Nvidia’s solid second-quarter earnings report showed how good it is to be the seller of the best pickax in the town with the big gold mine.
All of the big AI players are racing to pour hundreds of billions of dollars into ever-larger data centers to out-gigawatt their competitors in the space. And those data centers are usually stacked to the rafters with Nvidia’s powerful GPUs.
Meta, OpenAI, and xAI are all likely handing over many of those billions to Nvidia. For the second quarter, Nvidia reported $41.3 billion in revenue in its “Compute & Networking” segment — which includes all the gear it sells for data centers. That’s up 56% year on year.
Concentration of revenue
In Nvidia’s 10-Q filing, the company gave us a peek at who is buying all this gear, and how much — but without naming names.
“Customer A,” the largest single customer in this segment buying directly from Nvidia, represented 23% of sales for the second quarter, which works out to about $9.5 billion.
Which company could this be? We know that both OpenAI and Meta are currently building what might be the two largest AI data centers in the world. Meta’s $10 billion (or maybe it costs $50 billion?) Manhattan-sized “Hyperion” data center, currently under construction in Richland Parish, Louisiana, has been described by Meta CEO Mark Zuckerberg as “2GW+” but may scale up to 5 gigawatts, with several other multi-gigawatt projects on the horizon.
And let’s not forget OpenAI’s Stargate project, a $500 billion partnership with Nvidia, Oracle, and SoftBank that’s currently under construction in Texas. OpenAI recently announced a plan to develop 4.5 gigawatts of data center capacity at the Abilene, Texas, Stargate site.
The number of GPUs that will be needed to fill these data centers is absolutely enormous. We don’t have an easy way of knowing if OpenAI and Meta buy directly from Nvidia, but it’s hard to imagine that Customer A’s hardware won’t end up in one of the two megaprojects.
We’d be remiss not to mention Elon Musk’s xAI. xAI built its Tennessee “Colossus” data center in record time, and Musk has thrown out some crazy numbers regarding the number of GPUs he wants for his data center’s expansion.
Last month, Musk tweeted:
“The @xAI goal is 50 million in units of H100 equivalent-AI compute (but much better power-efficiency) online within 5 years”
Customers B and C aren’t messing around either, spending $6.6 billion and $5.7 billion, respectively, last quarter.
Together, Customers A, B, and C add up to 53% of second-quarter revenue for the Compute & Networking segment, or about $21.9 billion.
Is having so much revenue coming from so few customers a bad thing?
Not necessarily! The argument goes that having tight relationships with deep-pocketed customers allows a company to sell more stuff with less friction and more efficiency.
But the potential downside of being reliant on a handful of big spenders is that huge revenue streams could disappear overnight if one day a big customer starts using its own chips, or decides they are spending too much on AI data centers.
And, of course, there’s always the risk of a mercurial CEO just deciding to stop using your products for any reason.
For the time being, this doesn’t seem to be a problem. Everyone wants Nvidia’s latest chips, it can’t make enough of them, and everyone is planning to buy as many as they can get their hands on.
Source link
