GB300

Where the book uses it

• Chapter 4 — One Rack

  How a datacenter is actually built, from a converted factory in Memphis to a greenfield campus in Texas, and what the clock from first ground to first power really measures.

Sources that use it

• Brad Smith, “Microsoft’s $15.2 billion USD investment in the UAE,” Microsoft On the Issues, 3 November 2025 (export licenses “Approved in September” 2025 “enable us to ship the equivalent of 60,400 additional A100 chips, in this instance involving Nvidia’s even more advanced GB300 GPUs,” on top of “the equivalent of 21,500 Nvidia A100 GPUs” already accumulated in the country; $15.2 billion committed 2023–2029). In the Sources
• Meta Engineering, “Meta’s infrastructure evolution and the advent of AI,” 29 September 2025 (Hyperion to come online beginning in 2028 and scale to 5 GW; the Catalina reference design of 72 Blackwell GPUs at about 140 kW in the accelerator portion of a six-rack pod; fleet deploying GB200, beginning GB300, also AMD MI300 and MTIA; “Once finished, the Hyperion cluster will have the ability to scale up to a capacity of 5 gigawatts”; “Prometheus, our 1-gigawatt cluster, is currently underway”; “We accomplished this by building this cluster across several of our traditional data center buildings as well as several weatherproof tents, and adjacent colocation facilities”; “We are also evolving our software stack, including Twine and MAST, to support long-distance training across a geographically distributed set of data centers”). In the Sources
• NVIDIA, “GB300 NVL72” product page and NVL72 AI-factory reference architecture, components page (“Full rack requiring up to 142 kW”; 72 Blackwell Ultra GPUs; liquid-cooled; each GPU with eighteen fifth-generation NVLink links, “one per in-rack NVSwitch via the copper backplane,” so that “a single rack may form a fully connected L1 Domain with 72 GPUs,” read 20 September 2026), accessed 29 August and 4 September 2026. In the Sources
• SpaceXAI, “Grok 4.5,” 16 July 2026 (the page titled “Introducing Grok 4.5 | SpaceXAI”, describing it as “SpaceXAI’s smartest model”, as held 19 September 2026; a model “trained across tens of thousands of NVIDIA GB300 GPUs”, the page giving no parameter count; the Pricing section: “Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens. The model also achieves roughly 2x the token efficiency of comparable leading models, solving tasks in under half the number of steps”; the same page claims “twice greater token efficiency than the latest leading models at the same tasks” and, in a chart caption, that Grok 4.5 “resolves tasks with 15,954 output tokens on average, about 4.2× fewer than Opus 4.8 (max) at 67,020” on SWE Bench Pro, so the page states the claim three times against three denominators; the page dates itself “Jul 16, 2026”; read live 19 September 2026, unchanged). In the Sources
• IREN Limited, Form 10-Q filed 5 February 2026 (accession 0001878848-26-000015), Exhibit 10.3, Microsoft Partner Statement of Work with IE US Hardware 3 Inc., signed by Scott Guthrie, 31 October 2025 (total contract value $9,666,845,337.60; GB300s at Childress, Texas; five-year term; countersigned for the supplier by Kent Draper on 2 November 2025; §2.2, “Microsoft will not have any obligation to exercise this option”; §2.2: “Supplier shall use good faith efforts to procure” a “side letter or other agreement” under which, “upon the occurrence of the insolvency” of the supplier, its financiers and colocation provider “will agree to give Microsoft the option to continue to meet Supplier’s payment obligations under the financing documents or colocation services agreement”), §2.2 “Step-in right”. In the Sources

And 12 more: search the Sources.

Related terms

• GB200 NVL72
• Blackwell
• Rack
• 122 days
• Abilene
• Accelerator
• Colossus
• CPU
• GPU
• Greenfield
• H100
• Lien