Slowing Moore's Law: How It Could Happen

Weak points in the networks powering technological progress: chip factories
transhumanism, politics, predictions, survey
2012-03-162017-10-09 finished certainty: likely importance: 7

requires enor­mous com­put­ing pow­er; enor­mous com­put­ing power requires fur­ther pro­gres­sion of ; fur­ther Moore’s law relies on large-s­cale pro­duc­tion of cheap proces­sors in ever more-ad­vanced ; cut­ting-edge chip fabs are both expen­sive and vul­ner­a­ble to state actors (but not non-s­tate actors such as ter­ror­ists). There­fore: the advent of brain emu­la­tion can be delayed by global reg­u­la­tion of chip fabs.

Here I will defend the fol­low­ing chain of claims:

  1. that changes in Moore’s Law can affect the cre­ation of vari­eties of soft­ware, and AI soft­ware in par­tic­u­lar

  2. the semi­con­duc­tor indus­try is extremely cen­tral­ized, phys­i­cally

  3. these cen­tral­ized facil­i­ties are highly vul­ner­a­ble

    • But coun­ter­in­tu­itive­ly, not to small groups such as ter­ror­ists.
  4. and thus, Moore’s Law can be steered by affect­ing these facil­i­ties

Why might one be inter­ested in this top­ic? Chip fab risks and costs may turn out to be a major lim­it­ing fac­tor on Moore’s law, and a fac­tor that has gone almost entirely unrec­og­nized (eg. who has ever heard of “Moore’s Sec­ond Law” gov­ern­ing chip fab­s?). The progress of Moore’s law is of major impor­tance to any future fore­casts of eco­nomic growth, progress of both research & appli­ca­tion of & (“big data” vs clever algo­rithm­s), soci­o­log­i­cal devel­op­ments, and all fore­casts of a tech­no­log­i­cal Sin­gu­lar­i­ty. Any­one inter­ested in a bet­ter under­stand­ing of the top­ics, more accu­rate fore­casts, or per­haps even affect­ing events direct­ly. The fragility of chip fabs is of inde­pen­dent inter­est as it is not gen­er­ally appre­ci­at­ed, and has poten­tially global con­se­quences—­for exam­ple, a con­flict between China & Tai­wan (which is home to many cut­ting-edge chip fabs) would, at the least, inter­rupt semi­con­duc­tor chip deliv­er­ies to the rest of the world, a con­tin­gency which may not have been ade­quately pre­pared for.


In par­tic­u­lar: here are 2 main pro­posed routes to a :

  1. (WBE)/upload
  2. de novo (AGI)1.

WBE is gen­er­ally believed to be rel­a­tively straight­for­ward engi­neer­ing, but is expected to take 1018 FLOPS2 (su­per­com­puter esti­mate: by 2019) to 1043 FLOPS3 (su­per­com­put­er: by 2111) to do in com­par­i­son with the lat­ter which either takes a sim­i­lar amount (if it is as inef­fi­cient as emu­lated brains and winds up effec­tively being a syn­thetic upload) or poten­tially much less if there are effi­cient algo­rithms and opti­miza­tions unavail­able to evo­lu­tion through brains or unavail­able period with brains. Hence, we expect that uploads will be made more likely by pow­er­ful hard­ware and de novo AGI be made more likely by pow­er­ful software/algorithms/mathematical insight. This obser­va­tion imme­di­ately sug­gests that any slow­ing in hard­ware devel­oped will reduce the prob­a­bil­ity of uploads com­ing before de novo AGI (what­ever that prob­a­bil­ity is), and vice ver­sa, any slow­ing in software/math will reduce the prob­a­bil­ity of de novo AGI com­ing before uploads (what­ever that prob­a­bil­ity was). A slow­ing in both may not change the prob­a­bil­i­ties or rel­a­tive dates of them, but rather post­pone the appear­ance of either or both.

Why might we want to affect the nat­ural devel­op­ment of either approach? WBEs are fre­quently regarded as being poten­tially highly untrust­wor­thy and dan­ger­ous, con­sti­tut­ing an .

Regulating Moore’s law

If an orga­ni­za­tion (such as a sin­gle­ton world gov­ern­ment like the UN Secu­rity Coun­cil) wanted to towards de novo AGI (the desir­abil­ity of uploads is con­tentious, with pro and con), then they might with­hold sub­si­dies from hard­ware research & devel­op­ment, reg­u­late it, or even actively oppose it with penal­ties rang­ing from the finan­cial to the civil to the mil­i­tary. (Or per­haps neo-Lud­dites wish to delay com­puter progress in the inter­ests of eco­nomic jus­tice.) This has been infor­mally pro­posed before; for exam­ple, Ber­glas 2009–20124.


Of course, from the util­i­tar­ian per­spec­tive, such a move would have enor­mous con­se­quences inas­much as it would prob­a­bly con­sti­tute a price floor on pro­cess­ing pow­er, dis­ad­van­tag­ing low-power appli­ca­tions like embed­ded proces­sors or sup­plies to the devel­op­ing world; and many busi­ness arrange­ments are founded directly or indi­rectly on Moore’s law hold­ing for another decade or three. Poverty kills, and any delay in eco­nomic growth also kills. Since we can­not observe the con­se­quences of either de novo AGI or brain emu­la­tions before they are cre­ated5, and once one is cre­ated we will cease to care about their tim­ings (from an exis­ten­tial risk point of view), we must resort to argu­ments about what might affect their rel­a­tive appear­ances; even in math, argu­ments or proofs come with non-triv­ial error rates—so how much more so in a legal or eco­nomic argu­ment? To not be banal, we would need to make a very solid case indeed—a case I can­not make nor will make here. Let’s assume that the case has been made and exam­ine an eas­ier ques­tion, how fea­si­ble such strate­gies are at accom­plish­ing the goal at all.

General Feasibility

How fea­si­ble is this? The rel­e­vant mea­sure of pro­gress, Moore’s law, has oper­ated reli­ably since the 1950s, over more than 60 years, and has even sped up at times. This would sug­gest it is dif­fi­cult to slow down or reverse this par­tic­u­lar kind of tech­no­log­i­cal progress.

On the other hand, at no point in those 60 years has any­one seri­ously attempted to slow down Moore’s law, and any attempt to do so will have been through ordi­nary com­mer­cial meth­ods, which are highly lim­ited in what coer­cion can be applied. In par­tic­u­lar, that 60 year period has been his­tor­i­cally unusu­ally favor­able for tech­no­log­i­cal devel­op­ment, with no major wars in the rel­e­vant nations like Japan, Tai­wan, or Amer­ica (his­tor­i­cally unusu­al—and a war involv­ing China was, and still is, per­fectly plau­si­ble). Had Moore’s law sur­vived a mul­ti­lat­eral ther­monu­clear exchange, that would be evi­dence for robust­ness. Had Moore’s law sur­vived legal assault like peer-­to-peer file­shar­ing has, that would be evi­dence for robust­ness. But in fact, Moore’s law has been heav­ily favored by gov­ern­ment sub­si­dies to research as Amer­i­can com­puter & semi­con­duc­tor capa­bil­i­ties were seen as a major advan­tage over the Soviet Union dur­ing the Cold War; and they have con­tin­ued, as with the end of the Cold War, com­put­ers also became a key part of the Amer­i­can econ­omy and self­-im­age. (‘We may not have man­u­fac­tur­ing’, the elites say, ‘but we’re still the best Inter­net and com­puter econ­omy in the world! Just learn you some com­puter stuff and there is a good job wait­ing in Econ­omy 2.0 for you!’ There is enough truth to this.) Where there is suf­fi­cient polit­i­cal will, even the most dra­matic and acces­si­ble tech­nol­ogy can decline—wit­ness the end of Ming Chi­na’s extra­or­di­nary sea expe­di­tions led by or Toku­gawa-era Japan’s . Through­out his­to­ry, there have been long peri­ods of tech­no­log­i­cal stag­na­tion or even regres­sion: to draw a straight or expo­nen­tial line from a few points using over­lap­ping chronolo­gies is just cher­ry-pick­ing.

To shut down Moore’s law, one can attack either the poten­tial for future improve­ments or the exis­tence of cur­rent capa­bil­i­ty:

  1. If all research ended today, then Moore’s law would quickly die: some effi­ciency can be ironed out of exist­ing pro­duc­tion capa­bil­i­ty, but these price drops would quickly hit an asymp­tote where new dis­cov­er­ies would be required to make any notice­able improve­ments.
  2. If research con­tin­ued, but all chip fabs were destroyed, the knowl­edge would be ster­ile as every chip would be pro­duced in small hand-­made quan­ti­ties; and the com­plex­ity of every addi­tional advance would make each batch more expen­sive. (Price cuts must be real­ized at ; see lat­er.)

Targets for Regulation

When exam­ing a sys­tem to speed it up or slow it down, one wants as much lever­age as pos­si­ble, to manip­u­late cen­tral nodes. In a highly dis­trib­uted net­work, manip­u­la­tion may be dif­fi­cult to impos­si­ble as no node is espe­cially impor­tant: the sys­tem can only be manip­u­lated by manip­u­lat­ing a large frac­tion of the nodes. But many net­works are not very dis­trib­ut­ed, and may looks more like a but­ter­fly or fol­low the bow tie orga­ni­za­tional archi­tec­ture seen in cel­lu­lar metab­o­lism:

…a myr­iad of nutri­ent sources are catab­o­lized, or ‘fan in’, to pro­duce a hand­ful of acti­vated car­ri­ers (e.g. ATP, NADH and NADPH) and 12 pre­cur­sor metabo­lites (e.g. glu­cose 6-phos­phate, fruc­tose 6-phos­phate, phos­pho­enolpyru­vate and pyru­vate), which are then syn­the­sized into roughly 70 larger build­ing blocks (e.g. amino acids, nucleotides, fatty acids and sug­ars). The pre­cur­sors and car­ri­ers can be thought of as two ‘knots’ of sep­a­rate bow ties that are both fed by catab­o­lism, but whereas the for­mer ‘fan out’ locally to the biosyn­the­sis of uni­ver­sal build­ing blocks, the lat­ter fan out to the whole cell to pro­vide ener­gy, reduc­ing power and small moi­eties.

While the Inter­net is famously dis­trib­uted and does not at first glance fol­low any bow tie net­work, the semi­con­duc­tor indus­try is amaz­ingly cen­tral­ized: <14 com­pa­nies make up the major­ity of global man­u­fac­tur­ing. (like or ) does not actu­ally pos­sess man­u­fac­tur­ing facil­i­ties; they focus on research, design, and licens­ing. The facil­i­ties require myr­i­ads of skills, resources, and tools, and also as in a bow tie, the out­puts of these few facil­i­ties are shipped world-wide to be used in count­less appli­ca­tions in every field of endeav­our. The chip fabs look very much like the knot of a bow tie and where we might want to start.

Fab costs and requirements

The costs they bear to build each chip fab­ri­ca­tion plant is astound­ing, and increas­ing even as rev­enue growth slows (squeez­ing out many old com­pa­nies)6; the basic equip­ment alone begin in the hun­dreds of thou­sands of dol­lars, lith­o­g­ra­phy machines were $40 mil­lion a piece in 2009, and the most expen­sive sin­gle pieces of equip­ment (like step­pers) can reach prices as high as $50 mil­lion dol­lars. The soft­ware licens­ing and engi­neer­ing costs that go into a cut­ting-edge proces­sor are equally stag­ger­ing; Brown & Lin­den 2005 (see also Andreas Olof­s­son):

Cost reduc­tion via off­shore invest­ments in low-wage coun­tries was not a fea­si­ble strat­egy because fab­ri­ca­tion is so cap­i­tal-in­ten­sive that labor typ­i­cally accounts for 16% of costs (in­clud­ing depre­ci­a­tion) in U.S. fabs pro­duc­ing 200mm wafers, and less than 10% in the newer 300mm fabs, which under­cuts the major labor cost advan­tage of most indus­tri­al­iz­ing coun­tries.30…The eco­nomic char­ac­ter­is­tics of each step of the process dif­fer sig­nif­i­cant­ly. Design is skill inten­sive, and requires expen­sive EDA () soft­ware, which is typ­i­cally licensed per design engi­neer. Fab­ri­ca­tion requires a huge fixed invest­ment (cur­rently on the order of $2 bil­lion [c. 2004]) to build a plant (called a fab) that holds a wide vari­ety of expen­sive equip­ment and that meets extreme require­ments of clean­li­ness. Assem­bly also requires expen­sive equip­ment, but the over­all costs of plant and equip­ment are much lower than for the fab, as are the aver­age skill require­ments. Over­all, worker skill require­ments go down along the value chain (i.e., design is more skil­l-in­ten­sive than man­u­fac­tur­ing, which is more skil­l-in­ten­sive than assem­bly)…The avail­abil­ity of out­sourc­ing (for­eign or domes­tic) is par­tic­u­larly impor­tant for small com­pa­nies and start-ups because of the rel­a­tively large fixed cost of EDA tools, which are typ­i­cally licensed per engi­neer. One con­sul­tant esti­mated that the min­i­mum annual soft­ware expense for a small com­pany is $10 mil­lion.85 For the indus­try as a whole, EDA expense runs close to 1% of rev­enue. In that case, a com­pany earn­ing less than $1 bil­lion in rev­enue would be below the effi­cient scale for in-­house design. Only the nine largest fab­less com­pa­nies met that cri­te­rion in 2004. One con­sul­tant esti­mated that out­sourc­ing even within the United States would save a small start-up that does fewer than five designs a year up to two-thirds the cost of doing the work in-­house.86

…Chip design is highly skil­l-in­ten­sive, since it employs only col­lege-­trained engi­neers. A cou­ple of medi­um-­size chip designs will employ as many elec­tri­cal engi­neers as a fab for a year or more (although the skills are not directly trans­fer­able). A com­plex chip design like Intel’s Pen­tium 4, with 42 mil­lion tran­sis­tors on a 180nm linewidth process, engaged hun­dreds of engi­neers for the full length of the five-year pro­jec­t.[“Comms held Pen­tium 4 team together”, EE Times, Novem­ber 1, 2000. “Linewidth” refers to the size of the fea­tures etched on a wafer dur­ing the fab­ri­ca­tion process. Each semi­con­duc­tor process gen­er­a­tion is named for the small­est fea­ture that can be pro­duced.]

…The soft­ware effort itself has increased by 375%. Accord­ing to one soft­ware exec­u­tive, a typ­i­cal chip in 1995 went into a stand-alone prod­uct and required 100,000 lines of code. In 2002, a typ­i­cal chip for a net­worked pro­gram­ma­ble prod­uct requires a mil­lion lines of code.60 [Jerry Fid­dler, chair­man of Wind River Sys­tems, cited in “Keynoter says chip value is in its intel­lec­tual prop­er­ty,” EE Times, June 14, 2002.] The soft­ware, plus the greater com­plex­ity of chips them­selves, has caused design val­i­da­tion hours to grow by 90% for each mil­lion tran­sis­tors. By com­par­ison, the growth lev­els for the actual design engi­neer­ing jobs of logic and phys­i­cal design for each mil­lion tran­sis­tors are a rel­a­tively mod­est 17% and 52%, respec­tive­ly. This is largely because, as chips have got­ten more com­plex, the process of chip design has become more auto­mat­ed.61

The facil­i­ties are even more impres­sive: Fab 11X has 400,000 square feet of a quar­ter-mile on a side. Chip fabs use upwards of 40 miles of for their (: the pipes expand in diam­e­ter each gen­er­a­tion) with the inter­nal mono­rail trans­porta­tion can be 3 miles long, and the clean rooms must be con­structed with cus­tom pre-­cleaned con­struc­tion mate­ri­als. (Ce­ment con­sump­tion is so high that Intel just builds cement plants on their sites.) Chip fab energy con­sump­tion is mea­sured in megawatts, 55–65 megawatts in one case. Intel cheer­fully notes about its Fab 42 con­struc­tion:

First of all, Intel is using the largest land-based crane in the world—one that can pick up and place mas­sive roof trusses that weigh approx­i­mately 300 tons each. The crane is so large it had to be deliv­ered on trucks to the site in pieces—ap­prox­i­mately 250 truck loads in total. Addi­tion­al­ly, Fab 42 will require 24,000 tons of steel rebar and 21,000 tons of struc­tural steel. And to make room for the fab, 875,000 cubic yards of dirt had to be exca­vat­ed. When all is said and done, approx­i­mately 10.5 mil­lion man hours will be required to com­plete the pro­ject.

A fab cost ~$1.5b in 1998, $2b in 2004, $3b in 2007, and $5b by 2010. Jurvet­son wrote in 2004 that

Another prob­lem is the esca­lat­ing cost of a semi­con­duc­tor fab plant, which is dou­bling every three years, a phe­nom­e­non dubbed Moore’s Sec­ond Law. Human inge­nu­ity keeps shrink­ing the CMOS tran­sis­tor, but with increas­ingly expen­sive man­u­fac­tur­ing facil­i­ties—cur­rently $3 bil­lion per fab.

Ross 2003 did­n’t fore­see the com­ing explo­sion in prices in describ­ing Moore’s Sec­ond Law, or Rock’s law as he prefers:

Some­times called Moore’s Sec­ond Law, because Moore first spoke of it pub­licly in the mid-1990s, we are call­ing it Rock’s Law because Moore him­self attrib­utes it to Arthur Rock, an early investor in Intel, who noted that the cost of semi­con­duc­tor tools dou­bles every four years. By this log­ic, chip fab­ri­ca­tion plants, or fabs, were sup­posed to cost $5 bil­lion each by the late 1990s and $10 bil­lion by now.

Not so. VLSI Research esti­mates that fabs cost $2 bil­lion apiece, the same as in the late 1990s, even as their pro­duc­tiv­ity has gone up. “In the 1980s, the fabs increased their yield; in the 1990s, they started [in­creas­ing] their through­put,” Hutch­e­son says. (Through­put refers to the num­ber of wafers a fab pro­duces in a given time.) Wafer through­put rose from 20 per hour in the early 1990s to about 40 to 50 an hour today.

Any­how, the focus was wrong­head­ed; what mat­ters is not the cost of the fab but the value of its prod­uct. If a $100 bil­lion fab made so many tran­sis­tors per penny that it could under­cut the prices of a $10 bil­lion com­peti­tor, it would be eco­nom­i­cal (if, of course, you could get the seed cap­i­tal together from a coali­tion of com­pa­nies-or con­ti­nents).7

Intel’s Fab 32 cost an esti­mated $3b in 2007 (clean rooms: 184,000 square feet; total: 1 mil­lion square feet), revised to $3.5b by 2011. A 2009–2010 upgrade to an Intel fab, Fab 11X, cost $2.5b (on top of the $2b upgrade in 2007). The ‘first stage’ of New York 1.3 mil­lion square foot fab will cost >$4.6 bil­lion dol­lars ($1b report­edly sup­plied by New York State); Glob­al­Foundries CEO San­jay Jha esti­mated in 2017 that a 7nm-­ca­pable chip fab would cost $10-$12b and the 5nm $14–18b. Intel’s Fab 42 (be­gun 2011–2012) is pro­jected at >$10b. Fab 15 in Tai­wan is esti­mated at >$9.3 bil­lion, and they are prepar­ing to start a fab in 2015 pro­jected at >$26 bil­lion; a $20b esti­mate for their next fab was repeated in 2017. Con­struc­tion of a Ger­man chip fab has been blamed for con­tribut­ing to the finan­cial hob­bling of for­merly com­pet­i­tive AMD8, and involved com­pa­nies are resort­ing to col­lab­o­ra­tions to cover the cap­i­tal costs, even for the largest play­ers (eg. Intel & build­ing a $3b+ Flash fab together, or still able to build mem­ory chip fabs with inter­nal financ­ing its mem­ory fac­to­ries—as a gov­ern­men­t-backed rep­re­sent­ing 1⁄5th of the world’s 15th largest econ­o­my). The trend shows lit­tle sign of abat­ing for a vari­ety of rea­sons91011, and vastly out­paces infla­tion. At cur­rent rates, it is not impos­si­ble that the total cost of a bleed­ing-edge CPU/GPU chip fab may pass the $100b (in­fla­tion-ad­just­ed) mark some­where in the 2020s or 2030s—well before a num­ber of the Bostrom-Sand­berg esti­mates for hard­ware power reach­ing brain emu­la­tion lev­els. These bil­lions of dol­lars of expen­di­tures are devel­oped & man­aged by hun­dreds of thou­sands of employ­ees: TSMC has >38k and Intel >104k.

To put these finan­cial & infra­struc­ture invest­ments in per­spec­tive (par­tic­u­larly the pro­jected TSMC invest­ment of >$26b for one new fab), the —one of the largest, most expen­sive (after ), and intri­cate sci­en­tific pro­grams ever under­tak­en, try­ing mul­ti­ple path­ways to the atomic bomb in par­al­lel—­cost $24 bil­lion in 1945 or $25–30b in 2012 dol­lars, with 130k employ­ees.

One can’t help but think that even if pos­si­ble, no one will engage in such cap­i­tal expen­di­tures because it will be bad busi­ness. (In Jan­u­ary 2014, Intel halted devel­op­ment of its Fab 42—“touted as the most advanced high­-vol­ume semi­con­duc­tor-­man­u­fac­tur­ing facil­ity in the world” and “among the world’s largest con­struc­tion projects in recent years”—after ~$1b of con­struc­tion was com­plet­ed.) A semi­con­duc­tor con­sul­tant shows a 2012 esti­mate about the cost per gate of the smaller processes (which may require new chip fab­s):

Fig­ure 1: Cost per gate

3. Nex­t-­gen­er­a­tion 20-nm pla­nar CMOS will have a range of addi­tional tol­er­ance con­trol chal­lenges com­pared to 28-nm. One likely impact is that cost per gate at 20-nm will be higher than at 28-nm. With the poten­tial for increased cost per gate, addi­tional com­paction will need to be done, which will lengthen design com­ple­tion times. Cost per gate at 14-nm can also be higher than that at 28-nm.

…New libraries will need to be devel­oped, IP tran­si­tioned to the FinFET struc­tures, test chips run, and pro­duc­tion vol­umes ramped up. At 14-nm, com­plex chips will cost $200 mil­lion to $500 mil­lion to design, and re-spins will cost $20 mil­lion to $50 mil­lion. The cost of fail­ure will increase dra­mat­i­cal­ly. What’s more, 14-nm FinFETs are not likely to be in high­-vol­ume pro­duc­tion out­side of Intel until 2016 to 2017. High­-vol­ume pro­duc­tion will require lower power con­sump­tion and lower cost per gate than ear­lier gen­er­a­tions of tech­nolo­gies. After 14-nm, there will be a range of new chal­lenges (EUV, 450-mm, car­bon nan­otubes, etc). The semi­con­duc­tor indus­try must be real­is­tic that the sup­ply chal­lenges are becom­ing more dif­fi­cult, and there will be a length­en­ing of the time to migrate to smaller fea­ture dimen­sions.

Con­sis­tent with squeeze on rev­enue and esca­lat­ing cap­i­tal costs is the observed dis­tri­b­u­tion of man­u­fac­tur­ing. “Resource Allo­ca­tion & Sched­ul­ing in Moore’s Law Twi­light Zone”, Benini July 2012, pg2; we see 20 man­u­fac­tur­ers at the ancient 130nm, but just 5 man­u­fac­tur­ers at 22/20nm:

“Mar­ket vol­ume wall: only the largest vol­ume prod­ucts will be man­u­fac­tured with the most advanced tech­nol­ogy”

The scale thesis

Given all this, a nat­ural objec­tion is that chip fabs are only cen­tral­ized like this because it’s slightly bet­ter than the decen­tral­ized alter­na­tives. There’s no point in reg­u­lat­ing chip fabs because any seri­ous slow­down will sim­ply encour­age decen­tral­iza­tion and small­er-s­cale chip fabs. I con­tend that the above fig­ures are so extreme that this can­not be the case, and we have excel­lent rea­sons to believe that this cen­tral­iza­tion trend is robust and fun­da­men­tal, and dri­ven by basic eco­nomic facts and par­tic­u­larly bru­tal economies of scale; hence, chip fabs will con­tinue cen­tral­iza­tion as long as pos­si­ble, and any attempt to decen­tral­ize may well itself drive prices up and slow down proces­sor devel­op­men­t—ex­actly as intend­ed.

are one of the more robust obser­va­tions in man­u­fac­tur­ing: the more you make & for longer, the more expe­ri­ence or intel­li­gence builds up in your facil­i­ties & humans, and the cheaper or bet­ter they can make them. observes the curve in the 61–62 areas com­piled in the Per­for­mance Curve Data­base. The curve may be related to Moore’s law (eg it is seen in Korean semi­con­duc­tor pro­duc­tion).

One early exam­ple of the expe­ri­ence curve effect in com­put­ing is given on pg171 of Tur­ing’s Cathe­dral (2012) by while dis­cussing the choice of com­mod­ity com­po­nents in the con­struc­tion of the (one of the influ­en­tial early imple­men­ta­tions of a ):

“If the 6J6, which was the twin , had not existed dur­ing the [WWII] war and had not been widely used, I don’t know what we would have used for a tube,” says Willis Ware. Not only did the wide­spread use of the 6J6 mean that it was avail­able inex­pen­sive­ly, but it was found to be more reli­able as well. One of ’s last assign­ments at the Sta­tis­ti­cal Research Group at Colum­bia had involved the reli­a­bil­ity of muni­tions. “There had been a lot of acci­den­tal explo­sions of rocket pro­pel­lant units on air­planes in which the explo­sion would take the wing off a plane,” he explains. “And this would hap­pen in a very rare and erratic fash­ion. So we had some excel­lent peo­ple in sta­tis­tics there, includ­ing no less than , who founded while work­ing with our group. Sta­tis­ti­cal think­ing had become a part of my way of think­ing about life.” It turned out that the most reli­able tubes were those pro­duced in the largest quan­ti­ties-­such as the 6J6. As Bigelow described it, “We learned that tube types sold at pre­mium prices, and claimed to be espe­cially made for long life, were often less reli­able in regard to struc­tural fail­ures than ordi­nary tube types man­u­fac­tured in larger pro­duc­tion lots.”60

That higher qual­ity did not require higher cost was not read­ily accept­ed, espe­cially since IBM, who had used the 6J6 as the com­put­ing ele­ment in its pop­u­lar model 604 elec­tronic cal­cu­la­tor, had recently estab­lished its own exper­i­men­tal tube pro­duc­tion plant in Pough­keep­sie, New York, to develop spe­cial com­put­er-qual­ity tubes at a much higher cost. There was intense debate over whether the choice of the mass-­mar­ket 6J6 was a mis­take. Of the final total of 3,474 tubes in the IAS com­put­er, 1,979 were 6J6s. “The entire com­puter can be viewed as a big tube test rack,” Bigelow observed.61

“It was con­sid­ered essen­tial to know whether such minia­ture tubes as the 6J6 have rad­i­cally infe­rior lives com­pared to other types, to an extent ren­der­ing their use in design a major blun­der; and accord­ingly a crude life-test set up was devised and oper­ated to get some sort of a sta­tis­ti­cal bound on their reli­a­bil­i­ty,” Bigelow reported at the end of 1946. Four banks of 6J6 tubes, twenty in each bank, for a total of eighty tubes, were installed in a test rack so they were ori­ented up, down, and in the two hor­i­zon­tal posi­tions (cath­ode edge-­wise and cath­ode flat). The entire rack was mounted on a vibrat­ing alu­minum plate, and the tubes left to run for three thou­sand hours. “A total of six failed, four within the first few hours, one about 3 days and one after 10 days,” was the final report. “There were four heater fail­ures, one grid short and one seal fail­ure.”62

Financial fragility

This leads to an inter­est­ing ques­tion: if a chip fab were destroyed, how well would the com­pany weather it? It is dif­fi­cult to answer this, but I will note that Intel’s 2010 rev­enue was $54b, TSMC’s was $14b and Glob­al­Foundries’s was $3.5b. It is not clear that chip foundry com­pa­nies could sur­vive the destruc­tion of one or two of their fabs now, much less how finan­cially robust they will be after another cost dou­bling or two.

Or will the dou­blings con­tin­ue? If it ceases to become prof­itable to build chip fabs capa­ble of build­ing faster chips, or to upgrade the fabs, this sug­gests that Moore’s law may come to an end on its own with­out any kind of inter­ven­tion. One ana­lyst is already fore­cast­ing the death of new fabs and hence, a slow­ing or end to Moore’s law12. Some sim­ple eco­nomic mod­els put the shut­down between 2015 and 202513.

(The R&D efforts needed to sup­port fabs may be sim­i­larly increas­ing: “…R&D costs are rapidly increas­ing. In 2009, around $30 bil­lion, or 17% of rev­enue, went to R&D across the indus­try–a 40% increase over 1999.”.)

Effects of fab disruptions

Can we make any esti­mates about the fragility of the net­works sur­round­ing the chip fabs as well as the fabs them­selves?

Semi­con­duc­tor chips are per­haps the epit­ome of an : sophis­ti­cated prod­ucts where a sin­gle mis­take any­where in a process with dozens or hun­dreds of steps ren­ders the prod­uct val­ue­less. A sin­gle defect in a 4-core proces­sor may—at best—dis­able one of its cores, forc­ing the man­u­fac­turer to sell it for half or less, per­haps divert­ing it from high­-end con­sumers like gamers to mass-­mar­ket game con­soles or other cus­tomers. (For exam­ple, the has 8 cores but ships with just 7 func­tion­ing because the man­u­fac­tur­ing yield of func­tion­ing Cells is so low.)

Case studies

Sumitomo Chemical fire

An obscure inci­dent sug­gests what we might see in a dis­rup­tion; from “Real Chip Short­age Or Just A Pan­ic, Crunch Is Likely To Boost PC Prices”:

The dis­rup­tions now rag­ing in world com­puter chip mar­kets started when an explo­sion at a Sum­it­omo Chem­i­cal Co. fac­tory in the town of Niihama, Japan, on July 4 wiped out the source of 60% of the world sup­ply of an epoxy resin called cresol…When news of the Japan blast sur­faced in such indus­try pub­li­ca­tions as Elec­tronic Buy­ers’ News and InfoWorld, spec­u­la­tors started rush­ing to buy up the world’s sup­ply of DRAM chip­s…­Dataque­st’s Giu­dici said big com­pa­nies such as Inter­na­tional Busi­ness Machines Corp. and Apple Com­puter Inc., which typ­i­cally buy the chips under long-term con­tracts, are likely to get through the cur­rent price swings with only small price increases of around $30 per machine…Even these big com­pa­nies are pay­ing sharply higher prices when they buy on the spot-­mar­ket, though, he explained. DRAM prices for these large com­pa­nies have jumped from $37 on the aver­age in May to an aver­age of $55 today. “Some of these com­pa­nies have paid as high as $70 in the past week,” Giu­dici said. The hun­dreds of smaller com­pa­nies that have flour­ished in the cur­rent com­puter boom are likely to be forced to add $100 to $150 to the cost of each machine because they must buy on the spot-­mar­kets, Giu­dici said. Those hit hard­est, how­ev­er, are peo­ple like Ahmad who want to upgrade. Barry Lebu, pres­i­dent of 50/50 Micro­elec­tron­ics Inc. in Sun­ny­vale, Calif., said DRAM spot mar­ket prices are aver­ag­ing $89 per megabyte, up from $39 just four weeks ago: “Some are as low as $69 and some are hit­ting $119, but the aver­age price is $89.”

and “Indus­try ana­lyst still unsure of the sig­nif­i­cance of Hynix fire: Extent of fire dam­age still unknown at Hynix facil­ity in Chi­na, Jim Handy of Objec­tive Analy­sis fills us in on his take of the events”:

There are strong sim­i­lar­i­ties between this inci­dent an another fire in 1993. In July of that year a fire and explo­sion in a Sum­it­omo Chem­i­cal plant removed over 90% of the world’s sup­ply of a cer­tain epoxy that was almost uni­ver­sally used to attach DRAM dice to their pack­ages. The impact of this event was to gal­va­nize a DRAM short­age that was already devel­op­ing at that time. The short­age lasted until the end of 1995, longer than any short­age in the his­tory of the DRAM mar­ket. The dura­tion of that short­age was not the result of the fire—other fac­tors were at play. Still, the indus­try very quickly tran­si­tioned from the mild onset of a short­age to a very solid short­age as a a result of the inci­dent, even though abun­dant alter­na­tives to Sum­it­o­mo’s epoxy were iden­ti­fied within a week.

John C. McCal­lum’s “Graph of Mem­ory Prices Decreas­ing with Time (1957–2016)” shows an inter­est­ing abrupt plateau ~1993 for RAM, last­ing to ~1997, sug­gest­ing that the RAM short­age sparked by the Sum­it­omo Chem­i­cal fire had long-last­ing effects. The Sum­it­omo Chem­i­cal fire has passed into obscu­rity with­out much recent notice, so it’s hard to see how much impact it really had.

Toshiba NAND Flash memory

As it hap­pens, his­tory recently gifted us with a beau­ti­ful pair of more recent, bet­ter doc­u­mented exam­ples which enables us to answer: yes, progress is frag­ile.

In one exam­ple, a 13-minute power out­age in a sin­gle Toshiba/WD fab in June 2019 cost an esti­mated ~16% of global NAND flash mem­ory out­put that quar­ter. On 2010-12-08, the power grid serv­ing a Toshiba-San­disk chip fab (spe­cial­iz­ing in Flash mem­o­ry) suf­fered a brief fall in volt­age beyond what the fab’s local were designed for. This resulted in a fab-wide power inter­rup­tion of less than a tenth of a sec­ond (~0.07). Toshiba reported it took the plant 2 days to return to 100% oper­a­tion. This cut pro­duc­tion for the quar­ter by 20%, and world­wide pro­duc­tion by 7.5% (ac­cord­ing to iSup­pli’s Yang, quoted by the WSJ1415). Cov­er­age of this inci­dent men­tioned pre­vi­ous inci­dents in 2007 and 200016, and was rem­i­nis­cent of another inci­dent on 1993-07-04 where a fire at the only plant man­u­fac­tur­ing a cresol epoxy resin used in man­u­fac­tur­ing chip cases pro­voked an imme­di­ate dou­bling of DRAM prices until the plant was repaired over 4 months later, and a TSMC mal­ware inci­dent in 2018.

(Given the costs involved, one might expect reli­able UPSes to be default, but also remem­ber that this fabs can be using any­where up to the dozens of megawatts range of elec­tric­i­ty, which might be dif­fi­cult to com­pletely UPS. With­out a seri­ous and plau­si­ble risk, it would be unre­al­is­tic to expect the indus­try to invest mil­lions in sub­stan­tial local elec­tri­cal power capac­i­ty. And backup power sys­tems them­selves have been the source of errors—a gen­eral truth in engi­neer­ing highly reli­able com­plex sys­tems is that the com­plex­ity added to pre­vent errors is itself a major source of error.)

Kryder’s Law

The Octo­ber struck at the hub of a quar­ter of global hard drive man­u­fac­tur­ing capa­bil­i­ty. West­ern Dig­i­tal restarted one flooded plant in Decem­ber but oth­ers would not be online until March 2012. Shipped quan­ti­ties were not pro­jected to recover until Q3 2012 and it took until Sep­tem­ber 2012 for the vol­ume to actu­ally recov­er; the same source pre­dicted that the 2011 prices would only be matched in 201417.

The floods enable us to give a quan­ti­ta­tive mea­sure of how much progress was set back. Hard dri­ves fol­low : unit prices halve every 12 months. As it hap­pens, on 2011-03-28 I pur­chased one of the cheap­est hard dri­ves at that point, a 1.5tb Sam­sung hard drive for $51. By Kry­der’s law, in March 2012, I would be able to pur­chase 3tb for ~$50 (and ~$25 in March 2013). But as a mat­ter of fact, the cheap­est drive was $120 for 2tb. An equiv­a­lent 1.5tb drive is now not $25 as one would have pre­dicted before the floods, but a cool $100. And the cheap­est 3tb drive is $200. So not only did one crank of Kry­der’s law (a halv­ing) not hap­pen, a pre­vi­ous turn of the law was reversed, dou­bling prices; so we can esti­mate that Kry­der’s law has been set back by 2 years by some “slow-­mov­ing”18 flood­ing con­ducted by brain­less nat­ural forces.

What is strik­ing is the long-term effect of the Thai­land floods. If one were a skep­tic of the “expe­ri­ence curve effects” pre­vi­ously men­tioned, one might rea­son that R&D was con­stantly ongo­ing and so one would hope for a “catch-up effect” where the high prices were merely tem­po­rary and a sort of “super Kry­der’s law” oper­ates briefly to restore the orig­i­nal trend-­line as the pen­t-up R&D improve­ments are imple­ment­ed. Yet, indus­try fore­casts did­n’t esti­mate a decline until 2014, and we can watch the devi­a­tion from Kry­der’s law in real­time: 6 months after March 2012 (Au­gust 2012), close to a year after the floods, a 3tb hard dri­ve’s price had fallen from $200 to… $150. Exactly as pre­dicted from half a crank of Kry­der’s law. A year later (Au­gust 2013), the results are even more dis­mal: now 3tb costs… $130. No catch-up effect had yet occurred and it seems unlikely that chip fabs are remark­ably more robust. Extrap­o­lat­ing Kry­der’s law and the absence of any catch-up effect, we can make some pre­dic­tions out to August 2014:

  1. March 2013, a 3tb drive will cost: ~$113 (25% off $150) or 26.55gb/$; wrong (ac­tu­al: 22.34gb/$)
  2. August 2013: $75 (50% off $150) or 40gb/$; wrong (ac­tu­al: 23.25gb/$)
  3. March 2014: $56 (63% off $150) or 53.57gb/$; wrong (ac­tu­al: 30gb/$)
  4. August 2014: $38 (75% off $150) or 78.95gb/$; wrong (ac­tu­al: 33gb/$)

The first pre­dic­tion for March 2013 was blown: it required 24–30GB/$, while the cheap­est hard drive at Newegg was 22.4GB/$. So the first pre­dic­tion was off sub­stan­tial­ly, and on the low end, sug­gest­ing that if any­thing, I under­es­ti­mated the slow­down in hard drive growth. This pat­tern repeated itself through the last pre­dic­tion­s—where I had hoped for as much as 84gb/$, I must set­tle for a measly 33gb/$. Look­ing for expla­na­tions, I learned that the hard drive indus­try has seen a wave of merg­ers & con­sol­i­da­tions (just like the chip fab indus­try), going from scores of man­u­fac­tur­ers to . This con­sol­i­da­tion was partly respon­si­ble for the flood dis­as­ters by con­cen­trat­ing facil­i­ties, but has other impli­ca­tions: fewer com­peti­tors means less com­pe­ti­tion, less pres­sure to under­cut the oth­ers, fos­ters cartel-­like behav­ior, and sug­gests declin­ing prof­itabil­ity or dimin­ish­ing returns since the merg­ers may be dri­ven by economies of scale. Regard­less, the abrupt halt of Kry­der’s law seems to have caught oth­ers by sur­prise too, such as backup provider which wrote on 2013-11-26:

We are two years removed from the hor­rific flood­ing that caused the Thai­land Drive Cri­sis that cre­ated a world­wide short­age of hard disk dri­ves. Prices for hard dri­ves spiked and have remained stub­bornly high, only return­ing to pre-cri­sis lev­els in the last cou­ple of months. The mar­ket and com­pet­i­tive forces that over the past 30 years have pre­dictably dri­ven the cost per giga­byte of stor­age down took a vaca­tion…In addi­tion, the cost per giga­byte also declined in an amaz­ingly pre­dictable fash­ion over that time. Begin­ning in Octo­ber 2011 those 30-years of his­tory went out the win­dow…­Cost per GB for Hard Dri­ves: In Sep­tem­ber 2011, our cost per giga­byte was $0.044. That low water mark would not be achieved again until Sep­tem­ber 2013. In that two-year peri­od, our cost ran as high as $0.064 per giga­byte…When the Drive Cri­sis start­ed, indus­try pun­dits esti­mated that the hard drive mar­ket would take any­where from 3 months to 1 year to recov­er. No one guessed two years. Was the delay sim­ply an issue in rebuild­ing and/or relo­cat­ing the man­u­fac­tur­ing and assem­bly facil­i­ties? Did the fact that the two indus­try lead­ers, Sea­gate and West­ern Dig­i­tal, had to inte­grate large acqui­si­tions slow down the recov­ery and sub­se­quent inno­va­tion? What about the dra­matic shift towards tablets and away from Desk­tops and Lap­tops, has that changed the hard drive mar­ket and the declin­ing cost per giga­byte trend line forever? What­ever lies ahead, we’ll adapt.

Almost 4 years lat­er, in July 2017, it has become clear to Back­blaze that not only has the recov­ery never hap­pened, but post-2011, the improve­ment curves for hard dri­ves have dras­ti­cally wors­ened:

Up through the 4 TB drive mod­els, the cost per giga­byte of a larger sized drive always became less than the smaller sized dri­ves. In other words, the cost per giga­byte of a 2 TB drive was less than that of a 1 TB drive result­ing in higher den­sity at a lower cost per giga­byte. This changed with the intro­duc­tion of 6- and 8 TB dri­ves, espe­cially as it relates to the 4 TB dri­ves. As you can see in the chart above, the cost per giga­byte of the 6 TB dri­ves did not fall below that of the 4 TB dri­ves. You can also observe that the 8 TB dri­ves are just approach­ing the cost per giga­byte of the 4 TB dri­ves… the 6 TB dri­ves have been in the mar­ket at least 3 years, but are not even close to the cost per giga­byte of the 4 TB dri­ves. Mean­while, back in 2011, the 3 TB dri­ves mod­els fell below the cost per giga­byte of the 2 TB dri­ves they “replaced” within a few months. Have we as con­sumers decided that 4 TB dri­ves are “big enough” for our needs and we are not demand­ing (by pur­chas­ing) larger sized dri­ves in the quan­ti­ties needed to push down the unit cost? Approach­ing Zero: There’s a Lim­it: The impor­tant aspect is the trend of the cost over time. While it has con­tin­ued to move down­ward, the rate of change has slowed dra­mat­i­cally as observed in the chart below which rep­re­sents our aver­age quar­terly cost per giga­byte over time.

“Back­blaze Aver­age Cost per GB for Hard Dri­ves; By Quar­ter: Q1 2009–Q2 2017”; shows 2011–2012 price spike due to Thai­land floods fol­lowed by slower cost declines 2013–2017 than his­tor­i­cal­ly.

The change in the rate of the cost per giga­byte of a hard drive is declin­ing. For exam­ple, from Jan­u­ary 2009 to Jan­u­ary 2011, our aver­age cost for a hard drive decreased 45% from $0.11 to $0.06 – $0.05 per giga­byte. From Jan­u­ary 2015 to Jan­u­ary 2017, the aver­age cost decreased 26% from $0.038 to $0.028 – just $0.01 per giga­byte. This means that the declin­ing price of stor­age will become less rel­e­vant in dri­ving the cost of pro­vid­ing stor­age.


State-actors: Why Not Terrorism

We have seen how dif­fi­cult fabs are to make, how few they are, how even small dis­rup­tions spi­ral into global shifts. Does this imply that reg­u­la­tion could be accom­plished by any mod­estly capa­ble group, such as a imi­ta­tor or a souped-up ITS?


  1. the size, scale, and remote­ness from neigh­bor­ing build­ings of chip fabs implies both that secur­ing them against con­ven­tional para­mil­i­tary assault is easy, and as a per­cent­age of con­struc­tion & oper­at­ing costs, triv­ial. Secur­ing them against con­ven­tional mil­i­tary assault (cruise mis­siles, artillery strikes, etc.) is highly non­triv­ial but also unnec­es­sary as no ter­ror­ist groups oper­at­ing in the rel­e­vant coun­tries has access to such equip­ment and per­son­nel. Ter­ror­ists are which are easy to access and impos­si­ble to defend.

    This may change in the future due to tech­no­log­i­cal advance­ments like , which have been fore­cast by strate­gists like to rad­i­cally level the mil­i­tary play­ing field, but I sus­pect they and other changes may be a wash or even tilt the play­ing field against small actors (as they have famously done in Afghanistan & else­where).19

    If ter­ror­ists can­not go after the chip fab, they can go after sup­port­ing infra­struc­ture: attack­ing instead the power plants or power grid. But such an attack is still sophis­ti­cated beyond almost all ter­ror­ist attacks ever done and is also eas­ily defended again­st: plants use up to 60 megawatts of pow­er, and 1 megawatt of pow­er-pro­duc­ing capac­ity can be installed at $1–7m20 depend­ing on sources, so in the worst case the power vul­ner­a­bil­ity can be elim­i­nated by chip fabs set­ting up their own on-site power sup­ply with an invest­ment of <$420m (). Of course, the invest­ment is not waste—one does­n’t need to buy huge quan­ti­ties from the local grid if one has one’s own plant stand­ing idle. Power plants are not an invest­ment the semi­con­duc­tor indus­try would make for no rea­son (espe­cially if a local grid is avail­able), but the first suc­cess­ful attack on an elec­tric­ity grid and shut­down of a chip fab would imme­di­ately moti­vate a shift away from the grid by the remain­ing chip fabs (sim­i­lar to how 9/11 elim­i­nated the abil­ity of ter­ror­ists to hijack any Amer­i­can planes, because the pas­sen­gers will attack them sui­ci­dally rather than coop­er­ate). Even bean coun­ters won’t risk ruin­ing a >$20b invest­ment because they refused to pay for a <$400m local power plant!

  2. Ter­ror­ist groups are from the point of view of get­ting things done: they change their ide­o­log­i­cal posi­tions, they refuse nego­ti­a­tions or com­pro­mis­es, they fail to claim respon­si­bil­ity for attacks, they cause ‘back­fire effects’, and have a sta­tis­ti­cal record of vic­tory in the sin­gle per­cent­age point range. In gen­er­al, their moti­va­tion seems to be not any sort of ratio­nal choice of meth­ods (spec­tac­u­lar bomb­ings being in the first place) but rather a mar­ginal & unpop­u­lar form of social bond­ing. Any per­son seri­ously con­cerned about arti­fi­cial intel­li­gence would be pretty stu­pid (and heed­less of the many good rea­sons to err on the side of cau­tion in con­se­quen­tial­ist rea­son­ing) to con­clude that ter­ror­ism is the best strat­e­gy. Since vio­lence is a last resort, it will be resorted to only by those who think the threat is most press­ing, which for rea­sons sim­i­lar to the in auc­tions, will gen­er­ally be those who have made the largest errors in rea­son­ing. (This has been for­mal­ized as .)

    Ter­ror­ists moti­vated by such abstruse goals & rea­son­ing will likely be even more dys­func­tion­al; if highly edu­cated and intel­li­gent West­ern­ers in STEM fields can vocif­er­ously dis­agree on whether the cre­ation of AI is even pos­si­ble dur­ing the 21st cen­tury (much less dan­ger­ous), and mate­ri­al­ism still be a live issue in , it is highly unlikely that any notice­able num­bers of peo­ple will change their minds until an AI has actu­ally both been cre­ated and also demon­strated its harm­ful­ness. And even then it is unclear how well peo­ple would be able to coor­di­nate: the “nuclear taboo” took decades to devel­op. The only peo­ple who under­stand or care about these issues are geeks, and they are famously bad at coor­di­na­tion21 or agree­ment22 (“like herd­ing cats”). Con­trast the utter fail­ure of —who are moti­vated by griev­ances backed by past cen­turies of extra­or­di­nar­ily well-­doc­u­mented envi­ron­men­tal dam­age and oper­at­ing in a dom­i­nant West­ern ide­o­log­i­cal par­a­digm of —to do any­thing more than petty arson & van­dal­ism, with what our hypo­thet­i­cal ter­ror­ist group would have to accom­plish.

  3. Tran­s-­na­tional ter­ror­ist groups like Al-Qaeda have demon­strated in the 2000s that they are not resilient to allied gov­ern­ment sup­pres­sion; this point is most vividly illus­trated by con­sid­er­ing video of Osama bin Laden in his Pak­istani safe house before his assas­si­na­tion, watch­ing him­self on TV, and vainly writ­ing end­less emails to his remain­ing sub­or­di­nates and wish­ful plot­ting.

    Where such groups cur­rently suc­ceed, it is by a kind of ‘fran­chise’ strat­egy where the local mem­bers and nation­al­ist griev­ances take dom­i­nance. At best, facil­i­ties could be attacked in one coun­try and its neigh­bors, which would cer­tainly delay progress but on a time-s­cale mea­sured in years rather than decades, as the indus­try routes around the dam­age by rebuild­ing else­where and increas­ing capac­ity in exist­ing chip fabs.

  4. There is no pos­si­ble ben­e­fit to the ter­ror­ist group suc­ceed­ing

    This is a highly spec­u­la­tive util­i­tar­ian strat­e­gy, which could win no fer­vent adher­ents (no sol­dier jumps on a grenade for the sake of a cost-ben­e­fit equa­tion), as suc­cess is mea­sured in just the sta­tus quo—no faster chip­s—and one of the few sce­nar­ios in which one could have evi­dence of suc­cess would be the cre­ation of de novo AGI, at which point any poten­tial chip fab ter­ror­ist has more press­ing con­cerns. Sim­i­lar­ly, abstract ideals like Com­mu­nism have always been able to promise their adher­ents more con­crete ben­e­fits like groupies or high posi­tions after the war, but what can be promised in this case?

All of the­se, how­ev­er, imply that a nation or coali­tion of nations could: nations have large con­ven­tional mil­i­tary forces, (some­what) func­tion­ing deci­sion process­es, inter­na­tional pull and com­bat capa­bil­ity (espe­cially in the case of the USA’s Air Force or any nation pos­sess­ing cruise or bal­lis­tic mis­siles), and value the sta­tus quo in which they are top of the food chain. Some gov­ern­ments are even so far­sighted as to work on .

Covert fabs?

The spe­cial con­struc­tion, power & water con­sump­tion, and sheer scale of chip fabs sug­gest that it would be dif­fi­cult to covertly build a com­pet­i­tive under­ground chip fab; the Toshiba inci­dent sug­gests that weak non-nu­clear may be viable; the fragility of clean rooms sug­gest chip fabs could with­stand lit­tle over­pres­sure dam­age; the sen­si­tiv­ity of the chips & equip­ment to mechan­i­cal dam­age (which will likely increase as smaller nanome­ter processes are used) sug­gest that atten­u­ated shock­-waves23 may cause out­sized con­se­quences. Worse, for com­mer­cial rea­sons, all exist­ing facil­i­ties appear to be located within US allies and within short fighter range of Amer­i­can mil­i­tary bases (for exam­ple, all of Intel’s facil­i­ties).

We can con­trast chip fabs with par­al­lel exam­ples of suc­cess­ful black­-­mar­ket pro­duc­tion:

  1. coun­ter­feit elec­tron­ics are not pro­duced at covert fac­to­ries, but by legit­i­mate fac­to­ries which are either devoted to the task and con­doned by local author­i­ties, or by above-board fac­to­ries which are divert­ing part of their pro­duc­tion to knock­offs of their real pro­duc­tion dur­ing what is osten­si­bly down­time (“shadow shift pro­duc­tion runs”)

  2. illicit drugs:

    drug lab­o­ra­to­ries are numer­ous and pro­duce many tons of prod­ucts but they are gen­er­ally small and rel­a­tively low-tech: the process of one of the most chal­leng­ing nar­cotics to pro­duce, hero­in, was dis­cov­ered more than a cen­tury ago, and the chal­lenge stems more from the painstak­ing han­dling and manip­u­la­tion by an expe­ri­enced heroin chemist dur­ing the explo­sive24 acid­i­fy­ing step. Metham­phet­a­mine can be cooked up in a bath­room fol­low­ing the “shake and bake” recipe, and is pop­u­lar in part for this rea­son. Mar­i­juana grow­ing is triv­ial (al­beit grow­ing the finest mar­i­juana is very chal­leng­ing). Cocaine pro­cess­ing was sim­i­larly devised around a cen­tury ago and has mul­ti­ple processes, nev­er­the­less, the processes can be pretty sim­ple: in one, gaso­line dis­solves the cocaine from the leaves, sul­fu­ric acid & water are stomped into the gaso­line, and then a base like lye is added to form a cocaine paste. (Sul­fu­ric acid, lye, water, and oils have been avail­able for cen­turies or mil­len­ni­a.) It’s worth not­ing that for both heroin and cocaine, requir­ing the purifi­ca­tion of sub­stan­tial amounts of bulk raw mate­ri­als (poppy sap and coca leaves respec­tive­ly), even these rel­a­tively prim­i­tive processes take place in labs in law­less cartel-pro­tected areas like the Golden Tri­an­gle.

  3. bioweapons:

    weaponized bac­te­ria pro­duced in a covert lab might be an exam­ple, but so far they have not been pro­duced. The best known exam­ple of recent biowar­fare, the , remain mys­te­ri­ous but there seems to be gen­eral agree­ment they were pro­duced at or were related to the large well-­funded sophis­ti­cated legit­i­mate Amer­i­can mil­i­tary biowar­fare lab­o­ra­tory at . The sarin (toxin dis­cov­ered 1938) stand as some­thing of a coun­ter­point; they had facil­i­ties for pro­duc­ing sarin which were described as ‘sophis­ti­cated’, but while they were very inter­ested in more com­plex weapons like nuclear bombs and bac­te­ria, they only made sarin and VX gas, a pack­age-bomb, and an attempt to man­u­fac­ture 1000 -style assault rifles ended in fail­ure with one work­ing pro­to­type (although the AK-47 design is famous for being sim­ple and pro­ducible in even prim­i­tive coun­tries). It may be that bioweapons are much more demand­ing of “tacit knowl­edge” than gen­er­ally assumed.

  4. Nuclear weapons are an inter­est­ing enough exam­ple they deserve their own sec­tion.

Case-study: suppressing nuclear weapons

The task of slow­ing Moore’s law can be analo­gized to the task of slow­ing or stop­ping :

  • objects: nukes :: proces­sors

  • eco­nomic val­ue: , , :: any infor­ma­tion-pro­cess­ing task

  • trend: we can analo­gize in a few ways:

    1. num­ber of pos­ses­sors :: coun­tries with cut­ting-edge chip fabs
    2. mega­ton­nage of weapons :: speed of proces­sors
    3. num­ber of weapons :: num­ber of proces­sors

The trend brings out how mod­est our aims are with Moore’s law. What would a rogue state have to do to start or con­tinue a Moore’s law of nuclear weapons? Let’s take North Korea as an exam­ple. North Korea has a few nuclear bombs of a low kilo­ton­nage. To con­tinue any of the 3 anal­o­gous trends it would have to:

  1. spread nuclear weapons to one coun­try in the next 18 months (South Kore­a?), 2 coun­tries the next 18 months (Viet­nam? Cam­bo­di­a?), 4 coun­tries the next inter­val (Thai­land? Laos? Japan? Mon­go­li­a?), …
  2. dou­ble the kilo­ton­nage to say 50kt in the first peri­od, 0.1mt the next, 0.2 the next, 0.4, 0.8, 1.6, 3.2, and within 4 dou­blings it would be test­ing a and then have to begin design­ing giga­ton bombs of the sort that nei­ther the United States nor Rus­si­a—the pre-em­i­nent nuclear weapons design­er­s—ever dreamed of.
  3. begin dou­bling its weapon count, which would involve vast scal­ing up of its local ura­nium min­ing activ­i­ties and of course many new nuclear reac­tors for pro­cess­ing said ura­nium ore. The orig­i­nal nuclear weapons pro­gram was esti­mated to cost sev­eral per­cent­age points of North Kore­a’s annual GDP for decades, and while it would no doubt be cheaper to man­u­fac­ture bombs once the research is done, such dou­blings would quickly drive North Korea (more) bank­rupt.

All of this is remark­ably demand­ing; to give an idea, Liber­man 2003 briefly dis­cusses and dis­misses as absurd the idea that the Amer­i­can gov­ern­ment (!) could build a cut­ting-edge chip fab for mil­i­tary use (em­pha­sis added):

…the Depart­ment of Defense can attempt to achieve tem­po­rary solu­tions, such as build­ing its own next gen­er­a­tion gov­ern­men­t-owned chip fab­ri­ca­tion facil­i­ty, but this is likely to be both expen­sive and inef­fec­tive. If the best research and design capa­bil­ity shifts to China along with man­u­fac­tur­ing, this approach will not work past the next gen­er­a­tion or two of semi­con­duc­tor chip pro­duc­tion25. In addi­tion, such tem­po­rary solu­tions are not only unwork­able over time if the U.S. wishes to retain the best capa­bil­ity that is required for defense and intel­li­gence needs, but will be far more expen­sive than the solu­tions pro­posed above. This is because the oppor­tu­nity to lever­age off the com­mer­cial sec­tor (an approach which the DoD and intel­li­gence com­mu­nity rely upon at pre­sent) for new advances and cost sav­ings will be lost.

This anal­ogy is amus­ing, but more inter­est­ing is how well nuclear pro­lif­er­a­tion has suc­ceed­ed. At the dawn of the Nuclear Age, it was gen­er­ally believed that pro­lif­er­a­tion was inevitable: as the tech­nol­ogy dif­fused (eg. dif­fu­sion purifi­ca­tion tech­niques) and ever more trained engi­neers came into being and more nations acquired nuclear weapons, still more nations would acquire nuclear weapons, and so on and so forth. In par­tic­u­lar, it was “obvi­ous” that nuclear strikes would become stan­dard in war­fare: the US would use a nuke on Rus­sia to fore­stall Rus­si­a’s acqui­si­tion, or they would be used in Korea26, or in Viet­nam, or in the Mid­dle East, or ter­ror­ists would steal one from Rus­sia (or Pak­istan or…) or… But curi­ously enough, 67 years after Hiroshi­ma, not only have there never been any nuclear attacks post-WWII, pro­lif­er­a­tion has actu­ally been an incred­i­ble suc­cess, with the nuclear club num­ber­ing a mere 8 mem­ber­s—with mul­ti­ple coun­tries aban­don­ing their nuclear weapons or nuclear pro­grams27.

Lit­er­a­ture on the devel­op­ment of nuclear bombs afford another use­ful per­spec­tive: the in the devel­op­ment of advanced nuclear weapons. remarked that just know­ing that it was pos­si­ble for the USA to build & det­o­nate a nuclear bomb by 1945 made it easy for any pow­er­ful coun­try to develop one28, and MacKen­zie & Spinardi 1995 cite (pg10–11) sev­eral civil­ian exam­ples of what might be called rein­ven­tions of the atomic bomb (to which we could add the Ger­man team’s belated real­iza­tion of their crit­i­cal mass mis­take once they learned of Hiroshi­ma), observ­ing that the mere fact of suc­cess of var­i­ous pro­grams encour­ages addi­tional pro­grams29. But in no case do they come up with sophis­ti­cated designs along the lines of suit­case nukes, effec­tive tri­tium-­boosted mul­ti­-stage hydro­gen bombs, or com­pact ICBM-capable designs, and this holds true of nations attempt­ing to join the nuclear club: their bombs are all appar­ently rel­a­tively crude and low ton­nage, and incom­pa­ra­ble to the cut­ting-edge of Amer­i­can or Russ­ian designs.

So the les­son from nuclear pro­lif­er­a­tion sup­ports our the­sis: high­ly-ad­vanced tech­nolo­gies on the cut­ting-edge are much eas­ier to sup­press and con­trol than prim­i­tive slow designs which rep­re­sent the state of the art from pos­si­bly many decades ago.

How­ev­er, the nuclear exam­ple also empha­sizes the things coun­tries will not do. In the early Cold War, when the USA was still the sole nuclear power and the USSR was expected to take another 5–10 years for its active nuclear weapons pro­gram to suc­ceed, there were a num­ber of pro­pos­als that the USA elim­i­nate the threat: launch a uni­lat­eral strike on the USSR, deliver an ulti­ma­tum to cease nuclear devel­op­ment, or form a world gov­ern­ment with exclu­sive con­trol over nuclear weapons, etc. At this point, there was near-cer­tainty that nuclear weapons were real & an exis­ten­tial threat to coun­tries and human­i­ty, that the USSR under Stalin would seek nuclear weapons, that any nuclear stand­off would be pre­car­i­ous & dan­ger­ous, and that there would likely be pro­lif­er­a­tion to other coun­tries (not just allies of the USSR & USA like China or Eng­land). The USA chose inac­tion and an arms race, declin­ing any more extreme mea­sures (although it came close dur­ing a war for a minor coun­try called Korea, which strate­gi­cally makes no sense). Given that the case for pre­empt­ing the Cold War & MAD was in every way clearer & bet­ter than for chip fabs or AGI in gen­er­al, a for­tiori we can pre­dict with ~100% con­fi­dence that there will never be any con­ven­tional mil­i­tary attacks on chip fabs with the pur­pose of delay­ing AGI devel­op­ment.

Hardened non-covert fabs

Hard­en­ing does not seem to be an answer. Leav­ing aside the pre­vi­ous obser­va­tion that it’s unclear whether com­pa­nies (or even nation-s­tates) can sur­vive the con­struc­tion costs of unhard­ened fabs, even basic hard­en­ing is expen­sive. Linger et al 2002 argues that con­trary to the usual esti­mates (that ordi­nary under­ground con­struc­tion costs 3–5× nor­mal con­struc­tion cost­s), Nor­we­gian con­struc­tion reports (with their long expe­ri­ence in under­ground con­struc­tion) for power sta­tions and waste treat­ment plants indi­cate that under­ground con­struc­tion pre­mi­ums are more like 1.25× and may even­tu­ally save mon­ey. An Army overview agrees30. These hard­en­ing cost-es­ti­mates, how­ev­er, are not for struc­tures robust against nuclear or bunker-­bust­ing bombs, which require over­bur­dens up to 2000 feet. The com­plex is smaller than what might be needed to con­tain & defend a post-2012 chip fab against US assault; a 1997 page men­tions “dupli­cat­ing this facil­ity today with all of its mis­sions would cost about 18 bil­lion US dol­lars” (based on its 1965 con­struc­tion cost of $142m, which infla­tion-ad­justed is ~10 times larg­er). So opti­misti­cal­ly, hard­en­ing would cost another 25% (>$5b based on the TSMC Fab 15 esti­mate) or more than 500% (>$125b like­wise).

The China question

Like any good mer­can­tilist devel­op­ing East Asian coun­try, China has set its heart on mov­ing up the value chain of chip man­u­fac­tur­ing and heav­ily sub­si­dized local com­pa­nies31. While it may dom­i­nate prod­uct assem­bly and the sim­plest chip fab­ri­ca­tion, its ful­l-strength proces­sors like the is far from state-of-the-art (although this may give it an edge in power con­sump­tion at super­com­puter scale). Asia is a major semi­con­duc­tor con­sumer, and a great deal of gen­eral man­u­fac­tur­ing has already shifted there; both Tai­wan and China are poten­tial com­peti­tors32.

If we accept the Loong­son archi­tec­t’s pro­jec­tion of par­ity with Intel etc in 20 years (2031), this means that the cur­rent sit­u­a­tion of all cut­ting-edge chip fabs in reach of US power may face a seri­ous bar­ri­er: chip fabs located in main­land Chi­na. China is one of the few mil­i­tary pow­ers that can give the USAF a run for its mon­ey, and the other con­tender, Rus­sia, is entirely out of the chip fab busi­ness. What hap­pens in, say, 2041 with the Chi­nese chip fabs churn­ing out mature opti­mized designs? At this point, the WBE Roadmap pro­jec­tions sug­gest that brain emu­la­tion could be fea­si­ble by the 2040s and so the tar­get chip fabs may well have all been built in Chi­na.

Would China buy into any plan which the West some­how man­aged to agree on?

I con­sider the pos­si­bil­ity remote for many rea­sons when I project for­ward from the China of 2012:

  1. Eco­nomic growth is the chief man­date of the Chi­nese gov­ern­ment, and any­thing that might sab­o­tage that is opposed.

  2. Ren­t-seek­ing and cor­rup­tion are com­mon, par­tic­u­larly with ‘pres­tige’ projects or extremely large busi­ness inter­ests; if chip fab costs are brought under con­trol, they will still be enor­mous pool of cap­i­tal backed by even more enor­mous cor­po­ra­tions. Such enti­ties will be able to defend them­selves and defeat any such agree­ments.

  3. Inter­na­tional agree­ments are seen as a tool of an inter­na­tional sys­tem stacked against Chi­na. Any requests or “med­dling” in Chi­nese affairs is greeted with reflex­ive nation­al­ism.

    Con­verse­ly, inter­na­tional agree­ments are best seen as a form of legal war­fare; an expo­nent of this view was the wide­ly-read 199 text .

  4. China feels enti­tled to any­thing the West did dur­ing its own devel­op­ment; if the West could pol­lute with car­bon emis­sions, China may pol­lute (one of its main ratio­nale for ignor­ing things like the Kyoto agree­ment or sab­o­tag­ing cli­mate-change talk­s). A sim­i­lar argu­ment may be used with chip fabs.

  5. Spec­u­la­tive­ly, there may be more cul­tural com­fort in China with the pos­si­bil­ity of ‘robots’; at least, some Japan­ese have ascribed Japan­ese invest­ment into robots as being due to such cul­tural traits (eg. Murakami 2005), and this may carry over to Chi­na. With lit­tle anx­i­ety about the issue, why would they bother espe­cially when the costs are so con­crete and near?

All of these traits can change, and I fully expect many to change before 2040 (for exam­ple, I expect nation­al­ism & xeno­pho­bia to decrease as the pop­u­la­tion ages, and respect for inter­na­tional law will increase as China increas­ingly becomes the coun­try that ben­e­fits from a tidy sta­tus quo), but all of them, on top of the orig­i­nal unlike­li­ness?

  1. Any actual AI is likely to be a blend of approach­es—w­hole brain emu­la­tion and AGI form a con­tin­uum from a brute-­force mol­e­cule-by-­mol­e­cule emu­la­tion of a dis­sected human brain to a first-prin­ci­ples learn­ing algo­rithm capa­ble of generic cross-­do­main learn­ing & action which bears as much rela­tion to human think­ing as a heli­copter flies like an alba­tross or a sub­ma­rine (pace Dijk­stra) swims like a shark. A blend might be some­thing like a vast neural net­work based on human imag­ing stud­ies but with entire chunks of the brain replaced with sim­pli­fied pro­grams and hard­wired assump­tions imposed on each neu­ron in the inter­est of opti­miza­tion & com­pu­ta­tional tractabil­i­ty, where the sim­pli­fied chunks are those parts of brain func­tion­al­ity we have under­stood to the point where we can reim­ple­ment them by hand.↩︎

  2. Ana­log net­work pop­u­la­tion mode↩︎

  3. Sto­chas­tic behav­ior of sin­gle mol­e­cules↩︎

  4. After writ­ing this essay, I dis­cov­ered Ber­glas had briefly pointed out that proces­sors are a weak point:

    Try­ing to pre­vent peo­ple from build­ing intel­li­gent com­put­ers is like try­ing to stop the spread of knowl­edge. Once Eve picks the apple it is very hard to put it back on the tree. As we get close to arti­fi­cial intel­li­gence capa­bil­i­ties, it would only take a small team of clever pro­gram­mers any­where in the world to push it over the line. But it is not so easy to build pow­er­ful new com­puter chips. It takes large invest­ments and large teams with many spe­cial­ties from pro­duc­ing ultra pure sil­i­con to devel­op­ing extremely com­plex log­i­cal designs. Extremely com­plex and pre­cise machin­ery is required to build them. Unlike pro­gram­ming, this is cer­tainly not some­thing that can be done in some­one’s garage.

    So this paper pro­poses a mora­to­rium on pro­duc­ing faster com­put­ers. Just make it ille­gal to build the chips, and so starve any Arti­fi­cial Intel­li­gence of com­put­ing pow­er.

    We have a prece­dent in the con­trol of nuclear fuel. While far from per­fect, we do have strong con­trols on the avail­abil­ity of bomb mak­ing mate­ri­als, and they could be made stronger if the polit­i­cal will exist­ed. It is rel­a­tively easy to make an atomic bomb once one has enough plu­to­nium or highly enriched ura­ni­um. But mak­ing the fuel is much, much hard­er. That is why we are alive today. If some­one pro­duced a safe and afford­able car pow­ered by plu­to­ni­um, would we wel­come that as a solu­tion to soar­ing fuel prices? Of course not. We would con­sider it far too dan­ger­ous to have plu­to­nium scat­tered through­out soci­ety…(It might turn out that it is actu­ally the patent trolls and attor­neys that are our sav­ior. Intel­li­gence devel­op­ment would pro­vide a rich source of triv­ial patents and aggres­sive lit­i­ga­tion…)

    Sotala & Yam­pol­skiy men­tions “Ber­glas (per­sonal com­mu­ni­ca­tion) has since changed his mind, and no longer believes that it is pos­si­ble to effec­tively restrict hard­ware or oth­er­wise pre­vent AGI from being cre­at­ed.”↩︎

  5. In dis­cus­sions of tech­no­log­i­cal fore­cast­ing like AI pro­jec­tions or or , the demands of pessimists/skeptics that the events in ques­tion already have hap­pened before they can be dis­cussed are quite irri­tat­ing. I am reminded of the sar­cas­tic com­ment by 2 researchers on cli­mate change:

    “Should we trust mod­els or obser­va­tions?” In reply we note that if we had obser­va­tions of the future, we obvi­ously would trust them more than mod­els, but unfor­tu­nately obser­va­tions of the future are not avail­able at this time.

  6. “National Secu­rity Aspects of the Global Migra­tion of the U.S. Semi­con­duc­tor Indus­try”, Lieber­man 2003:

    A fun­da­men­tal change in the semi­con­duc­tor indus­try has been, in very sim­pli­fied form, that the price to per­for­mance curve has reduced rev­enue in the indus­try dra­mat­i­cally over the last decade. Dur­ing the early 1960s, and con­tin­u­ing until about 1994, the com­pound annual growth rate in rev­enue of the indus­try was 16%. From 1994 to the pre­sent, the growth rate has been approx­i­mately 8%.13 This sit­u­a­tion is com­bined with the very large costs asso­ci­ated with the devel­op­ment of new 300mm fab­ri­ca­tion facil­i­ties (“fabs”), as well as the increas­ing com­plex­ity and cost of research and design as the indus­try must develop meth­ods other than the tra­di­tional scal­ing meth­ods (mak­ing all aspects of the chips smaller and small­er) in order to increase per­for­mance. These fac­tors, and the cur­rent reces­sion, are dri­ving the indus­try to con­sol­i­da­tions.

    …The num­ber of state-of-the-art U.S. chip man­u­fac­tur­ing facil­i­ties is expected to sharply decrease in the next 3–5 years to as few as 1–2 firms that now have the rev­enue base to own a 300mm wafer pro­duc­tion fab, and likely less than a hand­ful of firms.15 Although the U.S. cur­rently leads the world semi­con­duc­tor indus­try with a 50% world mar­ket share, the Semi­con­duc­tor Indus­try Asso­ci­a­tion esti­mates that the U.S. share of 300mm wafer pro­duc­tion capac­ity will be only approx­i­mately 20% in 2005, while Asian share will reach 65% (only 10% of this from Japan).16 The remain­ing state-of-the-art U.S. chip-­mak­ing firms face great dif­fi­culty in attain­ing the huge amounts of cap­i­tal required to con­struct nex­t-­gen­er­a­tion fabs. This sit­u­a­tion stands in con­trast to that in Chi­na. To ensure that they develop the abil­ity to build the nex­t-­gen­er­a­tion fab­ri­ca­tion facil­i­ties, the Chi­nese cen­tral gov­ern­ment, in coop­er­a­tion with regional and local author­i­ties, has under­taken a large array of direct and indi­rect sub­si­dies to sup­port their domes­tic semi­con­duc­tor indus­try. They have also devel­oped a num­ber of part­ner­ships with U.S. and Euro­pean com­pa­nies that are cost-ad­van­ta­geous to the com­pa­nies in the short­-term.

  7. A caveat that will becom­ing increas­ingly impor­tant in com­ing decades.↩︎

  8. “The rise and fall of AMD: How an under­dog stuck it to Intel; Remem­ber when AMD could com­pete with Intel in both speed and price?”, ars tech­nica:

    AMD’s suc­cess­ful Athlon chip­s—Ars named the Athlon its “CPU of the Year” in 1999—had finally put the screws to archri­val Intel, and in 2000 the com­pany earned nearly $1 bil­lion in profits…AMD has been on a notable drop for nearly a decade now. To put it mild­ly, 2012 was a rough year: AMD lost over $1 bil­lion, effec­tively wip­ing out its $471 mil­lion profit in 2010 and its $491 mil­lion profit in 2011-its two most prof­itable years in the last decade. Over the last 15 years, AMD has sus­tained a net loss of nearly $7 bil­lion, and the com­pany has been down­graded by credit rat­ing agen­cies, burned by lower demand for PCs (and hence, for its prod­uct­s), and even called “un-in­vestable” by one Wall Street ana­lyst…How­ev­er, just weeks after the K7 debuted on June 23, 1999, Raza left AMD…As Raza tells the story today, his boss insisted on build­ing a fab in Dres­den, Ger­many, over Raza’s objec­tions. (That fab, which still oper­ates today as part of AMD spin-off Glob­al­Foundries, was com­pleted in the spring of 2000.) “The trou­ble in the entire eco­nomic model was that AMD did not have enough cap­i­tal to be able to fund fabs the way they were fund­ing fabs,” Raza said. “The point at which I had my final con­flict was that [Sanders] started the process of build­ing a new fab with bor­rowed money pre­ma­ture­ly. We did­n’t need a fab for at least another year. If we had done it a year lat­er, we would have accu­mu­lated enough prof­its to afford the fab in Ger­many. He laid the foun­da­tion for a fun­da­men­tally inef­fi­cient cap­i­tal struc­ture that AMD never recov­ered from. I told him: don’t do it….” Both Raza and Bar­ton recalled, inde­pen­dently of one anoth­er, one of Sanders’ mantras: “Real men have fabs.” Raza called this com­ment “simul­ta­ne­ously a sex­ist remark and the most stu­pid thing you can say,” and he saw the fab deci­sion as one of Sanders’ “sig­nif­i­cant acts of irre­spon­si­bil­i­ty.”

  9. Kim 2008, “Brick and Mor­tar Chip Fab­ri­ca­tion”

    Tech­nol­ogy scal­ing has pro­duced a wealth of tran­sis­tor resources and cor­re­spond­ing improve­ments in chip per­for­mance. How­ev­er, these ben­e­fits come with an increas­ing price tag, due to ris­ing design, engi­neer­ing, and val­i­da­tion costs of mod­ern chips [15]. The result has been a steady decline in unique appli­ca­tion-spe­cific inte­grated cir­cuit (ASIC) designs that enter pro­duc­tion [21]. This ini­ti­ates a vicious cycle. Fewer unique chips means that fabs have fewer cus­tomers across which to amor­tize their costs, lead­ing to even higher costs for those who do man­u­fac­ture chips. The cycle com­pletes as higher chip man­u­fac­tur­ing costs exclude even more poten­tial man­u­fac­tur­ing cus­tomers.

  10. Kim 2008, con­tin­ued:

    While Moore’s Law has fueled the semi­con­duc­tor indus­try, it has also fueled this spi­ral of increas­ing costs and shrink­ing fab cus­tomer bases. As tran­sis­tors have shrunk, the cost of fab­ri­cat­ing a semi­con­duc­tor device has grown com­men­su­rate­ly. While the fab­ri­ca­tion cost per tran­sis­tor has steadily declined [62], mul­ti­ple other expenses have bal­looned, con­tribut­ing col­lec­tively to the grow­ing total. For exam­ple, small fea­tures are more sus­cep­ti­ble to process vari­a­tion than larger ones, increas­ing the range of vari­a­tion and the pro­por­tion of faulty chips. In addi­tion, the smaller the tran­sis­tor, the more of them that can fit in a given amount of sil­i­con. The result is that cir­cuit com­plex­ity has been increas­ingly out­-strip­ping designer pro­duc­tiv­i­ty, in a phe­nom­e­non referred to as Moore’s Law’s corol­lary of “com­pound com­plex­ity” [143].

    The indus­try has dealt with these chal­lenges by increas­ing the engi­neer­ing effort that goes into each chip. This effort man­i­fests itself as larger design teams, or longer prod­uct cycles, and often both at once. The vast major­ity of this engi­neer­ing effort is incurred once per chip design, and does not vary with the num­ber of chips pro­duced. Accord­ing­ly, this expense is called the non-re­cur­ring engi­neer­ing cost (NRE) of a chip. Indus­try ana­lysts esti­mate that the NREs for a typ­i­cal 90nm stan­dard cell ASIC can range from $5M up to $50M [113].

  11. Kim 2008, con­tin­ued:

    Main­tain­ing a par­tic­u­lar­quires larger and larger batches of chips. This is because the sin­gle NRE is shared evenly across the pop­u­la­tion of chips pro­duced. The larger the pop­u­la­tion, the smaller the impact of the NRE on indi­vid­ual chip cost…The result of this sit­u­a­tion is that only high­-vol­ume chip man­u­fac­tur­ers, or those who can sell smaller batches at high prices, can afford to be in the chip busi­ness. More­over, at the same time that com­plex­ity and engi­neer­ing effort have been soar­ing, the com­mer­cial mar­ket has been demand­ing and reward­ing short chip design cycles. This is due to shrink­ing prod­uct life­times and the increas­ing com­pet­i­tive impor­tance of being the first to mar­ket with a new pro­duc­t…One of the inputs is the assumed NRE. The NRE includes all engi­neer­ing effort…with an engi­neer’s time cost­ing upwards of $380,000 per year [141], the engi­neer­ing cost is nearly always a sev­en-­fig­ure number…NREs also encom­pass the cost of tools, IP licenses if nec­es­sary, and pho­tolith­o­graphic masks. ASIC design tools typ­i­cally cost more than $300,000 [146]. Mask cost has been roughly dou­bling every tech­nol­ogy node, result­ing in a com­plete set of 90nm masks cost­ing between $1M and $3M [146]…In 2000, the cost to test each tran­sis­tor was 10% of the cost to man­u­fac­ture it. How­ev­er, as tran­sis­tors become cheaper and test­ing becomes more dif­fi­cult, it is pro­jected that by 2015 it will cost more to test a tran­sis­tor than to make it [73].

  12. “Moore’s Law reaches its eco­nomic lim­its”, Finan­cial Times 2012:

    “The high cost of semi­con­duc­tor man­u­fac­tur­ing equip­ment is mak­ing con­tin­ued chip­mak­ing advance­ments too expen­sive to use for vol­ume pro­duc­tion, rel­e­gat­ing Moore’s Law to the lab­o­ra­tory and alter­ing the fun­da­men­tal eco­nom­ics of the indus­try,” wrote Len Jelinek, chief ana­lyst for semi­con­duc­tor man­u­fac­tur­ing at the iSup­pli research firm, last month.

    Mr Jelinek pre­dicted that Moore’s Law would no longer drive vol­ume chip pro­duc­tion from 2014, spark­ing intense debate in Sil­i­con Val­ley. His rea­son­ing is that cir­cuitry widths will dip to 20nm (nanome­tres or bil­lionths of a metre) or below by that date. But the tools to make them would be too expen­sive for com­pa­nies to recover their costs over the life­time of pro­duc­tion.

    The costs and risks involved in build­ing new fabs have already dri­ven many mak­ers of logic chips (proces­sor or con­troller chips) towards a “fab­less” chip mod­el, where they out­source much of their pro­duc­tion to chip “foundries” in the Far East.

    The 14 chip­mak­ers who were in the game at the 90nm level have been reduced to nine at the cur­rent 45nm lev­el. Only two of them—In­tel and —have firm plans for 22nm fac­to­ries. Intel argues that only com­pa­nies with about $9bn in annual rev­enues can afford to be in the busi­ness of build­ing new fabs, given the costs of build­ing and oper­at­ing the fac­to­ries and earn­ing a decent 50 per cent mar­gin. That leaves just Intel, Sam­sung, , and .

  13. “The Eco­nomic Limit to Moore’s Law”, Rupp & Sel­ber­herr 2011

    …The reduced growth model (10) can be inter­preted in the fol­low­ing way; as long as fab costs increase with the same rate as they did in the past, the num­ber of tran­sis­tors per chip also increases at the same rate as in the past. How­ev­er, as soon as fab costs hit an eco­nomic bar­rier given by εg(t), fab costs can only increase at the same rate as the GWP [Global World Pro­duct] does. Con­se­quent­ly, tran­sis­tor counts will also grow at a reduced rate…The pre­dic­tion of the time at which we run into eco­nomic lim­i­ta­tions is very sen­si­tive with respect to the fab costs para­me­ter ε; choos­ing ε = 0.03%, a growth reduc­tion is pre­dicted around 2025, whereas the choice ε = 0.01% shows first signs of reduced growth already in 2015. Thus, with joint fund­ing of large fabs, an eco­nomic growth cap­ping can be shifted many years into the future so that we might face lim­i­ta­tions imposed by physics first.

  14. “Power Blip Jolts Sup­ply of Gad­get Chips”, Decem­ber 10, 2010, WSJ:

    Toshiba said the power out­age could cause a 20% drop in its ship­ments over the next two months or so of chips known as NAND flash mem­o­ry, which are used to store music, pho­tos and data in prod­ucts such as Apple Inc.’s iPhone and iPad. Toshiba, which makes the chips in part­ner­ship with Sil­i­con Val­ley com­pany San­Disk Cor­p., rep­re­sents about a third of the mar­ket as the sec­ond-largest sup­plier of the chips after Sam­sung Elec­tron­ics Co. After the next cou­ple months, the out­age isn’t expected to have a [sub­stan­tial] impact on world-wide ship­ments of flash mem­o­ry. Some big buy­ers of the chips, such as Apple, have long-term sup­ply arrange­ments with mul­ti­ple chip mak­ers. But the tem­po­rary dis­rup­tion comes as demand for NAND flash is surg­ing, notably from com­pa­nies hop­ing to offer new tablet com­put­ers to chal­lenge the iPad. Mar­ket watch­ers say some com­pa­nies could face tight sup­plies and higher prices just as they are try­ing to ramp up pro­duc­tion. “I don’t think it could come at a worse time,” said Krishna Shankar, an ana­lyst at ThinkE­quity.

    Toshiba’s trou­bles started early Wednes­day when, accord­ing to power sup­plier Chubu Elec­tric Power Co., there was a sud­den drop in volt­age that caused a 0.07-sec­ond power inter­rup­tion at Toshiba’s Yokkaichi mem­o­ry-chip plant in Mie pre­fec­ture. Even the briefest power inter­rup­tion to the com­plex machines that make chips can have an effect com­pa­ra­ble to dis­con­nect­ing the power cord on a desk­top com­put­er, since the com­put­er­ized con­trols on the sys­tems must effec­tively be reboot­ed, said Dan Hutch­eson, a chip-­man­u­fac­tur­ing ana­lyst at VLSI Research in San Jose, Calif. For that rea­son, chip com­pa­nies typ­i­cally take pre­cau­tions that include installing what the indus­try calls unin­ter­rupt­ible power sup­plies. Part of Toshiba’s safe­guards did­n’t work this time because the volt­age drop was more severe than what the backup sys­tem is designed to han­dle, a com­pany spokesman said. Power out­ages fre­quently cause dam­age to chips, which are fab­ri­cated on sil­i­con wafers about the size of din­ner plates that may take eight to 12 weeks to process, Mr. Hutch­e­son said. Wafers that are inside pro­cess­ing machines at the time of an out­age are often ruined, he added, though many that are in stor­age or in tran­sit among those machines are not. In some cas­es, a shut­down of the air-pu­ri­fy­ing and con­di­tion­ing sys­tem that keeps air in a chip fac­tory free of dust also could con­t­a­m­i­nate chips. Mr. Hutch­e­son com­pared the sit­u­a­tion to cut­ting off the power to an arti­fi­cial heart machine in the mid­dle of an oper­a­tion. “You lose the patient,” he said. On the other hand, he said that Toshiba’s esti­mate of the impact is a worst-­case sce­nario that may wind up to be sub­stan­tially less. Toshiba esti­mated that its ship­ments of NAND flash mem­ory could decline by as much as 20% through Feb­ru­ary as a result of the out­age. Based on the com­pa­ny’s share of the mar­ket, such a reduc­tion would trans­late into a 7.5% cut in world-wide ship­ments over that peri­od, but a much smaller per­cent­age for all of 2011, esti­mated Michael Yang, an ana­lyst at the tech­nol­ogy mar­ket research firm iSup­pli.

  15. Con­tin­ued:

    “The impact for the year is noth­ing,” Mr. Yang said, though he added it could tem­porar­ily raise NAND prices. Such prob­lems are far from unprece­dent­ed. In August 2007, a power out­age at Sam­sung’s mem­o­ry-chip plant near Seoul forced the com­pany to tem­porar­ily halt some pro­duc­tion lines. Mr. Yang said that cased prices for chips to rise for a few months. Tim­o­thy Luke, an ana­lyst at Bar­clays Cap­i­tal, esti­mated the out­age would reduce 2011 NAND flash chip sup­plies by 3% to 5%, pos­si­bly boost­ing prices to the ben­e­fit of rivals such as Sam­sung and Micron Tech­nol­ogy Inc. A San­Disk spokesman declined to com­ment on the out­age. One mit­i­gat­ing fac­tor is that chip demand is typ­i­cally lighter in Jan­u­ary and Feb­ru­ary than other parts of the year, which could reduce the chances of a short­age. On the other hand, demand for NAND chips has been ris­ing at an unusual rate, dri­ven largely by sales of smart­phones and tablets. ISup­pli in Sep­tem­ber pre­dicted that global ship­ments of NAND flash-mem­ory chips by vol­ume would jump 70% next year.

    Another instance of power out­age dam­ag­ing a frac­tion of global out­put hap­pened again to Sam­sung on 2018-03-09.↩︎

  16. “The Split Sec­ond Dis­rup­tion to the Sup­ply Chain”:

    In 2005, Pro­fes­sor Yossi Sheffi from the Mass­a­chu­setts Insti­tute of Tech­nol­ogy wrote a sem­i­nal work enti­tled The Resilient Enter­prise that exam­ined dis­rup­tions in cor­po­rate sup­ply chains. The very first chap­ter in that book was enti­tled “Big Lessons from Small Dis­rup­tions.” It unfolds a story about how a St. Patrick’s Day (2000-03-17) light­ning strike in Albu­querque, New Mex­i­co, started a fire in Fab­ri­ca­tor No. 22 of a Phillips NV chip man­u­fac­tur­ing plant which led to unfore­seen long-term con­se­quences. Alert plant employ­ees and auto­matic sprin­klers put the fire out in less than ten min­utes. Sheffi wrote, “A rou­tine inves­ti­ga­tion showed that the fire had been minor. Nobody was hurt and the dam­age seemed super­fi­cial. The blaze did not make head­lines in Europe, did not appear on CNN, and did not even appear in the Albu­querque news­pa­pers.”…The fire had directly ruined only eight trays of wafers but smoke from the fire spread beyond the imme­di­ate area and, along with soot spread by work­ers and fire­fight­ers, con­t­a­m­i­nated much of the rest of plant. A minor fire had turned into a major dis­as­ter. Sheffi reported that Phillips noti­fied its 30-­plus cus­tomers about poten­tial delays in chip pro­duc­tion but pre­dicted the delay would only be about a week…as soon as it [Nokia] real­ized that the delay was actu­ally going to be weeks or months, it took action…Er­ic­s­son was­n’t quite so lucky. Sheffi reported that Eric­s­son exec­u­tives received the same tele­phone call from Phillips as Nokia but they reacted very dif­fer­ent­ly. They believed that the delay would be a short one…Ac­cord­ing to Shef­fi, Eric­sson’s lack of a Plan B cost the com­pany around half a bil­lion dol­lars.

  17. iSup­pli attrib­utes the very slow recov­ery to, in part, long-term con­tracts lock­ing in PC man­u­fac­tur­ers and the hard drive indus­try con­sol­i­dat­ing into 2 oli­garchi­cal man­u­fac­tur­ers. One might be tempted to argue that the Thai hard drive floods are a poor case-s­tudy because of these idio­syn­cratic fac­tors; but that’s miss­ing the point: until a dis­rup­tion hap­pens, how can one know—e­spe­cially an out­sider—what will or will not be a prob­lem? Are there no long-term con­tracts for proces­sors, and is not Intel largely dom­i­nant? And what are the idio­syn­cratic fac­tors rel­e­vant to proces­sors? There must be some, even if we would not expect them to be the same as in the Thai hard drive floods.↩︎

  18. “Thai­land Flood­ing Crip­ples Hard-­Drive Sup­pli­ers”, The New York Times:

    The slow-­mov­ing flood­wa­ters, which are an accu­mu­la­tion from this year’s unusu­ally strong mon­soon rains in north­ern Thai­land, are grad­u­ally drain­ing into the sea. At what is known as the Bang Pa-In Indus­trial Estate, trucks have deliv­ered mas­sive pumps. Work­ers said they would start try­ing to remove water from the area on Mon­day. The flood­wa­ters descended to this area an hour north of Bangkok in early Octo­ber. Efforts to defend indus­trial areas with sand­bags and other bar­ri­ers were futile.

    “There was no way we could have held back the water,” said Sam­ruay Pakubol, a welder at an auto­mo­tive parts fac­tory here. Now out of work, he takes pas­sen­gers on a wooden boat down the streets of the indus­trial zone. Work­ers have caught and killed croc­o­diles swim­ming in the area, he said.

    Dale Schudel, man­ag­ing direc­tor of IntriPlex Thai­land, a com­pany that makes com­po­nents for hard-disk dri­ves, said his fac­tory in nearby Ayut­thaya had flood­wa­ters almost six feet deep. But pump­ing out the water, which will take about two to three weeks, is only the begin­ning of the cleanup. Mr. Schudel described the water as “highly cor­ro­sive.” “I think you have to ask your­self, if any fac­tory in the world were sub­merged in that much water, how much dam­age would there be?” he said.

  19. I think drones are an inter­est­ing issue, but right now they’re look­ing like a mas­sive shift toward government/corporate pow­er: in 20 years, you may be able to afford an awe­some drag­on­fly drone which you’d like to pilot into a chip fab and upload some evil (note I’m not even both­er­ing to think about the sce­nario in which you load the drone with 20kg of high explo­sives and crash it into a tar­get), but your drone won’t be able to get through the thou­sands of patrolling secu­rity drones spread out over the entire instal­la­tion! Enthu­si­asts right now may be play­ing around with nifty quadropters etc… and the US gov­ern­ment is prob­a­bly play­ing around with drones that never come down, have radar & surveillance/detection capa­bil­i­ties that us civil­ians can’t even guess at, and with net­worked fleets of them.↩︎

  20. is help­ful as usu­al, and a quick Google turns up a ran­dom lob­by­ing group giv­ing some num­bers (mul­ti­ply by 1000 to get megawat­t-hours):

    Accord­ing to a 2008 Elec­tric Power Research Insti­tute study, a con­ven­tional com­bined-­cy­cle nat­ural gas plant costs about $1,000 per kilo­watt of capac­ity con­struct­ed. A coal-­fired plant costs more than $2,500 per KW hour to build. The cost of a new nuclear plant is more than $4,000 per KW of capac­i­ty. Wind gen­er­a­tion costs are about dou­ble nat­ural gas instal­la­tion costs. A new solar plant in Florida is pro­jected at about $6,600 per KW.

  21. It’s a lit­tle like try­ing to get Amer­i­can lib­er­tar­i­ans to engage in effec­tive col­lec­tive action like lob­by­ing or —if they were inclined to be com­mu­nal and bow to hier­ar­chy, they’d not be lib­er­tar­i­ans but Repub­li­cans or some­thing!↩︎

  22. Less­Wrong is no excep­tion. It always amuses me that, as much as it gets called a cult, when you actu­ally ask LWers what they believe in sur­veys (2009, 2011), you find that the stereo­typ­i­cal Less­Wrong shib­bo­leth­s—Strong AI the­sis is true, AI will be accom­plished dur­ing the 21st cen­tu­ry, is the right eth­i­cal par­a­digm like lib­eral lib­er­tar­i­an­ism is the right polit­i­cal par­a­digm and athe­ism for reli­gion, is more true than other quan­tum mechan­i­cal inter­pre­ta­tions, is a good idea etc.—are actu­ally held by bare majori­ties and some­times pretty small minori­ties. For exam­ple:

    • lib­er­tar­i­an­ism (32%) is rarer than lib­er­al­ism (35%)
    • con­se­quen­tial­ists are only 62% of respon­dents (a big sur­prise to me) and Many Worlders 56%
    • ‘prob­a­bil­ity cry­on­ics will work’ aver­aged 21%; 4% of LWers are signed up, 36% opposed, and 54% merely ‘con­sid­er­ing’ it
    • AI is feared less than pan­demics (26% vs 17%)
    • the median Sin­gu­lar­ity is 2080 (not a Kurzweil­ian 2030–2040)
  23. This is another instance where detailed domain knowl­edge would be help­ful. Shock waves dis­si­pate fast: basic geom­e­try sug­gests they weaken as roughly the cube of the dis­tance because the energy is being dis­si­pated over a spher­i­cal vol­ume. More pre­cise­ly, from an old nuclear weapons text­book, I learned it was an inter­me­di­ate power between inverse square (a fixed energy spread­ing over a 2D area) and inverse cube (over 3D vol­ume) because they hit bound­ary lay­ers and reflect. Either way, a blast may need to be very close to a key part of a chip fab to mat­ter. This leads to some inter­est­ing trade­offs in tar­get­ing: Soviet nukes tended to have higher mega­ton­nage than Amer­i­can nukes, despite the lat­ter’s pre­sum­able tech­ni­cal supe­ri­or­i­ty, because their tar­get­ing was inac­cu­rate and the higher mega­ton­nage made up for the inac­cu­racy via overkill; Amer­i­can mis­siles favored mul­ti­ple war­heads on , because the high accu­racy meant that a tar­get could be brack­eted between sev­eral small nukes and the con­verg­ing shock­waves much more effec­tive.↩︎

  24. in men­tions that heroin syn­the­sis was intro­duced to one region by an old Chi­nese chemist with 7 dis­ci­ples—so many, pre­sum­ably, because he had only one arm left and needed assis­tants.↩︎

  25. Note that this claim is dou­ble-edged: if it will become impos­si­ble in 2 gen­er­a­tions, then we can infer that it must be pos­si­ble ‘now’ (2003) at some very high price, and given Moore’s sec­ond law, may have been pos­si­ble at a rea­son­able price in the past (<1980s?).

    The is a known exam­ple: among their many facil­i­ties like the , they in their Army facil­i­ties at (op­er­ated by National Semi­con­duc­tor Cor­po­ra­tion). There are many com­pelling appli­ca­tions for an NSA chip fab to the point where it’s hard to guess what its pri­mary mis­sion might be:

    1. cus­tom fab­ri­ca­tion of rel­a­tively low tech but extremely secu­ri­ty-sen­si­tive chips: per­haps encryp­tion chips for US gov­ern­ment use. (Nu­clear bomb com­po­nents have been sug­gested but would­n’t fall under their dual man­date of secur­ing US com­mu­ni­ca­tions and break­ing for­eign com­mu­ni­ca­tion­s.) Mak­ing cus­tom, lega­cy, and obscure chips is the offi­cial expla­na­tion:

      The orig­i­nal idea was to ser­vice NSA’s own needs in the intel­li­gence are­na, par­tic­u­larly for old com­po­nents that exec­u­tives of com­mer­cial semi­con­duc­tor firms are either unable or unwill­ing to pro­duce [due to clo­sures of chip fabs, eg. SVTC]. Now the facil­ity is ramp­ing up to ser­vice the needs of other fed­eral agen­cies and pro­grams, and is offer­ing a com­plete one-stop shop­ping cen­ter for design, fab­ri­ca­tion, assem­bly, and test. This is accord­ing to a sales pitch pre­sented by Leland Miller, NSA’s mar­ket­ing direc­tor for micro­elec­tron­ics, at the Octo­ber Asso­ci­a­tion of Old Crows (AOC) con­fer­ence in Vir­ginia Beach, Va. The man­u­fac­tur­ing process is bulk CMOS, using 6-inch wafers, to fab­ri­cate fea­ture sizes of 1, 0.8, and 0.5 microns (1000/800/500nm) in either two or three metal lay­ers. This is due to be upgraded to 0.22 micron (200n­m), six-level metal tech­nol­ogy by July 2000. Typ­i­cal pro­duc­tion amounts to 1,000 wafer starts a mon­th, Miller says. Yields are com­pa­ra­ble to the indus­try aver­age, he adds—around 15% for large die and 80–90% for small die. Out­put has risen sharply from 350,000 to 450,000 die over the past year, Miller notes. It was about 3,000 die when the facil­ity began at the 1-mi­cron lev­el. About 150 prod­uct designs are in the fab­ri­ca­tion process simul­ta­ne­ous­ly. The NSA facil­ity has con­cen­trated on the agen­cy’s own needs, par­tic­u­larly on old 1-mi­cron prod­ucts, but has ser­viced other gov­ern­ment cus­tomer­s…In addi­tion to fab­ri­ca­tion and pack­ag­ing of pro­gram­ma­ble gate arrays in plas­tic mul­ti­chip mod­ules, flipchips, and ball grid arrays, the facil­ity also can do cus­tom designs of appli­ca­tion-spe­cific inte­grated cir­cuits using any start­ing point from block dia­grams to fin­ished lay­outs. Typ­i­cal deliv­ery times are 12 weeks from sub­mis­sion of a data­base tape, but spe­cial orders can be han­dled in three weeks. Func­tional and para­met­ric tests are con­ducted using stan­dard com­mer­cial test equip­ment.

      This seems like the most likely expla­na­tion: it’s much eas­ier for a small plant to keep up if it’s just mak­ing old out­dated chips.

    2. cus­tom fab­ri­ca­tion of high­ly-spe­cial­ized chips for use in code-break­ing and sur­veil­lance: well-de­signed ASICs can offer over com­mod­ity CPUs or GPUs.

    More intrigu­ing is the obser­va­tion by David Honig that one pos­si­bil­ity is that the NSA is engaged in (men­tioned in one NSA job ad): e-beam is far too slow for prof­itable com­mer­cial chip pro­duc­tion, but it has the major advan­tage that it can pro­duce smaller chip fea­tures. For com­par­ison, the Wikipedia arti­cle says e-beam sys­tems have worked at “~20nm since at least 1990” (with the state of the art being <10nm) while Intel only began com­mer­cially sell­ing chips 22 years later in April 2012.

    There would be lit­tle rea­son for the NSA to pro­duce its own CPUs and GPUs with e-beam, and there are plenty of com­mod­ity ASIC mak­ers if the NSA sim­ply wanted a highly par­al­lel Bit­coin-style chip for brute-­forc­ing SHA hash­es, since in both cases out­sourc­ing would be cheaper (one sim­ply buys a lot of SHA-cracking ASICs since the prob­lem is embar­rass­ingly par­al­lel); see for exam­ple, NSA’s IBM designed/built WindsorGreen/WindsorBlue super­com­puter. But there is some­thing that expen­sive tiny e-beam-­made chips could be supe­rior at: run­ning ser­ial prob­lem­s—where one can­not sim­ply par­al­lelize over many CPUs/GPUs/ASICs—faster than any­one else can. What NSA prob­lems might ben­e­fit from a few cus­tom expen­sive very fast ser­ial chips, I do not know; noth­ing in the Snow­den leaks I’ve seen has jumped out as a can­di­date. 3. reverse-engi­neer­ing for­eign-pro­duced chips: either to try to break their sys­tems, or look for back­doors in Amer­i­can-used chips (or per­haps for back­doors which can then be exploited by the NSA). 4. test out back­doors of their own on sim­ple test­bed chips, either for use in com­mer­cial plants’ chips or to test out back­door-tests for #3↩︎

  26. On a per­sonal note, my grand­fa­ther served in Korea dur­ing the Korean War. He was never in any dan­ger because he was deployed well away from the front lines. We don’t know what he did, but he insists it is clas­si­fied, and hinted it involved nuclear mate­r­i­al. (He spent much of his later career involved in gas masks.) This would be con­sis­tent with the dur­ing the Korean War.↩︎

  27. pg 413–414 of :

    Though the United States and the USSR con­tin­ued to develop nuclear tech­nol­ogy at a break­neck pace, they began, how­ever hyp­o­crit­i­cal­ly, to pay homage to nuclear dis­ar­ma­ment in con­fer­ences and state­ments. At the same time a grass­roots move­ment began to stig­ma­tize the weapons. Demon­stra­tions and peti­tions attracted mil­lions of cit­i­zens, together with pub­lic fig­ures such as Linus Paul­ing, Bertrand Rus­sell, and Albert Schweitzer. The mount­ing pres­sure helped nudge the super­pow­ers to a mora­to­rium and then a ban on atmos­pheric nuclear test­ing, and then to a string of arm­s-­con­trol agree­ments. The Cuban Mis­sile Cri­sis in 1962 was a tip­ping point. Lyn­don John­son cap­i­tal­ized on the change to demo­nize Gold­wa­ter in the Daisy ad and called atten­tion to the cat­e­gor­i­cal bound­ary in a 1964 pub­lic state­ment: “Make no mis­take. There is no such thing as a con­ven­tional nuclear weapon. For nine­teen per­il-­filled years no nation has loosed the atom against anoth­er. To do so now is a polit­i­cal deci­sion of the high­est order.”205

    As the world’s luck held out, and the two nuclear-free decades grew to three and four and five and six, the taboo fed on itself in the run­away process by which norms become com­mon knowl­edge. The use of nuclear weapons was unthink­able because every­one knew it was unthink­able, and every­one knew that every­one knew it. The fact that wars both large (Viet­nam) and small (Falk­lands) were not deterred by the increas­ingly inef­fec­tual nuclear threat was a small price to pay for the indef­i­nite post­pone­ment of Armaged­don.

    …One hope­ful sign is that nuclear pro­lif­er­a­tion has not pro­ceeded at the furi­ous rate that every­one expect­ed. In the 1960 pres­i­den­tial elec­tion debates, John F. Kennedy pre­dicted that by 1964 there might be “ten, fif­teen, twenty” coun­tries with nuclear weapons.206 The con­cern accel­er­ated when China con­ducted its first nuclear test in 1964, bring­ing the num­ber of nations in the nuclear club to five in less than twenty years. Tom Lehrer cap­tured pop­u­lar fears of run­away nuclear pro­lif­er­a­tion in his song “Who’s Next?” which ran through a list of coun­tries that he expected would soon become nuclear pow­ers (“Lux­em­bourg is next to go / And who knows? Maybe Monaco”).

    But the only coun­try that ful­filled his prophecy is Israel (“‘The Lord’s my shep­herd,’ says the Psalm / But just in case-we bet­ter get a bomb!”). Con­trary to expert pre­dic­tions that Japan would “unequiv­o­cally start on the process of acquir­ing nuclear weapons” by 1980 and that a reuni­fied Ger­many “will feel inse­cure with­out nuclear weapons,” nei­ther coun­try seems inter­ested in devel­op­ing them.207 And believe it or not, since 1964 as many coun­tries have given up nuclear weapons as have acquired them. Say what? While Israel, India, Pak­istan, and North Korea cur­rently have a nuclear capa­bil­i­ty, South Africa dis­man­tled its stash shortly before the col­lapse of the apartheid regime in 1989, and Kaza­khstan, Ukraine, and Belarus said “no thanks” to the arse­nals they inher­ited from the defunct Soviet Union. Also, believe it or not, the num­ber of non­nu­clear nations that are pur­su­ing nuclear weapons has plum­meted since the 1980s. Fig­ure 5-22, based on a tally by the polit­i­cal sci­en­tist Scott Sagan, charts the num­ber of non­nu­clear states in each year since 1945 that had pro­grams for devel­op­ing nuclear weapons.

    The downslopes in the curve show that at var­i­ous times Alge­ria, Aus­tralia, Brazil, Egypt, Iraq, Libya, Roma­nia, South Korea, Switzer­land, Swe­den, Tai­wan, and Yugoslavia have pur­sued nuclear weapons but then thought the bet­ter of it-oc­ca­sion­ally through the per­sua­sion of an Israeli air strike, but more often by choice.

    Pinker focuses on the nuclear taboo, but can we ignore entirely the global restric­tions on nuclear tech­nol­ogy trans­fer and uni­ver­sal state reg­u­la­tion of any­thing to do with nuclear tech­nol­o­gy? The par­al­lels to semi­con­duc­tor chip tech­nol­ogy are clear.↩︎

  28. pg124–125, (em­pha­sis added):

    In the prob­lem of decod­ing, the most impor­tant infor­ma­tion which we can pos­sess is the knowl­edge that the mes­sage which we are read­ing is not gib­ber­ish. A com­mon method of dis­con­cert­ing code­break­ers is to mix in with the legit­i­mate mes­sage a mes­sage that can­not be decod­ed; a non-sig­nif­i­cant mes­sage, a mere assem­blage of char­ac­ters. In a sim­i­lar way, when we con­sider a prob­lem of nature such as that of atomic reac­tions and atomic explo­sives, the largest sin­gle item of infor­ma­tion which we can make pub­lic is that they exist. Once a sci­en­tist attacks a prob­lem which he knows to have an answer, his entire atti­tude is changed. He is already some 50% of his way toward that answer.

    In view of this, it is per­fectly fair to say that the one secret con­cern­ing the atomic bomb which might have been kept and which was given to the pub­lic and to all poten­tial ene­mies with­out the least inhi­bi­tion, was that of the pos­si­bil­ity on its con­struc­tion. Take a prob­lem of this impor­tance and assure the sci­en­tific world that it has an answer; then both the intel­lec­tual abil­ity of the sci­en­tists and the exist­ing lab­o­ra­tory facil­i­ties are so widely dis­trib­uted that the qua­si­-in­de­pen­dent real­iza­tion of the task will be a mat­ter of merely a few years any­where in the world.

  29. pg 39–40:

    To put it at its most ele­men­tary, while observ­ing oth­ers rid­ing bicy­cles does not enable one to learn the skills of the cyclist, it nev­er­the­less shows that cycling is pos­si­ble. Know­ing that older broth­ers or sis­ters have learned to ride can encour­age younger sib­lings not to con­clude from early fail­ures that the task is impos­si­bly hard.

    …The con­fi­dence—in­deed over­con­fi­dence—of wartime Anglo-Amer­i­can physi­cists (in­clud­ing Con­ti­nen­tal refugees) in the ease of devel­op­ment of a nuclear weapon does not seem to have been widely shared by their French, Ger­man, or Soviet col­leagues, and the gov­ern­ments of the last two coun­tries were uncon­vinced prior to 1945 that the task was fea­si­ble enough to be worth the kind of resources the Amer­i­cans devoted to it (see, e.g., Hol­loway 1981; Gold­schmidt 1984, p. 24).24 Trin­i­ty, Hiroshi­ma, and Nagasaki were dra­matic demon­stra­tions that the task was not impos­si­bly hard, and this proof (as well, of course, as the per­ceived threat to the Soviet Union) explains the sud­den shift in the USSR in 1945 from a mod­est research effort to an all-out, top-pri­or­ity pro­gram (Hol­loway 1981).

    As we have seen, the British test explo­sion in 1952, although no threat to France, con­tributed to the lat­ter’s weapons pro­gram by sug­gest­ing that devel­op­ing an atomic bomb was eas­ier than had pre­vi­ously been assumed. Like­wise, the Chi­nese explo­sion in 1964 showed other devel­op­ing coun­tries that the atomic bomb was not nec­es­sar­ily the pre­serve solely of the highly indus­tri­al­ized world. Fur­ther­more, pro­found ques­tions over the fea­si­bil­ity of early hydro­gen bomb designs helped delay the Amer­i­can move from an atomic to a hydro­gen bomb (Bethe 1982). By con­trast, all sub­se­quent hydro­gen bomb pro­grams could pro­ceed with con­fi­dence in the basic achiev­abil­ity of their goal, and, in words used in another con­text by a group of weapons design­ers (Mark et al. 1987, p. 64), “The mere fact of know­ing [some­thing] is pos­si­ble, even with­out know­ing exactly how, [can] focus … atten­tion and efforts.”

  30. tech­ni­cal report M-85/11 “Lit­er­a­ture Sur­vey of Under­ground Con­struc­tion Meth­ods for Appli­ca­tion to Hard­ened Facil­i­ties”:

    An eval­u­a­tion of a nuclear power plant con­cept [79] revealed that locat­ing the facil­ity under­ground with a cut and cover tech­nique would be 11% more expen­sive than an above-­ground plan. The increased cost was attrib­uted to direct con­struc­tion costs being 70% high­er, the need for spe­cial equip­ment for ven­ti­la­tion and other func­tions, and the addi­tional time required to build the under­ground struc­ture. More costs are incurred from hard­en­ing under­ground tun­nels to resist blasts or seis­mic loads. A design cost study [85] esti­mates that hard­en­ing a tun­nel to resist a seis­mic load of 0.5 g would increase con­struc­tion costs by 35%…Un­der­ground sub­way sta­tion con­struc­tion costs were also com­pared to show that an under­ground sta­tion using a tun­neled earth exca­va­tion tech­nique with an 85-ft (25.5-m) over­bur­den would cost about 25% more than one con­structed by cut and cover meth­ods with a 20-ft (6-m) earth cov­er, an 47% more than a cut and cover sta­tion with a 6-ft (1.8-m) cov­er…Ref­er­ence 77 com­pares the cost of sit­ing a nuclear power plant under­ground. The inves­ti­ga­tion found that a cut and cover buried facil­ity would cost 14–25% more and a mined rock plant 10–18% more than a sur­face power plant. A sec­ond report [86] states that costs for sit­ing a nuclear power plant under­ground in rock are about 25% more. Ref­er­ence 80 exam­ines the costs of under­ground homes and large pub­lic build­ings. Based on life-­cy­cle cost fig­ures of 5 case stud­ies: ‘It does appear clear, how­ev­er, that the use of earth­-shel­ter­ing does not increase con­struc­tion costs in any notable way, and may in fact rep­re­sent a decrease in some cases’ [80]. An exam­ple earth­-shel­tered house is cited as cost­ing 28% more to con­struct, but 12–20% less to own and oper­ate over the 30-year life of the home.

  31. Lieber­man 2003:

    The Chi­nese gov­ern­ment is suc­cess­fully using tax sub­si­dies (see below) to attract for­eign cap­i­tal from semi­con­duc­tor firms seek­ing access to what is expected to be one of the world’s largest mar­kets. This strat­e­gy, which is sim­i­lar to that employed by the Euro­pean Union in early 1990s, is a means of induc­ing sub­stan­tial inflows of direct invest­ment by pri­vate firms. Indeed, much of the fund­ing is Tai­wane­se, dri­ven by the tax incen­tives and their need for mar­ket access, espe­cially for com­mod­ity prod­ucts such as DRAMs. The strat­egy does not rely on cheaper labor, as that is a small ele­ment in semi­con­duc­tor pro­duc­tion.

    The Chi­nese are, how­ev­er, able to increas­ingly draw on sub­stan­tially larger pools of tech­ni­cally trained labor as com­pared to the U.S., from the large cohorts of domes­tic engi­neer­ing grad­u­ates.17 Impor­tant­ly, the out­put of Chi­nese uni­ver­si­ties is sup­ple­mented by large num­bers of engi­neers trained at U.S. uni­ver­si­ties and mid-­ca­reer pro­fes­sion­als who are offered sub­stan­tial incen­tives to return to work in Chi­na. These incen­tives for sci­en­tists and engi­neers, which include sub­stan­tial tax ben­e­fits, world-­class liv­ing facil­i­ties, exten­sive stock options taxed at par val­ue, and other ameni­ties, are prov­ing effec­tive in attract­ing expa­tri­ate labor. For exam­ple, the Chi­nese cen­tral gov­ern­ment has under­taken indi­rect sub­si­dies in the form of a sub­stan­tial rebate on the val­ue-added tax (VAT) charged on Chi­ne­se-­made chip­s.19 While many believe this is an ille­gal sub­sidy under GATT trade rules, the impact of the sub­sidy on the growth of the indus­try may well be irre­versible before-and if-any trade action is tak­en. There are a vari­ety of other doc­u­mented mea­sures adopted by the Chi­nese gov­ern­ment.20…Cur­rently the Chi­nese gov­ern­ment is pro­vid­ing a 14% rebate on VAT to cus­tomers who buy Chi­ne­se-­made semi­con­duc­tor chips, essen­tially pro­vid­ing a large sub­sidy of their domes­tic indus­try in clear vio­la­tion of GATT rules.14 Thus, U.S.-­made chips would pay a 17% VAT, and Chi­ne­se-­made chips would pay a 3% VAT. Given the tight price com­pe­ti­tion of chips and the grow­ing impor­tance of the Chi­nese chip mar­ket, this is a very [im­por­tant] step towards end­ing U.S. pro­duc­tion. [This VAT appar­ently was repealed.]

    …There are a vari­ety of other doc­u­mented mea­sures adopted by the Chi­nese gov­ern­ment.20 The devel­op­ment of spe­cial gov­ern­ment funded indus­trial parks, the low costs of build­ing con­struc­tion in China as com­pared to the U.S., and their appar­ent dis­in­ter­est in the expen­sive pol­lu­tion con­trols required of fab­ri­ca­tion facil­i­ties in the U.S. all rep­re­sent fur­ther hid­den sub­si­dies. The aggre­gate effect of these indi­vid­ual “sub­si­dies” may be only a few tens of per­cent­age points of decrease (lit­er­al­ly, only 20–30%) in the man­u­fac­tur­ing costs of the chips, but in such a cost-­driven indus­try, this dif­fer­ence appears to play an impor­tant role in dri­ving the entire off­shore migra­tion process for these crit­i­cal com­po­nents. Essen­tial­ly, these actions reflect a strate­gic deci­sion and rep­re­sent a con­certed effort by the Chi­nese gov­ern­ment to cap­ture the ben­e­fits of this enabling, high­-tech indus­try, and thereby threat­en­ing to be a monop­oly sup­plier and thus in con­trol of pric­ing and sup­ply.

  32. Brown & Lin­den 2005:

    From 1984 to 2004, the share of semi­con­duc­tor sales in Asia, includ­ing Japan, has risen from 38 to 63% of the world total.10

    …The shift of capac­ity from Japan and the United States to the rest of Asia (pri­mar­ily South Korea and Tai­wan) is strik­ing. Japan and the United States accounted for 80% of fab capac­ity in 1980, but only 49% of capac­ity in 2001…In 2001, approx­i­mately one-third of U.S.-owned capac­ity was located off­shore as shown in Table 5. The off­shore fabs were pri­mar­ily in Japan and Europe, which reflects the rise of joint ven­tures to share risk as the cost of fabs increased. Con­verse­ly, about 22% of the fab capac­ity located in North Amer­ica was owned by com­pa­nies based in other regions (not shown).

    …The prospects vary greatly by the insti­tu­tional envi­ron­ment in each loca­tion. As dis­cussed above, Tai­wan’s fab­less sec­tor, which did not arise as an out­growth of U.S. design off­shoring, is nearly a gen­er­a­tion behind U.S. rivals in terms of inno­v­a­tive prod­ucts. For now, local firms in India have gen­er­ally avoided the fab­less mod­el, but in China there is a small but increas­ing num­ber of fab­less firms tar­get­ing world mar­kets. Although for now Chi­nese firms lack expe­ri­enced engi­neers and man­agers and are be behind Tai­wan in their devel­op­ment of inno­v­a­tive prod­ucts, this will grad­u­ally shift in the years ahead. It is too early to know where this process will end.