The popular topic in the memory industry is record high DRAM pricing, but today I'd like to examine cost. First, a factor in the relative cost situation in the industry. Then, a ramification of the low cost of DRAM production in general.
In past articles I've mentioned that each weakening of one yen in the dollar/yen exchange rate gives Micron (NASDAQ:MU) a cost savings of $5 million to $7 million per quarter, so for a 2 yen drop Micron makes an extra penny per share per quarter. This is just the tip of the iceberg when it comes to currency issues in the memory industry.
69% of Micron's revenue and the vast majority of its profit comes from DRAM. There are three remaining major DRAM manufacturers: Samsung (OTC:SSNLF), Micron, and Hynix.
Hynix makes about 280k wafers per month of DRAM, with half coming from its M10 fab in Icheon, Korea, and the other half coming from its fab in Wuxi, China, well known for its fire last year. Hynix has stated that it will have to replace the antiquated M10 shell next year in order to be able to continue with any future die shrinks. Hynix's non-equipment costs in Icheon, including construction, staff, electricity, etc. are denominated in the Korean won, while in Wuxi the costs are in the Chinese yuan.
Most of Samsung's DRAM production is in the world's biggest DRAM production plant in Giheung, Korea. Last decade it was part of a $33 billion chip producing expansion. Samsung also makes a minor amount of DRAM in Austin, Texas. Samsung's $7 billion Xi'an, China fab doesn't make DRAM at all, and in NAND only makes very expensive 3D NAND. Most of Samsung's DRAM non-equipment expenses are in won, with a minor portion in US dollars.
Micron produces around 320k - 330k wafers of DRAM per month. 100k wafers are in Hiroshima, Japan (formerly Elpida). 80k wafers are in Taiwan (formerly Rexchip). 120k wafers are from partially owned Inotera, also in Taiwan. The remaining 20k - 30k wafers are from Virginia. Much of the expense is split between Taiwanese dollars and Japanese yen. Micron also has NAND facilities in Singapore and the US.
Let's compare these currencies and how these costs have trended over time:
The Korean won has gotten consistently stronger since the recession and is hitting new highs. Here are the 1 and 5 year USD/KRW charts which show a declining position for Samsung and Hynix vs. Micron as their won-denominated costs rise in relation to Micron's dollar costs.
Micron entered into an agreement to buy bankrupt Elpida a little over 2 years ago. Since then, Japanese costs have plummeted compared to Korean costs.
It's a similar story for the Taiwanese dollar. Rexchip and Inotera have seen steadily improving costs when compared to Samsung and Hynix.
The dollar has gotten gradually weaker against the Chinese yuan over the long term, leading to a gradually worsening relative cost situation for Hynix's Wuxi fab.
Samsung and Hynix are seeing reduced profits from a continually strengthening won. This is a key factor in return on invested capital calculations when they are deciding whether to expand production at their current facilities.
Memory In Processor, or Processor In Memory?
In a recent article, Seeking Alpha contributor Russ Fischer argued that Intel (NASDAQ:INTC) would integrate DRAM onto PC CPUs and mobile application processors. I believe this is precisely the opposite of the truth.
My initial retort was that integrating DRAM and the CPU would result in a larger die which would mean lower yields (meaning that the percentage of chips with unacceptable flaws or contamination increases with the size of the chip, a universal truth in the chip industry), and hence higher costs. It's also clear that Micron and Intel are already cooperating on the Hybrid Memory Cube which already tackles the problem of the "Memory Wall" at much lower cost than integrating the processor and memory. The HMC results in 15 times the bandwidth of DDR3 with 70% less energy usage, while also having a smaller physical footprint. This solution is already planned for a supercomputer that will use about 300,000 DRAM chips. This product is a result of a close collaboration with Intel, and in fact goes beyond the HMC. While it's true that both yield issues and the HMC argue against Intel's integration of large amounts of DRAM in large numbers of chips, the real story is a bit more complicated and much more interesting.
Microprocessor designer Russell Fish (no, not Russ Fischer, just a funny coincidence), inventor of one of IEEE's "25 microchips that shook the world" argues that CPU transistors cost hundreds of times as much as memory transistors. In Fish's words:
"All attempted to include main memory and CPUs on a single chip. All essentially succeeded in proving that the Memory Wall could be breached, but none made the slightest ripple in the commercial world. As product with commercial prospects they failed for the same reason an Itanium with cache the size of main memory would fail. The cost would have bankrupted King Solomon's Mines."
The problem is that CPUs are complex and have 9 or more metal layers, while DRAM is a simpler low power part with just 3 metal layers. Building DRAM on the CPU is like building a Hyundai in a Rolls Royce factory -- you can do it, but you won't like the cost. Fish's newest innovative solution is to do just the opposite and include simplified processors on a DRAM chip. If you're thinking that's something Micron should be doing then you're right, and they are. Micron's new product is called the Automata Processor, which I wrote about in December. While Fish's design is a von Neumann architecture chip, Micron's is a non-von Neumann chip, meaning that Micron's AP is not direct competition for Intel's CPUs, but rather for its coprocessors and for the FPGAs of companies like Xilinx (NASDAQ:XLNX). Perhaps Micron will expand to von Neumann Processor-In-Memory chips in the future.
Intel may integrate small amounts of memory (beyond just cache memory) in a small number of its high end chips, but it's simply too expensive to do on a large scale. It's far more likely that we'll see the opposite happen -- Micron will integrate processors into its memory chips resulting in processors that are low power, high speed, and yet still cheap. While this solution will initially benefit Big Data, the ultra-low power nature of the solution may ultimately benefit the Internet of Things, where the "Power Wall" considerations are paramount.
DRAM gets continuously cheaper to make on a per bit basis, and now Micron's relative cost level is improving when compared to its two remaining competitors. This low cost (and of course the record high selling prices) will lead to profits that will continue to surprise the analysts, and in the longer term will lead to new product opportunities where Micron can leverage its cost and experience advantages.
Disclosure: The author is long MU. The author wrote this article themselves, and it expresses their own opinions. The author is not receiving compensation for it (other than from Seeking Alpha). The author has no business relationship with any company whose stock is mentioned in this article.