches zero because language is inherently unpredictable, so every last improvement costs more than the one before it.
10Where the frontier sits now
Since GPT-4, labs no longer publish model size or training compute, which leaves the public unable to judge whether the spending is sound.
Outside estimates put the latest frontier models around a billion dollars of compute and well past where the original paper expected returns to flatten.
11Counting the bills
AI capex runs into the hundreds of billions a year while the leading lab still loses about two dollars for every dollar of revenue.
Investor money for AI buildouts is starting to arrive slower than the spending, an early sign the climb may not be fully funded.
12Why they keep climbing
The labs keep buying memory because they've decided the top of the ladder is worth any price, and your RAM comes from the same factories.
Memory markets have crashed before, so a failed AGI bet could flood the market, but nobody knows which direction this cycle breaks.
13The computer on your desk
Mac unified memory is fixed at purchase, so buy enough for the work you expect over the machine's whole life, including local AI.
PC builders can add RAM later, but kits that cost about $200 a year ago now cost around $1,250, so the upgrade path has its own price.
14Own it while you can
Relying on cloud AI or a leased machine hands the pricing decision to someone else, who can change it next month.
Subscription AI is heavily subsidized today, so treat current plan prices as temporary rather than as the real cost of the service.
If buying outright is hard, interest-free installments, off-lease used laptops or trade-in credit can still get you an owned machine at today's prices.
Glossary
Terms worth knowing.
RAM
Random access memory, the working memory a computer uses to hold whatever it is actively processing. It empties when the power goes off.
Flash storage
The memory inside SSDs and USB drives that keeps files when the power is off. Data centers buy huge amounts of it to store training data and outputs.
DRAM
Dynamic random access memory, the standard chip type used for RAM in PCs, phones and servers. Its contract price is the benchmark for the memory market.
High bandwidth memory (HBM)
Stacked, very fast memory placed right next to an AI chip to keep it fed with data. It uses far more wafer capacity per gigabyte than ordinary RAM.
Silicon wafer
A thin disc of silicon that a fab etches with circuits and cuts into individual chips. Wafer capacity is the hard limit on how many chips get made.
Unified memory
Apple's design where the CPU and GPU share one pool of memory built into the chip package. It is fast but cannot be upgraded after purchase.
Parameters
The adjustable numbers inside an AI model that are tuned during training. Each typically takes two bytes of memory, so model size maps directly to memory needed.
Scaling laws
Empirical rules showing that model error falls predictably as compute, data and model size grow, but with diminishing returns for each tenfold increase.
Petaflop day
A unit of training compute: a quadrillion calculations per second sustained for a full day.
Loss
The measured error of a model's predictions during training. Lower loss means the model guesses the next word more accurately.
AGI
Artificial general intelligence. OpenAI defines it as highly autonomous systems that outperform humans at most economically valuable work.
Contract price
The negotiated price large buyers pay memory makers for bulk supply, as opposed to retail prices on store shelves. Retail prices tend to follow it.
Resources
Things they pointed at.
00:00productApple Upgrade program
02:11productSamsung T9 portable SSD
05:05linkKioxia
05:50linkTrendForce DRAM contract price reports
07:02productCrucial (Micron)
07:49channelCheeky Pint podcast
13:15bookScaling Laws for Neural Language Models (Kaplan et al., 2020)
18:34linkEpoch AI training compute estimates
22:34channelBG2 podcast with Brad Gerstner
25:40productMac Studio
28:25productApple Card Monthly Installments
28:42linkeBay (used off-lease laptops)
Quotables
Lines you could clip.
02:11
“One terabyte of solid state storage, external hard drive, just increased in value 100%.”
concrete personal receipt, instantly relatable→ TikTok hook↗ Tweet quote
15:29
“You used a million times the compute that you initially started with and you only cut the error in half.”
“Every person on earth, one math problem every second, no mistakes, no eating, no sleeping for 400 years just to match one second.”
vivid scale analogy, no setup needed→ TikTok hook↗ Tweet quote
26:30
“You don't have to run AI yourself for somebody else's AI data center to affect what you can afford.”
the thesis in one line→ newsletter pull-quote↗ Tweet quote
28:06
“A day of reckoning is coming when you're going to have to pay what it is actually worth.”
contrarian take on subsidized AI subscriptions→ IG reel cold open↗ Tweet quote
30:40
“Maybe you can afford a Mac Studio today, but you might not even be able to afford a Mac mini three years from now.”
stakes framed as a personal budget→ TikTok hook↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
You can get your hands on some of the most expensive Apple products right now for only a couple of dollars a month. And it's all because of a new program, Apple Upgrade, that was launched in July of 2026. It's a great program for anyone who wants to purchase Apple products, but they've just been too expensive for them.
And that might be more of us than it was before. And the reason why is pretty crazy. So let's talk about it.
If you take the MacBook Pro, for example, it's a $2 ,000 machine. You're going to pay somewhere around $39 to $59 a month for it for three years.
And at the end of those three years, you do not own it. Although Apple will give you the opportunity to purchase it. But my question is in three years time, will you want a macbook pro that you purchased three years ago but what really got my attention and made me record this video is what happened a month prior a lot of people started looking at the prices for the new macs and ipads and started asking apple why are the prices going so high and this is what apple had to say apple says that the reason for max and ipad pricing rising is because of a shortage of memory chips but let's get their direct quote the rapid expansion of AI data centers has created an extraordinary surge in demand for memory and storage.
We have never seen a component price increase this much this quickly. And then they said that they've shielded customers from the component price surges so far. We have now reached the point where we need to begin raising prices on a number of products, including today's increases for iPad and Mac.
We know this is not welcome news and we are working. tirelessly to find solutions and so even though apple didn't say this directly i do believe that the shortage of memory chips due to their being purchased by data centers and one company in particular is purchasing a lot of it has a very large part to play in the apple upgrade program where they have to lease out computers because the prices are just getting outrageous and even the people who normally buy apple products are starting to think they're a little too expensive to purchase and so perhaps maybe they'd be willing to rent one or lease one for a couple of years even if we start looking at storage items that are smaller than the mac you'll see that the same pattern is taking place there so let's take for instance something like my solid state external hard drive a samsung t9 one terabyte very small inexpensive purchase when i made it it was listed for 95 on amazon during black friday last year and when i purchased it last year uh i spent a
somewhere around 109 bucks, give or take a few dollars with tax. But October of this year, that same drive was listed for $242. They didn't add anything to it.
They didn't give you any extra capabilities with it. They didn't add any storage with it. One terabyte of solid state storage, external hard drive, just increased in value 100%.
But even if we push the market numbers aside and we simply look at what Samsung listed, for. When they first launched this drive, it was only $140.
But if you go to purchase it from Samsung now, it's listed at $288. So from my perspective as a consumer, a drive that I spent a little over $100 for is now being sold for almost three times that price from the manufacturer a year later. That's more than double and it's almost three times the price increase.
And I know the anecdotes are not evidence. So we're going to look at this a little more. thoroughly to find out exactly what's going on and while we're doing that there's one question that we have to ask and answer is ai destroying the reason that we all own computers if you want to understand that it helps to understand two separate things that we call memory so say i'm editing a video The T9 is what stores the footage.
I can unplug it tonight and the footage will still be on my external hard drive tomorrow. It's always there. It's always safe.
The T9 uses what we call flash storage, but my computer also needs memory to play that footage while I edit. The working memory inside of every PC, every Mac studio, every computer, so to speak, is called RAM, random access memory. AI data centers need to buy enormous amounts of both for two different but very specific jobs.
Now, here's where this impacts whether you'll be able to afford a computer of your own in the future. An AI processor needs data fast enough to keep it busy. It needs to hold on to some numbers while it's calculating others.
And one way to feed it is with high bandwidth memory. Making one type of that memory takes roughly three times as much silicon wafer capacity per bit as making ordinary PC memory on the same process. So think of it this way.
With the same amount of silicon, that factory can make roughly three gigabytes of ordinary PC memory for every gigabyte of that AI memory. So a chip maker can run its factory all day, fill a very profitable AI order, and still have less capacity left for the RAM in your next computer. And the next part is storage.
And this has its own problem. Kioxia, which makes flash memory, said its 2026 production was already spoken for.
You can still buy a drive in a store, but if you're a company looking to place a new order with Keoxia, somebody else is already in line for the flash memory it's making this year. Now that doesn't tell us how many dollars AI added to the price of a single Samsung drive, but it does tell us what a buyer is up against when a factory already has customers lined up for an entire year's output.
Look at what happened to the ordinary RAM. Trendforce found that contract prices for conventional dynamic random access memory or DRAM rose between 93 % and 98 % in the first quarter of 2026. That means the price companies paid for that memory almost doubled in three months, and it's still climbing.
TrendForce expects DRAM prices to go up another 10 to 15 % this quarter. And this year, data centers are expected to take more than 70 % of the high -end memory being made.
HP says memory is now about 35 % of what it costs to build one of its PCs. Microsoft says about $25 billion of what it's spending this year comes from parts getting more expensive. And here's the one thing that stopped me.
In October of last year, OpenAI signed early deals with Samsung and SK Hynix. for up to 900 ,000 wafers of memory a month. That could be somewhere around 40 % of all the DRAM made in the world for one company.
Then last December, Micron, one of the biggest memory makers in the world said it was shutting down Crucial. That's the brand a lot of us bought our RAM and SSDs from for years. Micron said it was getting out so it could keep up with the demand from AI data centers.
And before anybody says it's just tariffs, tariffs have pushed up the price of a lot of things. But when Apple and Micron explain why memory costs more, they both point to AI data centers. And then just last month, Sam Altman went on a podcast called Cheeky Pint and talked about people running AI on their own computers.
And he said, from the world's point of view, it's quite terrible to have so much of the world's limited memory sitting in people's computers when those computers only use it about 5 % of the time. People ran with that like he was saying you shouldn't be allowed to own RAM. And I get it.
After the backlash, he walked it back and even changed his ex -account name to Ram Altman. So he was taking it in good spirits. But he said he wants people to put as much memory in their computers as they want.
And the world needs to build a lot more of it. And I get why people are suspicious. I've seen comments with hundreds of likes saying this whole shortage is on purpose to push everybody to the cloud.
When the CEO of the company asking for up to $4 ,000 40 % of the world's memory calls your RAM ownership quite terrible. I understand why people feel that way, but I haven't seen any proof that anybody planned it.
What I have seen is memory makers chasing the biggest orders. And right now that's the data centers. And the question is why do data centers really need that much memory?
There's a passage in the Bible where Jesus says, for which of you intending to build a tower, siteth not down first and counteth the cost, whether he have sufficient to finish it. So let's do that. Let's sit down and count what it's going to cost these companies to get where they say they're going.
And to do that, the first question we have to ask is what are they trying to build? Every AI chatbot you've used is doing one job underneath it all. It's guessing the next word.
You ask it a question and it guesses the first word of the answer. Then the next one, then the next one. And when it guesses well, you get a good answer.
When it guesses wrong, you get nonsense or an answer that sounds confident. but it isn't true. And here's why that guess matters so much.
To guess the next word in a math problem, the model has to follow the math. To guess the next line of code, it has to understand the code. So the better it gets at guessing, the more it has to understand to do it.
That's what these companies are chasing. OpenAI defines AGI as highly autonomous systems that outperform humans at most economically valuable work. In January 2025, Sam Altman wrote, we are now confident we know how to build AGI.
And when OpenAI released GPT -6 Astra in September, its president Greg Brockman said, Welcome to the AGI era. And to be fair, AGI is something that almost every AI company is aiming for.
And AGI isn't the finish line for some of these companies. They want to build something bigger and more powerful, more intelligent than AGI that can help solve almost any problem. But how much compute are they actually going to need for these machines?
And when we talk about compute, What does that actually mean? So before we go any further, I want to slow down on these two words that we keep hearing, computing power.
When somebody says a company needs more computing power, it sounds like a dial you turn up. But computing power is a pile of physical things somebody has to buy. Computing power is how many math problems the computer can do.
And to get more of it, a company has to buy three things. Number one, they have to buy chips. Every AI chip starts as a silicon wafer, a thin disc that gets cut into pieces in a factory.
More computing power means more chips and more chips mean more wafers. The second thing they need is memory. Every one of those chips needs its own fast memory sitting right next to it, holding the numbers it's working on.
That's the high bandwidth memory we talked about just a few moments ago. And it takes about three times as much wafer to make as that memory in your computer. And here's how much memory one of these models takes.
A model is made of billions or trillions of tiny settings called parameters. Think of them as dials and each dial takes two bytes of memory or 16 bits. So a model with 5 trillion of them needs about 10 terabytes of fast memory just to hold it.
That's more than 50 of the newest AI chips before it answers a single question. The third thing they need is electricity. They need power.
Every one of those chips has to be powered and cooled all day, every day, nonstop. So when you hear they need 10 times more computing power, I want you to hear more chips, more wafers, more memory, and more electricity. All of it at the same time.
Now let's go back and talk about something that we've talked about before, because it's going to be very relevant in all of our discussions. Back in 2020, a team at OpenAI that included Dario Amadei, who later started Anthropic, ran an experiment. They trained AI models over and over, giving each one more computing power, and they measured how wrong each one was.
They called that the loss. I'm going to call it the error. And it's important to note that lower is better.
So the less error or the less loss that a model makes, the better. And the paper measured computing power is something called a petaflop day. A petaflop is a quadrillion math problems every second.
That's one with 15 zeros after it. Think of it this way. If every person on earth did one math problem every second, it would take all of us about a day and a half to do what that computer does in one second.
Now let it run for a whole day. That's one petaflop. Here's the rule they found.
If you give the model 10 times more computing power, the error drops by about 11%. And they tested this over and over and the rule has held true. So let's walk through this one step at a time.
Using the paper's own formula, one petaflop of computing gives you an error of 2 .66. So step one, 10 times the computing power. The error drops to 2 .37.
Step two, now you need 10 times that, which is 100 times where you started. The error drops to 2 .11. Step three, a thousand times where you started.
Now the error drops to 1 .88. Step four, 10 ,000 times the compute you started with and you're down to 1 .68. Step five, a hundred thousand times the compute you started with and you're down to 1 .5.
Step six. A million times the computing power you started with and the error is only at 1 .33. You used a million times the compute that you initially started with and you only cut the error in half.
Look at what's happening. Every step costs 10 times more than the step before it, but every step takes a smaller bite out of the error. The first step knocked off 0 .29.
Step six only knocked off 0 .16. And remember, what every one of those steps really means. Every time you multiply the computing power by 10, you're buying more chips, more memory, and more electricity.
Based on how fast the industry's been growing, each of these steps has taken roughly two to three times more chips, more memory, and more power than the step before. with faster chips and longer training runs making up the rest and it never reaches zero and after spending all of that money on compute it never reaches zero and they know this in the very same paper they said themselves the error has to level off before it gets to zero because language has some unpredictability built into it there's always going to be some error left and every time they want to shave off a little bit more the bill multiplies by 10.
They even took a guess at where it might level off somewhere around step four with the model of about 1 trillion parameters. But so far based on industry guesstimates and so forth, they've surpassed that as far as the size of their models and also the ability to cut the error down. So where are they right now?
So here's the problem. Back in 2020, that paper showed everything, how big the models were, how much computing power they used, the formulas, all of it. Three years later, when OpenAI released GPT -4, they stopped doing that.
The report said that because of competition and safety, they weren't sharing how big the model was, what hardware they used, or how much computing power went into it. But in the same report, they said they'd predicted how well GPT -4 would do using smaller models trained with up to 10 ,000 times less computing power. So they know where they are on the ladder.
We don't. And I get part of the reason. If a competitor knows exactly how big your model is and how much you spent, they know exactly what to be.
But think about what that means for the rest of us. If we had those numbers, we could say you've got this many parameters and this much computing power. So you're this close or this far away to get the rest of the way you need this much more money.
And then somebody could say, we don't have that money. This is something we have to walk away from or find a cheaper way to do. And without those numbers, nobody on the outside can say whether this is a good bet.
or a bad one. All we've got is what the companies tell us. So let's put the models on the ladder with what we do know.
GPT -4 came out in 2023. Researchers at Epic AI estimate it was trained with about 230 ,000 petaflop days. That's past step five.
When OpenAI trained GPT -6 Astra, NVIDIA's CEO said it was trained on more than 100 ,000 of NVIDIA's newest Blackwell chips. To give you an idea of how much math that is, let's start small. Say you do one math problem every second.
Now say, again, that every person on Earth does the same thing. That's about 8 billion math problems every second. petaflop is a quadrillion math problems every second for all eight billion of us to match what a petaflop does in one second we'd have to keep going for about a day and a half as we already established one of those blackwell chips does about two and a half petaflops per second so to keep up with just one chip all of us would need about three and a half days to do what it does in one second Now line up 100 ,000 of those chips.
Even at the speed these chips actually run in practice, every single second of training will take all of humanity about 400 years. Every person on earth, one math problem every second, no mistakes, no eating, no sleeping for 400 years just to match one second. Now do that every second for months.
Researchers at Epic AI estimate the whole training run came to about 940 septillion calculations. For all of us to do that by hand, it would take about 3 .7 billion years. That's longer than the Earth has been here probably, or close to it if you're an evolutionist.
And all of that costs somewhere around a billion dollars just to get better at guessing the next word. on the latter that's about 11 million petaflop days step seven so the researchers guess the error will level off around step four and the industry's already somewhere around step seven and the models keep getting better anyway so did the paper hold up the basic idea did.
More computing power kept making better models, and OpenAI used that same math to predict how GPT -4 would turn out. But even Ilya Sutsaker, one of the people who built all of this at OpenAI, said last year that the age of scaling from 2020 to 2025 was great. But now we're back to an age of research.
In other words, just making it bigger might not be enough anymore. And yet, if you ask OpenAI, they'll tell you we're already in the AGI era.
So let's count the bills. Bill number one, chips and memory. Elon Musk Colossus data centers in Memphis now have close to a million of NVIDIA's AI chips.
Even if each one only had 80 gigabytes of fast memory, that's 80 million gigabytes. That's the RAM from about 5 million laptops in one place. And most of those chips have even more.
Bill number two, electricity. A typical data center uses as much electricity as 100 ,000 homes. The biggest ones being built right now use 20 times that.
as much as 2 million homes. OpenAI set a goal of 10 gigabytes of computing capacity in the US by 2029, and it says it's already past it. Last year, an investor named Brad Gerstner asked Sam Altman on his podcast how OpenAI could commit to about 1 .4 trillion in computing when its revenue was reported at around 13 billion.
Altman said they were doing well more than that. And then he said, Brad, if you want to sell your shares, I'll find you a buyer. So what's changed since then?
Their revenue has grown fast. OpenAI made about $13 billion in all of last year. By the end of this September, it was bringing in money at a pace of about $50 billion a year.
But they're still losing money. In just the three months from April to June, OpenAI brought in 6 .7 billion and lost 12 .3 billion so can they afford it not yet altman said last year they're betting that revenue keeps growing and that the thing that could hurt their plan is not getting enough computing power and they're not the only one spending this year alone amazon Google, Microsoft, and Meta are on track to spend somewhere between $720 to $740 billion building out data centers.
But investors are starting to get nervous. In September, OpenAI went out to raise $30 billion. So far, they've gotten about $11 billion at a lower value than they were hoping for.
They're still trying to raise money and they're still worth more than they were in March. But for the first time, the money isn't lining up as fast as the spending. So why would a company that's losing billions every quarter sign up to spend hundreds of billions more.
I don't believe they're buying all this memory just to have it. I think they've sat down and counted. They know every step costs 10 times more than the last one.
They know there's always going to be some error left and they've decided whatever's at the top of the ladder is worth it. That's the only explanation I can come up with for why they want the memory this badly. Here's why I walked you through all of that.
The memory in those data centers comes from the same factories as the memory in your next computer. And they're not buying it once. Every step up the ladder means buying more chips, more memory, and more electricity than the step before.
As long as they keep climbing, your next computer is competing with them for it. Now, remember what Altma said about your RAM? Think about the logic behind it.
The memory in a data center works all day long. The memory of your computer just sits there for most of the time. So if you believe that AI is about to do most of the work that people do, it makes sense to put the memory in the data center.
But that only works if AGI actually shows up. So what if it doesn't? What if they spend hundreds of billions of dollars climbing that ladder?
and the last bit of error just won't move. Then these companies are sitting on warehouses full of chips and memory they bought for a prize that never came. And memory has crashed before.
Back in 2022 and 2023, prices fell so hard that Samsung cut its production by about half. Nobody knows yet which way this thing is actually going to go. Which brings me back to the computer on my desk.
I'm getting ready to buy a Mac studio myself. And part of the reason I want one is so I can run an AI model on my own computer one day. I can't run state of the art models right now on a local machine.
I get it. But the efficiency of the smaller models are increasing. And one day, possibly in the next two years or less, we will be able to run state of the art intelligence.
on our local machines. If you look at the Base Max Studio, for example, that one comes with 36 gigabytes of memory for about $2 ,500. What got my attention was that even the base model costs 500 more than the one in replace.
Now, say you had $2 ,000 set aside for that computer. A generation ago, that covered the base model. But today you're $500 short because you've added a single thing to it and you could be buying it to edit videos like I do.
You don't have to run AI yourself for somebody else's AI data center to affect what you can afford. So here's something I want you to think about before you buy any computer. On a Mac, the memory is built into the chip.
It's unified memory. So whatever you pick when you buy it, that's what you've got for as long as you own it. You can't upgrade it.
If you build a PC, the RAM goes to slots so you can add more later when prices come down. But PC builders are getting hit too. Just last year, a 64 gigabyte kit of RAM went for about $200.
Today, it's around $1 ,250.
That's what concerns me about where this could go. If the computer I can afford doesn't have enough memory to do the work I bought it for, I can still do some of that work online. But now the company running the service gets to decide what it costs me next month.
And if I lease a more capable computer instead, I can do the work at home, but I still have to come up with the money to keep the machine when the lease ends. So to be fair, for a lot of people, paying for AI online? chat gpt claw subscriptions is the cheapest way in right now and i don't suggest that anyone cancels their subscriptions just yet some plans are 20 bucks a month some are 200 some are even 500 but that price only stays put until they decide to change it Because right now, these plans are heavily subsidized.
People are getting thousands of dollars in AI tokens, only spending a couple of bucks for it, and they are complaining, not realizing that a day of reckoning is coming when you're going to have to pay what is actually worth. worth if you rather own it you can still finance it apple car lets you pay off a mac over 12 months with no interest so on a 2500 mac that's about 208 dollars a month and when you're done it's yours or you can buy it used that's another thing that a lot of people don't look at Companies lease laptops for three years and then sell them off and you can find good ones for a few hundred bucks on sites like eBay and other places.
If you don't need the newest thing, that might be the smartest way to own a computer right now because the price of RAM is only going to increase over the next few years if you ask me. And today's prices are as good as it's going to be for a while, in my personal opinion. So if you haven't guessed it by now, here's what might surprise you.
I think that owning is worth more right now. Last month, Apple raised its trade -in values and a Mac studio can get you up to $1 ,700 on a trade -in. And if things keep going the way they have been going for a while.
and the AI companies continue to purchase the memory chips and the RAM, and there continues to be a very heavy demand for it, then what's going to happen to any computer that you purchase today, or specifically the RAM inside of your computer, it is going to appreciate, it is going to increase in value. So no matter what financial situation you find yourself in, If you work online, if you have an online business, if you do anything where you need a machine that's capable of handling your work, then you're going to need memory.
And if your machine is even a little bit out of date or you're feeling a little bit of bottleneck, now is the time to try to find a way. to get a machine that can handle your workload for the next couple of years comfortably at today's prices. Because if you wait until tomorrow to try to fix that problem, you are in for a very sad awakening when you've realized how high the prices are going to be compared to today.
If you bought a Ram or an S -turn drive last year or even two years ago, I wonder what you're seeing in the market. Do me a favor and just look it up to see what it costs now and compare it to what you actually paid for to see if the price went up or went down.
And I'm pretty sure you're going to find out what I found out. I want you to think about this before I end the video. If data centers keep buying up all of the memory and the price of memory continues to climb, take in the budget that you have today that you can spend.
What type of device will you be able to afford in three years? Maybe you can afford a Mac studio today. but you might not even be able to afford a mac mini three years from now and maybe not even an iphone itself maybe you're going to have to use apple's upgrade program to lease your device because they see how bad it's actually going to get i don't know these are just some thoughts that i'm having but if you got value out of the video make sure you hit the like button subscribe to the channel and as always take care have a great day and i'll see you in the next video
The Hook
The bait, then the rug-pull.
The title says buy before it's too late, and the opening frames it through Apple's new leasing program: a $2,000 MacBook Pro for $39 to $59 a month that you never own. The real story is the month before, when Apple blamed rising Mac and iPad prices on AI data centers eating the world's memory supply.
Frameworks
Named ideas worth stealing.
11:19list
The three things computing power buys
Chips (cut from silicon wafers)
Memory (high bandwidth memory beside every chip)
Electricity (power and cooling, nonstop)
Computing power sounds like a dial, but every 10x increase is a purchase order for more chips, more wafers, more memory and more electricity at once.
Steal forexplaining AI infrastructure costs to a non-technical audience
14:09model
The scaling ladder
1 petaflop day: error 2.66
Step 1, 10x: 2.37
Step 2, 100x: 2.11
Step 3, 1,000x: 1.88
Step 4, 10,000x: 1.68
Step 5, 100,000x: 1.50
Step 6, 1,000,000x: 1.33
Using the 2020 scaling-law formula, each rung costs ten times the last while removing a smaller slice of error. The industry is estimated to be around step seven.
Steal forvisualizing diminishing returns in any compute-heavy bet
03:38concept
Two things we call memory
Flash storage: keeps files with the power off
RAM: holds what you're working on right now
AI data centers buy both, for different jobs, which is why SSD and RAM prices are rising together.
Steal forhardware explainers
09:14concept
Count the cost
Borrowing Luke 14:28, the video sits down and tallies what reaching AGI would cost in chips, memory, power and dollars before judging whether the labs can afford it.
Steal forstructuring any 'is this bet worth it' analysis
CTA Breakdown
How they asked for the click.
VERBAL ASK
31:09subscribe
“If you got value out of the video make sure you hit the like button, subscribe to the channel.”
Soft, single-line ask after the closing thought experiment. The paid community link lives only in the description.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Meta gave away a frontier-class open model, Apple supplied the hardware to run it, and the case that follows is that AI ends up a device you buy once, not a subscription you rent.
Corey McClain downloads a free open-source clone of After Effects, then hands it to a Claude Code agent with zero training on the app and watches it build a real edit.
OpenAI rebuilt voice mode around GPT-5.6 and GPT-6 Astra, gave paying users up to unlimited daily hours, and made the case that talking may replace typing as the default way to use AI.
A creator runs OpenAI's new writing-style feature and his own custom voice skill against the same AI-written passage, then diffs the results line by line.
00:35
01:22
01:43
02:10
02:57
03:21
03:50
04:08
04:56
05:19
05:43
06:07
06:54
07:13
07:42
08:05
08:53
09:16
10:04
10:28
10:51
11:15
12:02
12:26
12:50
13:13
13:37
14:01
14:25
14:48
15:36
15:59
16:23
16:47
17:10
17:34
17:58
18:22
19:09
19:33
19:56
20:20
20:44
21:07
21:55
22:19
22:42
23:06
23:30
24:17
24:41
25:04
25:28
26:16
26:39
27:03
27:27
27:50
28:38
29:01
29:23
29:49
30:13
30:33
31:00
28:11
10:45
17:44
17:58
08:03
10:08