As certain governmental agencies and others have discovered, demand has a way of creating its own supply. This is a universal truth. In the early 1970s, American demand for French Bordeaux wines outstripped the capacity of the vineyards in France's Bordeaux region, where (legally) all such wines originate. No problem. France simply exported more Bordeaux wines to the U.S. than its total production of those wines! (Yes, you read that right.) And France continued to do so for years.
The American wine frenzy spilled over to benefit California's Napa Valley wineries. There was more demand for Napa wines than could be produced. No problem. Fully loaded 18-wheelers rumbled north from California's fecund but déclassé Imperial Valley vineyards through the Central Valley at night, bumper-to-bumper at 80 MPH. Unloading in black of night at various Napa wineries, they returned southward, empty, by light of day. Suddenly and magically, Napa wine supply met demand.
The Italians satisfied American demand for Chianti wines in a more original manner. Lacking grapes, the Italians turned to garbage and chemical vats. Using rotten bananas (which worked better than fresh) as a principal ingredient, ersatz Chianti was chemically brewed. A lot of this stuff was shipped to the U.S., where it sold very well indeed. Salute!
Alcohol is the defining ingredient of wine, and alcohol is not heavily used -- or abused -- by programmers. Nosirree. You favor caffeine, and many of you well-paid programmers get your caffeine not from ordinary coffee but from gourmet coffee, hand-picked from Hawaiian slopes by certified virgins (half of them male). Here in the Silicon Valley, there are chains of outlets providing fine coffees for immediate consumption or packaged for home brewing.
Demand for fine coffees skyrocketed, and demand does have a way of creating its own supply. In my November/December 1995 column, I mentioned a grocer who sold coffee at three prices -- all beans from the same bag. A certain coffee-bean wholesaler had anticipated me. For years, this wholesaler has been purchasing the meanest and cheapest of beans and reselling them as fine varietals. Among his customers were the two largest Silicon Valley gourmet coffee chains! And the chains' owners, managers, and (most importantly) customers didn't spot the substitution! A shipping paperwork foulup finally revealed the fraud.
Wine industry spokespersons, the honest ones, will admit that there were "certain abuses" in the past, but that these have, naturellement, been corrected. And the gourmet coffee merchants will (must!) admit the mistakes of the very recent past, but insist that corrections have been made (presumably in the last five minutes). Would you like another cup of this truly superb brew while you review my brochure showing the magnificent Montana oceanfront property I have for sale?
If the elusive power user actually exists (that is, if this Jabberwock ain't really a Boojum), then he or she reads Dr. Dobb's. The power user compiles 100K-line C++ programs while receiving faxes, reviewing e-mail, using Pagemaker to write a massive technical manual, running PhotoShop to do color separations for the manual, recalculating a spreadsheet, and playing a networked version of Quake -- all at once. As if this didn't place enough of a strain on your computer system, along come host-based high-resolution color ink-jet printers that eat another 8 to 12 MB of DRAM.
You need more memory! I wrote about this in the September/October 1996 Dr. Dobb's Sourcebook, when the cost of 32 MB of SIMMs had dropped to $210 at the local Fry's. As of the weekend before Christmas, that had dropped to $88 for 32 MB of EDO SIMMs. That's a 58 percent price drop in just six months!
Which is a very good thing, because you need 48 MB -- and you really oughta spring for a full 64 MB -- just to run Windows 95. I'm not joking even a little bit. This (48 MB of DRAM) is the most cost-effective way to speed up your Windows system. Forget dumping money into wide SCSI to speed swapping, just install enough DRAM so that no swapping occurs!
If you're not a power user and/or you don't have a host-based (GDI) printer, then you can get by with 32 MB or even a tad less. I'm limping along with a mere 36 MB at the moment -- I'm not a power user but I do have a 720 DPI Epson Stylus Color printer. I must confess that I never wholeheartedly embraced the legend of the power user.
(Now that enough memory to run NT, or even UNIX, costs less than $200, I wonder how the advocates and marketers of these operating systems are going to explain their failure to take over the world in 1997.)
If you're in the market for a new motherboard to go with your new MMX CPU, here's something to think about. If the motherboard uses one of Intel's chipsets, the choices are the 430FX, 430VX, and 430HX. This is important: Do not even think about buying anything that's not based on the 430HX. The FX and VX are both limited to L2-caching a mere 64 MB of DRAM.
Some folk are stuffing 96 to 128 MB into motherboards using the FX or VX chipsets. This works...slowly. Win95 loads into high memory and, with more than 64 MB on board, that high memory is not cached at the L2 level. Instant slowdown.
The 430HX chipset assigns up to three additional bits to the L2 tag, allowing up to 512 MB of DRAM to be cached. The Intel Marl ATX motherboard, for example, uses just one of those three extra bits to assure that all of its maximum capacity (128 MB) can be L2 cached.
My editor would shoot me if I tried to print the same 37-year graphic again after only six months, so check Figure 2 on page 51 of the September/October 1996 Dr. Dobb's Sourcebook (issue #259) to see what happened ten years ago. While you're there, plot a new point of $2.75/MB at December 15, 1996.
If the U.S. government does not repeat its stupidities of ten years back, the DRAM market will now stabilize and show a gradual price decline over the next year, with prices dropping to about $1.80/MB by next Christmas.
Recent events suggest that our government is capable of learning from its mistakes. Perhaps it won't shoot itself in the foot with a 155mm howitzer in an attempt to protect domestic DRAM producers such as Micron or NEC (which has a DRAM fab in Marysville, California).
The problem is, our government is supposed to protect NEC from dumping by Samsung. Dumping is when the Samsungs (read: nondomestic DRAM suppliers) sell us DRAMs too cheaply. What's too cheaply? Below cost. What's "cost"? That happens to be an interesting question.
In recent months, I've seen announcements that this DRAM maker or that has cut back on DRAM production. Yet DRAM prices continue to drop.
I don't think any DRAM maker has, in fact, cut production.
There is this thing called profit, and its corollary, loss. Many folk think profit is a tangible thing you can wrap your hands around. I ran small businesses for a couple of decades and I can assure you that "profit" is blue smoke. In fact, governments collect taxes, and one of the things they tax is companies. Different government entities, by tacit agreement, tax different things such as inventory, sales, and profit. The Feds concentrate on profit.
This means the Feds have to define profit, else everybody would claim to be operating at a loss. In the long run, the Feds' definition of profit is pretty reasonable. But in the short term, profit is hard to pin down. The rules of capital-equipment depreciation, which assume the future viability of the capital equipment, contribute to the confusion.
Which brings us to DRAM production: If the bottom suddenly drops out of the DRAM market, just how valuable is the DRAM industry's principal capital equipment -- the DRAM fabs? In other words, the governmental rules that define "profits" do not accommodate short-term variations worth a damn. I say again, in the short term, profit is blue smoke.
Now, about DRAM production, profit per chip, and cutbacks:
Today's state-of-the-art in-production DRAM fab probably cost about $1.2 billion originally. A DRAM fab being started now will probably have cost $2 billion by the time it enters production two years hence. Let's use that $1.2 billion figure and assume the fab has a 5-year life. It costs the fab owner around $300 million per year over that 5-year period, including interest costs. For a fab started today, the cost is over $500 million per year.
But for an existing fab, we'll use that $300 million figure. The fab owner is out the $300 million whether or not the fab is in use. The profit on a given DRAM depends on the fab cost, raw materials cost, operating expenses (mostly salaries), the number of DRAMs produced (and sold), and their ASP (average selling price). Fab costs include the testers. Materials cost includes stuff like packages. Operating expenses include testing, sales, marketing, and corporate overhead. We'll include any out-of-pocket expense in one of those three categories. So nobody is going to sell DRAMs at a loss, hmm?
Wrong. That $300 million is a done deal. Maybe it was a good investment (if DRAM prices are high) or a bad investment (if DRAM prices are low), but it's a done deal. The question is, should the fab be used to produce DRAMs or not? If I'm managing that fab, the answer is based on whether I can sell the DRAMs for more than the materials cost plus operating expenses.
Let's grab some numbers out of the air to go with that $300 million per year fab cost: Full production is $300 million/yr materials cost and $100 million operating expenses. For this, we have 100 million DRAMs we can sell. If we can't sell any of those DRAMs, we're out $700 million a year. And if we price the parts much above the going price, we ain't gonna sell any.
If we can sell the DRAMs for over $7 ASP, then we make a profit by government accounting rules, and we amortize the price of the fab (we cover the payments to the bank to repay the loan to build the fab).
If we sell the DRAMs for $5 ASP, accounting rules assert that we are losing $2 per DRAM. But if I'm running the fab, no way am I gonna shut the fab down. Why? If I shut the fab down, I'm out $300 million per year. If I run the fab full tilt, I only lose $200 million per year. It's a no brainer. By keeping the production line open, I'm ahead $100 million per year.
I grabbed those numbers out of the air. The point is, it makes a great deal of sense to keep DRAM production going 100 percent even when, by government-imposed accounting rules, each DRAM is being sold at a loss.
That's why I don't believe announcements of DRAM production cutbacks. It's an OPEC situation. The industry as a whole would benefit if all DRAM makers cut back until prices rose to a profitable level (if they are not already there). But each individual DRAM producer makes the most money (loses the least money) by keeping its production lines running full tilt. (Happily for us DRAM consumers, production is not concentrated in any one country, else that country would impose mandatory cutbacks to assure profitability. As it is, the Japanese aren't about to cut production to benefit Korean DRAM producers and vice versa. The Japanese and Koreans aren't terribly supportive of each other.)
If it's my fab, determining production levels is an easy decision. I keep the fab running until price levels drop below material plus operating costs. Obviously, that hasn't happened yet. Because if it did, everybody would shut down production. My opinion is that DRAM sales are still profitable, just not at the obscene levels prevailing over most of the past 5+ years.
If only we didn't have to worry about our own government hovering like Banquo's ghost over the scene.
DDJ