Nvidia's 2027 rack hike is a floor above 15%, and memory makers set it

Run the pass-through backwards and the memory move gets a size. Nvidia's gross margin is 75 per cent ^[6], so production cost is about a quarter of the price. If a 15-point price rise is memory cost handed on dollar for dollar, the cost side has moved roughly 60 per cent, and gross margin on that system lands near 65 ^[13]. That is a company-wide margin held against a system-level price, so read it as a scale check rather than a P&L. Nvidia did not respond to requests for comment ^[4], and the reporting rests on people who asked not to be identified ^[1].

Anything more precise than the floor is interpolation. The increase varies by chip generation and by memory configuration ^[2], so a single figure such as 17 per cent is a midpoint inside a range that is open at the top.

Where it lands is easier to cost. Exponential View priced serving Kimi K3, 2.8 trillion parameters and 1.4TB of weights, on a 72-GPU GB200 NVL72 rack: $3-4 million to buy and install, 120kW running continuously, or about $7 million a year if rented ^[9]. The weights are free, the rack holding them is not ^[17].

That runs against the way the demand side has been growing. The same research puts token elasticity at 12 to 18 per cent more volume for each 10 per cent price cut, with Demirer et al measuring about 11 per cent, an elasticity of -1.11 ^[10]. Total spend rises on a price cut, but thinly. Cheaper tokens have to be funded from somewhere, and from early 2027 the hardware line moves the other way ^[1], at a level set by the three DRAM suppliers ^[5] that the hyperscalers' own accelerator programmes also depend on ^[8].