A platform engineer printed Nvidia's Vera Rubin spec page on September 1 and sized an inference rack around it.
22 TB/s of HBM4 bandwidth per GPU, 3.6 TB/s of NVLink 6.
Open the same page this morning and those numbers read 19.2 TB/s and 3 TB/s.
Wayback snapshots place the edit between September 1 and September 11, and no announcement went out with it.
The rack view is the one that ruins a capacity plan.
Vera Rubin NVL72 memory bandwidth fell from 1,580 TB/s to 1,400, and NVLink from 260 TB/s to 216.
The smaller HGX Rubin NVL8 shape cut furthest, 176 TB/s down to 130, a 26 percent drop for anyone who specced that configuration.
None of this looks like a mistake.
TechPowerUp reports the NVLink Nyquist target came down from 56 GHz to 47 GHz, which eases signal integrity on the SerDes and cabling and helps yields, and HBM4 moved from 11.7 to 9.6 Gbps per pin.
The engineering case is sound.
What was missing was any note saying the figures had changed, and the archived copy is the only proof they did.
Anyone who committed a September budget to that bandwidth number owns the difference, and the procurement deck still shows the slide they signed.
The cheaper way to buy certainty is to buy information first.
Rent RTX 3090 capacity at $0.15/hr and storage at $2.49/TB, benchmark the model on real silicon, then size the commitment against measured throughput.
Buy Now Pay Later, zero upfront, pay after one month, so the test runs before the invoice lands.
Did you size your last cluster on the published spec sheet or on a benchmark you ran yourself?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w44&utm_content=evening
Originally posted on linkedin_personal