- Ethereum crypto mining profitability is down 90% yr/yr, and demand in GPU space has been vastly understated by Nvidia, the impact of this will become evident in Nvidia’s gaming revenue.
- Nvidia management public position on crypto has flip-flopped repeatedly last six months to point it’s unclear if they even understand what they are facing.
- Behind the scenes moves like EULA change and new GeForce Partner Program indicate Nvidia hunkering down with defensive moves.
- 2017 was year of GPU in AI - 2018 Will be Year of Purpose Built ML/DL Compute Accelerating Hardware - Google TPU2 volume production, dozens of startups launching products, Bitmain has entered market, Intel etc.
- Stock now most overpriced of Nasdaq heavyweights post recent correction, and thus most clearly vulnerable considering its moat is going to be seriously questioned as investors wake up to what’s in this report. Near term PT $180... $140 before year is out.
Nvidia is heading into turbulent waters, and despite a serious breakdown in the NASDAQ, the bulls in the stock are utterly asleep at the wheel. A short position has rarely looked much better than this.
Let's start with the obvious…
The ALT Coin Mining Party is over, and we didn't even need to wait for Ethereum Proof of Stake or a Bitmain ASIC miner to end it; simple crypto mining mania herd behavior took care of it on its own.
Here is what mining performance on an on an NVIDIA GTX 1070 6 card mining rig (170 MH/s 850-watt power draw) looks like Yr/Yr.
Rig Assembly Cost
Profit Per Month (.15kw/h power cost)
May 1, 2017
March 28, 2018
Want to bet on Ethereum?
Go buy some coins on Coinbase because building a new mining rig makes no economic sense anymore.
But cryptos have been volatile for months, surely, Nvidia management and Wall Street are prepared for this? Plus, Nvidia is not AMD. It is a gaming/AI company, you are beating a dead horse here right??
This headline which has made the rounds extensively over the past few months is based on data from JPR RESEARCH, the leading market research firm on GPU space. They got at this number based on statistical modeling using their historical GPU industry data and forecasts.
Now, let's compare this number to the Ethereum network hashrate increase in 2017.
The Ethereum network hashrate increased from 5900 GH/S to 160,000 GH/S during 2017. At a conservative 25 MH/S per graphics card average hashrate it takes 40 GPUs to add 1 GH/S to the network. That's 6.2ml cards in 2017 for Ethereum alone. The network hashrate recently peaked at 270k GH/S. That's another 4.4ml cards since Jan 1, which covers the period in which GPU prices went haywire. That roughly works out to a TTM demand of 10.6 million cards or $2.65 billion in GPU revenue for Ethereum alone. Add in Monero, Zcash, and other alt coins as well and that number climbs higher.
I am pointing this out because it's mathematically irrefutable!
Once a GPU is mining, it's plugged into the Ethereum network. As more GPUs join the network, the hashrate rises approximately based on the computing power of the added GPUs. The difficulty of the network adjusts to this rising compute, and hence the amount of coins miners earn go down as more and more miners compete against each other with a smaller relative share of the network hashpower.
This Is Mining 101.
You can literally log into top mining pools like ethermine.org, which accounts for 25% of the Ethereum network hashrate, and see each miner's active rigs and each individual block mined.
Quantifying demand from an industry doesn't get any easier and more approximate than GPUs for cryptomining. (more on this later)
So, here is the problem, alt coin miners spent roughly $3billion on GPUs over the last 12 months and now have zero incentive to spend even $1 more. And if things get worse (Ether goes POS, Bitmain ASIC arrives this q) they have plenty of incentives to start selling their cards on secondhand market. (go on eBay and check the listings they have had huge burst past few weeks with prices of cards now heading south).
Are the GPU makers ready for this and warning investors about this risk?
Well, AMD which has been down the GPU mania rode before with Bitcoin and has been very cautious about the market over past few months and even updated its 10-K risk factors to include this:
In addition, the GPU market has seen elevated demand due to the application of GPU products to cryptocurrency mining. For example, our GPU revenue has been driven in part due to an increased interest in cryptocurrency mining. The cryptocurrency market is unstable, and demand could change quickly. For example, China and South Korea have recently instituted restrictions on cryptocurrency trading. If we are unable to manage the risks related to a decrease in the demand for cryptocurrency mining, our GPU business could be materially adversely affected.
Nvidia unfortunately is a different story. It appears to have bought into crypto mania hook line and sinker.
Don't believe me? Well, then just listen to the ever-changing tune that is Nvidia with respect to crypto mining.
CEO two quarters ago:
"Thanks. Cryptocurrency and blockchain is here to stay. The market need for it is going to grow, and over time it will become quite large. It is very clear that new currencies will come to market, and it's very clear that the GPU is just fantastic at cryptography. And as these new algorithms are being developed, the GPU is really quite ideal for it. And so this is a market that is not likely to go away anytime soon, and the only thing that we can probably expect is that there will be more currencies to come. It will come in a whole lot of different nations. It will emerge from time to time, and the GPU is really quite great for it.
What we've done, our strategy is to stay very, very close to the market. We understand its dynamics really well. And we offer the coin miners a special coin-mining SKU. And this product configuration - this GPU configuration is optimized for mining. We stay very close to the market. We know its every single move and we know its dynamics."
Now Nvidia CFO on q4 call:
"While the overall contribution of cryptocurrency to our business remains difficult to quantify, we believe it was a higher percentage of revenue than the prior quarter. That said, our main focus remains on our core gaming market, as cryptocurrency trends will likely remain volatile."
CEO at GTX March 27, 2017: "Crypto is not our business"
From Crypto Super Fans to washing their hands of the market in less than six months. Not exactly encouraging if you are relying on these guys to accurately forecast market demand going forward. And I will go one step further and point out that checks with some of the big four AIB partners peg Crypto demand at 50%+ recently, and I am not the only one who has heard this. But I guess Nvidia's official position till it guides the street down is if we aren't selling directly to a crypto miner, then it's not crypto demand, no matter how obvious it is to us where the card has ended.
But for all the public talk out of management, Nvidia's behind the scenes move seems to paint a different picture.
At the end of 2017, Nvidia modified its GeForce EULA to read this:
No Modification or Reverse Engineering. Customer may not modify (except as provided in Section 2.1.2), reverse engineer, decompile, or disassemble the SOFTWARE, nor attempt in any other manner to obtain the source code.
No Separation of Components. The SOFTWARE is licensed as a single product. Its component parts may not be separated for use on more than one computer, nor otherwise used separately from the other parts.
No Sublicensing or Distribution. Customer may not sell, rent, sublicense, distribute or transfer the SOFTWARE; or use the SOFTWARE for public performance or broadcast; or provide commercial hosting services with the SOFTWARE.
No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that blockchain processing in a datacenter is permitted.
Why are they worried about datacenter deployments of their gaming cards?
Here is a potential reason why:
This is a recent startup that launched which allows miners to rent their Nvidia GeForce cards out to deep learning researchers. Now, this is not a datacenter deployment, but in fact a clever workaround to leverage the distributed mining GPU compute universe which Nvidia's CEO loves to talk about.
"At the highest level the way to think about that is because of the philosophy of cryptocurrency - which is really about taking advantage of distributed high-performance computing - there are supercomputers in the hands of almost everybody in the world so that no singular force or entity that can control the currency."- Nvidia CEO, GTX Conference March 2018
Problem is if everyone has a supercomputer that can handle AI level workloads which the Nvidia GTX 1080TI can do quite well (and far cheaper on cost to compute than $8k-$10k Tesla V100 cards), that's not good news for Nvidia's crown jewel of selling high end GPUs to hyperscale cloud monsters.
Vectordash is a small startup created by couple college kids at the University of Maryland, and they are already seeing robust demand for their model.
You think Amazon and Google Cloud which are forbidden from deploying these consumer grade cards in datacenter are happy about this (Google actually may be, but we will address that later).
And if some broke college kids can put this together as a side project, what happens when some of the giant ether mining farms all over the world throw some cash and their GPUs at this space? Is Nvidia going to sue all of them even when most of them will do this with a view to cash in while they can and deal with the legal ramifications later?
Nvidia has had to resort to legal maneuvers to limit how its cards are used because they know a crypto crash can hit its booming demand for high end AI GPUs even though researchers have figured out their consumer cards are plenty capable of doing this work.
But these aren't the only moves Nvidia is making to exert influence over the GPU market.
Just a few days ago, there was big news on Nvidia's new controversial GeForce Partner Program.
Basically, Nvidia is pressuring big partners in the AIB market to disassociate their existing gaming brands from AMD. Why do this?
Well, in early 2019, AMD is slated to 7nm based single die gaming optimized GPUs. This should give AMD a manufacturing cost/performance edge over Nvidia. AMD also slated to launch a 7nm Vega based machine learning AI geared card before year-end. So, with crypto demand crash essentially guaranteeing 2017 is going to be messy year for GPU companies Nvidia is doing exactly what you would expect from a bully, making anticompetitive moves to hamper their competitor before they leapfrog them.
People close to the GPU space are quickly figuring this out, and soon, the rest of Wall Street will too.
This is why while Nvidia's CEO says they need to make more cards there website currently shows this:
All the GeForce cards are out of stock. This is despite the fact that at the peak of crypto demand the website always showed "notify me" and that GPU card prices have dropped meaningfully with new cards now easily findable on Amazon.com or Newegg.com.
What is Nvidia up to?
I think the answer to this question is pretty clear, but as Nvidia has said nothing we can simply speculate, and my conclusion is that Nvidia is now prepping for the crypto crash and has no desire to be sitting on new GeForce 10 card inventory competing against a tsunami of less than 1 yr. old second-hand market cards from miners.
Reports out of Digitimes in Taiwan also would seem to support this view.
GPU demand from the cryptocurrency mining industry is showing signs of slowdown recently, and Nvidia has started taking measures to minimize possible damage, according to some market sources.
Taiwan Semiconductor Manufacturing Company chairman Morris Chang has recently said that he expects the company to see an on-year revenue growth of 10-15% in the first half of 2018 thanks to the demand from the cryptocurrency segment. The market watchers believe many cryptocurrency miners are likely to turn to procure ASICs from suppliers such as Bitmain.
Bitmain is ready to release ASIC products in April eyeing cryptocurrencies that have been relying on GPUs for mining, and the move is expected to reduce miners' demand for graphics cards.
Policies from governments worldwide on cryptocurrencies and significant price changes also have weakened the returns for mining.
Seeing the trend, Nvidia has recently started placing restrictions on its downstream graphics card partners, forbidding them to publicly promote cryptocurrency mining activities or actively sell its consumer graphics cards to miners, the sources said. Nvidia hopes to shift its main sales target back to consumers in the gaming market, the sources added.
Nvidia also has further increased its GPU quotes recently, which the sources believe is meant to help cover the gap that may occur after GPU demand starts sliding.
Since profitability from graphics cards has been weakening, Nvidia and AMD have both been decelerating the developments of their new GPU architectures and prolonging their existing GPU platforms' lifecycle, the sources said, adding Nvidia's next-generation GPU architecture Turing will not enter the mass production until the third quarter.
Nvidia can publicly claim to want to ease the pain on gamers and AI researchers, but every strategic move they are making indicates they are now in fact protecting themselves from misreading the crypto market, threats to their datacenter high end cards, and future threats from AMD and other competitors. And Wall Street, to a large degree while celebrating Nvidia's other business lines as diversification/insulation, has missed the boat that this in fact has exposed them even more because their consumer cards offer the best utility for AI/Gaming and hence why they have experienced much more than realized demand from miners.
This means be prepared for meaningful noise (those numbers are coming down!!) in Nvidia's guidance next couple of quarters as they deal with their very fluid situation. By my own estimates, Nvidia is facing a multi-billion dollar-revenue headwind going forward when you account for Crypto and Tegra Nintendo Switch revenue tailwind in 2017. This would be enough to run for the exits on a beat and raise momentum darling that is up 12x over past two years, but Nvidia's issues don't stop there.
Nvidia's Much Bigger Problem
No matter what your position is on crypto, there is no denying that there is a very consensual view out there that Nvidia is a diversified growth monster. Cramer loves to talk about AI, Datacenter, Autonomous Cars, Gaming, and HPC. Their CEO was even out last week calling these all much bigger businesses than crypto. And while that may in fact be true over the long haul, there is no denying crypto slotted in right behind gaming and data center in 2017. But the way these businesses lines are viewed is misleading.
Nvidia has essentially two business lines:
Graphics - Gaming/Professional GPU Accelerated Computing - This essentially encompasses AI/Datacenter/Autonomous/HPC etc. as it's all about the GPU being used to boost computational performance largely for ML workloads.
Now, when you view Nvidia this way, you start to appreciate the risks that have quickly emerged for essentially the core part of their entire growth story.
The GPU was in the right place at the right time as far as being the optimal solution available to handle the explosion of ML workloads. And there is no denying Nvidia has executed very well on that opportunity unlike AMD and other potential competitors.
2017 was the year of the headlines being dominated by the GPU because of the convergence of crypto and AI mania, but in the background, a quiet hardware innovation explosion has been taking place that threatens to completely displace the GPU as the de facto accelerator for AI workloads.
To fully grasp the issues facing Nvidia in GPU Accelerated ML/DL Computing, let's take a little journey into the history of Bitcoin mining
A Brief History of Bitcoin Compute Mining
The CPU is how we started with Satoshi Nakamoto mining one million Bitcoins with allegedly nothing more than a couple Intel CPUs.
Then, in 2010, a man named Laszlo Hanecz started to use his GPU to mine Bitcoin. Laszlo was finding about a block a day with his CPU, but once he got his GPU cracking, he was finding multiple blocks an hour. Laszlo quickly accumulated a huge stash of Bitcoin, and then entered into the now infamous transaction of 10,000 Bitcoins for two Domino's Pizzas (at their peak last year that worked out to a $100 million pizza). But the bottom line here was there was a nice GPU mining performance boost over the CPU that was roughly 30x.
Then in less than a year, you had mining geeks experimenting with FPGAs which brought power efficiency over the GPUs, but for the most part the market remained in the purvey of the GPU.
Until 2013, when the first ASICs arrived offering 10-15x performance boosts over GPUs with lower power consumption. This was the end of the GPU in Bitcoin mining, and we have been on ASICs ever since. These ASICs have quickly evolved to the point that 4 years later they are on the equivalent of another computing planet. Today, Bitmain's Antminer S9 delivers 13TH/S of hashpower for Bitcoin mining. That's 13,000x a high-end GPU, and oh, it consumes 1,300 watts so its 1,000x more power efficient.
To put this all this in perspective, at the current Bitcoin network hashrate of 27,000 PH/S, you are talking the equivalent of 10 billion Nvidia/AMD GPUs.
One high end Nvidia GPU will take 38,460 years to mine a bitcoin. Basically, you are about 35x more likely to win last week's $500 million Mega Millions jackpot (1/302million) than find a bitcoin block with your single high-end GPU. Think about these stats for a second. If the Bitcoin network was powered by the world's best GPUs, we'd literally have destroyed the Planet mining a virtual currency.
This is why regardless of your views on crypto, one thing is clear; the GPU has no future in the space.
There are 4 very simple reasons as to why that is.
Interest Alignment Decentralization/Security fallacy Energy Waste Crypto Currency Aspirations for Main Stream Acceptance
If you buy an ASIC miner, it's an upfront investment into mining of your currency and support of that network. If the currency price falls or the network falters, that's a sunk cost you will never recover. However, with a GPU mining, alt coins that don't have ASICs, this alignment is not there. You simply switch to a more profitable coin, and when you run out of profitable coins, maybe you use your card for research or gaming. Yet there are all these crypto purists out there who think alt coins need to be ASIC resistant, so some evil empire like Bitmain doesn't come in and dominate the network.
The problem with that argument is it really doesn't hold water. There are between 10 million and 14 million GPUs on the Ethereum network today. The largest mining pool in that network, Ethermine, has 25% of the hash power. That pool dwarfs the size of any other alt coin. So, in theory, all those GPU resistant alt-coins are quite susceptible to a 51% attack. Furthermore, there are some huge GPU mining farms just as you have the likes of Bitmain in Bitcoin, so the decentralization argument falls apart. Now, the energy argument is common sense, so all that is left is the cryptocurrency reaching the mainstream viewpoint. The GPU is fine and dandy up until the point a coin achieves some level of success, then the compute is simply not power/cost efficient. So, if you are a GPU mining advocate for crypto/blockchain, you need an endless cycle of speculative coins to keep transitioning to. The problem is that dynamic only exists in a speculative bubble. When a bubble pops winners and the losers emerge. If you are a winning crypto/blockchain network with aspirations of being super secure and having billions of transactions on your network, you are going to be relying on ASICs or ASIC Cloud technology. There is no getting around that from a proof of work standpoint.
Anyway, the evolution in cryptomining provides a nice segway into what I really want to focus on which is the revolution going on in purpose-built hardware for ML/DL workloads.
2018: The Year of Purpose Built ML/DL Compute Accelerating Hardware
While Nvidia did get lucky with the GPU basically being the best available option for AI workloads, it also has executed quite well. One of the main reasons I got bullish on the stock several years ago was because of their software/ecosystem edge in the space. Notably, CUDA, Nvidia's API that allows developers to access the GPU instructions set and run their models. This is why despite the fact that there are two big GPU companies, you never hear about AMD in DL/ML. This is because Nvidia has spent the last decade developing CUDA (and being punished by Wall Street for quite a while for that investment), and this allowed them to capture the developer ecosystem around ML/DL. AMD's OpenCL platform and DL libraries have been playing catchup, and Nvidia has reaped the benefits. But despite CUDA's edge, the AI space had up until recently lacked a dominant development platform to unify around.
Google TensorFlow has recently emerged as the leading software platform for ML/DL developers. Naturally, as the leading platform, it provides seamless integration with Nvidia's CUDA ecosystem, and one could argue that's been great for Nvidia. TensorFlow has simplified things for AI researchers, and as the leading hardware player in the space, Nvidia has benefited the most. I naturally disagree. Nvidia's been essentially the ONLY HARDWARE PLAYER in the space, and TensorFlow is going to end up being the Trojan Horse for the enemies at their gate. Were as before hardware developers focusing in ML/DL acceleration solutions had a mishmash of things to choose from (MXNet, Caffe, Pytorch), they now have one platform to focus on as far as support and integration goes. We've gone from several hard to define moving targets to one static and clear target.
"Things like TensorFlow make our lives so much easier…Once you have a neural network described on TensorFlow, it's on us to take that and translate that onto our chip. We can abstract that difficulty by having an automatic compiler." Michael Henry, CEO MYTHIC
"Nvidia has spent a long time building an ecosystem around their GPUs, and for the most part, with the combination of TensorFlow, Google has killed most of its value," Feldman said at the conference. "What TensorFlow does is, it says to researchers and AI professionals, you don't have to get into the guts of the hardware. You can write at the upper layers and you can write in Python, you can use scripts, you don't have to worry about what's happening underneath. Then you can compile it very simply and directly to a CPU, TPU, GPU, to many different hardwares, including ours. If in order to do that work, you have to be the type of engineer that can do hand-tuned assembly or can live deep in the guts of hardware, there will be no adoption… We'll just take in their TensorFlow, we don't have to worry about anything else."- Andrew Feldman CEO Cerebras
So, as you can see the competition is quite happy with TensorFlow, and in a minute, we will get to what they are up to. But before we go down that path, let's address Google.
Google: AI's Elephant In The Room
Google gets very little headline press in the investment community as a pure play in AI, but reality is they are the 800-pound gorilla in the space. For all the talk about hardware, data is arguably the true moat in AI and Google has mountains of it. But whether we are talking speech recognition, Search, GO, home automation, Chess, Health, Drones, Cameras, Waymo, or autonomous robots, you can't argue that Google doesn't have the application lead in AI. There is nobody even close, and for that reason, Elon Musk has a very good argument about fearing a Google-dominated AI world. But we will leave that conversation for another day as I really just want to focus more on what investors should do when a company like this makes moves to address its internal needs in area so critical to its core business. The short answer to that is PAY CLOSE ATTENTION.
In late 2013/early 2014, Google made an internal determination that fast-growing computational demands of neural networks could require them to double their data center footprint in a few years. This was of course unacceptable, so Google set out to solve this problem. They started buying up GPUs for training and launched a high priority project to develop an accelerator for their AI inference workloads. 15 months later, the Google TPU was designed, tested, verified, and ultimately deployed in their datacenter.
Think about that for a second……
Compare this to the Bitcoin compute evolution time table. In 15 months, Google had an ASIC accelerator deployed and functioning in their datacenter, and it made economic sense. That's nothing short of incredible. And within three years of project launch, they have TPU2 deployed which can do training as well as inferencing. A year later, they have made the TPU2 available in their public cloud beta and are apparently now going to volume production on these chips with Broadcom.
"TPUs are behind every search query; they power accurate vision models that underlie products like Google Image Search, Google Photos and the Google Cloud Vision API; they underpin the groundbreaking quality improvements that Google Translate rolled out last year; and they were instrumental in Google DeepMind's victory over Lee Sedol, the first instance of a computer defeating a world champion in the ancient game of Go." - GoogleBlog
Behind every search query….
Google has essentially built a walled garden around its AI empire in record time, and yet Wall Street and the likes of Cramer don't seem to think this is an issue for Nvidia. If you are Google and are now fully self-reliant on your cost advantaged AI hardware, what do you think you are going to do with your cloud access to it? The short answer to this question is weaponize it AWS style. It's basically a matter of time before Google starts slashing prices because it can afford to subsidize this business to make life miserable for all other hardware competitors in the space. Anyway, whatever your view on Google may be, there is no denying the fact that they have plotted a course that is extremely negative for Nvidia. This is Apple-esque behavior with respect to component suppliers, and it guarantees that the other big hyperscale players with huge data moats and other giant profit machines/financial resources will also seek to make similar investments to be able to compete.
Now onto the up and comers looking to break into this space guys….
Groq is an ideal place to start ML/DL hardware startup wise because the team consists of 8/10 of the engineers that built Google's TPU in record time. They are backed by Social Capital and have started to release data on their chip due out later this year. One can presume they aim to bring a more powerful version of Google's TPU to the competition which of course makes the name quite interesting.
From the GROQ website:
Wave was founded over 5 years ago and has been gradually building a very different solution to ML/DL workload problem. Instead of focusing on acceleration in the datacenter, they have made a self-contained ML/DL solution in a box.
"With the exception of Google's TPUs, the vast majority of training is currently done on standard Xeon servers using Nvidia GPUs for acceleration. Wave's dataflow architecture is different. The Dataflow Processing Unit (DPU) does not need a host CPU and consists of thousands of tiny, self-timed processing elements designed for the 8-bit integer operations commonly used in neural networks."- ZDNET
They also have been poking Nvidia of late on their LinkedIn page…
This is stealthy startup that has created all kinds of buzz lately. The main reason is because tech Midas touch blessed VC Sequoia just invested $50 million. The company can also count Google Deep Mind founder and UBER's Chief Scientist as personal investors.
What do they do?
"Graphcore's IPU (Intelligence Processing Unit) is a new AI accelerator bringing an unprecedented level of performance to both current and future machine learning workloads. Its unique combination of massively parallel multi-tasking compute, synchronized execution within an IPU or across multiple IPUs, innovative data exchange fabric and large amounts of on-chip SRAM give unheard of capabilities for both training and inference across a large range of machine learning algorithms."- Graphcore
"The performance gain is substantial even at smaller batch sizes. When we scale up to eight C2 accelerator cards we only use a batch size of 64.
At any point in this space, using an IPU system is a substantial performance leap over existing technologies. For example, the best performance reported on a 300W GPU accelerator (the same power budget as a C2 accelerator) is approximately 580 images per second."- Graphcore
This is another stealthy startup that we know little about, but has reached a billion valuation. The interesting thing about Cerebras is their CEO and founding team come from the AMD acquired Seamicro and seem to have GPU expertise. Also, their CEO has been quite outspoken with respect to the GPU's inherent limitations with ML/DL and why it will be replaced by new hardware. I am watching this one closely.
This company just emerged from stealth last week announcing a $56 million round led by Google Ventures. Naturally, this is interesting as it's Google making a hardware related ML investment. What do they do? Again, a little mysterious but seems broader based approach with maybe software centric focus?
"Dave Munichiello, general partner at GV, noted that the market for dedicated AI hardware continues to grow rapidly. "This team uniquely understands how critical adaptability and flexibility are for ever-evolving artificial intelligence and machine learning approaches," he said. "Other platforms have been designed for AI and machine learning techniques that exist today. SambaNova's software-defined infrastructure anticipates and supports a rapidly evolving ecosystem. We firmly believe that over time this computing approach will lead the industry in distributed machine learning and data analytics infrastructure."- Siliconangle
Yeah, those bitcoin mining maniacs are now a player in the AI space. Bitmain has been working on a TPU solution for a while, and in November, they launched their first product. Their BM1680 chip which comes in a card or server solution promptly sold out. Customers have started to receive shipments the past few two weeks, and Bitmain has a full roadmap here with their next gen chip slated for fall availability. Suffice to say Bitmain has gotten insanely rich of the mining boom and may end up being exactly what China (and not the rest of chipland) desperately desires in the form of a homegrown formidable dedicated leading AI chip manufacturer.
This is Chinese AI Unicorn that aims to grab 30% market share of the ML/DL Chinese market in the next 3 years.
Here is what they are up to:
"Cambricon developed the Cambricon-1A, a chip for deep learning applications. "The Cambricon-1A processor is specifically designed for AI deep learning, and outperforms traditional processors in graphic and voice recognition by at least two orders of magnitude," Chen stated.
"It also has a high integration density several times that of traditional processors, making it possible to install AI chips on mobile devices."
Cambricon intellectual properties are already in use by Huawei Technologies Co Ltd, for its latest smartphone, the Mate 10, and by State-owned Sugon Information Industry Co Ltd, to improve server operations computing speeds, and reasoning abilities.
Cambricon also announced 3 new intellectual properties, which will support low-power applications in face recognition and driving applications.
"We will focus on both in-device AI and cloud AI. But we won't make chips for consumer electronic devices ourselves. We will sell our intellectual properties to hardware makers so that they can better integrate AI into their in-house chips," Chen said.
Then, you have Mythic, DeePhi Tech, Horizon Robotics, and many more AI hardware startups plays from those focusing on the edge and inference to total solutions. The point is in the vast majority of these cases, the companies have something ready to go or will have something ready to go this year. This isn't far off down the road dreaming type stuff, and the investment appetite in the private market has been huge lately which opens the door to many IPOS for pre-revenue scale companies in semi land the likes of which have not been seen in decades.
And then you have what's going on with the other big boys.
Microsoft's Project Brainwave is using Intel's Altera FPGAs to provide 'real time AI' datacenter acceleration. You can draw your own conclusions on how this pans out, but the bottom line is Microsoft is willing to stake a hardware differentiation approach much like Google vs eating the whole Nvidia pie. The story here being ultra-low latency!
Intel also has been saying that their Nervana deep learning accelerator chips will be revenue generating by the end of the year. Then, you have Amazon and Facebook and of course Apple. The former two are up to all kinds of things in the space none of which appear super clear yet. But I can say that we do know another hyperscale player has a dedicated AI accelerator chip coming from TSM in H2 2018, and if I was to guess it would be either Amazon or Facebook who has pulled this off already just to keep pace with Google. Apple has been working more on phone AI chips, and has started getting vocal with AI papers and a blog as they exit Willy Wonka mode for the collaboration driven competitive world of AI. Then, you have Qualcomm aspirations at the edge, Arm aspirations, Cavium, Xlnx, and many more. This is obviously a lot to digest, and I could probably spend another 50 pages writing about the things I have learned deep diving into the space over the past six months, but unfortunately, it's time to wrap up.
Ok enough dilly dallying around with research. Here is the bottom line. From a investor perception standpoint, it's never going to be as good as it was for Nvidia in 2017 in GPU accelerated computing. That you can be certain of. Whether we are talking Google TPU2, Intel Nervana, AMD's recent progress with TensorFlow on Github, the arrival of Bitmain, the end of crypto mania, one of the pure plays in AI startups taking off, rising Chinese competition etc. Nvidia's investors better be getting ready for a new normal. There is nowhere to go but down from here, and the range of possible outcomes is about as wide as they could ever be. So, simply short the stock. I am of the view that the shares will nearly get cut in half from their recent peak, and for sure be spending the rest of their time in 2018 in the sub $200 range. And if you are looking for a hedge here in chip land, take a gander at Broadcom. They have been dragged down by all the QCOM takeover mess, but in the background, they have been quietly building a custom AI ASIC business which could be a multi-billion revenue story in a very short time. You could also look at Google if you been paying much attention here. They look like the big winner with Facebook's AD empire under siege, Tesla's growing problems, and Nvidia's soon to be very well recognized issues. I kind of like that pair as well.
Anyway, if this report hasn't convinced you that selling Nvidia here is a no brainer, nothing will.