You Can't Afford Local AI: NO Affordable RAM Until 2027 & Beyond

0:00 This is the machine people buy to run AI models at home. One small box, one chip, 128 GB of memory. When Framework announced it, that exact configuration was $1999

0:11 and it sold out in days. I opened the same page this week. Same box, same chip, same memory. 3449. Nothing about it got faster. It got 72% more expensive and Framework says the

0:25 reason is the memory inside it. So, where did that money go? Samsung just posted the biggest quarterly profit in its history. 89 and a half trillion won. Call it $63 billion

0:35 in 3 months. Now, look at where almost none of it came from. Phones. The mobile division lost money that quarter. 700 billion won in the red. A first for Samsung. Every one of that profit came

0:46 from one division. The one that makes memory chips. The one that makes your RAM. 89.2 trillion of the 89.5. That division ran about a 70% operating margin. 70% on a commodity. On the most

0:59 boring, most part inside your computer. So, follow that margin to the other end of the trade because somebody is paying for it and it is you. This video answers three things. Who bought the memory,

1:11 whether that is a cartel, and what it now costs to own a machine that runs AI at all. Start on the shelf. A 32 GB DDR5 kit cost about $95 before this started. Today, it is 380 to 590. And there is a

1:25 receipt you can check yourself. TrendForce publishes a weekly spot price for a mainstream DDR4 chip holding exactly 1 GB. In November of 2025, it traded at $12.18.

1:37 On the 7th of August, it hit 42.45. Up 249%. 64 GB now runs about $1018. Four to six times what it was inside a year. That 70% margin and your checkout

1:50 total are the same fact from opposite ends. If that sounds like a tracker artifact, take it from a company that buys the stuff. Framework publishes what it pays, $10 a gigabyte in December, 12

2:02 to 16 by February, 13 to 18 by March. In July, it reported LPCAM2 arriving at more than double its previous inventory. The same squeeze reached graphics memory. Consumer GDDR6 went from about

2:15 $2.50 a gigabyte to about $7.50 since autumn of 2025. Nvidia has raised what it charges board partners three separate times this year. And the 5090 still carries a 1999 list

2:27 price. It has not sold out in months, but a memory shortage does not stop at the enthusiast aisle. It reaches every object you own with a chip inside it, starting with the one in your pocket.

2:38 Xiaomi re-priced nine phones in a single month, up 13% on a Redmi. Lu Weibing, who runs Redmi, said memory for the same specification now costs nearly four times what it did a year earlier.

2:49 Raspberry Pi has raised prices twice for the same reason, $60 of it on the 16 gigabyte board. That board now sells around $205 against $120 at launch. Consoles went, too. On the 1st of

3:02 August, Xbox rose $100, 150 on the larger drive, and Microsoft wrote the reason down, storage and memory more than two and a half times higher, with another doubling expected by the autumn

3:13 of 2027. Storage is the same story with a different chip. Phison's chief executive said a 1 terabit NAND part went from $4.80 to $10.70 in 6 months, and that

3:23 every manufacturer had told him 2026 was already sold out. That was November. So, the whole PC market re-prices around it. Dell went up 15 to 20%. Its operating chief saying he had never seen memory

3:35 costs move this fast, and IDC now expects average PC prices 4 to 8% higher this year with shipments actually shrinking. And renting the compute instead does not

3:45 dodge it. In January, on a Saturday, Amazon raised the price of its H200 instances by 15%, while the four biggest cloud companies budget north of $600 billion of capital spending this year.

3:57 So, is this scarcity or is it a decision? In June, a group of consumers and small businesses filed in the Northern District of California and argued decision. Their case rests on a

4:07 structure the industry agrees on. Samsung, SK Hynix, and Micron are roughly 90% of global DRAM revenue by the complaint's count. Three companies, one part every computer needs, barriers

4:18 to entry measured in tens of billions. In a market that shape, you do not need a meeting to land on the same answer. The mechanism they point out is the wafer. High bandwidth memory for AI

4:28 accelerators comes off the same lines as the memory in your desktop and burns considerably more wafer per usable bit. So, one bit of HBM does not cost one bit of DDR5. It costs several.

4:40 Then comes the arithmetic that made it a headline. Up 171.8% year-over-year by the third quarter of 2025, another 50% in the fourth, 93 to 98% more in the first quarter of 2026.

4:54 Compound those and you land near eight times. And here is the sentence the case turns on, quote, "Not one of the three defendants expanded or announced plans to meaningfully expand commodity DRAM

5:05 production capacity. A price signal that violent is an instruction to build." The complaint argues the instruction went unanswered, which is the question worth sitting with. If a sevenfold price rise

5:15 will not make you build more factories, what will? Either the incentive is broken or building was not the plan. I want to be straight about the other side because it is stronger than the headline

5:25 suggest. Antitrust law does not ban companies from reacting the same way to the same conditions. Plaintiffs have to show agreement, not similarity, and two earlier federal DRAM cases died on that

5:36 exact line. What makes this one different is the history behind it. Between 1998 and 2002, these same firms ran a real cartel and the Justice Department proved it. Samsung paid $300

5:49 Hynix 185 executives went to prison. Nothing is proven this time and the filing is dated the 25th of June. But while that gets argued, the supply side moved. On the

6:00 4th of August, a Taiwanese trade paper reported that the whole of 2027 is already allocated DRAM and high bandwidth memory both. July and August are when buyers normally lock in the

6:10 following year. This year the queue closed early. The buyers are the ones you would guess, cloud providers and AI companies signing agreements that run 3 to 5 years, some putting deposits on

6:21 memory that does not exist yet. The report says some firms resorted to begging for high bandwidth supply and paying far more for it. SK Hynix's chief executive had already said it in July.

6:32 Kwak Noh-Jung told Reuters, "We forecast that next year will be the worst year in the industry's history from the supply perspective." That is the head of the second largest memory maker on the

6:41 planet describing his own market as historically broken. On the day his company listed on the Nasdaq. So what happens to consumer memory? Not zero. The reporting says significantly reduced

6:52 and that distinction matters. Phone and PC makers get smaller allocations than this year competing for what is left after the servers are fed. And I want the caveat on the frame not

7:02 buried. This chain starts with one trade paper citing industry insiders. It fits everything else we can verify but none of the three companies has confirmed it. Reportedly is doing real work in that

7:13 sentence. What is not in dispute is TrendForce's published outlook from the 30th of July. DRAM ran about 1 to 2% short across 2026 and they expect that gap to widen in

7:24 2027 with the shortage structural into 2028. So if the three of them will not build fast enough, the obvious hope is a fourth. There is one. China's CXMT is now the fourth largest DRAM manufacturer

7:37 in the world and it stopped being a rounding error a while ago. By the end of this year, analysis from SemiAnalysis puts CXMT near 350,000 wafer starts a month against Micron's estimated 385.

7:50 Its operating margin hit 70% last quarter, within a few points of the incumbents. But, on the 30th of July, seven United States senators led by Chuck Schumer wrote to Apple demanding a

8:00 commitment to stay away from CXMT and YMTC with an answer due by the 21st of August. Both firms sit on a Defense Department list of companies suspected of ties to People's Liberation Army. So,

8:12 the one pool of spare capacity on Earth is the pool Washington is working to wall off, which leaves building it here, and the calendar is brutal. On the 7th of August, SK Hynix approved

8:23 54 trillion won, about 38 billion dollars, for two new fabs. Read the dates. Cheongju breaks ground in February of 2027. First clean room December of 2028.

8:35 Yongin breaks ground that July. First clean room June of 2029. Micron's first Idaho fab lands wafers in the middle of 2027, inside the year the shortage peaks. A memory fab takes 18 to 24

8:48 months, and money does not compress concrete. So, nothing lands in time, which makes this a question about the next few weeks. If you are specking a machine to run models locally, the

8:57 binding constraint is capacity. How much memory you hold rather than how fast you move it. And you can see what happens when capacity gets scarce enough. On Alibaba, there are listings for a 5090

9:09 carrying 96 gigabytes, three times what Nvidia ships at about 5,858 dollars in volume. Nvidia does not make that card. Somebody is desoldering memory and reflowing denser modules onto

9:21 retail boards. So, can you still run AI at home? Yes, but price it the way the market prices it now. On that same framework page, 32 gigabytes is 1269, 64 is 1959, and 128 is 3449.

9:35 You are not really buying a computer anymore. You are buying memory with with a computer attached. So, here is my call, and it is not a hedge. Buy the memory

9:44 now, and buy capacity over speed. The forecasters on record point one way. TrendForce has contract prices climbing through this year. Gartner sees no relief before late 2027.

9:55 Intel's Lip Bu Tan says the industry told him nothing improves until 2028. The counter argument is one signal, and it is not a forecast. The memory maker equity index is down 38% from its high,

10:08 while chip spot prices set records, and that divergence is what the top of a cycle looks like. If that bull is right, you will pay less by waiting. I would still buy now, and buy the bigger kit.

10:19 The thing worth naming is not any one company. It is the idea underneath that in a three supplier market restraint pays better than capacity. Once that is true, no meeting is required for the

10:29 result to look like a conspiracy. The shortage stopped being a problem to solve, and became the product. Which leaves one question I cannot answer for you. The only spare memory capacity on

10:39 the planet is being built in the country we are working hardest to lock out of our supply chain. So, which do you want more? Memory you can afford, or the wall?