Thirty-nine dollars. That's what the memory inside an iPhone 17 Pro cost Apple — one 12-gigabyte DRAM package, quietly soldered in, barely a rounding error against a $1,099 phone.
That same package in the iPhone 18 Pro could cost $145.
Nothing about the chip changed. The chip got scarcer, and scarcity is now the product.
For two years, the thing everyone feared running out of was the GPU. Nvidia was the choke point, the kingmaker, the company you couldn't buy your way past. That fear had a price, and the market just repriced it.
Investors reportedly switched their attention from GPUs to memory chips — and Nvidia reportedly lost market value in the pivot. The bogeyman changed costume. The boring gray memory chip is the thing that's actually short.
Last week we mapped how the memory crunch was reshaping silicon. This week the story has a confirmed victim and a confirmed number: Apple's price hikes are now explicitly pinned to component costs, and Meta's defensive chip build just got a September production date.
Here's the mechanism, stated plainly, because the plainness is the point.
Memory isn't short because demand for phones exploded. It's short because three companies — Samsung, SK Hynix, and Micron, who together control more than 95 percent of global DRAM — decided to make a different, more profitable kind of memory.
That kind is high-bandwidth memory, or HBM: DRAM dies stacked twelve to sixteen layers deep, wired vertically, feeding AI accelerators at close to a terabyte per second. Mobile DRAM manages about 68 gigabytes per second. The gap is why AI chipmakers pay the premium, and why the memory makers happily take it.
The margins tell the whole story. HBM commands gross margins around 60 percent. Standard mobile DRAM earns roughly 40. Per wafer, HBM brings in three to five times the revenue of conventional DDR5.
So the makers aren't failing to produce your phone's memory. They're choosing not to.
Now the part that makes this mathematically irreversible inside your buying window — the 3-to-1 conversion ratio.
Producing a gigabyte of HBM takes three to four times the wafer capacity of producing a gigabyte of mobile DRAM. Micron has reportedly disclosed the consequence: every wafer committed to HBM deletes the equivalent of three wafers' worth of conventional memory from the market.
This is not a knob that turns both ways on demand. HBM runs on dedicated lines — separate tools, separate masks, a separate packaging step called CoWoS that SK Hynix has reportedly already committed through 2026. You cannot swing that capacity back to phone memory next quarter because the physical equipment doesn't do both jobs.
Every AI chip built physically subtracts phone memory from the world. That's the sentence to remember.
Every AI chip built physically subtracts phone memory from the world. That's the sentence to remember.
The result shows up in the contract prices. Conventional DRAM jumped roughly 90 to 95 percent in the first quarter of 2026, per TrendForce, with another 58 to 63 percent projected for the second. Tim Cook called it a "hundred-year flood" — something he hadn't seen in over 40 years in the industry.
So who eats that cost?
Not the memory makers — they're the ones getting rich. Not Apple, not entirely: Cook told the Wall Street Journal that price increases are "unavoidable" despite efforts to absorb chip costs that have quadrupled since last year. He also called the situation "unsustainable," which is a CEO's way of telling you the bill is coming to your door.
It flows downhill. It always does.
The companies with the deepest pockets buy their way around it. Meta is a case study in doing exactly that.
In September, Meta is reportedly on track to start producing the latest version of its own AI chip — the MTIA program, designed with Broadcom, built at TSMC — specifically to lower its dependence on Nvidia GPUs. And critically, it's reportedly buying its RAM directly from Samsung.
That's the same defensive crouch we described in our OpenAI piece: everyone in AI is now reacting to a scarcity they didn't create. OpenAI is building an inference chip with Broadcom. Anthropic is reportedly weighing its own with Samsung. When the giants build and buy their way to the front of the memory line, someone's standing at the back.
Apple appears to be at the back for the component that matters here.
TechInsights estimates that $39 DRAM package climbing to $145 in the iPhone 18 Pro — a single component, up more than $100. By one estimate cited to the WSJ, Apple would need to add $270 to the next iPhone Pro just to hold its margin intact.
Within 24 hours of Apple raising prices across Macs, iPads, HomePods, and the Vision Pro on June 25, IDC's Nabila Popal revised her iPhone 18 Pro forecast upward — to as much as $200. The read was simple: Apple spared the iPhone in that round, and the scale of the other hikes signaled its ability to absorb memory costs was exhausted.
That's your $200. It isn't a gouge. It's the consumer end of a wafer allocation decision made in three fabs you'll never see.
Now zoom out, because this isn't just an Apple problem — it's the shape of the whole market.
Omdia forecasts the global average selling price of a smartphone jumping from US$467 in 2025 to US$565 in 2026. That's a 21 percent rise, which Omdia says is the largest annual increase in both percentage and dollar terms the industry has ever recorded.
At the same time, shipments fall 12.2 percent — 152 million fewer phones. Yet the total value of the market grows 6.1 percent. Fewer phones, more money. That's premiumisation by force.
The clearest casualty is the cheap phone. Sub-A$145 handsets are being abandoned by the global brands to smaller regional players, because at these component costs, a budget phone barely earns its shelf space. A meaningful volume recovery isn't expected until 2028.
So here's the buyer's actual decision, stripped of comfort.
If you're eyeing an iPhone 18 Pro Max, or any flagship launching this fall, the price is going up, and the reason is structural rather than seasonal. Memory price rises are expected to moderate in the second half of the year — but moderate is not reverse. Omdia is clear the costs stay structurally elevated.
Waiting for prices to fall is a losing bet through 2026. According to one analysis, the one thing that could genuinely reset the curve is China's CXMT bringing on enough new capacity to become a pricing anchor — and even the optimistic case for that is roughly 18 months out and hostage to equipment restrictions.
Which leaves the honest advice: if the phone you have works, holding it is not a hardship. But if you're going to buy in this cycle, buy on the earliest terms you can, because the calendar is not on your side.
The GPU was never the thing that was going to cost you personally. The gray chip was. Reportedly, it took the market two years — and Nvidia's share price — to admit it.