SpaceX's second data center brought clusters online for a quarter of typical industry cost, as separate xAI site missed 99.9 percent uptime goal

Two prices sit inside the same build. The company's own boast puts the first two computing clusters of its second data center at roughly a quarter of what the industry typically pays, without saying which expenses that estimate covers ^[3]; The Information's account of how the speed was achieved puts mobile gas turbines, Tesla Megapack batteries and more than 100 portable chillers on the Macrohard site ^[4], and describes Tennessee and Mississippi facilities running for months with no backup cooling or power at all ^[6]. Redundancy is a large share of what a conventional data center dollar buys, so the plainest reading of the discount is that it was financed by the plant that only earns its keep on the bad day, and the internal target quantifies the bad day precisely: 99.9 percent allows 8.76 hours of downtime a year, about 44 minutes a month ^[3], and the site consistently sat under it ^[5].

Now the money the uptime is protecting. Google's reported deal runs about $920m a month for roughly 110,000 Nvidia GPUs ^[13] and Anthropic's Colossus lease as much as $1.25bn a month, cancellable ^[14], which is $2.17bn a month or $6.51bn a quarter ^[1] against the $2.6bn of AI revenue actually reported for the second quarter of 2026 ^[12]. That is 2.5 times the last reported quarter ^[2], so either the contracts bill later than they were signed or the capacity behind those headline numbers is not yet delivered. Annualised, the two named contracts are $26.04bn, leaving under $2bn of the roughly $28bn book for Reflection AI ^[4]; Anthropic alone is about 54 percent of it ^[5], and Anthropic is the one holding the exit.

What this reshuffle is not doing is also visible in the staffing. A Falcon and Dragon senior production manager and a Starbase product-support director have moved onto the data center team ^[10], a Neuralink engineer has been brought in, and the open roles are civil engineering ^[11] - commissioning and construction skills, drawn from programmes that presumably wanted to keep them. Sites beyond the Memphis hub, including proposed Texas locations, have been pushed back ^[7], while June's disclosed 1.4 gigawatts leaves 0.6 gigawatts to add, 43 percent above the current base, to reach more than two gigawatts by the end of 2026 ^[8][7].

This is probably wrong, but the cheap build may still pencil: training runs that checkpoint lose hours to an outage rather than a contract, and $2.6bn in a quarter against $737m a year earlier, up 3.5 times ^[6], is revenue a fully redundant schedule would not have booked at all. I would drop the read entirely if the outages turn out to be commissioning teething, because temporary chillers and turbines can be swapped for permanent plant without repricing the build - the Bastrop foundry for turbine blades and vanes, which Musk says could pull deployment forward by up to 18 months ^[18], is an argument that the fix is procurement rather than architecture, though the NAACP's complaint about unpermitted turbines at Memphis ^[19] is an argument that permanent plant has its own queue. One caution on the record: this account describes Nicolls as running SpaceX's AI data centers in one sentence and overseeing xAI's infrastructure in another ^[1][2], and the equity has already marked something, at $143.36 on September 1, some 36 percent under June's $225.64 peak and 37 percent above August's $104.83 low ^[17][8].