🚀 How Colossus Flips the Script on Modern Data Centers 🏭
Colossus turns the “data center” into a compute factory by rewriting almost every rule underpinning modern hyperscale campuses. Here's what makes the Memphis-based super-computer so disruptive.
India’s cloud and AI economy is at an inflection point. If you're in the data, cloud, or AI space, you’ve felt the shift.
I’m Gaurav Jeyaraman, and I’ve curated this newsletter as your front-row seat to India’s digital infrastructure boom. Timely, sharp, and built for those who want the inside track.
🎯 🚨 The Big Story 🚨 🎯
Everyone’s talking about how Grok 4 is smashing records on benchmark tests. Here, we break down how a new-age data center in Memphis is the actual driving force. But first, some context:
🚀 Context: Grok 4’s Leap 🚀
xAI launched Grok-1 in November 2023 and pushed four full releases in just 20 months, landing Grok 4 this July!
Humanity’s Last Exam (HLE): Grok 4 set a new record 50.7 % with tools (26.9 % text-only), outscoring Gemini 2.5 Pro (21.6 %) and GPT-4 (o3) (~21 %). Its multi-agent variant, “Grok 4 Heavy,” pushes the tool-enabled score to 44.4%.
GPQA (graduate-level physics): Grok 4 posted 87.5 %, edging Gemini 2.5 Pro (86.4 %), GPT-4 (o3) (83.3 %), and Claude Opus (79.6 %).
LMArena crowdsourced tests: Across 4k+ community prompts, Grok 4 now ranks #1 in Math, #2 in Coding, #3 on “Hard Prompts,” vaulting from 8th to 3rd overall.
🧮 Why?…Chips > Chatter 🗣️
Here, Canadian-American venture capitalist Chamath Palihapitiya, speaking on the All-In Podcast, explains why Musk’s 200,000+ GPU cluster is the key.
As Booz Allen's Bassel Haider puts it, "Compute is no longer a resource you rent by the hour; it is a sovereign asset you construct, cool, and guard like a refinery."
🤖 From Factory Floor to AI-Core in just 19 days 🏭
Colossus redefines the data center as an AI factory, offering a roadmap to industrialize intelligence across infrastructure, investment, and oversight, leaving the competition eating dust.
📍 Dec 19, 2024, Memphis, TN: An unused Electrolux plant transforms into “Colossus,” a 100 MW, 100,000 GPU AI factory in just 19 days.
📍 July 10, 2025, xAI HQ: Elon Musk’s team unveils Grok 4, while teasing a 100,000-GPU upgrade using NVIDIA’s new GB200 NVL72 racks.
Why this matters: The cluster was stood up in just 122 days; the GPU fleet itself was cabled and live in 19 days, versus the 3-4 years a conventional DC-build normally takes.
The speed came from repurposing a 785,000 sq ft Electrolux factory, showing how brown-field industrial conversions can significantly out-pace green-field shells while slashing embodied carbon.
📍 Location, location, location: With 100 MW of instantly available power from TVA, direct access to Mississippi River water for low-cost cooling, and FedEx’s global superhub just 20 minutes away, Memphis offered unmatched advantages in power, efficiency, and logistics.
Some cool features include:
100,000 NVIDIA H100 GPUs in Phase 1 (doubling to 200,000 GPUs this year), packing roughly 1 exaflop of AI compute into a single building, dwarfing today’s Top-500 HPC machines.
Each rack holds 64 GPUs and draws well over 120 kW, an order of magnitude denser than typical cloud racks (10-15 kW), forcing a wholesale re-think of power, cooling, and floor loading.
Colossus is the world’s largest direct-to-chip liquid-cooled AI cluster, using servers, CDUs, and rack manifolds that were designed for liquid first, removing fans and enabling tool-less, quick-disconnect service. This doubles compute density and reportedly cuts electrical overhead by up to 40%, pushing PUE close to 1.05.
Racks are grouped into 512-GPU pods linked optically; pods stitch together over the leaf–spine Ethernet fabric. The design lets xAI lift in additional pods without touching the global topology, a model others are already copying.
Instead of the HPC-standard InfiniBand, Colossus runs an NVIDIA Spectrum-X 400 GbE fabric with a dedicated NIC per GPU, plus 400 GbE for the host CPU, delivering ~3.6 TB/s per server. This signals that merchant Ethernet can now meet exascale-AI latency demands, unlocking larger vendor ecosystems and driving costs down for future builds.
Unlike colocation halls that chase many tenants, Colossus is purpose-built for one workload: training Grok and future xAI models (plus overflow from Musk’s other companies). Compute, network, storage and power are tuned to that single objective, creating an “AI factory” blueprint others (OpenAI–Microsoft, Meta, Google) are now emulating.
Bottom line: Colossus isn’t just a bigger data center, it’s proof that density, speed, and AI-specific design can crush the legacy cost-and-time curve. Whether operators embrace or reject its environmental trade-offs, the project sets a new baseline that future GPU campuses will be measured against.
Expect at least five G20 nations to break ground on Colossus‑equivalent clusters by 2027, citing national AI sovereignty - Bassel Haider, Data & AI, Booz Allen
🎥 Clip du Jour 🎥
David Sacks on the race for AI: Will the US technology stack, become akin to the US Dollar as the world’s reserve?
💰 Big Bets 💰
A look at major investments—deals, capex announcements, hyperscaler expansions, sovereign cloud plays.
💼 NTT DATA has announced a major $1.5 billion investment in India to add 400 MW of data centre capacity over the next three years.
🏢 Singapore has marked its largest data centre REIT listing in eight years with the debut of NTT DC REIT.
☁️ Zilliz, the company behind popular open-source vector database tech, is expanding its cloud footprint with a new Azure region in Central India to better serve customers in the country.
⚡ Google has unveiled a $25 billion investment to expand data centers and AI infrastructure across the U.S.’s largest electric grid, aiming to meet soaring AI demand.
🌐 Singapore-based CapitaLand Group has announced a major push to become one of India’s top three data centre operators by 2027–28, earmarking nearly $1 billion for expansion.
🏗️ Blackstone has announced a $25 billion investment to build data centers and natural gas power plants in Pennsylvania.
💧 Google has signed a landmark $3 billion deal with Brookfield Asset Management to secure up to 3 GW of hydropower for its U.S. data center operations.
🧠 Meta CEO Mark Zuckerberg has announced plans to invest hundreds of billions of dollars into building massive AI data centers, aiming to accelerate the race toward artificial superintelligence.
🗞️ ICYMI 🗞️
📊 Data center capacity in India is projected to reach 3 GW by 2030, driven by rapid digital adoption, supportive government policies, and a thriving investment climate.
📍 The Mumbai Metropolitan Region MMR led India's land market in the first half of 2025, with Bollywood-linked deals and data centre investments drawing significant attention.
🌊 Amogy has raised $80 million to advance its ammonia-to-power tech for ships and data centers, a major leap in clean energy innovation.
📈 Data center bandwidth has surged by 330%, driven by explosive AI workload growth and massive investments in next-gen infrastructure.
🚨 A minor theft was reported at a closed data center near Raebareli, Uttar Pradesh, with no critical infrastructure impacted.
🗣️ Voices From the Ground 🗣️
🎥 Interview: CtrlS Chairman Sridhar Pinnapureddy, speaking to DCD, discusses India’s booming data center market and the company’s regional expansion plans.
📖 Interview: AWS’s Satinder Pal Singh on why GenAI can be a game changer for India.