During a recent earnings call, Tesla CEO Elon Musk, the world’s richest man, summed up the shortage of Nvidia enterprise GPUs in a few sentences. 

“We’re using a lot of Nvidia hardware. We’ll… take Nvidia hardware as fast as Nvidia will deliver it to us.” Musk said. 

He later added: “I don’t know if they could deliver us enough GPUs…but they can’t. They’ve got so many customers.” 

Musk is just one person in a giant line of customers waiting for Nvidia’s GPUs for AI. 

Nvidia CEO Jensen Huang said that ChatGPT was the iPhone moment for AI. The Nvidia GPU is the iPhone of enterprise hardware. 

Nvidia isn’t prioritizing its limited supply of red-hot GPUs to customers only with wads of cash. Nvidia screens customers for plan, purpose, and workloads before supplying GPUs. 

The goal is to ensure the workloads align with the GPU’s capabilities and that the customer is a good steward of Nvidia’s hardware. 

Customers need to meet specific Nvidia guidelines to be considered for a GPU, said Colette Kress, the chief financial officer at Nvidia, during an analyst meeting at the Citi Global Technology Conference this week. Here are some of those: 

Do you have a PO or purchase order? 

That’s Nvidia’s polite way of saying, “Show me the money.” 

Kress said that a PO provides insight into the company’s plans and Nvidia’s role in those plans. She said the PO helps Nvidia start planning for the customer from its end. 

“We worked with many large companies for many years. They do help us understand the planning process and work, and that is one piece of our process. It says ‘help us in terms of that'” Kress said. 

Intel’s chief financial officer, Dave Zinsner, has said that understanding a customer’s plans opens the door to upselling more products — software or infrastructure — to meet specific requirements. 

Nvidia also sells its AI software, CPUs, and network hardware alongside its GPUs. 

Are you ready to receive the GPU?  

Nvidia wants to know about the computing setup to handle the superhot speeds of GPUs and cooling requirements to handle the heat wave caused by the H100s. 

Setting up data centers is not a quick process; it takes time and planning, Kress said. Some customers are also looking to add computing and networking in the latter stages of the setup. 

“We are looking in terms of exactly when you expect to need us to provide in … your data centers,” Kress said. 

What size are your models, and what compute do you need? 

Nvidia works very closely with some companies to understand their strategic AI plans and the size of the models and compute requirements. 

The company’s A100 and H100 GPUs have been used to train models like GPT-3.5 and GPT-4, which have billions of parameters. 

But for some smaller models, Nvidia will recommend other GPUs. Kress gave the example of L40S based on the model and setup in their data centers. 

“You can take an OEM or ODM server — we probably have 100 of them coming out — that they will be able to put in four L40S [cards] inside of that configuration. That’s a great … server for a small model in terms of training, but also doing the inferencing.” Kress said.  

Where are Nvidia GPUs Going?

 Nvidia is allocating GPUs globally across customers and sectors. 

Nvidia is allocating the most GPUs to cloud services providers, bringing in the largest chunk of the company’s revenue. Google is offering customers the A3 supercomputer, which has up to 26,000 H100 GPUs and 26 exaflops of computing power. AWS only in late July announced its first H100 EC2 instances, and Microsoft soon followed suit with the ND H100 v5-series Azure virtual machines, which came out of beta. 

After cloud providers, the second largest allocation of GPUs goes to consumer internet companies and large enterprises. 

“CSPs are also selling to the enterprises as well as they stand up compute for research, standing up for large universities and also setting up for enterprises,” Kress said. 

Not Ignoring Small Customers

CoreWeave passed Nvidia’s test and is one of the lucky small cloud providers with a couple of H100 GPUs. The cloud provider offers only GPU computing, and Nvidia has taken a small stake in the company. 

“CoreWeave is … specialized in accelerated computing; that is their goal. CoreWeave has also quite some skills in terms of their speed of adoption, their speed in terms of setting things up,” Kress said. 

The cloud provider has worked with large customers on setting up computing infrastructures. 

“They are small, they did have some allocation, but … it’s very small,” Kress said. 

GPU Crunch Will Ease into 2025

Nvidia’s GPUs are flying off the shelf quickly, and that will remain the case until the end of the second quarter in 2024. The inventory levels will remain lean at between $4.3 billion and $5.2 billion from the end of the third quarter of 2024 until the middle of next year. 

“We’ve been fairly flat in terms of the inventory we have on hand at the end of each quarter,” Kress said. 

But Nvidia is working to alleviate the shortages by trying to increase the supply of CoWoS packaging — which brings the memory and chip together — to help out TSMC, which manufactures the GPUs. Nvidia is working with existing and finding new partners to help resolve the stress on TSMC to increase its capacity. 

“We expect [CoWoS] supply to increase each quarter even as we move into fiscal year 2025. And we do expect there to be certain large [partners] step up as we increase the overall CoWoS capacity,” Kress said. 

Nvidia has ramped up its purchase commitments for its GPUs to $15.3 billion in the first quarter of 2024 and $19.3 billion in the second quarter, which points to more GPUs hitting the market. 

“We also have long-term purchase commitments and prepayments with some of our providers to make sure that we can help them as they stand up capacity for them,” Kress said. 

Mark Liu, the chairman of Nvidia’s manufacturing partner TSMC, stated, “Currently, we can’t fulfill 100% of our customers’ needs, but we try to support about 80%. We think this is a temporary phenomenon. After our expansion of [advanced chip packaging capacity (CoWos)], it should be alleviated in one and a half years.”

Moore’s Law Is Dead, So Don’t Buy CPUs

Kress ruthlessly killed Moore’s Law yet again during the talk, indicating the time for CPUs is over and the era of GPUs is here. 

“Moore’s Law dying has really caused a view of “what do we do with our CPU servers? Is that an upgradable solution? Or is this the time to move to accelerated computing?” she asked.  

GPUs provide more bang for the buck and can do more computing in less space than a server room full of CPUs, Kress said. 

 “You have to be able to build the performance … using less energy, doing it faster,” Kress said, adding “GPUs … do that. So we think this is just the beginning.”  

In the meantime, most smaller-scale HPC users may have to be patient and wait for both GPUs and their results.