NVDA: AI Upside Greatly Outweighs Crypto Risks
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Tejas Raut Dessai
FOR IMPORTANT DISCLOSURES 203.901.1633/.901.1634
psagawa@ / email@example.com
April 23, 2018
NVDA: AI Upside Greatly Outweighs Crypto Risks
GPUs have enjoyed a significant boost in demand from cryptocurrency mining and many investors fear that NVDA could be at risk from changes in that market. We believe that these concerns have been overplayed for 3 reasons: 1. NVDA’s exposure is offset by AI and other drivers – Crypto-mining has risen to about 7.3% of sales (13.7% of the gaming segment), up around 500bp YoY. Datacenter has risen to 19.9% of sales, up 1034bp YoY and we expect 87% growth in FY19. Automotive is 5.7% of sales, and while it grew 14.6% in FY18, the rise of ADAS systems portends acceleration. 2. Crypto-mining GPU demand will not collapse – There are headwinds – weaker crypto-coin prices, a shift from Proof-of-Work to Proof-of-Stake, and competition from ASIC-based mining rigs – but we believe the impact will play out gradually. 3. Slack demand absorbed downstream – Graphics card prices are down 20-25% from an early Feb peak, but remain nearly 50% above list price, as GPU suppliers still struggle to meet volume demands. Slowing demand may hit retail prices further, but chip makers will be insulated for quite a while. Meanwhile falling prices should draw renewed demand from gamers who had been priced out, and from marginal crypto-miners, further supporting GPU demand. Our conclusion: NVDA and AMD GPU sales are likely to exceed expectations colored by overly cautious views of crypto-mining.
- We believe analysts are overly pessimistic on NVDA sales growth. Consensus projects NVDA revenues to slow from 41% growth in FY18 to 27% in FY19 and 15% in FY20. Amongst the reasons given are an “inevitable” deceleration in cloud datacenter investment, increased competition for graphics cards, and the risk of an “implosion” in GPU demand for crypto-mining.
- Datacenter sales will continue to surprise. Yes, torrid infrastructure spending by the leading cloud platforms (GOOGL, AMZN, MSFT, IBM, BIDU, BABA, Tencent, et al.) must slow, but not as steeply as many assume and the share spent on AI processing will grow, with NVDA poised to take 60% or more of that (ASICs and FPGAs will take the rest). As such, datacenter demand for NVDA GPUs should show little, if any, deceleration in 2018, and could even accelerate. While we believe that NVDA’s automotive opportunity – more Autonomous Driver Assistance Systems than fully autonomous robocabs – will take time to develop, the opportunity 2-3 years out remains substantial.
- NVDA’s exposure to crypto is modest. Estimates suggest that more than $3.5B was spent on crypto-mining computing over the past year, up more than 5x YoY. However, higher retail prices accounted for almost half of the growth, while GPU chip prices remained steady. Moreover, we believe that as much as 50% of purchases by crypto-miners were used equipment and existing stock, further reducing the impact for chipmakers. In addition, we believe that the popularity of AMD Radeon GPUs amongst the crypto community suggests NVDA held less than its average market share in this segment. As such, we believe its exposure to have been ~14% of its gaming GPUS and ~7% of total sales in 4Q18.
- Coin prices likely well above real breakeven cost. After a fall from a $19K Dec17 peak, Bitcoin has been stable at $7-8K since early Feb, up more than 5X YoY. Recent reports suggest the breakeven for miners has risen to as much as $8,800, but we are skeptical. First, high rig prices are temporary – GPU prices remain 45% above MSRP and ASIC rig maker Bitmain is believed to be making windfall profit. Second, estimates from just 4 months ago put breakeven MUCH lower, at $3-4,000, depending on power costs. This seems a case of a bottleneck – mining gear retailers – capturing surplus than a structural obstacle. Expect mark-ups to contract further if demand wanes. Finally, increased competition amongst miners for the most popular coins may have increased breakeven but should also prompt arbitrage to less popular coins and support for new coins.
- Proof of Stake gaining ground … slowly. A bigger risk is Ethereum’s gradual migration from “Proof of Work” (e.g. transaction fees paid to the miners that solve difficult algorithms the fastest) which drives GPU demand, to “Proof of Stake” (e.g. fees paid to users who have staked the most coins for the longest) which doesn’t. The debate between PoW and PoS proponents within the crypto community is vigorous – PoS greatly reduces processing demand and power costs, but dramatically advantages the biggest players with the longest standing, discouraging market participation. We believe that very few PoW currencies will follow Ethereum’s lead.
- ASIC threat is not new. Beijing-based crypto-miner Bitmain launched a custom-designed ASIC for Bitcoin mining in 2014 and has been selling it to 3rd parties since early 2016. The platform has been wildly successful – some reports suggest that as much as 80% of Bitcoin volume is validated by its servers – but it has been less successful with other coins (It will have a new Ethereum chip in July). While the ASICs are intrinsically more cost effective, Bitmain has reduced that advantage by raising its retail prices dramatically in pursuit of margins. Moreover, many are concerned that backdoors programmed into Bitmain’s solution are a major security risk, particularly given increased Chinese government scrutiny of the crypto-market. While Bitmain is already a huge player, we don’t see the dynamics of its competition with GPU-based alternatives to change in the foreseeable future.
- Slack demand absorbed downstream. Retail prices for graphics cards appropriate for crypto-mining hit a peak more than 90% above MSP earlier this year and have since fallen ~25% to a 45-50% premium. This would seem to indicate a damping of demand, but the cost has been borne by retailers, who have enjoyed the windfall of sky-high markups, rather than manufacturers, who have kept their prices steady and struggled to meet market demand volumes. If anything, this trend should revive demand from PC gamers, who have been priced out of graphics board upgrades. Lower card prices also bring in new marginal buyers for crypto mining.
- NVDA likely to continue beating expectations. Given our strong beliefs that NVDA’s datacenter business can sustain its stellar growth and that the downside risks from crypto-mining demand are overstated in analyst models, we expect the company to deliver $13.1B in FY19 sales, well ahead of the $12.4B consensus. Our simple earnings model suggests ample leverage to the bottom line, with the potential to exceed the $6.27 consensus EPS projection by more than 10%. We understand investor concerns over valuation but note that stocks like NVDA have historically been driven by upside surprises and upward revisions, rather than multiples and DCFs. AMD, with a higher exposure to mining, and XLNX, with less, have also been tagged with unnecessary pessimism.
Don’t Fear the Crypto
NVDA shares have been the object of an intense bull-bear tug-o-war that has played out over the past years, with the bulls enjoying the better of it with the stock up more than 140% YoY. Still, despite a thunderous upside surprise in 4QFY18 results delivered in February, a new bearish narrative has gained traction. As per usual, the story begins with valuation and pessimism over the sustainability of cloud datacenter investment, but folds in warnings of a Bitcoin apocalypse. Once the crypto-bubble pops, say the naysayers, NVDA will fall, and hard. We see a few things wrong with this story.
First, in NVDA’s most recent quarter, we calculate the YoY growth in GPU sales for crypto-mining at ~$220M. That is about 30% of NVDA’s total sales growth – 760bp out of 34%. While impressive, it is hardly the most important growth driver for company sales – that would be datacenter, which for all FY18 accounted for nearly 20% of sales while growing 132% for the year. Despite an acceleration of sales to datacenters in FY18, analysts, habitually following the “law of large numbers”, are forecasting a sharp deceleration for FY19 to less than 70% growth. We think this is extremely pessimistic. While we agree the big cloud platforms that dominate datacenter processor purchases are likely to slow down CAPEX from the torrid pace set in 2017, the share of spending on NVDA’s GPUs, which are the backbone for AI computing capabilities, will rise quickly. We expect NDVA datacenter sales growth of 85%, for FY19 sales almost $400M above consensus. That is a substantial margin against disappointment from crypto-mining.
Second, we don’t expect the crypto-mining market to collapse. Bitcoin’s breathless run to more than $19K in Dec., followed by stomach-turning drop to $6K in Jan. fueled apocalyptic projections for crypto-currency. With Bitcoin now back to more than $8K, one report suggests the breakeven for miners is 10% above that – we are skeptical that it is so high. A big factor is inflated retail rig prices, that had risen to a 90% markup over MSRP – these are coming down. Another is increased competition amongst miners – arbitrage to less popular coins and support for new coins will partly mitigate this.
Another worry baked into the bear case is Ethereum’s shift from “Proof of Work” to “Proof of Stake”. PoW allows miners to earn validation fees based on the completion of complex computing algorithms, thus driving demand for powerful GPU powered mining “rigs”. PoS shifts the basis for awarding fees to the size and seniority of coins “staked” as investment in the system, eliminating the arms race for more powerful rigs. While this is an obvious net “bad” for GPUs, Ethereum’s transition will be gradual and there is substantial resistance to other crypto-coins following the precedent, damping the risks to NVDA. Finally, a threat from custom ASIC solutions – China’s Bitmain is the most prominent – is nothing new. ASICs will continue to take share, but not at an accelerated pace.
Even if crypto demand for GPUs slackens, most of the pain will be taken downstream. The retail markup on NVDA PC cards skyrocketed along with the price of crypto-coins to more than 90% only to fall back to about 45% today, but NVDA’s prices have been rock-steady, even as it struggled to meet demand. If anything, a lower retail price brings gamers (and marginal miners) back to the market, a buffer against the gloom and doom scenario.
We believe that NVDA’s FY19 sales are likely to continue the trend of significant upside quarterly surprises with leverage to EPS. AMD, which has a greater share of crypto-mining, has more risk but is also likely to beat expectations. XLNX has less exposure than either and will flourish on datacenter and 5G.
Nvidia Skeptics Abound
Nvidia is one of the hardest stocks for value investors to abide. It trades at 36.6 times the forward consensus for a hefty 3.94 5-year PEG ratio. The only way that this makes sense is that the consensus is wrong. We believe that this is the case.
In its FY18, ending in January, Nvidia delivered 40.6% sales growth and 87.5% EPS growth, topping estimates by an average of 21% per quarter (Exhibit 1). This was an acceleration from the 38% sales growth delivered in FY17, driven by a 138% increase in GPU sales to datacenters, which use the parts to train and run AI models. Datacenter made up 19.9% of sales for the full year (Exhibit 2). The big drivers of Nvidia’s data center demand were big cloud platforms – capital spending by Alphabet, Amazon, Microsoft, IBM, Facebook, Alibaba, Baidu and Tencent was up 24.5% in 2017 to $46.5B (Exhibit 3). Within that, we believe that the proportion of processor spend targeted at AI increased by 9120bp, and thus, Nvidia’s big year.
Analysts are concerned that it is all downhill fast from here. The CAPEX expectations for the big five American platforms portend a 518bp slowdown to 17.7% spending growth over the next 3 years (Exhibit 4). Surely this means a proportional drop for Nvidia’s datacenter segment, suggesting a sharp deceleration to less than 70% growth in FY19 and an asymptotic approach to something in the 10% range thereafter. This line of thinking relies on the assumption that the growth in investments to accommodate both the shift to the cloud and the adoption of AI-based application architectures hit their peaks in CY17. Looking at graphs of the trailing 12-month datacenter and total YoY sales growth, along with consensus forecasts reveals the inflection points implicit in consensus (Exhibit 5-6).
Exh 1: NVDA Quarterly Revenue and Earnings Surprises
Exh 2: NVDA Annual Sales Breakdown by Segments, FY2015 – FY2018
Exh 3: Historical Capex Spending for Big Hyperscalers, 2012 – 2017
Exh 4: Consensus Expects a Sharp Drop in Cloud Capex Spending
Exh 5: NVDA Financial Snapshot, April 2018
Exh 6: Quarterly LTM NVDA Datacenter Sales and Y/Y Growth
Datacenter Sales Have Plenty of Runway
Cynically, we note that the sell side consensus expected that inflection point to have happened last year, scrambling to raise estimates every quarter when it didn’t come, the most recent blowout results reported just two months ago. Part of this rampant over conservatism comes from the semiconductor analyst community’s distrust of cloud demand.
We believe that most world-wide enterprise computing is destined to shift to the cloud, driven by compelling economics and performance advantages (see http://www.ssrllc.com/publication/37440/). We believe that cloud hosting will displace trillions of dollars in annual spending on private enterprise datacenters. Even with significant disinflation, we expect hosting to be a $450B market by 2025, dominated by the small handful of leading platforms that have already established their leadership. Moreover, those leading platforms will also continue to invest to support their own dominant consumer franchises, which may be leveraged in many new directions (e.g. autonomous robocabs, healthcare, consumer finance, etc.). It is not at all clear that 2018 is the year of the investment inflection point, and a quarter or two of slowing spending would more likely a be respite than a rapprochement – CAPEX tends to be lumpy. For example, the top 7 global hosts have averaged 23% annual growth in aggregate over the past 5 years, but right in the middle was 2015, when spending was up just 4.6%. Investors taking that as inflection point would have been very wrong, as spending growth popped back up to 25.7% the very next year.
Furthermore, cloud platforms specifically buy GPUs for training and running AI models. We have written of this (http://www.ssrllc.com/publication/hyperscale-semiconductors-processor-diversity-coming-to-the-cloud/). Between now and 2025, we believe that the proportion of processing spending earmarked for AI acceleration will grow from 22% today to 45% or more (Exhibit 7-8). Within that, GPU’s face competition
Exh 7: Annual Cloud Processor Spending Share by Processor Type, 2017 – 2025
Exh 8: Annual Cloud Processor Spending Forecast by Processor Type, 2017 – 2025
from high performance FPGAs, which can be reprogrammed to optimize processing for specific applications, and from ASICs, which are hardwired as optimal solutions for those applications. We believe that all three solution types will find their place in the platforms of the Cloud/AI era, but that GPUs are likely to stabilize at roughly 70% of the total spend. Nvidia, whose CUDA programming rubric is widely adopted by AI developers, is in the driver’s seat for that market. While repeating FY18’s 136% growth in datacenter revenues seems too much to ask for, the consensus deceleration from 40.3% topline growth to 27.6% this year implies a very sharp fall off. We believe consensus embeds less than 70% growth in data center revenues, still accounting for half of the overall expected sales expansion. This is far too pessimistic.
The Joy of Crypto-mining
The rest of Nvidia – gaming, professional visualization, automotive and OEM/IP – would grow less than 18%. In FY17, these segments collectively grew 21.9%, so this trajectory is significantly less severe (Exhibit 9). Part of this was 35% growth in Gaming. Part of this was the successful launch of the Nintendo Switch console, which featured an Nvidia GPU, but a bigger factor was gaming processors bought to mine crypto-coins.
The headlines on the crypto-market have been lurid. The price of a Bitcoin topped out at more than $19,000 in December of last year, only to collapse to $6,000 in January. With that, retail prices for GPU gaming cards dropped 25% amidst “I told you so” admonishments from a posse of investors and business executives who considered the markets for blockchain-based assets a tulip-style bubble. Against this backdrop, the retail price of GPU gaming cards dropped 25% in a month, prompting the same gloomy predictions for Nvidia and AMD, the primary suppliers (Exhibit 10).
Exh 9: NVDA Quarterly Sales and Growth by Segment, 1FQ17 – 4FQ18
Exh 10: NVDA Graphic Cards Average Retail Market Price Trends