The Memphis Cluster: The Socio-Technical Cost of the xAI Colossus
- Bryan White
- Jan 17
- 17 min read

I. Introduction: The Era of the Gigafactory
In the summer of 2024, the global race for Artificial General Intelligence (AGI) materialized physically on the banks of the Mississippi River. In an industrial corridor of Southwest Memphis, Tennessee, historically defined by heavy manufacturing and logistics, a new kind of industrial behemoth emerged with unprecedented speed. This facility, known as "Colossus," represents the flagship infrastructure of xAI, the artificial intelligence company founded by Elon Musk. Retrofitted into a former Electrolux manufacturing plant in a record-breaking 122 days, Colossus was designed to be the largest single-site AI training cluster in the world.1
The facility’s purpose is singular and demanding: to train the "Grok" series of Large Language Models (LLMs). These models, characterized by parameter counts numbering in the trillions, require computational throughput that dwarfs the capabilities of traditional academic or governmental supercomputers. By January 2026, the project had expanded from a singular high-performance cluster into a sprawling campus straddling the Tennessee-Mississippi border, comprising Colossus 1, Colossus 2, and a third facility provocatively named "MACROHARDRR".1 Together, these facilities target a total power capacity of nearly two gigawatts—roughly equivalent to the electrical consumption of two million homes—to power a fleet of over 555,000 Graphics Processing Units (GPUs).1
This report offers a deep-dive analysis of the Colossus project, dissecting it not merely as a feat of computational engineering, but as a complex socio-technical system. We will explore the architectural evolution of its silicon substrate, the thermodynamic necessities driving its transition to liquid cooling, the atmospheric chemistry of its controversial power generation, and the regulatory collision that culminated in a landmark Environmental Protection Agency (EPA) ruling on January 15, 2026.4 Through this lens, Colossus serves as a case study for the friction between the digital ambitions of the twenty-first century and the physical, legal, and ecological realities of the American South.
II. The Computational Substrate: Silicon, Architecture, and the Blackwell Transition
To understand the environmental and infrastructural footprint of Colossus, one must first comprehend the machine's anatomy. The training of modern foundational models is a problem of massive parallelism, requiring tens of thousands of processors to operate in a synchronized state, performing matrix multiplications at speeds measured in exaflops (quintillions of floating-point operations per second).
A. The H100 Hopper Architecture
The initial phase of Colossus, known as Colossus 1, was built upon the NVIDIA H100 "Hopper" Tensor Core GPU.5 This processor represented a paradigm shift in high-performance computing (HPC) specifically tailored for the "Transformer" neural network architecture that underpins models like Grok.
The H100 introduced the "Transformer Engine," a specialized hardware component designed to dynamically manage numerical precision. In deep learning training, not all calculations require high-precision 32-bit floating-point numbers (FP32). The Transformer Engine automatically switches between 16-bit (FP16/BF16) and 8-bit (FP8) formats on a layer-by-layer basis. By processing data in 8-bit formats where possible, the chip essentially doubles its throughput and halves its memory footprint without degrading the accuracy of the final model.6
For Colossus 1, xAI deployed approximately 100,000 of these units.8 Each H100 GPU draws up to 700 watts of power.6 When aggregated, the thermal and electrical load of the silicon alone—excluding cooling and networking overhead—reached tens of megawatts immediately upon activation.
B. The Transition to Blackwell and "MACROHARDRR"
As the project expanded into 2025 and 2026, the hardware strategy shifted. The training of Grok 3 and Grok 4 demanded even greater memory bandwidth and compute density. Consequently, Colossus 2 and the MACROHARDRR facility were designed to house the NVIDIA Blackwell GB200 platform.1
The Blackwell architecture differs fundamentally from the Hopper generation in its integration. The GB200 is not just a GPU; it is a "superchip" that couples two Blackwell GPUs with a Grace CPU on a single board, connected by a high-speed interconnect known as NVLink-C2C. This architecture addresses a critical bottleneck in AI training: the speed at which data moves between the processor and memory.
The Blackwell GPUs utilize HBM3e (High Bandwidth Memory, Enhanced), which delivers data transfer rates approaching 8 terabytes per second. For the MACROHARDRR expansion, xAI is deploying the GB200-NVL72, a rack-scale design where 72 GPUs function as a single, coherent accelerator. In this configuration, the boundaries between individual servers dissolve, and the entire rack operates as a massive, unified graphics card with petabytes of shared memory.10 This density, however, creates a thermal flux so intense that traditional air cooling becomes physically impossible, necessitating the liquid cooling solutions discussed in Section IV.
C. Scale and Throughput
By January 2026, the Colossus complex housed a heterogeneous mix of hardware:
Legacy Cluster: ~100,000 H100/H200 GPUs.
Blackwell Expansion: ~455,000 GB200/GB300 GPUs.1
The theoretical peak performance of this aggregate system is difficult to quantify using standard metrics, but estimates suggest it exceeds several hundred exaflops of AI compute (FP8). This scale allows xAI to reduce the training time of foundational models from months to weeks, a critical advantage in the highly competitive AI sector.
III. The Nervous System: Spectrum-X, RDMA, and the Ethernet Renaissance
A supercomputer is defined less by its raw processing power than by its ability to communicate. If the GPUs are the brain's neurons, the network is the synaptic web connecting them. In traditional supercomputing, this web has almost exclusively been built using InfiniBand, a specialized, low-latency networking standard. However, xAI made a divergent engineering choice for Colossus, opting for an Ethernet-based solution using the NVIDIA Spectrum-X platform.8
A. The "Elephant Flow" Problem
AI workloads generate distinct traffic patterns compared to cloud computing. Cloud traffic typically consists of millions of small, independent data flows (mice flows). AI training, conversely, involves "elephant flows"—massive, sustained bursts of data occurring simultaneously across thousands of GPUs during the synchronization phase of training.
In a standard Ethernet network, these elephant flows can cause catastrophic congestion. If two massive data streams attempt to occupy the same buffer space on a switch, packets are dropped. In AI training, a single dropped packet can stall the entire cluster, as thousands of GPUs wait for the missing data to be re-transmitted. This phenomenon, known as "tail latency," can reduce the effective efficiency of a multi-billion dollar cluster to a fraction of its theoretical peak.12
B. Spectrum-X and Adaptive Routing
To solve this without abandoning the ubiquity and cost-effectiveness of Ethernet, Colossus utilizes the Spectrum-X networking platform. This system introduces technologies previously reserved for InfiniBand into the Ethernet protocol.
The core innovation is Adaptive Routing. In traditional Ethernet (ECMP), a data flow is assigned a single path through the network topology. If that path is congested, the data waits, even if other paths are empty. Spectrum-X enables packet-by-packet dynamic routing, spraying data across all available paths and reassembling it at the destination. This ensures that the network fabric is utilized evenly, preventing "hot spots" and flow collisions.8
C. 1:1 Networking and RDMA
The architecture of Colossus features a 1:1 networking ratio, meaning for every GPU, there is a dedicated 400Gbps or 800Gbps network interface card (NIC).13 This massive bandwidth is essential for Remote Direct Memory Access (RDMA).
RDMA allows a GPU in one rack to write data directly into the memory of a GPU in another rack without involving the operating system or the CPU of either server. This "zero-copy" networking reduces latency to microseconds. By implementing RDMA over Converged Ethernet (RoCE) with the Spectrum-X congestion control, xAI achieved a network fabric that behaves like a lossless supercomputer interconnect while retaining the scalability of Ethernet. This architectural choice was pivotal in allowing the cluster to scale to 100,000+ nodes, a scale where traditional InfiniBand topologies often face management complexity limits.12
IV. Thermodynamic Engineering: Breaking the Thermal Wall with Direct-to-Chip Liquid Cooling
The defining physical constraint of the post-H100 era is heat. As transistor density increases, the thermal flux (heat energy generated per unit of surface area) of the silicon package has surpassed the cooling capacity of air. The Colossus project serves as a premier industrial example of the phase transition from air cooling to liquid cooling in hyperscale data centers.
A. The Physics of the Thermal Wall
Air is a thermally inefficient medium. It has a low specific heat capacity (approximately 1.005 J/g°C) and poor thermal conductivity. In a traditional data center, cold air is forced through the server chassis to convect heat away from heatsinks. However, as rack power densities approach 50 kilowatts (kW) and beyond, the volume of air required to maintain safe operating temperatures becomes unmanageable. The velocity of the air would need to be so high that the fans themselves would consume a significant percentage of the rack's power budget and generate deafening acoustic noise.14
The NVIDIA Blackwell racks used in Colossus 2 can exceed 100 kW per rack. Attempting to cool this density with air is akin to trying to cool a nuclear reactor core with a desk fan; the physics simply do not allow for the necessary rate of heat rejection.
B. Direct-to-Chip (DLC) Liquid Cooling Mechanics
To breach this thermal wall, xAI partnered with Supermicro to implement a massive Direct-to-Chip (DLC) liquid cooling architecture.2
In this system, air heatsinks are replaced by "cold plates"—blocks of copper or aluminum mounted directly onto the GPUs and CPUs. These cold plates contain internal microchannels, often machined to tolerances of less than 100 microns, which maximize the surface area available for heat transfer.16 A dielectric fluid or treated water solution is pumped directly through these plates.
The thermodynamic advantage is overwhelming. Water has a heat transfer coefficient approximately 3,500 times greater than that of air.14 This allows the liquid to capture up to 98% of the heat generated by the components at the source.18
C. The Coolant Distribution Unit (CDU)
The heart of this hydraulic system is the Coolant Distribution Unit (CDU). The CDU acts as a heat exchanger and a pump station. It circulates the clean, controlled coolant loop through the servers and transfers the absorbed heat to a secondary "facility water" loop.2
This separation is critical. The water touching the $30,000 GPUs must be ultra-pure and chemically balanced to prevent corrosion or fouling. The facility water, however, carries the heat out of the building to evaporative cooling towers. This architecture allows Colossus to run its coolant at higher temperatures (warm water cooling), further improving energy efficiency by reducing the need for mechanical chillers (refrigeration) to super-cool the water before it enters the racks.18
V. The Energy Nexus: Combustion Turbines, The Brayton Cycle, and the Quest for Gigawatts
While liquid cooling solved the thermal challenge, the electrical challenge required a more controversial solution. The grid interconnection process for gigawatt-scale loads is notoriously slow, often taking years for utility providers like Memphis Light, Gas, and Water (MLGW) and the Tennessee Valley Authority (TVA) to build the necessary substations and transmission lines. Unwilling to delay the training of Grok, xAI opted for a behind-the-meter power generation strategy that sparked a national legal battle.
A. The Aeroderivative Turbine Array
To bridge the gap between grid capacity and the facility's immediate needs, xAI installed a massive array of mobile combustion turbines—reports indicate between 18 and 35 units were operational at peak usage.4
These devices are typically "aeroderivative" gas turbines—essentially modified jet engines mounted on trailers. They operate on the Brayton thermodynamic cycle:
Compression: Air is drawn in and compressed to high pressure.
Combustion: Natural gas (methane) is injected and ignited, raising the temperature of the gas mixture.
Expansion: The high-velocity, high-temperature gas expands through turbine blades, spinning a shaft connected to an electrical generator.
These turbines are prized for their power density and ability to start up quickly. However, they are designed for temporary or emergency use, not as base-load power plants for permanent facilities.
B. Atmospheric Chemistry: The NOx Problem
The combustion of methane (CH4) is relatively clean compared to coal, producing primarily carbon dioxide (CO2) and water (H2O). However, the high temperatures required for efficient turbine operation trigger a side reaction involving atmospheric nitrogen (N2).
At temperatures exceeding 1,300°C, the nitrogen and oxygen in the air react to form nitrogen oxides (NOx), primarily nitric oxide (NO) and nitrogen dioxide (NO2). This process is known as thermal NOx formation.
N2 + O2 → heat → 2NO
2NO + O2 → 2NO2
Nitrogen dioxide is a reddish-brown gas and a potent respiratory irritant. More critically, in the presence of sunlight and volatile organic compounds (VOCs), NOx acts as a precursor to the formation of ground-level ozone (O3), the primary component of smog.20
C. The Cumulative Impact on South Memphis
The deployment of these turbines in South Memphis was not an isolated event. The area, particularly the 38109 zip code, has long been an industrial hub, hosting a Valero oil refinery, a TVA coal (now gas) plant, and a steel mill. Consequently, the local airshed is already burdened with elevated levels of particulate matter and ozone precursors.5
The "chimney effect" of xAI’s turbine array—clustering dozens of un-scrubbed exhaust stacks in a small area—created a concentrated plume of pollutants. Environmental groups argued that this effectively constituted the construction of a new, unregulated power plant in a community that arguably already exceeded the safe carrying capacity for industrial emissions.22
VI. Atmospheric Chemistry & Regulatory Conflict: The EPA Ruling and the Nitrogen Oxide Debate
The friction between xAI's "move fast" ethos and the regulatory framework of the Clean Air Act came to a head in late 2025 and early 2026. The central dispute revolved around the legal definition of "temporary."
A. The "Non-Road" Engine Loophole
Under the Clean Air Act, stationary sources of air pollution (like power plants) are subject to "New Source Review" (NSR), a rigorous permitting process that requires the installation of Best Available Control Technology (BACT) to minimize emissions. However, "non-road engines"—engines that are portable and move from location to location—are regulated differently, often with less stringent site-specific oversight.
xAI and its contractors utilized a regulatory interpretation, often called the "364-day loophole." This interpretation suggests that if a portable generator does not remain in the same location for more than 12 months, it retains its "non-road" status and does not require a stationary source permit.4 By deploying trailer-mounted turbines and theoretically retaining the ability to move them, xAI operated the array without the federal air permits typically required for a 100+ megawatt power plant.
B. The EPA Ruling of January 15, 2026
On January 15, 2026, the U.S. Environmental Protection Agency (EPA) issued a decisive ruling that effectively closed this loophole. The agency declared that the gas turbines used by xAI were not exempt from air quality permit requirements.4
The EPA’s rationale focused on the function of the equipment rather than its form. Although the turbines were mounted on wheels, they were powering a permanent, stationary structure (the data center) and effectively serving as a substitute for the electrical grid. Therefore, the agency revised its policies to state that operating such machines requires air permits even if they are used on a "portable or temporary basis".4
C. Implications and Fallout
This ruling rendered the unpermitted operation of the turbines illegal. While xAI had eventually secured permits for a subset of the turbines (15 units) from the Shelby County Health Department, the ruling implied that the larger array (up to 35 units) operated in violation of federal law.4
The decision represents a significant precedent for the data center industry. As power constraints become the primary bottleneck for AI deployment, other companies may attempt similar "behind-the-meter" generation strategies. The EPA’s ruling signals that the regulatory state will not allow technicalities regarding equipment mobility to bypass the substantive requirements of the Clean Air Act, particularly in communities already suffering from non-attainment of air quality standards.
VII. Hydrological Impact: The Maxson Wastewater Connection and Aquifer Protection
While air quality dominated the headlines, the hydrological impact of Colossus is equally significant. The massive evaporative cooling towers required to reject gigawatts of heat consume water on a scale comparable to agricultural irrigation.
A. The Memphis Sand Aquifer
Memphis sits atop the Memphis Sand Aquifer, a geological formation containing some of the purest artesian drinking water in the world. The aquifer is separated from the surface by a confining layer of clay, but breaches in this layer can allow surface contaminants to migrate downward.
Initial concerns arose that xAI would draw its cooling water—estimated at up to 5 million gallons per day (MGD)—directly from the aquifer or the municipal supply that relies on it. Drawing such massive volumes could lower the water table and potentially accelerate the downward migration of contaminants from nearby industrial sites, such as the coal ash ponds at the TVA Allen plant.24
B. The Greywater Solution: Membrane Bioreactors
To mitigate this risk, xAI, in coordination with city officials, committed to funding a greywater recycling facility. This plant intercepts treated effluent from the nearby T.E. Maxson Wastewater Treatment Plant—water that would otherwise be discharged into the Mississippi River—and purifies it for industrial use.25
The treatment process utilizes a Membrane Bioreactor (MBR) system, a state-of-the-art wastewater technology.
Fine Screening: The effluent is first screened to remove any remaining particulates.
Biological Treatment: The water enters anoxic and aerobic reactor tanks where microorganisms digest organic pollutants.
Membrane Filtration: Instead of using gravity to settle solids (as in traditional treatment), MBRs force the water through semi-permeable membranes with pores small enough to block bacteria and suspended solids.
Chemical Polishing: The permeate is treated with Sodium Hypochlorite (bleach) and other agents to prevent biological fouling in the cooling towers.26
This facility is designed to treat up to 13 million gallons per day, providing enough cooling water for Colossus, the TVA plant, and Nucor Steel, theoretically insulating the drinking water aquifer from the industrial demand.27
VIII. Acoustic Ecology: Noise Pollution in the Southaven Expansion
As xAI expanded Colossus into Southaven, Mississippi (the "MACROHARDRR" facility), a new environmental stressor emerged: noise. The physics of sound propagation from gas turbines differs significantly from other industrial noises due to its frequency profile and persistence.
A. The Decibel Reality
Residents in Southaven, particularly in the Colonial Hills subdivision, reported a "constant loud humming noise" resembling a leaf blower that persists 24 hours a day.29 Sound pressure measurements taken by residents showed indoor levels fluctuating between 40 and 60 decibels (dB), with outdoor levels exceeding 70 dB at the property boundary.29
To contextalize, 70 dB is roughly the volume of a vacuum cleaner or freeway traffic. While not instantly damaging to hearing, continuous exposure to this level triggers physiological stress responses, disrupts sleep patterns, and is classified by the World Health Organization as a cause of cardiovascular and cognitive impairment.
B. The Physics of Turbine Whine
The noise generated by aeroderivative turbines is often tonal, meaning it is concentrated at specific frequencies related to the rotation speed of the turbine blades (blade pass frequency). Tonal noise is psycho-acoustically more annoying than broadband noise (like wind) because the human brain struggles to filter it out.
The "whirring" described by residents 29 suggests high-frequency components from the air intake compressors, which can travel significant distances, especially at night when thermal layers in the atmosphere can refract sound waves back toward the ground. Despite local noise ordinances capping industrial noise, enforcement has been complicated by the political and economic weight of the project, with local officials describing the noise as "temporary construction noise" despite its persistence.29
IX. Socio-Economic Dimensions: Environmental Justice, Tax Incentives, and the Boxtown Community
The Colossus project exists within a specific socio-economic geography. South Memphis is a predominantly Black, low-income community that has historically borne the brunt of the region's industrial development—a dynamic often termed "environmental racism" by sociologists and activists.
A. The Economic Trade-Off
Proponents of the project, including the Greater Memphis Chamber and Mayor Paul Young, frame Colossus as a transformative economic engine. The "Gigafactory of Compute" promised a capital investment of over $6 billion and the creation of hundreds of high-tech jobs.5
To capture this value, the city structured a Payment-In-Lieu-Of-Taxes (PILOT) arrangement or similar tax incentive structure, though reports indicate xAI paid a significant portion of its taxes early, becoming the second-largest taxpayer in the county after FedEx within a year.31 City officials established a $3.25 million "special revenue fund" derived from xAI’s tax payments, earmarked specifically for community improvements in the surrounding neighborhoods of Boxtown and Whitehaven.32
B. Community Skepticism and the "Facts Over Fiction" Campaign
Despite these financial infusions, community trust remains low. Activist groups like Memphis Community Against Pollution (MCAP) argue that the health costs of the pollution—manifesting in asthma and other chronic conditions—outweigh the fiscal benefits. They contend that the jobs created are specialized technical roles unlikely to be filled by local residents without significant workforce development.33
The tension was exacerbated by a mailer campaign titled "Facts Over Fiction," which was distributed to residents. The fliers downplayed the environmental risks of the gas turbines, claiming they were "cleaner tech" and regulated. Activists and local politicians, such as State Rep. Justin Pearson, characterized these mailers as "lying to us about xAI's methane gas pollution" and criticized the opacity of the permitting process, which involved non-disclosure agreements (NDAs) for city officials.22
This friction highlights the "Promethean Bargain" offered to de-industrialized regions: the acceptance of environmental externalities in exchange for participation in the new digital economy. For Boxtown, the hum of the turbines and the smog precursors are the visceral price of Memphis's entry into the AI age.
X. Conclusion: The Future of the AI-Energy-Environment Triad
The xAI Colossus datacenter stands as a monument to the staggering velocity of the artificial intelligence revolution. In less than two years, the project evolved from a dormant factory into the nerve center of the global AI arms race, pushing the boundaries of what is physically possible in silicon, networking, and thermodynamics.
The technical achievements of Colossus are undeniable. The deployment of 100,000+ liquid-cooled GPUs, the pioneering use of Spectrum-X Ethernet for AI, and the construction of the world's largest greywater recycling for compute are feats of engineering that will influence data center design for decades.
However, the "122-day miracle" of its construction exposed deep fissures in the regulatory and social frameworks that govern industrial development. The reliance on unpermitted gas turbines, the exploitation of regulatory loopholes, and the collision with environmental justice communities reveal that the constraint on AGI is no longer just compute—it is energy, permitting, and social license to operate.
The EPA’s ruling of January 2026 serves as a corrective mechanism, signaling that the urgency of technological progress does not grant immunity from environmental law. As xAI expands into the multi-gigawatt era with the MACROHARDRR facility, the lessons of Memphis will likely define the industry's future. The sustainable path to AGI will require not just faster chips and smarter networks, but a harmonic integration with the energy grid, the water cycle, and the communities that host the physical infrastructure of the digital mind.
Table 1: Technical Specifications of the xAI Colossus Complex (Jan 2026)
Feature | Colossus 1 | Colossus 2 | MACROHARDRR (Bldg 3) |
Status | Operational | Operational | Retrofit / Deployment |
Primary GPU Architecture | NVIDIA H100 (Hopper) | NVIDIA GB200 (Blackwell) | NVIDIA GB200 / GB300 |
Cooling Methodology | Liquid-Assisted / DLC | Full Direct-to-Chip (DLC) | Full Direct-to-Chip (DLC) |
Network Fabric | Spectrum-X Ethernet | Spectrum-X Ethernet | Spectrum-X Ethernet |
Est. Power Capacity | ~500 MW | ~1 GW | ~500 MW |
Primary Power Source | Grid + Temporary Turbines | Grid + Perm. Substation | Grid + On-site Gas Plant |
GPU Count (Est.) | ~100,000 | ~250,000+ | ~200,000+ |
Table 2: Environmental Impact Metrics
Metric | Value / Description | Source / Context |
Water Demand (Peak) | ~5 Million Gallons/Day | Evaporative cooling demand 24 |
Greywater Plant Capacity | 13 Million Gallons/Day | Maxson WWTP diversion capacity 27 |
Noise Levels (Southaven) | 40-60 dB (Indoor) / 70+ dB (Outdoor) | Resident measurements of turbine noise 29 |
Turbine Emissions | Nitrogen Oxides (NOx), VOCs | Precursors to ground-level ozone (smog) |
Heat Rejection | ~1.8 Gigawatts (Total Site) | Thermal energy released to atmosphere |
(Note: Data derived from available reports, regulatory filings, and press releases as of January 17, 2026.)
Works cited
xAI Colossus Hits 2 GW: 555,000 GPUs, $18B, Largest AI Site | Introl ..., accessed January 17, 2026, https://introl.com/blog/xai-colossus-2-gigawatt-expansion-555k-gpus-january-2026
Inside the 100K GPU xAI Colossus Cluster that Supermicro Helped Build for Elon Musk, accessed January 17, 2026, https://www.supermicro.com/CaseStudies/Success_Story_xAI_Colossus_Cluster.pdf
Elon Musk’s xAI to invest $20 billion in Mississippi data centre, accessed January 17, 2026, https://timesofindia.indiatimes.com/technology/tech-news/elon-musks-xai-to-invest-20-billion-in-mississippi-data-centre/articleshow/126427391.cms
Elon Musk's xAI datacenter generating extra electricity illegally ..., accessed January 17, 2026, https://www.theguardian.com/technology/2026/jan/15/elon-musk-xai-datacenter-memphis
Tech, Toxins, and Memphis: Evaluating the Environmental Footprint of the xAI Facility - Tennessee Bar Association, accessed January 17, 2026, https://www.tba.org/?pg=Hastings2025AIX
H100 GPU - NVIDIA, accessed January 17, 2026, https://www.nvidia.com/en-us/data-center/h100/
NVIDIA H100 versus H200: how do they compare? - CUDO Compute, accessed January 17, 2026, https://www.cudocompute.com/blog/nvidia-h100-vs-h200-how-will-they-compare
Inside xAI Colossus, the 100000-GPU Supercluster Powered by NVIDIA Spectrum-X, accessed January 17, 2026, https://www.naddod.com/blog/xai-colossus-100-000-gpu-supercluster-powered-by-spectrum-x
Colossus Unbound: xAI's Memphis Expansion Targets 1 Million GPUs in the Race for AGI, accessed January 17, 2026, https://markets.financialcontent.com/prnews/article/tokenring-2026-1-2-colossus-unbound-xais-memphis-expansion-targets-1-million-gpus-in-the-race-for-agi
The Colossus Awakening: xAI's 555,000-GPU Supercluster and the Global Race for AGI Compute - FinancialContent, accessed January 17, 2026, https://markets.financialcontent.com/wral/article/tokenring-2026-1-15-the-colossus-awakening-xais-555000-gpu-supercluster-and-the-global-race-for-agi-compute
Spectrum-X | Ethernet Networking Platform for AI - NVIDIA, accessed January 17, 2026, https://www.nvidia.com/en-us/networking/spectrumx/
NVIDIA Ethernet Networking Accelerates World's Largest AI Supercomputer, Built by xAI, accessed January 17, 2026, https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus
xAI Colossus | Supermicro, accessed January 17, 2026, https://www.supermicro.com/en/featured/xai-colossus
Liquid Cooling vs Air: The 50kW GPU Rack Guide (2025) - Introl, accessed January 17, 2026, https://introl.com/blog/liquid-cooling-gpu-data-centers-50kw-thermal-limits-guide
Understanding direct-to-chip cooling in HPC infrastructure: A deep dive into liquid cooling, accessed January 17, 2026, https://www.vertiv.com/en-us/about/news-and-insights/articles/educational-articles/understanding-direct-to-chip-cooling-in-hpc-infrastructure-a-deep-dive-into-liquid-cooling/
Supermicro Total Rack Scale Liquid Cooling Solutions, accessed January 17, 2026, https://www.supermicro.com/manuals/brochure/Brochure-Liquid-Cooling-Solutions.pdf
accessed January 17, 2026, https://www.vertiv.com/en-us/about/news-and-insights/articles/educational-articles/understanding-direct-to-chip-cooling-in-hpc-infrastructure-a-deep-dive-into-liquid-cooling/#:~:text=Liquid%20cooling%20leverages%20the%20higher,the%20capabilities%20of%20traditional%20methods.
Rack-Scale Liquid Cooling Solutions - Supermicro, accessed January 17, 2026, https://www.supermicro.com/en/solutions/liquid-cooling
New images reveal Elon Musk's xAI datacenter has nearly doubled its number of polluting, unpermitted gas turbines - Southern Environmental Law Center, accessed January 17, 2026, https://www.selc.org/press-release/new-images-reveal-elon-musks-xai-datacenter-has-nearly-doubled-its-number-of-polluting-unpermitted-gas-turbines/
Elon Musk's xAI facility is using gas turbines in South Memphis, we're taking action, accessed January 17, 2026, https://www.selc.org/news/resistance-against-elon-musks-xai-facility-in-south-memphis-gets-stronger/
Elon Musk's xAI facility is polluting South Memphis - Southern Environmental Law Center, accessed January 17, 2026, https://www.selc.org/news/elon-musks-xai-facility-is-polluting-south-memphis/
In South Memphis, Elon Musk's Colossus Operated Gas Turbines Without Appropriate Permits, Residents and Activists Claim - Inside Climate News, accessed January 17, 2026, https://insideclimatenews.org/news/17072025/elon-musk-xai-data-center-gas-turbines-memphis/
Elon Musk's xAI gets permit for methane gas generators | Tennessee - The Guardian, accessed January 17, 2026, https://www.theguardian.com/us-news/2025/jul/03/elon-musk-xai-pollution-memphis
xAI Supercomputer - Protect our Aquifer, accessed January 17, 2026, https://www.protectouraquifer.org/issues/xai-supercomputer
Inside Memphis' Battle Against Elon Musk's xAI Data Center - Time Magazine, accessed January 17, 2026, https://time.com/7308925/elon-musk-memphis-ai-data-center/
Colossus Water Recycling Plant PER Phase - TDEC Dataviewers, accessed January 17, 2026, https://dataviewers.tdec.tn.gov/dataviewers/BGWPC.GET_WPC_DOCUMENTS?p_file=554491025485987891
xAI's new facility will save Memphis 5 billion gallons of water annually - Teslarati, accessed January 17, 2026, https://www.teslarati.com/xai-new-facility-will-save-memphis-5-billion-gallons-water-annually/
Permit Details - TDEC Dataviewers - TN.gov, accessed January 17, 2026, https://dataviewers.tdec.tn.gov/dataviewers/f?p=2005:34051:::NO:34051:P34051_PERMIT_NUMBER:SOP-24025
Southaven Residents Raise Concerns About Noise From xAI ..., accessed January 17, 2026, https://www.memphisflyer.com/southaven-residents-raise-concerns-about-noise-from-xai-turbines/
Southaven residents fear pollution, complain of noise, from Elon ..., accessed January 17, 2026, https://mississippitoday.org/2025/11/24/southaven-residents-fear-pollution-complain-of-noise-from-elon-musks-xai-data-center-turbines/
Elon Musk's xAI becomes Memphis' 2nd largest taxpayer in just one year: report - Teslarati, accessed January 17, 2026, https://www.teslarati.com/elon-musk-xai-memphis-2nd-largest-taxpayer-one-year/
City of Memphis to invest $3.2M of xAI tax dollars back into local neighborhoods, accessed January 17, 2026, https://www.localmemphis.com/article/money/business/memphis-32m-xai-taxes-into-neighborhoods/522-b062ee6f-d9cc-40be-9816-cfd6b614076c
South Memphis Residents Skeptical of Musk's xAI Economic Growth Claims as Pollution Concerns Grow - NCRC, accessed January 17, 2026, https://www.ncrc.org/south-memphis-residents-skeptical-of-musks-xai-economic-growth-claims-as-pollution-concerns-grow/
Elon Musk's xAI accused of pollution over Memphis supercomputer - The Guardian, accessed January 17, 2026, https://www.theguardian.com/technology/2025/apr/24/elon-musk-xai-memphis



Comments