Dr Fazal Ali

Coun­tries that in­vest in build­ing da­ta cen­tres low­er the bar­ri­ers to de­vel­op­ing more of their own AI com­pa­nies. Con­verse­ly, coun­tries that do not in­vest in hy­per­scale AI in­fra­struc­ture will be nowhere in the AI val­ue chain. They will re­main tool im­porters, and their work­force will stag­nate as users of AI as­sem­blages and as par­tic­i­pants in da­ta-dri­ven in­no­va­tion in the new Dig­i­tal Cre­ative In­dus­tries.

Land, po­lit­i­cal sta­bil­i­ty, re­la­tion­ships with mem­bers of the Five Eyes Se­cu­ri­ty Al­liance, which par­tic­i­pates in joint sig­nals in­tel­li­gence (SIG­INT) col­lec­tion and shar­ing, a skilled work­force, and a sta­ble en­er­gy grid are crit­i­cal fac­tors in at­tract­ing in­vest­ment. Screen economies and dig­i­tal cre­ative in­dus­tries that in­no­vate with da­ta are now close­ly linked to in­vest­ment in AI ar­chi­tec­tures and Q-com­merce.

How­ev­er, da­ta cen­tres need un­in­ter­rupt­ed pow­er at a scale and speed that wind and so­lar can­not match. This has sparked dis­cus­sions about or­bital com­pute in­fra­struc­ture and a com­ing wave of space-based da­ta cen­tres, mov­ing from con­cept to ear­ly de­ploy­ment, fu­elled by ve­loc­i­ty and ven­ture cap­i­tal. This will re­quire pack­ing light­weight servers in­to space-qual­i­fied mod­ules pow­ered by bat­ter­ies and so­lar en­er­gy and us­ing high-band­width com­mu­ni­ca­tions back to Earth.

Gem­i­ni has al­ready pi­lot­ed a one-kilo­watt (kW) satel­lite with five em­bed­ded GPUs. This ear­ly space plat­form car­ried da­ta-cen­tre-grade com­pute in­to or­bit, en­abling Gem­i­ni to run in­fer­ence. How­ev­er, call­ing it a da­ta cen­tre is mis­lead­ing. It is an ear­ly hu­man in­stance of putting one of the first satel­lites that car­ry da­ta-cen­tre-grade GPUs in­to space.

Flocks of or­bital, space-based AI in­fra­struc­ture of­fer sev­er­al struc­tur­al ad­van­tages. These in­clude un­con­strained en­er­gy scal­ing and cost com­pet­i­tive­ness rel­a­tive to ter­res­tri­al sys­tems. We must now as­sess whether space-based com­pute can de­liv­er pre­dictable per­for­mance, re­peat­able de­ploy­ment, cred­i­ble re­li­a­bil­i­ty, and com­pet­i­tive eco­nom­ics, even af­ter ac­count­ing for launch ca­dence, re­place­ment ro­ta­tions, in-space cool­ing, and da­ta move­ment costs.

Da­ta cen­tres in­volv­ing satel­lites or­bit­ing Earth are there­fore not a fan­ci­ful idea. We are al­ready in that fu­ture. The next gen­er­a­tion of da­ta cen­tres will car­ry a ten-kW rack-scale sys­tem in­to space, with mul­ti­ple ad­vanced AI chips and more ro­bust cool­ing in­fra­struc­ture.

That star-date is the point from which AI be­he­moths will be able to have part of their ter­res­tri­al in­fra­struc­ture or­bit­ing the plan­et in a flock. The re­al step-change, how­ev­er, is planned for 2028, when a three-ton, 200kW sys­tem de­signed to fit in­to Star­ship’s de­ploy­ment launch for­mat is de­ployed.

Af­ter that, re­gard­less of any fail­ures or de­lays that may push the launch date to 2030, we will have ef­fec­tive­ly suc­ceed­ed in de­ploy­ing a space-based da­ta cen­tre ca­pa­ble of han­dling large-scale in­fer­ence work­loads.

A flock of 88,000 satel­lites or­bit­ing the Earth with these GPUs will de­liv­er about 20 gi­gawatts of com­pute ca­pac­i­ty, ded­i­cat­ed to in­fer­ence work­flows. At the heart of this hu­man am­bi­tion is the no­tion of “in­tel­li­gent fail­ure.”

Elon Musk has re­de­fined “fail­ure” be­fore our very eyes. Fail­ure be­gins with de­sign­ing smart pi­lots that keep fail­ures slight. It is a cul­ture alien to West In­di­an civil­i­sa­tion. We must learn to fail small and re­cov­er fast.

We must not run from the end­less op­por­tu­ni­ties that fall at our feet; in­stead, we must fail small, adapt, it­er­ate, and re­cov­er with speed. There is no Big Bang! On­ly the next re­hearsal. In­tel­li­gent fail­ure turns on lim­it­less naivete and an ab­sence of ar­ti­fi­cial­i­ty.

Such fail­ure is al­ways in­tel­li­gent be­cause there is no for­mu­la to pre­cise­ly map the ter­rain in ad­vance. Fail­ure is bet­ter than nev­er try­ing. To do noth­ing is to nev­er fail. Nev­er fail­ing is do­ing noth­ing. This is an un­known cul­tur­al oa­sis for for­mer plan­ta­tion so­ci­eties that con­tin­ue to lean on the pil­lars of Ward Schools that cre­at­ed no space for fail­ure.

Da­ta cen­tre con­struc­tion on Earth faces one prin­ci­pal con­straint. En­er­gy! We are un­able to build the re­quired en­er­gy in­fra­struc­ture fast enough. In space, the con­straints and bot­tle­necks around al­low the quick build­out of large-scale so­lar in­stal­la­tions dis­ap­pear.

While da­ta cen­tres that run on so­lar pow­er need bat­tery stor­age or oth­er back­up pow­er for night­time use, so­lar gen­er­a­tion can be near con­tin­u­ous in the right or­bit. Or­bital da­ta cen­tres will use a dawn-dusk sun-syn­chro­nous or­bit that fol­lows the Earth’s ter­mi­na­tor line and will re­main in con­tin­u­ous sun­light.

This fu­ture is be­ing de­signed on the pos­si­bil­i­ty that launch costs will de­cline ma­te­ri­al­ly to about US$500 per kilo­gram. At that turn­ing point, or­bital com­pute will be eco­nom­i­cal­ly com­pet­i­tive with ter­res­tri­al al­ter­na­tives. Bring­ing in­fra­struc­ture costs in space down to about US$5M per megawatt, com­pared with US$12M to US$15M for ter­res­tri­al sys­tems, is an­oth­er tar­get.

Or­bital AI-As­sem­blages will be a fer­tile field for hy­per­scalers and emerg­ing AI-fo­cused neo­cloud providers. De­mand for or­bital com­pute is al­so be­ing dri­ven by the in­creased dif­fi­cul­ty of find­ing new en­er­gy sources for da­ta cen­tres. Ter­res­tri­al da­ta cen­tres near con­flict zones are vul­ner­a­ble to per­sis­tent drone at­tacks on hy­per­scale AI cam­pus­es and en­er­gy in­fra­struc­ture.

Dr Fazal Ali com­plet­ed his Mas­ter's in Phi­los­o­phy at the Uni­ver­si­ty of the West In­dies. He was a Com­mon­wealth Schol­ar who at­tend­ed the Uni­ver­si­ty of Cam­bridge, Hugh­es Hall; the Provost of the Uni­ver­si­ty of Trinidad and To­ba­go; the act­ing Pres­i­dent of UTT; and the Chair­man of the Teach­ing Ser­vice Com­mis­sion. He is the Pres­i­dent of NI­HERST and an ex­ter­nal ser­vices con­sul­tant with the IDB.