--- title: "Cluster: what it means · The AI Compute Race · DRYBURGH" source: https://dryburgh.com/vol1/glossary/cluster description: "Cluster: A datacenter's worth of chips wired to work as one machine. Chapter 1." --- # Cluster What does “Cluster” mean? ClusterA datacenter’s worth of chips wired to work as one machine. *Chapter 1.* From *The AI Compute Race*. [Read the book](https://dryburgh.com/vol1#buy) ## Where the book uses it - [Chapter 1 — The Stakes](https://dryburgh.com/vol1/notes/chapter-1)What the race is for: the belief that a lead in AI can compound into a lasting advantage, how long such leads have actually lasted, and what a trillion dollars a year is being bet against. ## Where the record uses it - Campus [Colossus 2](https://dryburgh.com/vol1/ledger/colossus-2) - Campus [Stargate UAE](https://dryburgh.com/vol1/ledger/stargate-uae) - Scorecard indicator [Largest evidenced operating site](https://dryburgh.com/vol1/scorecard#row-11) ## Sources that use it - Patrick Kennedy, ServeTheHome, “Inside the 100K GPU xAI Colossus Cluster that Supermicro Helped Build for Elon Musk,” 28 October 2024, reprinted by Supermicro as a case study, December 2024 (64 H100s per rack, eight racks per 512-GPU block, one coolant loop and CDU per rack, nine 400 GbE links per server; cabling crews and Megapacks observed on site; “the initial cluster of four 25,000 GPU data halls is up and running for around 100,000 GPUs at the time of our visit”; on the power work seen on the visit: “huge power cables”; on the fabric: “Many supercomputer networks use InfiniBand or other technologies, but this is Ethernet. Ethernet means it can scale.”; on the Megapacks: “the millisecond spikes and drops in power were stressful enough that putting the Tesla Megapacks in the middle to help buffer those spikes in power helped make the entire installation more reliable”; sponsored by Supermicro, the article says, and made with “special approval by Elon Musk and his team”) [In the Sources](https://dryburgh.com/vol1/sources/company#supermicro-colossus-case) - Leopold Aschenbrenner, *Situational Awareness: The Decade Ahead*, June 2024 (chapter IIIa, “Racing to the Trillion-Dollar Cluster”, the training-compute table’s ladder: about 1 GW in 2026, 10 GW in 2028 and 100 GW in 2030, each marked approximate; the same chapter: “total AI investment could be north of $1T annually by 2027”; in the same chapter under “AI revenue”, p. 81 of the PDF, the milestone is when “a big tech company (Google, Microsoft, Meta, etc.)” will “hit a $100B revenue run rate from AI”, and the June 2024 route to it is built on a seat: “there are around 350 million paid subscribers to Microsoft Office—could you get a third of these to be willing to pay $100/month for an AI add-on?”, with a footnote’s own caveat that “intermediate, pre-AGI models take a lot of ‘schlep’ to properly integrate into company workflows”; chapter I, “From GPT-4 to AGI: Counting the OOMs”, opens: “AGI by 2027 is strikingly plausible”; the PDF at the first address runs to 165 pages; read from the house’s full-site archive of the essay (Library S130), the chapter page held in the cite-check archive and the PDF held beside it; “$1T/year of total AI investment by 2027 would be dramatic—among the very largest capital buildouts ever—but would not be unprecedented”, in the same chapter; on power, in chapter IIIa: “An H100 is 700W”, and “let’s work with 1kW per H100-equivalent here”; chapter II, “From AGI to Superintelligence: the Intelligence Explosion”: “Automated AI research could probably compress a human-decade of algorithmic progress into less than a year (and that seems conservative). That’d be 5+ OOMs”; chapter I, “From chatbot to agent-coworker”: “By 2027, rather than a chatbot, you’re going to have something that looks more like an agent, like a coworker”; chapter II, under “Automating AI research”: “We don’t need to automate everything—just AI research”; chapter IIIa opens: “the race to AGI won’t just play out in code and behind laptops”, and “AI is a massive industrial process: each new model requires a giant new cluster, soon giant new power plants, and eventually giant new chip fabs”). The full PDF is at the first address, chapter I’s page at the second. [In the Sources](https://dryburgh.com/vol1/sources/other#sa-essay) - Mackenzie Hawkins and Haze Fan, “Moonshot’s Kimi Uses 20,000 Nvidia Chip Cluster From Alibaba,” Bloomberg, 31 July 2026, updated the same day (on people with knowledge of the companies’ operations: Moonshot “has a computing power agreement with Alibaba Group Holding Ltd. for the use of around 20,000 Nvidia Corp. chips”; “That hardware forms a substantial chunk of the overall computing capacity Moonshot uses for its Kimi models”; the chips “are from Nvidia’s earlier generation of Hopper products”; “Multiple people close to Moonshot said the agreement is specifically for H200 chips”, which Alibaba’s spokesperson called “completely groundless” while not denying the 20,000) [In the Sources](https://dryburgh.com/vol1/sources/press#bloomberg-moonshot-alibaba-jul26) - Ellie Holbrook, Robert Boswall, Jeremie Eliahou Ontiveros, Nicolas Bontigui and Dylan Patel, “What is So Hard About Behind-The-Meter Power For Datacenters? Part 1,” *SemiAnalysis*, 10 September 2026 (the free portion of a paid post, 7,872 of 9,577 words by the publisher’s count; tallies of the firm’s own energy model: “Our Energy Model now tracks 75GW of firm, binding orders in the supply chain only for for behind-the-meter AI compute - of which ~20GW alone ordered in Q2 2026”, the doubled word the post’s; “by the end of the year, ~3GW of operational US datacenter IT capacity will be powered behind-the-meter”; “Oracle’s 1.3GW IT Project Jupiter”; the chart beneath the 75GW sentence, read as an image, titled “Datacenters Have Booked About 75 GW of Onsite Generation” and subtitled “Generation booked for behind-the-meter datacenter power, US, GW nameplate, cumulative by announcement quarter, batteries excluded”, from the model’s “edition 24 Aug 2026”, with no figure for any quarter but the last; of orders without a home, buyers who “pay deposits, yet don’t have actual datacenter sites to deploy them”; and later in the post, of the 3 GW, “3GW of Datacenter IT capacity is already running with behind-the-meter power plants”; the post’s chart of orders by quarter, read as an image, subtitled “Quarterly order capacity since 1Q25 · stripes mark capacity without a downstream project”, prints 0.4, 6.3, 5.7, 7.8, 19.3 and 22.8 GW over the quarters from 1Q25 to 2Q26 and “11.8 QTD” for the third quarter to 24 August, a box reading “74.2 GW SINCE 1Q25”, “70% PROJECT-LINKED”, “30% STILL UNALLOCATED”, and the footnote “All tracker statuses, including Announced; dated by stored deal-announcement quarter”; the chart under its two 3 GW sentences, subtitled “Quarter-end operational US IT capacity by onsite power technology”, prints “2.9 GW” over 4Q26, a box reading “0.3 → 2.9 GW” and “9.0× BY Q4 2026”, and the footnote “Q3 2026 onward reflects the Energy Model outlook.”; both images held in the Library since 12 October 2026 and read again on 13 October 2026; the chart beneath the “$100B per GW per year” sentence, read as an image, titled “On a GB300 Cluster, Revenue Per Gigawatt Climbs 8× Up the AI Stack”, measured in “$B per GW per year” and sourced to the “SemiAnalysis Tokenomics Model, SemiAnalysis Inference Simulator”, with five bars: “Neocloud 5-yr IaaS average” (“typical Oracle, CoreWeave & Nebius pricing”) $12B at 1.0×, “B300 on-demand” $29B, “SpaceX + Anthropic” $31B and “SpaceX + Google” $48B both marked “deal”, and “OpenAI / Anthropic inference API” $100B at 8.3×, the chart dating its prices nowhere) (SemiAnalysis sells subscriptions, consulting and its own models of chip, memory and datacenter supply into the markets it measures; its closing section is for paid subscribers and was not read; held in the Library since its capture of 5 October 2026). [In the Sources](https://dryburgh.com/vol1/sources/research#semianalysis-btm-sep26) - SemiAnalysis, “The Great GPU Shortage – Rental Capacity – Launching our H100 1 Year Rental Price Index,” 2 April 2026 (a representative one-year H100 price rose almost 40 percent from a low of $1.70 an hour to $2.35 by March 2026; on-demand sold out across all GPU types; capacity due online through August–September 2026 already booked; customers renewing older H100 capacity at original two- or three-year rates; “H100s are getting renewed at the exact same rate they were signed at 2-3 years ago and some H100 contracts are being renewed for 4 years though 2028”; on hyperscalers that “backstop these deals”: “Hyperscalers can collect a slice of the project revenue by offering the backstop of their balance sheet without actually expanding the balance sheet”; of its own use: “SemiAnalysis as a company has, over the past 7 days, consumed billions of tokens costing around ~$5/M tok on average”; on who renews, as a kind: “most of the large AI Clusters that are ramping are captive to AI Labs”, which “are also preventing supply from re-entering the sub 4y contract market by simply renewing their existing H100 and H200 clusters”). [In the Sources](https://dryburgh.com/vol1/sources/research#semianalysis-shortage-apr26) ## Related terms - [One machine](https://dryburgh.com/vol1/glossary/one-machine) - [122 days](https://dryburgh.com/vol1/glossary/122-days) - [MLGW](https://dryburgh.com/vol1/glossary/mlgw) - [Training](https://dryburgh.com/vol1/glossary/training) - [TVA](https://dryburgh.com/vol1/glossary/tva) - [Algorithmic efficiency](https://dryburgh.com/vol1/glossary/algorithmic-efficiency) - [Application layer](https://dryburgh.com/vol1/glossary/application-layer) - [Arsenal behind every arsenal](https://dryburgh.com/vol1/glossary/arsenal-behind-every-arsenal) - [Capital expenditure](https://dryburgh.com/vol1/glossary/capital-expenditure) - [Clocks of diffusion](https://dryburgh.com/vol1/glossary/clocks-of-diffusion) - [Epoch AI](https://dryburgh.com/vol1/glossary/epoch-ai) - [Exclusive generation](https://dryburgh.com/vol1/glossary/exclusive-generation) From the glossary of The AI Compute Race. - [Campuses](https://dryburgh.com/vol1/ledger) - [Tests](https://dryburgh.com/vol1/tests) - [Scorecard](https://dryburgh.com/vol1/scorecard) - [Data](https://dryburgh.com/vol1/data/releases) - [Chapter by chapter](https://dryburgh.com/vol1/companion) - [The record at a glance](https://dryburgh.com/vol1/observatory) - [The machine](https://dryburgh.com/vol1/machine) - [One campus, four futures](https://dryburgh.com/vol1/hyperion) - [Glossary](https://dryburgh.com/vol1/glossary) - [People](https://dryburgh.com/vol1/people) - [Sources](https://dryburgh.com/vol1/sources) - [Notes](https://dryburgh.com/vol1/notes) - [Where the book was wrong](https://dryburgh.com/vol1/corrections) - [Where its sources were wrong](https://dryburgh.com/corrections?book=vol1) - [The Paper](https://dryburgh.com/vol1/paper) - [Consent](https://dryburgh.com/vol1/consent) - [Who builds and pays](https://dryburgh.com/vol1/parties) *The AI Compute Race: Five Races, One Machine* is a book by Lee S. Dryburgh, the first volume of THE COMPUTE ECONOMY, published by Dryburgh, an imprint of Contestra OÜ, Tallinn, Estonia; its record closes on 8 November 2026. [About the book](https://dryburgh.com/vol1)[Tell me when it is out](https://dryburgh.com/vol1#notify)