Somewhere in the last three years, a story about artificial intelligence became a story about a commodity part that costs a few dollars and sits on a strip of green plastic.
On August 22, 2026, Bloomberg reported that Nvidia has told its largest customers that prices for servers built around its AI chips will rise by more than 15% for systems shipping in early 2027. The increase covers both the Grace Blackwell platform now in the field and the Vera Rubin generation coming next. The customers on the receiving end are the entire demand side of the AI economy: Microsoft, Google, Oracle, Meta and Amazon, plus labs including OpenAI and Anthropic.
Nvidia is not raising the price because the GPU got more expensive to make. It is raising the price because of the memory bolted around it — DRAM, LPDDR and high-bandwidth memory — and because the three companies that make that memory have, for the first time in the history of the AI buildout, more leverage than Nvidia does.
The $2.1 Million Rack Where Memory Is the Line Item That Hurts
The scale of the problem shows up in the bill of materials for a single rack. According to figures reported by TrendForce, a Vera Rubin VR200 NVL72 rack carries a bill of materials of roughly $2.1 million, and memory now accounts for about 29% of it. Nvidia’s internal target for that share is reportedly 20%.
Nine percentage points does not sound like a crisis until it is multiplied by the number of racks the hyperscalers intend to buy. And the direction of travel is worse than the snapshot. Against the previous Grace Blackwell generation, the total cost of a Vera Rubin rack has risen roughly 2.1×. The memory cost inside it has risen 2.5×. Memory is not just expensive; it is the part of the machine getting expensive fastest.
| Vera Rubin VR200 NVL72 rack | Figure |
|---|---|
| Total bill of materials | ~$2.1 million |
| Memory share of that BOM | ~29% |
| Nvidia’s preferred memory share | 20% |
| Total cost vs. Grace Blackwell | 2.1× |
| Memory cost vs. Grace Blackwell | 2.5× |
| DRAM in one rack, in phone terms | ~4,500 smartphones |
Bill-of-materials figures as reported by TrendForce; the price increase itself was first reported by Bloomberg on August 22, 2026. Nvidia has not published a price list.
That last row is the one worth sitting with. A single AI rack holds about as much DRAM as 4,500 smartphones. There are a great many racks.
Nvidia Asked for Memory. It Was Offered 60% of It.
The most revealing move in this whole story is not the price rise. It is what Nvidia did to its own product.
Vera Rubin uses SOCAMM modules — a socketed, server-oriented take on the low-power memory that goes into laptops and phones — attached to each Vera CPU. The original design called for 192GB per module. In July, TrendForce reported that Nvidia had halved that to 96GB, cutting the memory in an NVL72 rack from roughly 55TB to 28TB.
The reason given was not thermal, or architectural, or a change of engineering mind. It was allocation. Based on preliminary 2027 production plans from Samsung, SK hynix and Micron, Nvidia is reportedly set to receive only about 60% of the low-power DRAM it estimates it needs.
So Nvidia redesigned the machine around the supply it could actually get. The HBM4 attached to the GPUs themselves — 20.7 terabytes of it per rack — was left alone, because that is the part you genuinely cannot cut. The system memory took the hit.
Why Nvidia Is Now Bidding Against Your Phone
Here is the collision at the centre of this story, and it is a genuinely new one.
High-bandwidth memory has always been a specialist AI part. Nobody buying a laptop was competing for it. But SOCAMM is built from LPDDR — the same low-power DRAM family that goes into smartphones, tablets and thin notebooks. By moving system memory onto LPDDR, Nvidia moved the AI buildout directly into the consumer memory market.
Counterpoint Research projects that the Rubin platform alone will require more than 6 billion gigabytes of LPDDR in 2027 — more than Apple and Samsung combined consume for every phone, tablet and laptop they ship.
That is one product line, from one vendor, outbidding the two largest consumer electronics manufacturers on Earth for the same parts. Data centres already absorb an estimated 70% of global memory production. Rubin is the generation that makes the overlap direct rather than indirect.
Three Companies, and 2027 Is Already Sold
Nvidia has no leverage here for a simple structural reason: there are three suppliers of consequence in the world, and every one of them is full.
Samsung, SK hynix and Micron have reportedly sold out their entire 2027 DRAM and HBM capacity. Customers are being allocated only 60% to 70% of the volumes they asked for. Prices for those shipments are, in many cases, still to be set closer to delivery — which in a market this tight is not a negotiation so much as a queue.
The supply-side answer is fabs, and fabs are slow. New capacity from Micron and SK hynix does not reach volume production until 2027 at the earliest. Gartner expects the crunch to run at least through the first half of 2027. Deloitte does not expect meaningful new capacity before 2029 or 2030. The chief executive of SK hynix has said he expects 2027 to be the worst supply year in the industry’s history.
None of this is a shortage in the ordinary sense of a factory breaking down. It is a deliberate reallocation. The memory makers moved wafer capacity toward high-margin HBM for AI accelerators, and the capacity that used to make the DDR5 in a gaming PC went with it.
You Already Paid Part of This Bill
The consumer half of this story started before the Nvidia memo, and most people experienced it without being told what it was.
On June 25, 2026, Apple raised prices on every Mac and iPad it sells, by up to $300. The MacBook Air went from $1,099 to $1,299. The 14-inch MacBook Pro rose $300 to $1,999, and the 16-inch to $2,999. Apple attributed the increases to component costs, memory and storage foremost among them. It was Apple’s first broad hardware price rise in years, from a company that historically holds a price point and adjusts the specification instead.
Apple was not alone or even first. Dell began lifting commercial PC prices by 10–30% in late 2025. Microsoft raised Surface prices. Lenovo, Acer and Asus have all described 2026 increases as unavoidable.
In February, Gartner put numbers on where this ends. Its forecast expects combined DRAM and SSD prices to climb roughly 130% by the end of 2026 against 2025 levels, pushing PC prices up 17% and smartphone prices up 13%.
| Gartner forecast, 2026 vs. 2025 | Change |
|---|---|
| Combined DRAM + SSD prices | +130% |
| Average PC price | +17% |
| Average smartphone price | +13% |
| Worldwide PC shipments | −10.4% |
| Worldwide smartphone shipments | −8.4% |
| Memory as a share of PC component cost | rising to 23% |
Gartner press release, February 26, 2026. Shipment declines are Gartner’s projection for calendar 2026.
The forecast that should worry anyone who buys a computer on a budget is the one furthest out: Gartner expects the sub-$500 PC segment to disappear entirely by 2028, as memory rises to 23% of component cost. A price increase spread evenly across a market is an inconvenience. A price increase concentrated in a fixed-cost component is a floor, and floors remove the bottom of the market first.
The Cost Curve Turned Over
For three years the reassuring number in AI has been the price of intelligence. Cost per million tokens fell relentlessly — a frontier-class model got roughly 80% cheaper in a single 48-hour stretch last July. The implied promise was that this is a technology whose unit economics improve on their own, the way they did for transistors, bandwidth and storage.
That promise held while the binding constraint was chip design, which improves. It does not hold when the binding constraint is fab capacity for a commodity, which improves at the speed of construction permits and five-year capital plans.
Every hyperscaler capex model for 2027 assumed a certain cost per rack. That number moved 15% in a memo, and the thing that moved it is not a component anyone can innovate around inside a year. Nvidia can design a faster GPU on an 18-month cadence. Nobody can design a new DRAM fab in less than about four years.
What To Watch
- Whether the increase actually holds. A price relayed through server builders in August 2026 for hardware shipping in early 2027 is a forecast wearing a memo’s clothing. Watch Nvidia’s gross margin guidance for whether the company expects to pass the cost on or absorb some of it after all.
- Whether the hyperscalers cut units or eat the cost. Microsoft, Google, Meta and Amazon can fund a 15% increase. The question is whether they buy the same number of racks at a higher price or the same dollar amount of racks — and the second answer would take roughly a seventh out of 2027’s compute growth.
- The SOCAMM configuration at launch. If Vera Rubin ships with 96GB modules, the memory shortage successfully shrank Nvidia’s flagship product. That is a more durable fact about this market than any price.
- Consumer DRAM contract prices into Q4. Reporting already suggests the consumer surge is cooling as buyers hit an affordability ceiling, even while AI demand keeps climbing. Demand destruction in phones and PCs is the only near-term mechanism that frees up capacity, and it works by making people not buy things.
- Anyone announcing a fab. The single most consequential AI announcement of the next year may not come from a lab at all. It may be a memory maker committing capital to capacity that arrives in 2030.
The Takeaway
Nvidia told its biggest customers that AI servers get more than 15% more expensive in early 2027, and the cause is memory rather than the accelerator. Memory is now about 29% of a $2.1 million Vera Rubin rack, its cost has risen 2.5× in one generation, and Nvidia halved the system memory in its own flagship because suppliers offered it roughly 60% of what it asked for.
Underneath that is a market with three sellers, no spare 2027 capacity, and no new fabs before 2029. The same shortage already moved a MacBook Air from $1,099 to $1,299 and is forecast to erase the sub-$500 PC by 2028.
The AI boom has spent three years being described as a race between models. The constraint that finally bit is a strip of low-power DRAM, and the companies that make it are, for now, the only ones in this industry who get to name their price.