Semiconductor Forecasts: Too Bullish on Smartphones, Too Bearish on the Cloud
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Tejas Raut Dessai
FOR IMPORTANT DISCLOSURES 203.901.1633/.901.1634
psagawa@ / email@example.com
April 25, 2019
Semiconductor Forecasts: Too Bullish on Smartphones, Too Bearish on the Cloud
Industry forecasts of 3.5-4% annual semiconductor demand embed critical assumptions about end markets that deserve scrutiny. We differ in 5 areas: 1. Smartphones – 2-3% CAGR for Smartphone and peripheral chips is too aggressive; 2. Cloud Investment – overall server chips down 5-6% over 5 yrs. is pessimistic; 3. Shift to SSD – 13-15% projected annual growth in solid state drives implies a major near term inflection point that may take a few years longer; 4. Automotive and Industrial – Forecasters believe chip demand for EVs, ADAS, and IoT can accelerate to 15% growth within a couple years – maybe, but there is a lot of risk on timing; 5. 5G – 6-7% forecast annual growth for mobile base stations and optical backhaul underplays the potential. For semiconductor investors, we see opportunities in the cloud, 5G deployment and automotive ADAS systems driving upside for GPUs, FPGAs, datacenter switching, optical networking, complex RF, and DRAM. We are more cautious for chip segments more exposed to consumer devices, SSDs, and IoT, such as SoCs, NAND Flash, image sensors, etc., with the biggest risk coming from a projected 17% upgrade cycle boost in chip sales for premium smartphones in 2020 that would add more than 200bp to overall semi market growth in that year.
- Market forecasters see global chip demand growing at a 3.5-4% pace. Semiconductors are a $475-500B global market that industry analysts expect to grow to $550-600B by 2023. For investors, this scenario is attractive, but we are concerned that the end market assumptions that underlie the forecasts may be aggressive in some cases and overly cautious in others, skewing the projected product mix and perhaps driving an overly optimistic total demand projection. We have 5 major areas of contention.
- Forecasts embed a major premium smartphone upgrade cycle in 2020. Smartphones priced at >$100 have been declining in units for two years, with further declines expected in 2019. Chip forecasts expect a sharp reversal of that trend in 2020, with the most expensive category tagged to grow 7.6% YoY and the basic level seen as up 5.2%. Chips for premium phones are expected to outpace that with 17.8% growth, with 14.4% growth for basics, driven by the unit growth, expected expansion in chip content from sensors, and from a recovery in memory pricing. Moreover, chips for smartphone accessories – like smartwatches (+18.5% in 2020), head-mounted displays (+51.8%) and other wearables (+19.6%) – are projected to add an incremental $1.1B to 2020 overall market growth. We believe that this massive demand snapback is quite unlikely, with negative implications for SoCs, modems, device RF, DRAM, NAND Flash, and various categories of sensors.
- Forecasts lack faith in the cloud. Chip demand for servers grew an astounding 46% in 2018, as the big cloud operators – AMZN, MSFT, GOOGL, FB, BABA, etc. – invested furiously in datacenter capacity. Given the lumpiness of spending patterns, it is not surprising that forecasts expect cloud capex to drop sharply in 2019 (although we believe it could still rise) but while industry forecasts project server spending to bounce back in 2020, it also expects a sharp inflection point the next year and for 2022 levels below 2018. We think the cloud has only tapped a small portion of its TAM and that server growth is likely to be far better than the -1.2% CAGR suggested by industry forecasts. This has strong positive implications for AI processors (GPUs, FPGAs, ARM CPUs, ASICs), memory (AI is very DRAM intensive), and high-speed datacenter networking.
- Forecasts anticipate an inflection point for solid state storage. SSDs were up about 20% in 2018, carried by that big cloud capex, and market forecasts expects sales to fall back in parallel with server chips in 2019. Thereafter, the projections are off to the races, doubling over the next three years on the presumption that lower flash memory prices will finally induce the paradigm shifts from disk drives to solid state. We too believe that transition will happen but have less confidence that it will kick off with such gusto in 2020. This puts further risk around expectations for NAND flash and leaves a bit more room for the disk drive makers.
- Forecasts expect big growth from EVs, ADAS and IoT. Collectively, chips for Advanced Driver Assistance Systems (ADAS), Electric Vehicles, industrial security and “other” industrial (together representing Internet of Things), are projected to go from 4.7% of total semiconductor industry demand in 2018 to 9.9% in 2023, on better than 20% annual growth. This is a BIG piece of the overall 3.7% annual market growth, and it could be aggressive. We are inclined to be bullish on ADAS, but less enthusiastic for near term EV and IoT investment.
- Forecasts are ho-hum about 5G. Sales of chips for mobile base stations are seen up more than 15% this year, but decelerate sharply thereafter, despite a global 5G roll-out that we expect to take years to play out. Optical transport, for which 5G is a substantial driver, is seen as a steady 7% growth market. We think these forecasts are cautious – it is curious to have a bullish IoT projection but a weak 5G rollout in the same demand model, as the wireless technology is a precursor to much of the most intriguing IoT functionality. We expect more robust global 5G investment to drive upside for FPGAs, base station radio components, and optical transmission and switching chips.
- Overall chip volume assumptions look a tad high, but there are bright spots. Smartphones account for nearly a quarter of all semiconductor revenues. Weakness there will be hard to mitigate, even with significant likely upside for hyperscale datacenter and 5G wireless network investment. This is a general negative for fabs, semi-cap equipment and supplies. We are more enthusiastic for DRAM than consensus, but less so for flash, which has much greater exposure to smartphones and may see delays in the transition to SSDs. We remain bullish for AI-friendly processing solutions, such as GPUs and, especially, FPGAs, which will see added benefit from 5G. Otherwise, 5G semis are more difficult to play, as more aggressive network rollouts are offset by a relatively meek premium smartphone upcycle. We are bullish on optical components, which will see strong demand in datacenters and in metro networks for 5G backhaul.
Chip Forecasters Have Bought the Apple Industry Narrative
The major tech industry analytics firms – Gartner, IDC, and Forrester – are not far apart in their projections of 3.5-4% 5-year semiconductor industry growth. Interestingly, all are calling for an upcycle in premium smartphones starting in 2020 and for what had been breakneck growth in cloud datacenter spending to proceed cautiously after hitting a bit of a wall in 1H2019. In this, the analysts are tacitly endorsing a clearly Apple-ian view of the world, a narrative where future user functionality advances will reside on the device in their pocket and in adjunct accessories (watches, earbuds, and other wearables), rather than in the cloud. We, and much of the rest of the tech world, are skeptical of that paradigm, and suggest that chip investors adjust their expectations accordingly.
There are implications in 5 different areas. First, calling for a better than 17% YoY jump in sales of chips for premium smartphones in 2020 requires a significant leap of faith after three consecutive years of disappointment. We think the market is saturated and that price points are too prohibitive to allow a significant reversal of the trend toward longer user replacement cycles, even with the launch of 5G in some markets. On the flipside, the “consensus” model shows little confidence in the cloud. Spending on hyperscale datacenters boomed in 2018, driving roughly 45% YoY growth in server chips for the year, and creating gargantuan compares for 2019. Most observers are calling for sales to fall this year, but the analysts compound that by forecasting a tepid recovery in 2020, and nothing but cyclical ebbs and flows through 2023. In a scenario where smartphone functionality is fairly static and innovation is accessed from AI-driven applications in the cloud, this forecast will be flat wrong.
Third, chip analysts have been calling an inflection point on solid-state storage for a few years. Datacenters are still buying hard disk drives, but their relative price gap vs. better performing NAND flash storage solutions continues to close. We concur that the shift is coming, but projections of 25% annual growth in SSDs starting in 2025 is aggressive. Moreover, bullish projections for flash memory are also predicated on that roaring recovery in the premium smartphone market. Fourth, forecasters are also counting on a strong pickup in automotive and industrial chip sales, driven by advanced driver assistance systems, electric vehicles, and sensors for physical security solutions and other Internet of Things applications. Again, we agree with the sentiment, but quibble with the timing – yes on ADAS (ADAS: The Near-Term Opportunity for Components) but wait a bit on EVs and IoT. Finally, industry forecasts call for a telecom capex bump in 2019 and 2020 for 5G deployment but a sharp deceleration thereafter. With hundreds of global carriers phasing in the technology over time, we see a much longer investment cycle than that. Moreover, as wireline broadband replacement rises (yes, we expect that to be a big thing) and as new applications take hold (AR, IoT, autonomous driving, etc.) we see a much more bullish scenario long term.
Our call is somewhat bearish for chips in general – smartphones are nearly a quarter of the global semi market – and particularly for NAND Flash and other components destined for personal devices, but also for contract fabs and for semi cap equipment. We are bullish on silicon for cloud AI (GPUs, FPGAs, DRAM, optical networking), for ADAS systems (GPUs, FPGAs, LiDAR, industrial image sensors), and for 5G infrastructure (FPGAs, optical networking, high end radio).
Chips and Dips
2019 is expected to be a down year for semiconductor sales, the first annual decline since 2015, driven by weak smartphone sales, a pull back in datacenter investment and soft pricing in memory. However, the bad news is expected to be short lived. Industry analysts, like Gartner and IDC, see a substantial bounce back in 2020, more than covering the 2019 decline and setting the market toward robust 3.5-4% annual growth off that relatively strong 2018 number toward a 2023 market projected at nearly $600B (Exhibit 1, 2).
Exh 1: Gartner Semiconductor sales forecast expects inflection point in 2020
The methodology used to build these forecasts is consistent across the firms that generate them. Multiple teams, each responsible for a specific corner of the chip market, prepare analyses of demand within their purview based on their own assumptions about the consumer and enterprise markets that ultimately consume the products containing semiconductors. These bottoms up projections are then rolled up into an omnibus model with little effort to reconcile those hidden assumptions about end markets. Thus, anomalies, like projecting a big smartphone upgrade cycle while also forecasting a weak buildout for the 5G networks that would ostensibly be prompting the sharp acceleration in the replacement rate of premium phones.
Beyond the inconsistencies born of the bottoms up methodology, the component forecasts also embed controversial assumptions about the broader tech market that may not be aligned to investors expectations. For example, the Gartner model assumes that big smartphone upgrade cycle while projecting cloud datacenter investment to stagnate. Implicit in this is a distinctly Steve Jobs-ian view of the world, whereby the device is the center of the user’s experience and the cloud is an adjunct resource. In this perspective, more smartphones at higher prices are inevitable. Of course, this conclusion is far from universally believed, and an alternative narrative, where users derive ever more of their functionality from applications hosted in the cloud and use their devices largely as access points for those applications, would have drastically different implications for semiconductor demand (Exhibit 3).
So, these hidden assumptions matter a lot. We reviewed the top-level semiconductor forecasts from a handful of industry analysis firms and identified 5 places where we disagree with the underlying narrative.
Exh 2: IDC Semiconductor sales forecast expect similar inflection point in 2020
1 – We Do Not Expect a Major Cyclical Recovery in Premium Smartphones for 2020
Global smartphone unit sales have been in decline since 2016, with the premium segment peaking a year earlier. In the spotlight was Apple, which returned to sales growth in 2017 with a price hike to offset volume declines but delivered disappointment again in 2018, when a second increase found the limits of consumer elasticity. This inflection point is tied to several factors. First, total smartphone penetration tops 36% of the world’s population, with some growth from a robust market for used phones, but a trajectory that suggests near saturation. Second, the average age of a smartphone is lengthening as consumes take longer to upgrade. This phenomenon is tied to the near elimination of carrier phone subsidies, to the rising price of premium devices, to the massive upgrade activity seen in 2015 with the introduction of Apple’s first “phablet” sized iPhones, and, most importantly, to the perception that new models are modest incremental improvements rather than bold innovation. While each subsequent round of product introductions seems to offer a better camera, a more vivid display and a faster processor, consumers report being satisfied with the cameras, displays and processor performance of the smartphone that is already in their product. 2019 gives every sign of being a third straight year of lower sales (Exhibit 4).
Exh 3: Comparative summary of assumptions in semiconductor industry forecasts
The major semiconductor forecasters think that this slide ends in 2020 (Exhibit 5). Gartner’s chip forecast sees premium segment unit growth of 7.6% in 2020. IDC doesn’t break out the segment but sees overall smartphone unit growth reversing course to low single digit growth. The driving factor is supposedly 5G, which is expected to juice those sluggish replacement rates and bring new buyers into the premium segment. We are quite skeptical. While we are bullish for 5G, the deployment will ramp over several years, with big markets like most of Europe, phasing in after 2020. Moreover, previous introductions of new network technology have not spurred massive upgrades, but rather have seen consumer interest build with time. We believe this will also be true, as the biggest initial impetus for 5G deployment is the cost for carriers of managing network congestion. Only after the technology can offer nearly ubiquitous coverage do we expect it to begin driving meaningful incremental upgrades. This is likely to be after 2020. Indeed, Apple, which accounts for more than 40% of unit sales in the premium smartphone segments, does not even plan to bring out a 5G capable iPhone until the very end of 2020.
Moreover, 2015, the peak of premium smartphone unit sales, was an odd year. Despite the popularity of phones with much larger displays in the Android world, Apple had remained resolutely behind its single form factor. The iPhone 6, introduced at the end of 2014, broke the big screen embargo and stimulated a massive upgrade cycle of pent-up demand, borrowing from future upgrade cycles as well. Analysts have since waited for the echo that has never come. 3 years later, we believe that there will be no echo, and that calls for an inflection point back to premium segment growth are likely misplaced.
Exh 4: Global Historical Smartphone Unit Shipments, 2009 – 2018
Exh 5: Industry Est. for Premium Segment Smartphone Sales and Unit Volume, 2017 – 2023E
The big cloud guys – Amazon, Google, Microsoft, Facebook, Alibaba, etc. – spent huge in 2018, driving sales for chips in the enterprise data processing segment up more than 45% YoY (Exhibit 6). Against a giant compare, spending was already looking softer in 4Q18 and forecasters expect it to decline in 1Q19 and for the rest of the year (Exhibit 7, 8). While there have been comments from management at some of the cloud operators that they would be taking a short break to digest the big 2018 investment, growth in the commercial hosting business remains very robust- better than 40% YoY – and engagement for the biggest consumer cloud franchises also continues apace.
We note that cloud datacenters operate with dramatically lower overhead costs – e.g. personnel, power, real estate, telecommunications – relative to private enterprise facilities. As such, the shift to the cloud is likely to facilitate the capture of more data and the use of more computing at lower overall costs. Some of this additional computing and storage will likely go to enable AI applications, which are powerful, but infrastructure hungry. We also note that total global spending on enterprise datacenters is a more than $3.5T annual market. With current cloud spending at less than $200B, there is a LOT of runway left to address. We have written on this. (Cloud software: A long Journey with Three Layers).
Exh 6: Consensus expectations of cloud capex decline sharply over next two years
The strong sustained growth of consumer internet franchises will also be a major driver of hyperscale datacenter investment. Digital advertising, streaming media, and e-commerce all are huge global markets with long runways and compelling value propositions to drive investment in datacenter infrastructure.
In this context, chip forecasts that call for server spending to drop between this year and 2022 are difficult to grok. Gartner’s forecast vacillates in an up one year, down the next pattern, that peaks in 2021 and begins to head down thereafter. While the shift to the cloud will entail a corresponding decline in spending on servers for private datacenters, we believe the net will be a significant positive for total processing capacity, as massive savings on the non-hardware overhead elements – more than half of total costs – enable new infrastructure heavy AI applications and ever greater consumption of computing. Expect major upside surprises for sales of processors and memory for the big cloud platforms (Exhibit 9, 10).
3 – Calls for a Major Solid-state Storage Breakout are Probably Too Early
Eventually, spinning magnetic disk drives will be obsolete, entirely replaced by faster, more reliable and lower power solid-state drives composed of NAND flash memory chips. This transition has already played out in devices, where few portable devices sport hard disks anymore and desk top computers continue their slow decline. Datacenters are still a different story. Despite the freefall in NAND flash prices over the past 18 months, the cost advantage of magnetic drives is still substantial – more than 5-1 on a per TB basis.
Exh 7: Industry Data-processing Server Semiconductor Sales Est., 2017 – 2023E
Exh 8: Industry Data-processing Server Semiconductor Units Est., 2017 – 2023E
Exh 9: Global Enterprise Data Center Spending Forecast, 2017 – 2022E
Exh 10: SSR Public Cloud Spending Forecast in all 3 Categories, 2017 – 2021E
Exh 11: Industry estimates for DRAM and Flash Unit Volumes, 2017 – 2023E
Exh 12: Industry estimates for DRAM sales project significant downturn
Exh 13: Industry estimates for NAND Flash memory sales project optimism
Forecasters see flash pricing firming in 2H19, contributing to a substantial cyclical turn from a 3-4% 2019 decline in semiconductor industry revenues to a nearly 12% rise for 2020. Curiously, the models, which show a sharp deceleration from 24% to 19% growth in unit volumes of solid-state drives for 2019 despite the big price drop, show stable 17% annual growth in units for 2020 and 2021 in the face of what is expected to be rising flash prices (Exhibit 11, 12, 13). This translates to better than 30% growth in chip revenues for SSDs in both years.
We suspect that this is optimistic. Hyperscale datacenter operators love SSDs for their speed and low power, particularly for AI applications, but deploy storage in parallel with server investment. It is a bit incongruous to project a dramatic deceleration in datacenter servers alongside a massive shift to NAND Flash drives. As for traditional datacenters, we note that there is still a market for tape storage, a 60-year old technology, for enterprise data archives – change can take longer than you think. Given that chips for SSDs are more than 5% of the total semiconductor market by value, the projection of better than 30% annual growth in the segment next year (and again in the year after that) is a bold call that likely has far more downside risk than upside potential (Exhibit 14).
Exh 14: Cost advantages of Disk Drives over SSD are still substantial at 5 is to 1
4 – Electric Vehicle and Internet of Things markets may be slower to develop
The semiconductor forecasters are believers in electric vehicles. Chips for EVs are projected to accelerate sharply from just over 20% growth in 2018 to nearly 40% growth in 2019 and holding to more than 30% growth in 2020 (Exhibit 15, 16). We are as green as the next, and we don’t follow the automobile market closely, but this acceleration smacks of an Elon Musk talking point. Given Musk’s penchant for hyperbole and his track record for making adamant calls that turn out to be years early, we are a tad skeptical. We are more enthusiastic about the strong ADAS surge projected for 2020 (ADAS: The Near-Term Opportunity for Components) – but the 30-40% growth expected there is to the high side of our own views.
Similarly, we are believers in the Internet of Things, a concept that has been bandied about for the better part of a decade and has always been just around the corner. The Gartner forecast sees a sharp acceleration to double digit growth in chip sales for industrial applications coming for 2020, with security, “other” and power management applications leading the way. These categories fit to the IoT concept, with a proliferation of sensors part of the spending spree. Strangely, Gartner also predicts a tepid rollout of 5G, which as a very low power, low cost, standardized, and eventually nearly ubiquitous networking option, would likely be a major prerequisite for a boom in IoT deployments. We see IoT as taking off later in the 5G adoption cycle, after a couple more years in the “trough of disillusionment” phase of the famous Gartner hype cycle (Exhibit 17).
Exh 15: Industry forecast for EV electronics growth is too aggressive, 2017 – 2025E
Exh 16: SSR Forecast for ADAS specific Semi by Component Type, 2017 – 2025E
Exh 17: IOT is still in the phase of inflated expectations on the Gartner Hype Cycle
5 – 5G is a Bigger Deal Than Chip Analysts Imagine
5G is a big advance from 4G. The standard accommodates much more usage within each slice of bandwidth, supports a much wider range of radio frequencies, allows simpler and cheaper capacity expansion via small cells, enables a palette of service speeds from extremely narrow to the broadest of broadband, and cuts latency delays to a small fraction of its 4G predecessor (Exhibit 18). With time, it will jump start a range of exciting new wireless use cases – the IoT mentioned above, augmented reality, vehicle-to-vehicle communication, low latency cloud gaming, and many others. We believe high-speed, unlimited use mobile services will also begin to erode fixed broadband for targeted neighborhoods with sufficient backhaul capacity. However, the immediate purpose for 5G is to relieve network congestion caused by increasingly heavy data usage. 4G LTE, with limited spectrum options, lower spectrum efficiency and more costly options for cell splitting, is a dead end for carriers who will be in a race to stay competitive with rivals who will also be rolling out 5G.
While the increasingly unlikely T-Mobile/Sprint merger would have stepped up 5G investment in the US, we still believe that the technology will see a long and robust global deployment ramp. In contrast, semiconductor forecasts see this year as the peak in network spending growth, projecting deceleration to single digit growth thereafter (Exhibit 19). We think this is pessimistic. First, 5G deployments are still in
Exh 18: Comparison of 5G network benefits vs. 4G standards
Exh 19: US carrier capex and consensus expectations show embedded pessimism
Exh 20: Testing equipment revenue has already been growing as a precursor to 5G
their early phases in the US and Asia and have hardly begun in Europe – peak macro base station deployments are still many months ahead (Exhibit 20). Second, as 5G usage builds, a second phase of small cell deployment – with a much higher percentage of the spending on semiconductors – will follow. We are very bullish for this next phase of spending. Chip forecasts are not.
With smartphones accounting for 24.8% of all semiconductor sales, shortfalls vs. consensus forecasts in that category will be difficult to cover. This alone leaves us to be pessimistic for the industry for 2020. Companies and product categories – e.g. FLASH memory (e.g. Toshiba, WDC, INTC), image sensors (e.g. SNY), modems (e.g. QCOM, MediaTek), RF (e.g. AVGO, SKYW), etc. – dependent on smartphones could be weaker than anticipated; volumes for semiconductor fabs (e.g. TSCM, Samsung) could disappoint; and capital equipment and supplies could follow that. FLASH could see a double whammy if the adoption of SSDs by hyperscale datacenters is slightly less torrid than anticipated. Cheap sensors (NXP, Infineon) may see a slower IoT ramp. The demand growth for chips used in the powertrains of electric vehicles (NXP, ON, STM) may be a bit off the expected pace.
To the positive side, datacenter components – e.g. processors for AI acceleration (NVDA, XLNX), power friendly CPUs (QCOM, IBM), DRAM, high-speed optical interfaces and switches, etc. – likely have meaningful upside to forecasts (Exhibit 21). We are also bullish on wireless network infrastructure components – with the caveat that some suppliers for base stations also will be hit by the weaker than expected smartphone market. The best pure play may be XLNX, as FPGAs are critical for base stations but not used in smartphones and other devices. We are also bullish on chips for ADAS (advanced driver assistance systems), such as industrial image sensors (ON) and on-board processors (NVDA, XLNX). We also suspect that analysts, giddy for the eventual inflection point on solid-state drives, may be underestimating demand for traditional disk drives. STX and WDC could have a better than expected rebound from their 2018 cyclical trough.
Exh 21: SSR Summary of Semiconductor Industry Outlook in Near-Term