Description

Preamble: NVDA was posted earlier as a short here by glgb913. I am writing from the other side, to present my long thesis on NVDA. The stock has rallied hard since our underwriting and pitching it at this time is seemingly difficult, especially given the YTD run-up and outperformance relative to the rest of the market. We are long-term focused patient investors, and I can’t find a better idea to write up before the submission deadline here, illustrating my variant perception of the business. I editorialize most of the technical concepts to help advance the debate on the business fundamentals. (Read the disclaimer and risks sections towards the end of this post as you conduct your own due diligence. This post is not a recommendation to buy or sell a security but rather intended for a discussion of key bull and bear debates, based on research of public information, including company websites and regulatory filings. This post does not constitute professional or investment advice. Past performance is not indicative of future results. Consult with a licensed professional if needed before making any investment decisions. Investing in public markets can be subject to severe short-term and long-term risks, including resulting in permanent loss of entire invested capital. This post represents my personal opinions and not that of any other entity. I and/or entities that I advise can change views and positioning, at any point of time for any reason, or lack thereof, without any notice. I and/or entities that I advise are not and will not be held liable for any outcomes, including any indirect or consequential damages. I can be 100% wrong. This article has no guarantee to be accurate, complete, or current, and cannot be relied on as such. Please conduct your own due diligence.)

I believe in a concentrated yet balanced portfolio, including some best-of-breed names along with some structural hedges with other ideas, targeting a risk-adjusted positive skew. Nvidia was underwritten as a key ingredient for our growth exposure over the next 5 years in this new compute cycle. I’d argue it is much less of a cyclical semi and has better structural moats than several software names. Given a product-cycle-driven and potentially catalyst-rich backdrop in a new compute regime, I go over my investment thesis, providing some context around the underlying tech. I address a set of key bear arguments and hope this is helpful to think through the key debates. 

Investment Thesis

(i) Nvidia GPUs are general-purpose accelerators that should continue to take share in datacenter.

First, here’s the context. Per an empirical law proposed by Gordon Moore (RIP), transistor count on semiconductor chips doubled every two years until recently. Another empirical law, Dennard Scaling was based on power density being constant when transistor size decreases, so clock frequencies scaled higher, and chips got faster. But that stalled after 2005 because of heat dissipated at small dimensions.

Then came the concept of multiple cores (chip components that execute instructions). Central Processing Units (CPUs) use multi-cores (128 cores in some latest ones in the market) and target single-thread performance. Graphics Processing Units (GPUs) use many smaller cores (some NVDA GPUs have 10,000+ now) and were built initially for gaming applications. But the main difference is that GPUs are designed to maximize throughput with single-instruction multiple-thread (SIMT) architecture, which makes parallel processing possible [1]. So now you can simultaneously render pixels on the screen to show you that cat, everything everywhere all at once, without having to wait in a serial fashion i.e. one pixel at a time.

Adopting a fabless model with foundry partners such as Taiwan Semiconductor (TSMC), Nvidia developed GPUs demonstrating high throughput, measured as 8-bit tera operations per second (TOPS), doubling every year from Kepler (2012) at 4 to H100 (2022) at ~4000, more than that predicted by Moore’s law (doubling every 2 years).

A ‘big bang’ for the field of AI happened in 2012, when a grad student team from U. Toronto won the ImageNet, a major computer vision contest, with a model trained on NVDA GTX 580 GPUs given their programmability and high floating-point performance on the Fermi architecture. (One of the researchers later became the cofounder of OpenAI.) NVDA then developed more general-purpose (GP) GPUs and improved chip architectures over the last decade. Tensor cores were introduced in Nvidia’s Volta (2017) to exploit sparsity, a feature of matrix multiplications involved in training large language models. Tensor cores offer lower precision (good enough outside of supercomputing), and hence higher throughput. Nvidia also developed transformer engines, which include high-level programming with linear algebra libraries such as TensorRT, for machine learning models, including LLMs such as GPT.

The period 1960-2010 saw a lot of ups and downs for AI, including the notable AI winter of the 1970-80s. The doubling period for compute measured in the number of operations (in units of petaflop/s x days) was every 2 years. After 2012, models got bigger and NVDA GPUs facilitated accelerating the compute doubling time to every 3.4 months [1]. Starting with use cases such as recommender systems and social network feeds, NVDA scaled to 30%+ market share (of revenues, much less in units) in datacenter processor revenues within just a decade of the early GPGPUs. 

NVDA has gained about 430 bps of datacenter revenue share/year in the last 5 years. NVDA’s dominant position in the accelerated compute market is likely to continue. Hyperscaler spending on accelerators is estimated to still be a low portion (<10%) of their capex and should inflect upwards in this new cycle. We expect NVDA to continue taking share from CPUs over this compute cycle. We model NVDA’s datacenter revenues to grow faster than the current consensus CAGR estimate for the next 5 years.

(ii) Nvidia is not just a chip maker; its unique co-design approach integrating hardware and software has resulted in a heterogenous-compute system ideal to ride the current inflection in AI cycle.

While Nvidia’s chips garner a lot of attention (given consumer gaming was a major use case previously), what’s more important is their systems-level solution with a full-stack approach to hardware and software for accelerated compute. Over the last decade, given its highly profitable gaming segment, NVDA invested in R&D at an average 24% of revenues which funded growth in GPGPUs. NVDA developed a series of GPU architectures that evolved over the years, advancing not only compute (throughput) but also storage (memory capacity and memory bandwidth). Think of memory capacity as the number of books your backpack can carry and memory bandwidth as how much you can open it so the books can be transferred fast to the locker. NVDA GPU memory bandwidth increased 8x in a span of 10 years, with typical CPUs still at an order of magnitude below GPUs. The ability to perform parallel processing and the memory advancements helped GPUs become the computing workhorse on which LLMs grew exponentially.

In an industry ever-focused on Moore’s law, which clearly slowed down in the last decade due to limitations of physics in manufacturing at small scales, Nvidia also focused on Amdahl’s law, which is about solving weakest links or bottlenecks in the system level solution. NVDA developed the NVLink fabric for fast interconnect between system components, such as CPUs and GPUs. A highly relevant acquisition from 2020 in this context was Mellanox, which provides InfiniBand high-bandwidth interconnects from system to system, and a host of other networking (ethernet switching) solutions.

Early on, NVDA abstracted the middle layer of software (the low-level language interacting with the hardware) into a unified architecture layer (across all generations of GPUs) called CUDA.  Software CUDA platform provides developers end-to-end workflow solutions and comes with high switching costs. AI/ML systems are fundamentally stochastic and not deterministic, whether from initializing weights and biases in a neural network or attention mechanisms in a transformer. This new software regime was dubbed as Software 2.0 by Andrej Karpathy (former Tesla Sr. Director that returned to OpenAI recently) at the AI Summit 2018. In contrast to Software 1.0, coding is <10% of machine learning systems; data workflow management is the pain point [2].  If you run the same code twice you may get different solutions, and managing data workflow becomes critical for debugging as well as explainability. 

As AI gets operationalized across enterprises, it becomes critically important for developers to have end-to-end solutions for managing workflows. The switching costs in this new regime are much higher than typical SaaS businesses, as more than performance or cost savings, what matters more is developer comfort and inertia. Over the years NVDA developed custom accelerated libraries, frameworks, SDKs, and a comprehensive compilation stack on the CUDA GPU computing platform. With a robust ecosystem of 4M developers, NVDA’s CUDA helps form a structural competitive advantage that is hard to replicate. CUDA remains a strong driver for GPU performance and ASP uplift at each new GPU architecture.

(iii) While NVDA’s likely dominance in training is understood, NVDA’s value proposition for inference workloads is underappreciated.

Inference is critical as it is performed real-time (i.e low latency at the prompt or the customer will search elsewhere, for example) in production unlike training --- which is performed as batch processing that can last several days. Per Amazon at their AWS:Reinvent 2022 “for every $1 spent on training, up to $9 is spent on inference.” So, as inference costs per query reduce over time and adoption rises, it is likely inference market can grow to be bigger than its current state. At GTC 2023, NVDA provided 4 optimized solutions for inference, L4, L40, H100 NVL, and Grace Hopper. NVDA’s CUDA/Triton inference engines offer critical capabilities such as model/hardware optimization, and quantization.

Currently, industry estimates are at about 60-70% of current inference workloads being performed on CPUs. From our research, we believe that as LLM and other multi-domain use cases gain traction, the market is going to shift away from CPUs as they cannot meet GPUs in price performance for handling the compute requirements of LLMs.

A more important consideration is sustainability, with news headlines around power utility infrastructure not being able to keep pace with data center growth, for example, in one of the major data center corridors [3]. As datacenters optimize their power envelope, NVDA’s performance/watt advantage over peers should stand out, and any energy savings would likely be reinvested for further acceleration of workloads in the datacenter.

(iv) OpenAI is talked about as the AWS-like critical infrastructure software play of this AI cycle; but I think the market is underestimating the potential for NVDA as an Operating System layer for the AI stack.

NVDA recently laid out AI foundation models as a service, including pre-trained large models of Nemo, Bio-Nemo, Picasso. It also developed the DGX Cloud to be deployed leveraging the infrastructure of cloud service providers (MSFT, GOOG, ORCL). In GTC 2022, NVDA talked at length about AI Operating System helping enterprises with adoption of AI/M, with CUDA as the low-level programming layer, pre-trained models and end-to-end data management solution including (i) integration with enterprise tensorflow/pytorch workloads, (ii) base/fleet command for edge/distributed workloads and (iii) domain-specific toolkits. With new AI foundation models and a cloud-agnostic AI cloud service DGX Cloud for custom GPU instances, NVDA’s full-stack solution facilitates an operating system layer for enterprises, with potential for monetization over est. 50M enterprise servers worldwide.

(v) Rebound in gaming and strength in auto segment: NVDA’s gaming revenues are guided to rebound to $2.5B/quarter run rate potentially later this year. In other segments, at GTC 2023, NVDA updated the auto 6-year design pipeline to $14B (vs $11B/$8B in previous two GTCs). NVDA-powered automotive computing systems for fleets of Jaguar Land Rover (build from 2025) and Mercedes (partnership announced in 4QFY23 call), should support incremental growth longer term.

My rebuttals to some key bear arguments:

(i) Bear argument: “Training demand is mostly pulled forward and there will be fewer large models, likely lowering compute required.”

Per OpenAI’s paper released with GPT-4 model (model size undisclosed) was already trained in Aug 2022. If the trends of Google’s Chinchilla and Meta’s LLaMA model papers were to continue, models are likely getting mostly smaller, with larger models maintained by only a few larger players. While we agree with that, larger models are not going away for sure; they tend to be more data efficient (fewer tokens needed to train) and are needed to derive smaller models. If the industry gets to using more quantities of smaller models, we believe that they may have to be trained on large corpora of data tokens to reduce model log-losses. Such tokens could include proprietary datasets, for e.g. Bloomberg recently released a research paper documenting the development of BloombergGPT. Nvidia’s DGX cloud is now available on hyperscaler clouds for enterprises to train their own models. The rise of domain-specific enterprise use-cases could continue supporting compute demand tailwinds.

(ii) Bear argument: “The bigger growth segment of AI market is likely inference, which is a competitive market, and NVDA’s edge in training is missing here.”

As discussed in the investment thesis points, inference is potentially a land grab for accelerators. NVDA is likely to gain a substantial share in this growing segment given (i) low latency, better performance per watt at scale, and resulting TCO advantage of its GPUs (ii) incumbency of training loads already present on NVDA GPUs, and (iii) various software and hardware optimizations available with the Nvidia CUDA/Triton inference engine.

(iii) Bear argument: “Datacenter capex is likely pressured given overbuild during Covid.”

While we acknowledge datacenter capex digestion is likely given the capacity build-up in the covid years, we think this dynamic is well understood from the last earnings at most datacenter plays. Datacenter capex dedicated to accelerated server spend is likely to inflect upwards and we believe capex estimates are likely to be revised up over the next 5-10 years of this compute cycle, as structural demand shifts may become clearer to hyperscalers.

(iv) Bear argument: “Hyperscalers will continue developing their own custom silicon solutions and NVDA’s growth and margins will be pressured.”

At GTC 2023, Jensen announced Google Cloud, Microsoft Azure, and Oracle OCI as hyperscaler partners for deploying DGX Cloud. Google Cloud was announced as a partner for deploying NVDA L4 inference chips for LLMs. This positive surprise should assuage concerns around Google TPUs taking away the shine from NVDA GPUs. While we wait to see how this debate eventually plays out, for those bears that argue hyperscalers could get competitive in chip-making and inhibit the rise of DGX cloud, our counter would be:

(i) GPUs are likely to perform better than custom silicon for a wider range of general-purpose workloads.

(ii) Hyperscalers would benefit from being “utilities” [4] drawing diverse workloads to the platform, selling their commoditized compute and storage to, for example, new crop of AI-native startups. For example, Snowflake, as a cloud-agnostic data warehouse that leverages hyperscale cloud infrastructure, grew while competing against Amazon Redshift, Microsoft Azure Synapse Analytics, and Google BigQuery.

(v) Bear Argument: “Newer software frameworks such as OpenAI Triton, Pytorch 2.0, or TensorFlow are a threat.”

I think it is unlikely existing workloads can be transferred easily given the end-to-end solution for debugging and explainability (similar to version control of Software 1.0) that’s critical in AI/ML workflows. With the several accelerant libraries added each year, e.g. CuDNN, CuLitho, CuOpt, and CuQuantum, CUDA’s value proposition has been increasing as an all-stop shop for AI/ML developers. CUDA is free for users and aimed to increase the adoption of NVDA hardware.

From our industry checks, the feedback we got was that (i) it seems hard to usurp Nvidia’s 15+ year advantage of providing the bells and whistles of AI/ML training and inference workloads, unless the new software competitors are proven to be much better and hardware is much cheaper. (ii) CUDA offers critical low-level programming without which it is hard, if not impossible, to get NVDA GPUs run optimally. Ultimately the debate can be settled perhaps by which hardware system-level scale-out solution has the best performance (by price and per watt).

(vi) Bear Argument: “AI hype cycle is no different than that we have seen for metaverse or crypto.”

While I agree there’s a lot of excitement for AI, it is hard to brush it off as a hyped-up craze when we consider the origins of this compute cycle and developments since 2018 which make automating creative output. ChatGPT Plus queries running slow at peak times in the day and faster at night reminds us of the dial-up modem days of internet of the 1990s. Per an AI Infrastructure survey [5] by run.ai, 89% of companies face compute allocation issues regularly. A recent article [6] from The Information mentioned that even Microsoft employees seem to be having a hard time getting access to GPUs.

Unlike metaverse or crypto, the AI use-cases are tangible and with massive reach, including a knowledge worker base of 1.5 B, and large and small enterprises adopting AI. While some normalization of demand is reasonable to expect over the next 5-10 years in this compute cycle, we also expect to see the rise of several new use-cases: for example, LLMs and edge inference could breathe a fresh life into iphone cycle (re-vamp of Siri, more custom filters likely with multi-modal GPT-4).  

Valuation

Nvidia’s datacenter revenues exceeded 50% of its total revenues in CY2022. Over 2017-2022, datacenter revenues grew at a 51% CAGR. Street models we see seem to grow this at 22% CAGR over the next 5 years. We model a base case CAGR of 30% for this estimate. We believe the higher CAGR is reasonable to expect when you consider the drivers within inference, as well as training. Hyperscaler capex grew at ~28% CAGR over the last 5 years. The trajectory of this capex going forward may have near-term uncertainty given overbuild during covid. But AI spend is likely a priority as evidenced from recent AI initiatives at major hyperscalers. We model hyperscaler capex as growing at 15% CAGR over 2022-27E, a relatively lower pace than that during 2017-22. Our modeled NVDA datacenter revenue comes to within mid-teens percentage of our hyperscaler capex forecast.

The split between training and inference revenues for Nvidia is undisclosed. For the next 5 years, the debate is on what inference market could be relative to that of training. NVDA CEO estimated at GTC 2023 that inference is likely to grow to be the same size as training, which informs our model assumption for inference. We model 70%/30% mix of CPU/accelerated for inference workloads now eventually shifting to a 30%/70% given our conviction from our research on the need for GPUs in inference, with some share loss factored to account for accelerators potentially gaining traction. For training, we are with street in assuming GPU workload accelerating at 25% CAGR, with other accelerators modeled as growing at a faster pace from their smaller base, and non-parallelizable CPU workloads growing at a slower pace.

My base case EPS 5-yr CAGR comes to be over 30% under this framework, reaching earnings power in the range of $13.5-$15. Assuming an exit NTM P/E of 35x, we underwrite about 15% IRR from current levels. With discounting at a rate of 10% for 2 years, we get to $390 2024E PT. If NVDA’s software solutions get monetized over the next five years to even low single-digit percentage of AI enterprise TAM (50M servers at $3K fee/year) presented in GTC 2022, there could be an additional upside. I discuss potential catalysts and risks to my thesis below.

Risks to my investment thesis, where I can be wrong:

(i) If rates stay higher for longer under potentially persistent high inflation, growth names could be under pressure and NVDA’s valuation multiple could contract. In the longer term, stock returns follow earnings, and our variant perception is on earnings potential, not a market multiple. We think continued execution in datacenter/auto and normalization of gaming could start an upwards earnings revision cycle for NVDA, which can potentially attract incremental buyers in any severe drawdown. From our research, it seems competitive risks are likely limited given NVDA’s hardware/software and ecosystem advantage. While we are not macro investors, portfolio hedge ideas based on growth or several other factors could be critical for portfolio construction and risk management.

(ii) As of 4QF23, Nvidia reported inventory at a high 212 days, with management guiding for normalization in the second half of this year. We will be closely following earnings to track this metric.

(iii) Gaming may take longer to normalize and pose a growth/margin drag.

(iv) AI/ML markets may be competitive and NVDA may cede share to other accelerators (Cerebras, Sambanova, Graphcore, Intel Habana) and peer GPU makers (AMD and Intel), and FPGAs (Alterra and Xilinx). Our base case factors a modest GPU share loss to these peers, but to the best of our knowledge, Nvidia benefits from its CUDA advantage, robust ecosystem, and fast pace of innovation (e.g. arriving at new use-case-specific products such as L4, L40, or H100 NVL) at scale.

(v) Nvidia had >20% of sales exposure to China as of last year. Nvidia developed products such as A800 and H800 to satisfy export control laws. Any export control restriction headlines may adversely affect sentiment.

References:

[1] OpenAI, “Two distinct eras of compute usage in training AI systems,” May 2018

[2] Sculley et al, “Hidden Technical Debt in Machine Learning Systems,” NeuRIPS 2018

[3] R. Miller, “Dominion Resumes New Connections, But Loudoun Faces Lengthy Power Constraints,” Data Center Frontier, Sep 2022

[4] B. Moore, “We will begin to think of the cloud providers as “utility” rather than “solution” providers,” from “Computing and data infrastructure in ’22,” Kleiner Perkins, Jan 2022.

[5] 2023 State of AI Infrastructure Survey, run.ai Accessed March 2023

[6] A. Holmes and K. McLaughlin, “Microsoft rations access to AI hardware for internal teams,” The Information, March 15, 2023

Disclaimer: This post is not a recommendation to buy or sell a security but rather intended for discussion of key bull and bear debates, based on research of public information, including company websites and regulatory filings. This post does not constitute professional or investment advice. Past performance is not indicative of future results. Consult with a licensed professional if needed before making any investment decisions. Investing in public markets can be subject to severe short-term and long-term risks, including resulting in permanent loss of entire invested capital. This post represents my personal opinions and not that of any other entity. I and/or entities that I advise can change views and positioning, at any point of time for any reason, or lack thereof, without any notice. I and/or entities that I advise are not and will not be held liable for any outcomes, including any indirect or consequential damages. I can be 100% wrong. This article has no guarantee to be accurate, complete, or current, and cannot be relied on as such. Please conduct your own due diligence.

Catalyst

Potential catalysts to watch for:

•  Datacenter : 

  • Nvidia navigating through supply chains for meeting H100/A100 demand

  • Adoption of L4/L40/H100 NVL for inference, and of Grace (CPU), Bluefield (DPU), Grace Hopper (CPU+GPU), DGX cloud

  • More disclosure around revenue model/partnership with CSPs for DGX cloud, foundation models

  • ASP uplift in the aggregate portfolio as higher ASP products come to the market

  • Rebound in demand from China hyperscalers for products A800, H800

  • Pickup in enterprise demand for Nvidia GPUs / Software / AI Operating system

• Other Segments

  • Rebound of gaming demand as supply chains get back to normal

  • Continued strength in auto / omniverse demand