Memory Takes 68 Cents of Every Cloud Capex Dollar - Nvidia Just Showed You the Denominator

The Reported Number

Nvidia's Q2 FY27 revenue came in at $96.2 billion, up 106% year over year and 18% sequentially. Data Center accounted for $89.0 billion of that, up 117%. Q3 guidance is $108.0 billion, plus or minus 2%, against a consensus that had been sitting near $92.2 billion for the quarter just reported.

Analysts were not modelling this conservatively and still came in low. That matters more than the beat itself, because the people whose job is forecasting this demand keep underestimating it, and your supply plan is downstream of the same estimates.

Where the Money Actually Goes

Here is the figure to carry into your next planning meeting. TrendForce projects DRAM and NAND Flash at 68% of major cloud service provider capital expenditure in 2027.

Sit with that for a second. Not GPUs. Not networking, power or cooling. Memory takes roughly two thirds of the budget at the largest buyers in the market.

That reframes what you're competing against. You aren't bidding against another industrial buyer for the same DIMM. You're bidding against an organisation that has decided memory is the majority of its infrastructure spend and has the balance sheet to act on it.

The Interconnect Note Worth Catching

Buried in the same release: Nvidia says Spectrum-6 switch systems supporting pluggable and co-packaged optics are arriving at gigascale AI factories.

Co-packaged optics moving into production deployment changes optical module sourcing from a commodity purchase into a qualification exercise. If your networking content follows the hyperscale roadmap even loosely, the ownership, test coverage and approved-alternate questions want answering before your next build freeze, not after.