We have no inside knowledge of Nvidia or Coreweave. But, we write this based on the 10K and 10Q from Nvidia, the news and research from FactSet. We bring to the table almost two decades in Corporate Banking for Ayesha and over two decades in Technology for Mayhem.
We know there have been numerous allegations about some big conspiracy theory surrounding Nvidia and we want to clear up some of the confusion.
Data Center vs Consumer Products
Nvidia has been driving significant innovation into technology that helps to facilitate an acceleration of complex, mathematically intensive data sets. Using graphical processors, Nvidia found that its own technology could outperform CPUs. From there the company built an innovation advantage by creating CUDA, which eventually become the software ecology that builds a relatively durable competitive moat around Nvidia’s data center offerings.
Data center sales are much more profitable than consumer products for Nvidia. With gaming in decline, Nvidia’s A100 and H100 sales to data center customers more than made up for the difference. Data center sales have become the largest driver of revenue and revenue growth for Nvidia, and these products are extremely high margin. Meaning that Nvidia spends less and makes more on these products dollar for dollar made than, say, video cards for gaming.
GPUs have Specific Advantages vs CPU for Specific Workloads
Why GPUs are Critical for Data-intensive Computing
Parallelism: Many data-intensive tasks can be broken down into smaller sub-tasks that can be executed simultaneously. GPUs excel at this kind of parallel processing.
High Throughput: The architecture of a GPU allows for more data to be processed per unit of time compared to a CPU, especially for tasks that can be parallelized.
Specialized Libraries: Libraries like CUDA for NVIDIA GPUs and OpenCL for other vendors have made it easier to offload data-intensive computations to the GPU.
Machine Learning and AI: Training models often involve a lot of matrix operations, which GPUs are highly efficient at handling.
Big Data Analytics: Processing and analyzing large data sets can be expedited using the parallel computing capabilities of GPUs.
Scientific Computing: Simulations involving complex calculations can be significantly sped up using GPUs.
Real-Time Processing: GPUs can process large streams of data in real-time, making them ideal for applications like video processing and gaming.
Cost-Efficiency: While GPUs may be expensive, the computational speed-up they offer for specific tasks can make them more cost-effective in the long run.
Empirical Evidence
Speed-Up in Machine Learning: In a study published in 2017, researchers found that using GPUs can speed up the training time of deep learning models by a factor of up to 50 compared to using CPUs (source).
Big Data Analytics: According to a paper published in the ACM Digital Library, GPUs can accelerate big data analytics tasks by 10x to 100x compared to traditional CPU-based systems (source).
By providing high throughput, parallelism, and specialized libraries, GPUs have become a critical component in data-intensive computing, machine learning, big data analytics, and more.
Nvidia’s Secret Sauce: CUDA
NVIDIA's Compute Unified Device Architecture (CUDA) is a parallel computing platform and application programming interface (API) model that allows software developers to use a CUDA-enabled graphics processing unit (GPU) for general-purpose processing. While other APIs like OpenCL (Open Computing Language), OpenACC, and Vulkan also offer GPU-accelerated computing, CUDA has certain advantages that have made it a popular choice in various domains. Here are some of the key advantages of CUDA:
Performance Benchmarks: Studies often show CUDA outperforming other APIs in specific tasks. For instance, a research paper comparing CUDA and OpenCL for financial computing found that CUDA was up to 30% faster (source).
Feature Comparison: A comprehensive study comparing CUDA and OpenCL stated that while OpenCL provides more flexibility by supporting multiple hardware vendors, CUDA provides a richer set of features and often better performance (source).
Active Development: NVIDIA actively maintains and updates CUDA, ensuring that it takes advantage of the latest hardware features.
Strong Community: Due to its widespread adoption, CUDA has a large community of developers, contributing to forums, open-source projects, and academic research.
Widespread Use: CUDA is used across various industries including finance for risk modeling, in sciences for computational biology, and obviously in machine learning and data analytics.
The Idea is that Nvidia has been Channel Stuffing
Channel stuffing is when you bring forward sales bookings or use accounting practices to show that the sales occurred in that quarter in order to beat the quarter’s estimates. Channel stuffing is by no means illegal unless you are faking the sales. It’s just a matter of timing. But, here’s the thing ultimately it catches up to somehow or the other since you’re reporting on a quarterly basis.
Channel stuffing is usually done by private businesses when they need to show results to their bankers. But, we're not saying that Nvidia is doing that. If anything the opposite is true. Huge companies, like Microsoft, Amazon, and Tesla are buying so many Nvidia GPUs that there is a significant backlog.
Nvidia’s Revenue Recognition
There’s been discussion surrounding Nvdia’s revenue recognition. We see nothing sinister there either.
They recognize product revenue when the product is transfered. This is the normal way accounting is done for products. When title of goods transfer, you book at as revenue in the income statement and get paid cash. So you increase your cash balance. If you do not get cash and sell on credit, you book it as accounts receivables. They do this for software as well.
As far as continuing obligations go, such as licensing - these are booked as deferred revenue or unearned revenue. As the obligation gets filled, they are moved over to revenue on the income statement. Most people do this on a yearly basis for contracts, or based on milestones.
Nvidia actually does it in a slightly weird way, in that they recognize all the revenue at the end of the contract. In my mind, that is actually more conservative and leads to lumpy revenues.
In either case, costs are recognized as and when they are incurred, which is exactly what they are doing.
In the case of service warranties, software support, subscription revenues - these are all booked as per usage. Also normal.
We believe two of the issues were:
Revenue growth higher than accounts receivable growth - this would happen due to mismatch in timing for revenue recognition and the deferred revenue.
Massive Gross Margin (c. 70%) because they book revenues first and cost later under the licensing. This is perhaps a misunderstanding. They actually book revenue later.
Where did Coreweave Come From?
Nvidia saw an opportunity to make money from the end product of their chips and the jumped on the train. They have been talking about Coreweave for the last several quarters in earnings calls and conferences. The first news article I found on Coreweave pivoting to data centers and using NVDA chips was from 2021.
Nvidia invested $100 million in Coreweave on 20 April 2023. Nvidia’s net operating cash flow for the first 6 months of 2023 was $9.2B. That’s 2.4% of their cash flows. Their cash balance as of July was $5.78B, after their investment.
There’s no secret here. A quick search of FactSet gave me this information within a minute. Nvidia has active investments in 37 companies. (The investment amounts listed in the table below are the aggregate funding round amounts and not the amounts Nvidia has invested).
Coreweave’s Recent $2.3B Financing Structure and Cash Raise
Coreweave has pledged their chips as collateral for a loan of $2.3B. As a banker, I would jump at this opportunity. In the world we live in, NVDA chips are probably better than gold. Okay I’m exaggerating but, you know what I mean. They are a solid form or collateral and we would consider them raw material inventory. As far as I am concerned, it’s like financing a bale of wheat for a flour manufacturer. Sometime, they would use pre-shipment financing and sometimes they would use post-shipment financing for extra funds. This would be like post shipment financing since the chips are already with Coreweave.
The big issues is: Where is this financing going? Apparently it’s going to buy more chips from Nvidia.
Well, that usually is what companies do with financing, they invest in more inventory or equipment for growth. The question people are raising is that this money will be used to pay Nvidia. And if it is, so what? It’s like you buying a Tesla on credit and taking finance from the bank after the purchase.
And if it’s going to Nvidia to buy more chips, that means that Nvidia get’s more revenue. Because I doubt Nvidia keeps selling to Coreweave on credit without getting some payment. I am hard pressed to see how Nvidia is a loser in all this other than Coreweave going bankrupt and NVDA losing their investment of $100 million.
Coreweave is raising cash based on a $8B valuation. They are out to raise 10% of this which would mean $800million. I suspect this is for Magnetar’s exit.
From what we understand, the bottom line is, Coreweave is a hyperscaler. Despite their beginnings and sponsors, the company was at the right place, at the right time. They have become so popular that many of their GPU-accelerated cloud instances are "out of stock.” So they have a business need to order as much as they can, which is where the need for financing comes in.
Coreweave’s Financiers
Given the size and nature of the financing, I can imagine that this would not be a straightforward arrangements. But, it doesn’t deviate much from what we would do with a syndicated loan or club deal where you would have several banks participating in a big loan. Most of the time one bank does not and should not take on the financing of loans over $200-$300million.
We have
Arrangers (who arrange the loan and negotiate the terms among all parties);
Bookrunners who underwrite the loans and then arrange for other banks to join with their portion of the funds;
Participating financiers or banks who would come in with just their portion of the loan. They just give out their loan amount and take back the repayments + interest so there’s no added hassle
Documentation banks who would be in charge of documentation, collateral, and monitoring the conditions throughout the loan
And we would have the account bank who would actually manage the funds and maintain the account for collection of interest, principal and then would distribute to the other parties.
I’ve tried to designate these roles to each of the parties in Coreweave’s process so you know what’s going on.
Magnetar Capital - Arrangers - Apparently they did something shady during the GFC. Perhaps they are shady. But, they are more a conduit for raising finance and after 15 years they are still around. I don’t really understand the accusations being made here. For what it’s worth, Magnetar just brought the deal to the table. They are arrangers. Each party to the deal will do their own due diligence.
Blackstone Tactical Opportunities - Bookrunner - This is an investment arm of Blackstone. They invest in high risk opportunities and can mobilize funds quickly. From what I can tell, they are not bound by a specific mandate and therefore, are nimble and can invest in anything that makes sense to them. There are plenty of small and big non-bank financial institutions who do this. And Blackstone’s arm here doesn’t necessarily have to feed back to their property arm. They are not a small company and they can very well maintain separate functions.
Coatue and DigitalBridge Credit -Participating financiers- these are the investors. So Blackstone here would act as the lead and these guys would be participating.
BlackRock, PIMCO, and Carlyle - Funds and Accounts manager - This is a bit funny to me. We usually have one bank doing this. There could be more than one bank if we have multiple currencies. So USD bank and foreign currency bank. I don’t think that is the case here. Each of these entities are big enough on their own to manage the entire loan. This is what I would want to find out. But, I doubt it’s anything shady. It could be that they want to keep up relationships with all three entities. Sometimes the reasons are just that simple.
Closing Thoughts
So that’s that. Is Nvidia overvalued? Perhaps. Can the price come down? Yes, sure. Every company has a bearish thesis but, to make stories up where there are none, is not great in the world of finance. Particularly when such stories involve unfounded claims.
We’re sure there are far better arguments to be short NVDA than to simply yell fire in a crowded theater.