Micron Reporting

With Micron’s quarter ending ultimo may, it is possible to get a mid-quarter update on the memory market situation and with that the general market.

Comparing Micron’s Revenue to COGS, it is obvious that the price increases have not stopped yet. Both NAND and DRAM prices increased after a short decline in NAND pricing.




Although 6.1% quarterly revenue growth is certainly respectable, it is also the lowest growth in this cycle, indicating that the memory cycle might be over. The year over year growth is still over 40%



The divisional revenue shows that the CNBU division and MBU division are pulling the wagon. Interestingly it looks like the Mobil business is starting to outpace CNBU that have been dominant over the last cycle. This could mean that the smartphone manufacturers in China are starting to accept the inflated DRAM and NAND pricing. SBU is declining to indicate that Micron is not supporting the open SSD market anymore but preferring to sell to the Data center directly.


Microns business is still dominated by the Computing market (Datacenter up, Graphics up, PC down) and the Smartphone dominated consumer market. All markets but communications show healthy growth.


Our conclusion is that the Datacenter market has cooled slightly while the smartphone market has decided to ignore the high memory pricing and boost manufacturing again. The memory cycle could be at its end. The Cloud market can quickly change to high growth again as the revenue is based on very few large customers and their infrastructure decisions.





The AI transformation of the Semiconductor Industry.

The overall market change.

What should have been the expected end of the current semiconductor cycle has transformed into a more complex market dynamic that has been experienced earlier. Since the beginning of the current cycle, most semiconductor product groups have generated significantly more revenue in growing end markets but a few have been on fire. We are now at a point where some product groups have stopped growing signalling an end to the cycle but one product group, in particular, has more fuel.

Two markets, in particular, has fuelled this growth cycle:

  1. The consumer market is driven by the semiconductor content growth in smartphones and the booming Chinese demand for cheaper local products.
  2. The computing market consisting of a mature PC market, a declining enterprise market and a data center cloud market on steroids.



From a product perspective, both markets have had a hunger for more memory, creating shortages and increased pricing. The memory market has generated more than double the revenue it did at the beginning of the current semiconductor growth cycle.



This development has favoured the large memory producers and forced Intel to abdicate the semiconductor throne to Samsung in Q2-17. Intel and Samsung combined are generating close to the same revenue as all the semiconductor companies outside the top 10 list.




The strategic rise of Semiconductor Memory Products.

A deeper dive into the market share data reveals how memories have overtaken processing products as the largest category dwarfed all other product categories.



Of the underlying product categories, only the main memory product groups, DRAM and NAND flash are gaining share, all other product groups are down.



The current DRAM boom is to a high degree driven by price increases. This is a result of the investment strategy of the main DRAM companies that all expected that NAND flash would be driving the next growth cycle in Smartphones and the Datacenter. The overinvestment in NAND flash and underinvestment in DRAM will take some time to correct. The market has not been able to get the DRAM needed and the response has been increased prices to a level where it is now limiting smartphone growth.




The Data Center is now dictating the semiconductor market

While the price of DRAM can suffocate Smartphone growth it can do little to limit the appetite of memory in the Datacenter. Amazon, Microsoft, Google, Apple and Facebook all invest heavily in data center infrastructure and with that – DRAM. The buying decision is a long-term investment decision and not consumer decisions as drove the memory markets in the early days. Amazon is not going to say no to an AWS customer because of high DRAM pricing.

The semiconductor revenue in the data center has grown quite dramatically in the current cycle and most of the growth has been in the Memory category.




As a result, Memory products are now more important than processing products in the Datacenter.



The memory revenue growth in the Data-center is close to a staggering 4x in only eight quarters.



This has caused a dramatic shift in the Semiconductor bill of material for a new cloud hyperscale datacenter.




Where the first call used to go to Intel for processing products, it might now go to a memory supplier.




As Semiconductor memories is a brute force investment game, it requires a lot of resources and guts – something only a few companies in the industry has.



Amazingly Nvidia has been able to generate similar growth rates as the memory companies in the data-centre. This is a piece of the puzzle that is transforming the Semiconductor industry.


Most of the top 10 Semiconductor companies now derive a significant proportion of their revenue from the Data Center Market.


The growth rates of the individual product areas give the last piece of the puzzle to how the semiconductor market will be transformed. It has long been known that datacenters are in the process of shifting their workload away from traditional processing storage and delivery of data to AI workloads of machine learning and inference. As a result, more of the total data in the Datacenter will stay in the datacenter and not be externally communicated. We believe the high growth rate of memory (even cleaned for pricing) and GPU’s combined with the low growth rate of processing is the first evidence that AI is now impacting the Semiconductor industry significantly and will change it forever.


The transforming semiconductor market drivers:

  1. Computing is now the largest and fastest growing market segment in the market
  2. The Datacenter market is over half the computing market and the only one that is growing. PC is stagnant and the Enterprise datacenter is in decline.
  3. The datacenter makes investment buying decisions that are less price sensitive than consumer decisions.
  4. The Datacenter is only at the beginning of it AI revolution and will need a lot more GPU and DRAM in the future.
  5. The current semiconductor cycle is not over for DRAM. This is not likely to change for several quarters.
  6. DRAM supply might not be able to fulfil Data center demand for quite some time, maybe even years.
  7. Semiconductor companies that are not aligned with this market, might struggle to grow and to attract investors.
  8. The Datacenter market is dominated by a few very large tech companies that are experiencing significant revenue growth in their cloud data center revenue. For Amazon, AWS is much more profitable than the rest of their business.
  9. The tech companies are much more likely to increase their investments dramatically than cutting them back.
  10. We believe the semiconductor market will never be the same again.


We supply the visual data for your semiconductor or market strategy. We track the semiconductor business and adjacent markets in detail. Contact Claus Aasholm for more details.








The Golden Age of Semiconductor Factory Ownership

As the costs of pursuing Moore’s law has spiralled out of control, more and more semiconductor companies have changed their manufacturing strategy to be less dependent on in-house process technology and to a higher degree, utilise the investments of Semiconductor foundries like TSMC and similar. Without the ability to compete with their process, fabless semiconductor companies must rely on their ability to design great chips, software and development systems. Despite having one less parameter to compete on, the fabless model has been promoted as superior to manufacturing models based on owned fabs.

The fabless model looks superior seen from the short-run financial perspective. With process costs embedded into the cost of sales and low capital investments, fabless companies can generate high free cash flow.  The long-run perspective could be different. Fabless companies could be commoditised or their products could become obsolete by being outperformed by products with superior process technology.

Although the fabless model is more often chosen out of the inability to participate in the intense capital investment game, there are many companies that have some or a significant proportion of manufacturing served at internal factories.

The 4 manufacturing models we track.

Understanding the different manufacturing models and how they perform over time an in different market conditions is an important element of Semiconductor Business Intelligence. Fab investments are B$ decisions as is M&A activity that might contain used factories that will require an operate, sell, close decision. Rather than making these decisions based on gut feel, it is important to get evidence and data support for decisions.

We track top 40 semiconductor companies (accounting for 85% of the semiconductor chips on the market) and have divided them into 4 categories.

Full Fab: Companies that are fully dependent on owning their own factories like Intel, Samsung and other memory companies.

Fab Heavy: Companies with over half the revenue manufactured in internal factories

Fab Light: Companies with less than half the revenue manufactured in internal factories

Fabless: Companies without internal semiconductor manufacturing capabilities.

As companies change manufacturing strategy over time, so does our categorisation. A good indicator of the fab- dependence is to track the revenue per dollar of Property, Plant and Investment on the balance sheet as seen below. Conglomerates are semiconductor companies that produce other products than semiconductors and need to be analysed differently. They are in general manufacturing companies though.




The current upcycle in the light of manufacturing strategy.

In the chart below it becomes quite obvious that the current upcycle has benefitted the fab owners which is not surprising given the dominance of memory. This has been dampened by the less than industry revenue growth of Intel Corporation. The large memory companies have all increased their revenue significantly based on price increases on DRAM and volume and price increases of NAND flash.




The Q1-18 revenue growth vs the same quarter last year shows that all categories have grown healthy with the fully fab’ed companies in the lead. Q1-18 growth vs last quarter was negative for Fullfab while the FabLight model was the only one that grew.




This semiconductor cycle has changed the manufacturing landscape significantly. Even though only a few companies now fully dependent on their own factories they represented more than 50% of the revenue in Q1-16. This has now grown to over 62% making the traditional manufacturing model the most common also.





A benefit of tracking the different manufacturing strategies is to be able to identify inflexion points and potential changes in supply and demand. We are able to track all significant financial parameters by manufacturing strategy.

The chart below shows the changes in revenue and in inventory compared to the beginning of the up cycle. The chart includes total inventory and revenue for conglomerates. The fabless model had an immediate spike in inventory despite negative revenue growth. The model requires careful planning and forecasting and gets influenced by foundry pricing. Part of the inventory increase could be a higher cost of products produced due to constrained supply from foundries. Similarly for the fab light model.



The fab heavy model seems quite robust in the upmarket. The companies using this manufacturing model are typically producing products that are less sensitive to process innovation like analogue, discrete and power products predominantly for the automotive and industrial markets.

You can





Fasten your Seatbelt, Semiconductors could hit a wall.


The evolving memory market

The top suppliers of Semiconductor NAND and DRAM have reported their Q1-18 results. For the first time in 9 quarters, there is a decline in one of the two major groups of memories. NAND flash declined by 2.8% while DRAM kept growing. While DRAM revenue was less than NAND revenue at the beginning of 2016, it has now grown to become the largest component of the memory market.

A key driver of this market change is the lack of supply that has increased prices of both DRAM and NAND. The investment response from the memory companies has been skewed in favour of NAND in the expectation that the Mobile and Datacenter markets would drive demand. The current reality is that both markets have demanded more DRAM than expected, far outpacing supply.




Memory Consumption

During this memory upcycle, all main memory market has grown, only the Datacenter market has taken a share of the market. From under 20% of the memory market in Q1-16, the Datacenter now consumes close to a third of all the memories produced. As memories for the data centre is bought based on a corporate investment strategy, it is not as sensitive to price increases as the other memory markets dominated by consumer decisions.

The mobile market, especially in China has suffered under the increased pricing of memory to the level where the Chinese government has tried to intervene. The PC and Graphics market consists of a declining PC element and a growing graphics element helped by the growing market for crypto mining equipment.




The shifting workload in the Data Centre

The impact of DRAM consumption in the Datacenter is obvious. The Hyperscale Cloud Data Center is changing workload from storing and delivering data to clients to increasingly working on Big Data and Artificial Intelligence tasks. These new tasks require processing of data inside the data centre and as a result, large banks of working memories in DRAM. The focus is shifting from moving data to mining data.



The Future of the Memory Market

While both the DRAM and the NAND manufacturing capacity has been to low to meet the demand for over two years, NAND supply now shows signs of reaching the market demand. The mobile market will still be under the influence of high DRAM pricing but is likely to increase if DRAM pricing declines. The key to the near-term future of the memory market is DRAM capacity.

The results of the suppliers of Semiconductor manufacturing equipment can give insights into the investments in the different product areas. The largest supplier of manufacturing equipment is Applied Materials, they have reported the following share of revenue:


Q1-18 Q1-17
Foundry 21% 41%
DRAM 31% 19%
NAND 37% 33%
Logic/Other 11% 7%

Although the investment in different product areas is not equal, it is obvious that the Semiconductor Companies are moving a significant part of their investments towards memory. From 52% of revenue a year ago to 68% today. NAND investments are still growing although the market indicates equilibrium. This could signal a sharp dive in NAND pricing. DRAM investments have increased significantly and will be in catch up mode for quite some time. The more worrying element is that foundry investments have declined as a share of revenue. Outside DRAM, it does not look like the investments in new capacity is well aligned with the market outlook.

Reporting from KLA-Tencor and LAM research show a similar picture:




If you are an investor, you might want to move your investments from NAND towards DRAM. If you are a semiconductor buyer, you might want to fasten your seatbelt. The next couple of years could be a bumpy ride as there are limited investments going into Foundry and Logic capacity causing potential droughts and price increases on almost all other products than NAND, that looks heavily overinvested.


Intels Bluff!

We just got busy with Intel’s indication that they could bid for the Broadcom, Qualcomm, NXP combination. We will update this article as our research progresses, but it is fair to say we have limited confidence in Intel’s interest in acquiring the combination. Intel has traditionally struggled with product and market areas that they could not grow to at least one B$/Qtr. Many product/market areas have been killed or sold off after a short period of intense investment.

We believe Intel is releasing the intention to bid as an attempt to prevent a third contender to the Semiconductor throne behind Intel and Samsung. The main problem for Intel is that it is the market leader in the fastest growing semiconductor market: The Cloud and Datacenter market – but Intel is not growing much compared to Nvidia and Samsung. Intel is not a leader in AI that now includes good memory solutions.


Revenue and Growth rates

If you are a reader of this blog, you know that we believe the acquisition that will match Intel’s strategy perfectly is Micron. The DRAM has gone from a commodity in the PC to a vital component in Artificial Intelligence in the Cloud datacenter, which is Intel’s target market number 1.




Intel is seeking meaningful revenue growth that it will not get from the Broadcom, Qualcomm, NXP combination. It is obvious though that the size of the new company would be problematic for Intel though.


Business Complexity

This analysis shows the size of each main product/market combinations for each of the 4 players in 2017. Intel’s business complexity would increase significantly – not something they are good at.







Did Trump just kill Micron Technology?

More to follow.

Is Softbank milking ARM?

There was a lot of concern back in 2016 when Softbank bought ARM holdings for a record 32B$. Was the Japanese bank going to treat ARM as a milking cow or was it a big bet on the future as stated by the CEO of Softbank, Masayoshi Son. With the closure of 2017, it is possible to get an overview of a full year in which ARM has been fully controlled by SoftBank. The investment strategy and statements from Softbank still indicate an agressive investment philosophy but what is the verdict of the financial data.




The available data confirms that Softbank is not treating ARM as a cashcow. Rather than trying to optimise EBITDA, ARM is allowed to increase operational costs significantly even though the revenue growth is single digits. R&D has increased by a staggering 60% while the SG&A costs has been boosted 44%. The headcount has increased by 22% while ARM’s semiconductor customers are cutting headcount by 5-10% a year.



The revenue of ARM has three main categories. The chart below excludes software to show the increaseing importance of royalties from chips sold, over licenses bought up front when customers engage with ARM and design with their architecture. Although licensing revenue is going down ARM is confident the new licenses has higher royalty potential that the earlier licenses.



Being acquired by SoftBank looks like a very good move for ARM. The agressive increases in operational costs shows that this is truly a future bet for SoftBank.


Other highlights and snacks from the lastest ARM presentation:

The ARM business model is very solid and perceived fair by the company’s clients. This is quite difficult as Qualcomm can testify. Their business model is constantly challenged by their customers and authorities. What is also interesting about it is that it give a view into the semiconductor market that is not available from the semiconductor companies.


As semiconductor devices increasingly become more integrated, more devices are containing processors. Arm now estimate that 70% of all integrated devices contain processors and that ARM has penetrated a third of that market.



Although the revenue from licensing has been decreasing the licensees has higher royalty potential:




Like the semiconductor companies we follow, ARM is also seeing a slowdown in the number of mobile units sold. The other markets have grown quite healthy and above the average non-memory market growth.



While ARM dominates the slow-growing market segment for mobile processors they have close to zero market share in the fast-growing server and data center market. This market, dominated by Intel and challenged by Nvidia, will be the next big target area for ARM.



The automotive market is changing from a power/discrete/microcontroller centric market to also include high-performance computing. ARM has a low market share but well positioned to get a significant share of the new opportunities. The chart below is a great overview of how ARM views the opportunity.



That ARM is interested in the IoT market is no suprise. However the company has a very interesting view on how the IoT market is going to be monetised. Semiconductor companies might enable this market but they will not derive the most value.



ARM have some interesting figures on the 4G to 5G deployments and the benefits that will be derived.


If you want direct access to our Semiconductor product, market and business data, please contact Claus Aasholm


Why would TSMC’s “collaborate” with Samsung on Bitcoin?

The ongoing transition of processing power from the PC to the Enterprise Datacenter and then to the hyper-scale cloud, is being joined by a new processing category. The meteoric rise of the value of Bitcoin has a significant impact on the Semiconductor Processing revenue. Although from a small base, the processing mining revenue has grown over 500% year over year, dominated TSMC customers and Nvidia. In Q4 the Cryto Mining share had risen to over 3% of the total processing market.



(Q4-17 numbers lack companies that have not yet reported – will increase PC and DC but not Crypto)


Already a key memory supplier to the Cryptocurrency mining market, Samsung is preparing to act as manufacturing foundry for the ACIS’s needed for processing. This might not be a major surprise but the statements that this is going to be a cooperation with TSMC does not make immediate sense. Samsung and TSMC are competitors in the foundry market and TSMC is without rivalry the largest manufacturer of processing ASIC’s to the Bitcoin mining market. Why should they “collaborate” with Samsung?



Bitcoin miners have traditionally been using Grapics Cards as they are better suited for mining that traditional server and PC processors. This has benefitted both AMD and Nvidia that has been the main suppliers of the cards to a level where there is a shortage. In order to protect its graphics customers, Nvidia has created and sold dedicated mining platforms during the last year but have stated that Bitcoins are best mined by specialised ASIC’s. This is confirmed by the rise of specialised mining rig manufacturers using proprietary ASICS’s. The main foundry for this activity has been TSMC that were willing to share their Crypto revenue figures last quarter but did not want to disclose them in Q4.  This is likely as a result of the pending “collaboration” with Samsung. It is obvious that this is not in TSMC’s interest and that they are not doing this voluntarily.

The reason for the cooperation is in our opinion driven by TSMC’s customers and more specifically by the largest Bitcoin mining Company, Bitmain Technology Inc. The company is claiming to deliver 70% of all the mining rigs to the Bitcoin miners and have had all of their BM1384/7 chips made on the TSMC 16nm process. Delivering in excess of 200K units a year with 100+ ASIC’s in each, they are likely behind most of TSMC crypto revenue. We believe Bitmain is now buying for more than 1.5B$ yearly run rate with TSMC and is now so powerful they can force TSMC to help Samsung 2nd source their ASIC’s. This could have a significant impact on both TSMC’s and Samsung’s revenue if the Bitcoin valuation stays high.



The top Bitcoin Mining ASIC companies:

Bitmain: Incorporated in Hongkong, the Beijing startup Bitmain Technologies Ltd is controlled by a trust in the Cayman Islands. Bitmain designs the silicon that goes into its bitcoin mining rigs, assemble the machines, then sells them to customers around the world. It also operates the machines for its own account, runs vast bitcoin mines that it rents out on contract to others, and, finally, manages several of the world’s largest mining “pools”—agglomerations of processing power so huge that they greatly improve the odds of successfully mining a bitcoin block. Bitmain manages Antpool and BTC.com, account for 28.9% of all the processing power on the global bitcoin network. Last four days results shown below.



Bitmain produces and sells hundreds of thousands Antminer rigs a year. According to Bitmain, they sell 70% of all mining rigs to the market, effectively providing 70% of all the processing power on the network. The latest Antminer rig is based on the BM1384/7 chip.

BitFury: Bitmain considers BitFury their main competitor. Although predominately a blockchain company, BitFury also delivers hardware for bitcoin mining: BitFury is basing their hardware on a proprietary 16nm ASIC manufactured by TSMC called BF8162C16

Canaan: Another Beijing company creating mining rigs based on their Avalon Blockchain ASIC

Article: Bitmain


“The state of the memory market” by Hynix

Hynix reported their Q4 results today and is giving the next insight into the memory market. Adding close to 1B$ in revenue by adding 100M$ in COGS, Hynix is still increasing pricing on DRAM. This is confirmed by the increase in operating income. More than 700M$ additional profit on 100M$ extra production cost.


The divisional data shows that NAND is now outgrowing DRAM. The NAND revenue growth is predominantly through increases in shipment quantities while the DRAM expansion is still price driven.



You can find the full Hynix factbook here.

Strategic Selection of Automotive Suppliers in the age of Disruption

Selecting Semiconductor suppliers involves intense collaboration between highly skilled technical purchasers, component engineers and R&D engineers. Specs, parameters and features need to be evaluated for products launches that are several years out for cars that need to last for decades. While all of these activities are vital for successful design and manufacturing it is important not to neglect the strategic elements of supplier selection. This is of particular importance in the automotive industry where supplier decisions have impacts spanning decades.

The time from products are introduced to the time they become obsolete are measured in months making the supplier selection even more important. Most suppliers last longer than their products (unless they get acquired)


The overall business model of the Supplier

Semiconductor companies are built for revenue growth. Like sharks, they need to move to stay alive. Because of the massive investments and rapid deterioration of competitive technology, the need for growth is relentless. At the same time, Semiconductor investors expect double-digit returns every year. When semiconductor companies struggle to grow top line they start to grow the bottom line. They shift their focus from investment to optimisation of profit. As a customer, you can experience flat or even increased pricing even though the manufacturing costs keep decreasing. The stalled semiconductor supplier is betting on the cost of switching suppliers is too high for the customer.

A semiconductor company that fails to grow is also in danger of being acquired. As a customer, acquisitions are not your friends. Acquisitions increase the probability that products lines get killed or become low priority and the supplier is bleeding people with valuable knowledge key to keeping the product line alive.

Select suppliers that has healthy growth and invest in OpEx & CapEx. You can contact Claus Aasholm or use our semiconductor factbooks to help with your strategic selection.

The leadership position of the supplier.

Apart from the business model, it is key that the supplier is an expert within your market. All semiconductor companies want to sell their products but not all can help you innovate and understands your market in detail. Align yourself with the wrong supplier and you will have to teach them about your business. The first parameter in your supplier selection should be the supplier’s revenue and market share.



While NXP still has a leadership position it is also obvious that their position is under pressure in particular from new players like Samsung and Qualcomm signalling the change in automotive electronics from body and powertrain to infotainment and ADAS.


The expertise of the supplier

A good proxy for expertise is to understand how dependent the supplier is on automotive products. It is rare that a high dependence on automotive is not followed by a high level of expertise in automotive.



While Intel might have great products it is safe to assume that Nexperia knows a lot more about the automotive market in general. It will probably also be safe to assume that Intel knows a lot more about ADAS that Nexperia which is why it is also important to understand the product categories that each semiconductor supplier sells.





The Expert/Leader map for Automotive.

A simple way of getting an overview of the expertise and the leadership position of the supplier is the Expert/Leader Map. If a supplier both has high revenue and a high reliance on automotive revenue they are an Expert/Leader. Companies with lower revenue and high reliance on are experts while companies with high revenue and low reliance are leaders.





Not only are these tools relevant for supplier selection, they can also be used for supplier negotiation. Companies outside the Expert/Leader box tend to respond more constructively in negotiations once they have been shown the Expert/Leader Map. They are well aware that buying from Expert/Leaders are a better deal from an expertise perspective and are often willing to compensate. Information and data are vital for supplier negotiations.

None of these tools is meant to be absolutes but should be used in conjunction with the operational purchasing tools.




Is it sensible to attempt to combine 90 business cultures into one?

The possible combination of 3 large Semiconductor companies is an extremely risky business operation. Apart from the operational elements the distinct cultures of Broadcom, Qualcomm and NXP would collide and create tension for many years to come. Anybody who has ever worked in a Semiconductor company knows that the individual cultures live for decades after the CEO has declared the merger successful and stated that only one culture prevails.


This specific combination is between three companies that have a predatory history and have bought many companies over time. Our research covers the last 10 years and shows that this business combination involves 87 other company cultures bringing the total to 90. Although it is likely that several of these cultures have not survived, the triple merger is certainly a risky operation.

We have visualised the complete merger story as a Metro Map that by the way has a similar size as the Washington Metro.




If you are interested in using our research, please contact Claus Aasholm for more information



Visualising the IoT opportunity in just two slides.

During their investor day 2017, Analog Devices showed two slides that brilliantly captures the IoT opportunity for semiconductor companies:

The first illustrates the number of devices per person from many people per device to many devices per person. A 4th wave should be added to illustrate that IoT really does not require a person as it is more about Machine to machine communication.


Although electronics become more advance and cheaper at the same time, revenue is driven by the exponential increase in number of intelligent units that needs semiconductors.



You can see all the slides from Analog Devices Investor day here:


Nvidia’s Earnings report and call

Nvidia reported another great quarter. Revenue was up and growth rates continued to be strong although slightly lower than earlier.

Operating income followed.

And for the first time Nvidia was able to generate over 1B$ in operating cashflow and in free cashflow.

The earnings release and conference call revealed the following:

  • The cryptocurrency mining impact on the business is hard to quantify but is reported as OEM revenue. Although significant in size, it is not going to distract Nvidia from the strategic business areas. It remains an opportunistic opportunity for the company. The revenue was 70m$ in Q3 vs 150m$ last quarter. Whenever a currency becomes too large or too difficult to mine it become viable to optimise it in an ASIC. 
  • The newly introduced Volta processor is delivering 10x performance in deep learning applications, far outpacing Moore’s law. Volta is already ramping significantly in all major cloud data centres.
  • The TensorRT platform suggests a Nvidia move into a more software-based approach in the hyper-scale data centre.
  • Three major Chinese customers have adopted the Volta V100. Alibaba, Baidu and Tencent will join Amazon, Facebook, Google and Microsoft in using the GPU for AI in their datacenters.
  • Volta has also penetrated the main server customers enabling server-based AI.
  • The attendance and downloads of AI-related conferences and software has increased by 5 to 10x over the last couple of years
  • Within the data centre, the main areas for GPU’s are
    • Supercomputing or high-performance computing. This is a 11B$ market in which Nvidia has a 15% (733m$) market share.
    • Deep Learning Training
    • Inference – the queries made by billions of internet users to the data centre. Today this is supported by CPU’s but GPU’s can increase the network speed by 100x
    • AI as a Service or Public Cloud AI Services – multi-billion dollar opportunity for Nvidia.
    • Vertical Industries: Automotive ADAS, Healthcare Diagnostics, Manufacturing, Robotics, Logistical industries
  • Nvidia sees the vertical industries as the largest opportunity and they are now ready to address them with the GPU capabilities in the cloud.
  • Volta is only in its infancy and the expectation is that it will continue to grow while the graphics part of the business is seasonal.
  • Cloud customers either have Volta capability or has announced it and the customer need is expected to be high.
  • GPU for graphics is sold one at the time, heavily influenced by seasonality and game launches. E-sports is becoming increasingly attractive and having the best gear drives down latency and impacts a players ability to win. The other element is high graphically attractive content. The last element is social – people want to share their brightest moments and that requires performance.
  • Volta is the single biggest processor that humanity has ever made. 20B transistors, 3D packing, the fastest memories available,  a couple of hundred watts replacing hundreds of CPU’s.
  • Nvidia is a one architecture company. This is not a limiting factor as there is so much software, numerical libraries, inference software, compilers and other. Jen-Hsun believes this is a massive advantage for Nvidia and makes the company perform at a level that requires much more people if you use multiple architectures.
  • Customers have many choices on what architecture to select. Jen-Hsun believes that the strong commitment that Nvidia has to a single architecture is convincing customers that Nvidia is a safe bet.
  • The world is experiencing computing problems on a scale we have not seen before and High-Performance Computing and AI have never been more attractive.
  • On the cooperation and announcements of AMD and Intel, Jen-Hsun believes Rodger leaving AMD is a great loss for the company. The Modern GPU is not a (G)PU’s anymore. They are not limited to graphical.

You can get the Q4-17 Strategic Factbook on Nvidia here