Google and Meta will use Nvidia networking gear while building rival chips, Nvidia says

The incentive is visible in Nvidia's revenue mix. Of the $96.2 billion quarter ^[7], $89 billion was data center ^[8], a 92.5% share ^[14]. Cryptopolitan reports that hyperscalers such as Microsoft account for more than half of quarterly revenue ^[19], and if that share applies to this quarter the floor is $48.1 billion ^[20]. Among those buyers, Meta rents Google's TPUs and ships its own MTIA designs ^[3], Google sells TPUs to outside customers ^[4], and Microsoft showed its Maia 300 inference chip ^[6]. Each GPU one of them stops buying has to be replaced in Nvidia's sales by something, and the cards and switches that wire up whatever the buyer picks instead are the obvious candidate.

We think the evidence supports a narrow version of the thesis: Nvidia has a place in the network of two of its sharpest custom-silicon rivals. Nvidia's lead sentence, as reported, puts both under Spectrum-X and related interconnect gear ^[17], and Cryptopolitan's FAQ goes further, describing both companies as standardizing on Nvidia's interconnect ^[18]. Meta's arrangement fits that word. Google's support for a transport protocol on upcoming cards fits it less. Two other readings fit the published facts. A hyperscaler can support a card or protocol at small cost and small commitment, which would make this an interoperability gesture. Or the adoption is a bridge until the buyers' own designs catch up. The account includes no order sizes for either company, so it cannot separate those readings from ours.

What would prove the narrow reading wrong is a growth rate. Total revenue rose 106% ^[7] from roughly $46.7 billion a year earlier ^[15]. If Nvidia ever reports a networking line and it grows well below that pace while Google and Meta add custom silicon, the layer is losing share. Management's forecast that the coming fiscal year will be supply-constrained ^[9] complicates any early read: when supply is the limit, a hyperscaler's choice of Nvidia gear says as much about availability as about preference.

The Vera Rubin rack is already in the field. Nvidia began shipping it in August ^[11], Supermicro announced its own Vera Rubin rack shipments on September 23, and a week after that CoreWeave said one of its customers had production workloads running on the rack ^[11]. The optics announcement involves partners making the hardware too. The module Nvidia announced with Cisco runs at 400 gigabits per lane, which at 12.8 terabits in total is 32 lanes ^[16]. Against 200-gigabit optics, the companies claim a fourfold gain in bandwidth density and roughly a 1.3-times gain in power efficiency ^[10]. TE Connectivity and Amphenol are partners on the components ^[10].