Complete Technical Paper Content β Ai Loan Approval
1. Executive Abstract, Problem Formulation & Background
An AI data center is a specialized data center facility designed for the computationally intensive tasks of training and running artificial intelligence (AI) and machine learning models. Unlike general-purpose data centers, they are often optimized for the parallel processing demands of AI workloads, typically using hardware such as AI accelerators (GPUs, TPUs) and high-speed interconnects. The global push to construct these specialized facilities accelerated dramatically during the AI boom of the 2020s. In 2026, major tech companies were estimated to spend $650 billion on AI data centers. Their purchases of memory chips, particularly High Bandwidth Memory, led to a global memory supply shortage amid a broader competition for semiconductors, power, and infrastructure. About 70% of global computer memory production in the 2026 fiscal year has been purchased for AI data centers. AI-focused data centers use around 60 kilowatts of electricity per server rack, while general-purpose data centers typically use around 10 kilowatts of electricity per rack. By 2026, many U.S. communities had organized against the local construction of AI data centers, which have a large power and water footprint and generate heat, noise, and other pollutants. By July 2026, local community resistance had blocked the construction of AI data centers worth some $130 billion.
Between January and August 2024, Microsoft, Meta, Google and Amazon collectively spent $125 billion on AI data centers. Citigroup forecasted that $2.8 trillion would be spent on AI data centers by 2030, while McKinsey and Company estimated that almost $7 trillion would be spent globally by that time. According to S&P Global, $61 billion has been spent on the data center market as a whole in 2025, while debt issuance for data centers was $182 billion during the same year. Large technology companies have offloaded the financial risks of building AI data centers by setting up special purpose vehicles or by contracting with neoclouds. For example, Meta's Hyperion was mostly funded by Blue Owl Capital, which did so using a bond offering from PIMCO. Those bonds were sold to a number of clients, including BlackRock. Meta did not borrow money itself and instead established a special purpose vehicle from which it would rent the data center. This deal was structured by Morgan Stanley for $30 billion, the largest known private capital transaction as of 2025. Neoclouds such as CoreWeave have gone into debt to buy computer chips from Nvidia for their data centers, and the chips themselves have been used for loan collateral. As of December 2025, CoreWeave took out three GPU-backed loans, collectively worth $12.4 billion, from private credit firms (Blackstone, Coatue, BlackRock, PIMCO) and from banks (Goldman Sachs, JPMorgan Chase, Wells Fargo). Thus, these companies provide an indirect connection between private credit and established banks. Data centers have also established asset-backed securities, and debt for data centers has its own derivative financial products. The real estate industry, including asset managers, public companies and private investors, has also invested in AI development.
Average AI data centers have an electricity footprint equivalent to 100,000 households, and use billions of gallons of water for cooling their hardware. In 2025, the International Energy Agency estimated that the larger AI data centers currently under construction could consume as much electricity as 2 million households. A 2024 report from the United States Department of Energy stated that data centers overall used 17 billion gallons of water per year in the United States, primarily due to "rapid proliferation of AI servers", and that this usage was forecasted to grow to nearly 80 billion gallons by 2028. Researchers estimated that AI data centers in the United States would emit 24β44 million metric tons of carbon dioxide and use 731β1,125 million cubic meters of water per year between 2024 and 2030. Peaking power plants, which have been proposed as a power source for AI data centers, emit sulfur dioxide and have historically been located disproportionately near communities of color in the United States. Reciprocating internal combustion engines, proposed as another power source for a data center, emit PM 2.5, nitrogen oxides, and volatile organic compounds.
The construction of AI data centers during the AI boom starting in 2025 has exhausted the global supply of the computer memory. To meet demand from AI servers, the three major DRAM manufacturers, Samsung, SK Hynix, and Micron, prioritized production of High Bandwidth Memory (HBM) and high-end server DRAM, contributing to a global memory supply shortage and intensifying competition for advanced memory chips. Because HBM consumes far more fabrication capacity than conventional memoryβproducing a single bit of HBM requires roughly three times the wafer capacity of DDR5. The resulting price increases and shortages have affected computer memory broadly, including HBM, DRAM, and NAND flash memory. Under an agreement with Samsung and SK Hynix, the Stargate project alone is set to receive 900,000 DRAM wafers per month, roughly 40% of total global DRAM production.
In the United States, both the Biden administration and second Trump administration supported the construction of AI data centers. In January 2025, then-president Joe Biden signed an executive order for federal government agencies to support AI data centers on federal sites built by private companies, study their effect on energy prices, and encourage their use of renewable energy. In April 2025, the United States Department of Energy suggested 16 possible sites, including Los Alamos National Laboratory, Sandia National Laboratories and Oak Ridge National Laboratory. In its July 2025 AI Action Plan, the second Trump administration supported increased production of AI data centers. Several US states have incentivized local data center construction. For example, in 2024, lawmakers in Michigan approved tax breaks for data center equipment and construction material. Some data center companies have also invested or promised to invest in the infrastructure of local communities. In December 2025, Democratic senators Elizabeth Warren, Chris Van Hollen, and Richard Blumenthal wrote to seven technology companies (Google, Microsoft, Amazon, Meta, CoreWeave, Digital Realty, and Equinix) that they will investigate the effects of those companies' operations on consumer energy bills, highlighting AI data centers in particular. That same month, 25 Democratic lawmakers wrote to the inspector general of the United States Department of Commerce over possible conflict of interest concerns involving secretary of commerce Howard Lutnick and AI data centers. Lutnick met with a co-founder of Fermi, which proposed an AI data center project financed by Cantor Fitzgerald and Newmark Group. AI Infrastructure Coalition (AIIC), an organization led by Brian O. Walsh, Kyrsten Sinema, and Garret Graves, has supported Trump's AI Action Plan, with Sinema becoming personally involved in a local meeting on an AI data center in Arizona. Venture capital firm Andreessen Horowitz, Big Tech companies (Google, Meta, Microsoft), telecommunications companies (Cisco, Corning Inc., Lumen Technologies), data center providers (Digital Realty, QTS), energy companies (Duke Energy, Entergy, ExxonMobil, NextEra Energy), and utility companies (PG&E, Pinnacle West Capital) are members of AIIC as of November 2025. AIIC has been supported by Hogan Lovells. According to Goldman Sachs in 2025 an addition half million workers will be needed by 2030 to build the power infrastructure, which is acute because of AI data center build out. In October 2026 Google says Gemini grew from 400 million to 1 billion users in three months.
Some analysts have expressed concerns about overbuilding of AI data centers, warning that their infrastructure risks obsolescence due to changes in demand and technology. Estimates for the lifespan of the GPUs used to power data centers range from one to eight years. By contrast, CPUs in more traditional data centers have a shelf life of about 5β7 years. Machine learning model training causes significant stress on computer chips.
Advocates have also linked data center construction to the AI bubble. Wisconsin Watch expressed concern that data centers could become stranded assets in the context of an AI bubble, leaving energy customers to pay for the costs. Journalists have warned that the large number of links between actors and organizations in the AI ecosystem as well as the large number of companies involved each indicate strong risks for the global economy and financial markets. A hedge fund founder questioned the economics of AI data centers, concluding that they need about $1 trillion of revenue for profitability. In November 2025, the North American Electric Reliability Corporation warned that building new data centers could negatively affect the electrical grid and cause power outages during extreme weather. The independent monitor of PJM Interconnection warned that its power grid cannot support new data centers and supported a federal moratorium on data centers. Machine learning developers, when training large language models, often use data centers at their full capacity, conflicting with other users (including households) during peak usage, which may lead to blackouts. Power disruptions in data centers can cause errors in machine learning model calculations ("silent data corruption"). Large technology companies who are building data centers have asked public officials and land owners to sign non-disclosure agreements and have appeared to use shell companies. In one case, when a joint OpenAI-Oracle data center was rejected by Saline Township in Michigan, landowners and developers responded by suing the town. The developers acquired the land from the resulting settlement. The Michigan Public Service Commission approved DTE Energy to provide power to the data center, part of Stargate. The approval was part of an ex parte motion, with no public input. In 2025, over 230 groups have signed a letter supporting a moratorium on constructing AI data centers in the United States, concerned by impact on the environment and energy bills. Signatories include Food & Water Watch, Greenpeace, Friends of the Earth and Physicians for Social Responsibility. Senator Bernie Sanders has also supported a moratorium on AI data centers. Ashley LaMont, of Honor the Earth, argued that data centers on tribal lands would not help Native Americans obtain data sovereignty. The NAACP has also expressed opposition to AI data centers.
Several tech companies, including Starcloud, Google, Nvidia, Blue Origin and SpaceX, have announced projects for or otherwise expressed interest in building data centers in outer space. Google announced Project Suncatcher, which plans to test whether its Tensor Processing Unit (TPU) chips (which are specifically designed for machine learning) can function in the environment of space. The company aims to deploy satellites with these chips in a data center by 2027. In November 2025, Starcloud, a startup supported by Nvidia, launched a satellite with an Nvidia H100 GPU that the company used to deploy and develop a large language model. CNBC reported that this model, NanoGPT, was the first model trained in space. Starcloud deployed a second model based on Google's Gemma.
2. Related Work, Literature Survey & Algorithmic Foundations
Fairness in machine learning (ML) refers to the various attempts to correct algorithmic bias in automated decision processes based on ML models. Decisions made by such models after a learning process may be considered unfair if they were based on variables considered sensitive (e.g., gender, ethnicity, sexual orientation, or disability). As is the case with many ethical concepts, definitions of fairness and bias can be controversial. In general, fairness and bias are considered relevant when the decision process impacts people's lives. Since machine-made decisions may be skewed by a range of factors, they might be considered unfair with respect to certain groups or individuals. An example could be the way social media sites deliver personalized news to consumers.
Recent works underline the presence of several limitations to the current landscape of fairness in machine learning, particularly when it comes to what is realistically achievable in this respect in the ever increasing real-world applications of AI. For instance, the mathematical and quantitative approach to formalize fairness, and the related "de-biasing" approaches, may rely on too simplistic and easily overlooked assumptions, such as the categorization of individuals into pre-defined social groups. Other delicate aspects are, e.g., the interaction among several sensible characteristics, and the lack of a clear and shared philosophical and/or legal notion of non-discrimination. Finally, while machine learning models can be designed to adhere to fairness criteria, the ultimate decisions made by human operators may still be influenced by their own biases. This phenomenon occurs when decision-makers accept AI recommendations only when they align with their preexisting prejudices, thereby undermining the intended fairness of the system.
P ( R = r | A = a ) = P ( R = r | A = b ) β r β R β a , b β A {\displaystyle P(R=r\ |\ A=a)=P(R=r\ |\ A=b)\quad \forall r\in R\quad \forall a,b\in A}
This means that the classification rate for each target classes is equal for people belonging to different groups with respect to sensitive characteristics A {\displaystyle A} . Yet another equivalent expression for independence can be given using the concept of mutual information between random variables, defined as
I ( X , Y ) = H ( X ) + H ( Y ) β H ( X , Y ) {\displaystyle I(X,Y)=H(X)+H(Y)-H(X,Y)}
In this formula, H ( X ) {\textstyle H(X)} is the entropy of the random variable X {\displaystyle X} . Then ( R , A ) {\textstyle (R,A)} satisfy independence if I ( R , A ) = 0 {\textstyle I(R,A)=0} . A possible relaxation of the independence definition include introducing a positive slack Ο΅ > 0 {\textstyle \epsilon >0} and is given by the formula:
P ( R = r | A = a ) β₯ P ( R = r | A = b ) β Ο΅ β r β R β a , b β A {\displaystyle P(R=r\ |\ A=a)\geq P(R=r\ |\ A=b)-\epsilon \quad \forall r\in R\quad \forall a,b\in A}
Finally, another possible relaxation is to require I ( R , A ) β€ Ο΅ {\textstyle I(R,A)\leq \epsilon } .
3. Dataset Ingestion, Preprocessing & Stochastic Augmentations
R β₯ A | Y . {\displaystyle R\bot A\ |\ Y.}
P ( R = r | Y = q , A = a ) = P ( R = r | Y = q , A = b ) β r β R q β Y β a , b β A {\displaystyle P(R=r\ |\ Y=q,A=a)=P(R=r\ |\ Y=q,A=b)\quad \forall r\in R\quad q\in Y\quad \forall a,b\in A}
This means that all the dependence of the decision R {\displaystyle R} on the sensitive attribute A {\displaystyle A} must be justified by the actual dependence of the true target variable Y {\displaystyle Y} . Another equivalent expression, in the case of a binary target rate, is that the true positive rate and the false positive rate are equal (and therefore the false negative rate and the true negative rate are equal) for every value of the sensitive characteristics:
P ( R = 1 | Y = 1 , A = a ) = P ( R = 1 | Y = 1 , A = b ) β a , b β A {\displaystyle P(R=1\ |\ Y=1,A=a)=P(R=1\ |\ Y=1,A=b)\quad \forall a,b\in A}
P ( R = 1 | Y = 0 , A = a ) = P ( R = 1 | Y = 0 , A = b ) β a , b β A {\displaystyle P(R=1\ |\ Y=0,A=a)=P(R=1\ |\ Y=0,A=b)\quad \forall a,b\in A}
A possible relaxation of the given definitions is to allow the value for the difference between rates to be a positive number lower than a given slack Ο΅ > 0 {\textstyle \epsilon >0} , rather than equal to zero. In some fields separation (separation coefficient) in a confusion matrix is a measure of the distance (at a given level of the probability score) between the predicted cumulative percent negative and predicted cumulative percent positive. The greater this separation coefficient is at a given score value, the more effective the model is at differentiating between the set of positives and negatives at a particular probability cut-off. According to Mayes: "It is often observed in the credit industry that the selection of validation measures depends on the modeling approach. For example, if modeling procedure is parametric or semi-parametric, the two-sample K-S test is often used. If the model is derived by heuristic or iterative search methods, the measure of model performance is usually divergence. A third option is the coefficient of separation...The coefficient of separation, compared to the other two methods, seems to be most reasonable as a measure for model performance because it reflects the separation pattern of a model."
Y β₯ A | R . {\displaystyle Y\bot A\ |\ R.}
P ( Y = q | R = r , A = a ) = P ( Y = q | R = r , A = b ) β q β Y r β R β a , b β A {\displaystyle P(Y=q\ |\ R=r,A=a)=P(Y=q\ |\ R=r,A=b)\quad \forall q\in Y\quad r\in R\quad \forall a,b\in A}
4. Neural Network Backbone, Layer Specifications & Loss Formulations
This means that the probability of actually being in each of the groups is equal for two individuals with different sensitive characteristics given that they were predicted to belong to the same group.
Assuming Y {\textstyle Y} is binary, if A {\textstyle A} and Y {\textstyle Y} are not statistically independent, and R {\textstyle R} and Y {\textstyle Y} are not statistically independent either, then independence and separation cannot both hold except for rhetorical cases. If ( R , A , Y ) {\textstyle (R,A,Y)} as a joint distribution has positive probability for all its possible values and A {\textstyle A} and Y {\textstyle Y} are not statistically independent, then separation and sufficiency cannot both hold except for rhetorical cases. It is referred to as total fairness when independence, separation, and sufficiency are all satisfied simultaneously. However, total fairness is not possible to achieve except in specific rhetorical cases.
Most statistical measures of fairness rely on different metrics, so we will start by defining them. When working with a binary classifier, both the predicted and the actual classes can take two values: positive and negative. Now let us start explaining the different possible relations between predicted and actual outcome: True positive (TP): The case where both the predicted and the actual outcome are in a positive class. True negative (TN): The case where both the predicted outcome and the actual outcome are assigned to the negative class. False positive (FP): A case predicted to befall into a positive class assigned in the actual outcome is to the negative one. False negative (FN): A case predicted to be in the negative class with an actual outcome is in the positive one. These relations can be easily represented with a confusion matrix, a table that describes the accuracy of a classification model. In this matrix, columns and rows represent instances of the predicted and the actual cases, respectively. By using these relations, we can define multiple metrics which can be later used to measure the fairness of an algorithm:
Positive predicted value (PPV): the fraction of positive cases which were correctly predicted out of all the positive predictions. It is usually referred to as precision, and represents the probability of a correct positive prediction. It is given by the following formula: P P V = P ( a c t u a l = + | p r e d i c t i o n = + ) = T P T P + F P {\displaystyle PPV=P(actual=+\ |\ prediction=+)={\frac {TP}{TP+FP}}}
False discovery rate (FDR): the fraction of positive predictions which were actually negative out of all the positive predictions. It represents the probability of an erroneous positive prediction, and it is given by the following formula: F D R = P ( a c t u a l = β | p r e d i c t i o n = + ) = F P T P + F P {\displaystyle FDR=P(actual=-\ |\ prediction=+)={\frac {FP}{TP+FP}}}
Negative predicted value (NPV): the fraction of negative cases which were correctly predicted out of all the negative predictions. It represents the probability of a correct negative prediction, and it is given by the following formula: N P V = P ( a c t u a l = β | p r e d i c t i o n = β ) = T N T N + F N {\displaystyle NPV=P(actual=-\ |\ prediction=-)={\frac {TN}{TN+FN}}}
False omission rate (FOR): the fraction of negative predictions which were actually positive out of all the negative predictions. It represents the probability of an erroneous negative prediction, and it is given by the following formula: F O R = P ( a c t u a l = + | p r e d i c t i o n = β ) = F N T N + F N {\displaystyle FOR=P(actual=+\ |\ prediction=-)={\frac {FN}{TN+FN}}}
True positive rate (TPR): the fraction of positive cases which were correctly predicted out of all the positive cases. It is usually referred to as sensitivity or recall, and it represents the probability of the positive subjects to be classified correctly as such. It is given by the formula: T P R = P ( p r e d i c t i o n = + | a c t u a l = + ) = T P T P + F N {\displaystyle TPR=P(prediction=+\ |\ actual=+)={\frac {TP}{TP+FN}}}
5. Training Protocol, Convergence Behavior & Experimental Evaluation
False negative rate (FNR): the fraction of positive cases which were incorrectly predicted to be negative out of all the positive cases. It represents the probability of the positive subjects to be classified incorrectly as negative ones, and it is given by the formula: F N R = P ( p r e d i c t i o n = β | a c t u a l = + ) = F N T P + F N {\displaystyle FNR=P(prediction=-\ |\ actual=+)={\frac {FN}{TP+FN}}}
True negative rate (TNR): the fraction of negative cases which were correctly predicted out of all the negative cases. It represents the probability of the negative subjects to be classified correctly as such, and it is given by the formula: T N R = P ( p r e d i c t i o n = β | a c t u a l = β ) = T N T N + F P {\displaystyle TNR=P(prediction=-\ |\ actual=-)={\frac {TN}{TN+FP}}}
False positive rate (FPR): the fraction of negative cases which were incorrectly predicted to be positive out of all the negative cases. It represents the probability of the negative subjects to be classified incorrectly as positive ones, and it is given by the formula: F P R = P ( p r e d i c t i o n = + | a c t u a l = β ) = F P T N + F P {\displaystyle FPR=P(prediction=+\ |\ actual=-)={\frac {FP}{TN+FP}}}
The following criteria can be understood as measures of the three general definitions given at the beginning of this section, namely Independence, Separation and Sufficiency. In the table to the right, we can see the relationships between them. To define these measures specifically, we will divide them into three big groups as done in Verma et al.: definitions based on a predicted outcome, on predicted and actual outcomes, and definitions based on predicted probabilities and the actual outcome. We will be working with a binary classifier and the following notation: S {\textstyle S} refers to the score given by the classifier, which is the probability of a certain subject to be in the positive or the negative class. R {\textstyle R} represents the final classification predicted by the algorithm, and its value is usually derived from S {\textstyle S} , for example will be positive when S {\textstyle S} is above a certain threshold. Y {\textstyle Y} represents the actual outcome, that is, the real classification of the individual and, finally, A {\textstyle A} denotes the sensitive attributes of the subjects.
Demographic parity, also referred to as statistical parity, acceptance rate parity and benchmarking. A classifier satisfies this definition if the subjects in the protected and unprotected groups have equal probability of being assigned to the positive predicted class. This is, if the following formula is satisfied: P ( R = + | A = a ) = P ( R = + | A = b ) β a , b β A {\displaystyle P(R=+\ |\ A=a)=P(R=+\ |\ A=b)\quad \forall a,b\in A}
Conditional statistical parity. Basically consists in the definition above, but restricted only to a subset of the instances. In mathematical notation this would be: P ( R = + | L = l , A = a ) = P ( R = + | L = l , A = b ) β a , b β A β l β L {\displaystyle P(R=+\ |\ L=l,A=a)=P(R=+\ |\ L=l,A=b)\quad \forall a,b\in A\quad \forall l\in L}
Predictive parity, also referred to as outcome test. A classifier satisfies this definition if the subjects in the protected and unprotected groups have equal PPV. This is, if the following formula is satisfied: P ( Y = + | R = + , A = a ) = P ( Y = + | R = + , A = b ) β a , b β A {\displaystyle P(Y=+\ |\ R=+,A=a)=P(Y=+\ |\ R=+,A=b)\quad \forall a,b\in A}
Mathematically, if a classifier has equal PPV for both groups, it will also have equal FDR, satisfying the formula: P ( Y = β | R = + , A = a ) = P ( Y = β | R = + , A = b ) β a , b β A {\displaystyle P(Y=-\ |\ R=+,A=a)=P(Y=-\ |\ R=+,A=b)\quad \forall a,b\in A}
Model Figures & Evaluation Curves (4 Images)
Deep Learning Execution Pipeline
(Preprocessing)
(Feature Extraction)
(Self-Attention)
(Loss Optimization)