AI: How to Predict Financial Time Series & Price Collapse in AI Infra
AI: How to Predict Financial Time Series & Price Collapse in AI Infra
The AI market is shifting into a different paradigm
Hi Fintech Futurists —
Today we highlight the following:
AI: Price Collapse in AI Infrastructure
AI: Model Review for Prediction of Financial Time Series
LONG TAKE: Stripe and Coinbase embrace Generative AI + our AI investment index
PODCAST: What really happened at Synapse according to Founder Sankaet Pathak
CURATED UPDATES: Machine Models, AI Applications in Finance, Infrastructure & Middleware
To support this writing and access our full archive of newsletters, analyses, and guides to building in the Fintech & DeFi industries, see subscription options here.
AI: Price Collapse in AI Infrastructure
We highlight for you a key article from the AI industry.
Here are the summary points.
Launch of ChatGPT and H100 GPUs: ChatGPT was launched in November 2022 using Nvidia's A100 GPUs. The H100 series, introduced in March 2023, promised 3x the performance of A100s at only double the price, sparking an AI arms race among startups aiming to surpass OpenAI.
Massive Investment in AI Hardware: Driven by the potential to build more powerful AI models, investors poured tens of billions of dollars into startups acquiring H100 GPUs, leading to a surge in demand and price hikes. Initial rental rates for H100s jumped from $4.70 per hour to over $8 as companies scrambled to train their models.
Nvidia’s Profitable GPU Leasing Market: Nvidia's 2023 strategy capitalized on the demand surge by marketing the H100s rental at $4 per hour, making data center investments highly lucrative. With GPU farms yielding over $100,000 per GPU annually, Nvidia positioned these assets for a payback period of less than 1.5 years.
Price Drop in 2024: By early 2024, H100 rental prices fell to about $2.85 per hour, with further drops to $1-$2 per hour observed in mid-2024 during auctions for short-term usage. This decline marked a rapid shift from Nvidia's initial projections, driven by increased competition and supply.
Impact on ROI for GPU Investments: For investors, the profitability of H100 rentals dropped significantly; at prices below $2.85 per hour, the internal rate of return (IRR) barely exceeds 10%. If prices fall below $1.65 per hour, infrastructure investments in H100s are expected to incur losses over a 5-year period.
Shift Toward Open-Weight Models: The emergence of high-performing open-weight AI models, like LLaMA 3, reduced the need for training from scratch, leading to a preference for fine-tuning these models.
Decline in Foundation Model Investments: The market saw a sharp drop in the number of small to medium AI model creators as open-weight models became a viable alternative. Today, there are fewer than 50 teams globally pursuing large-scale model training, down from the surge seen in 2023.
Excess GPU Capacity Hits the Market: As companies pivoted to fine-tuning or completed their model training, unused H100 capacity from long-term reservations flooded the market. Compute resellers like Runpod, Together.ai, and Vast.ai began offering these resources at discounted rates, driving down prices.
Emergence of Alternative GPUs: New GPU technologies from AMD and Intel, as well as cheaper inference solutions like Nvidia's L40S, started to challenge the dominance of H100s. These alternatives offer competitive performance at lower costs, contributing to the commoditization of GPU hardware and the decline in H100 pricing.
NVIDIA’s revenues and stock price are not suffering — they are the ones selling the shovels to everyone in the gold rush. But add this line of thinking to our previous note about the expected / required payback from consumer and enterprise applications, and whether the infrastructure in AI can meet profitable demand.
There is no question that this stuff is transformative, profound, and here to stay. So was the Internet. But we remind you of this graph.
As financial practitioners, what matters most are the applications. Below, we review the latest literature on applying state of the art AI to capital markets.
AI: Model Review for Prediction of Financial Time Series
Forecasting financial time series is notoriously challenging due to the high volatility and dynamic nature of markets. Factors such as macroeconomic events, geopolitical developments, and shifting investor sentiments contribute to the non-stationary behavior of market data, making accurate predictions difficult. Yet, the presence of well-documented capital market anomalies suggests that exploitable patterns exist.
Historically, models linking predictive signals to future returns have been linear (e.g., traditional factor models like Fama-French), limiting their ability to capture the nonlinear dependencies inherent in financial markets. As we've highlighted in our previous AI analysis, uncovering these complex structures within financial data necessitates machine learning techniques. The era of earning macroscopic alpha from basic econometric models is over.
For instance, a highly cited paper titled "Deep Learning with Long Short-Term Memory Networks for Financial Market Predictions (2018)” applied LSTM networks — recurrent neural networks (RNNs) designed for sequence learning — to forecast S&P 500 constituents from 1992 to 2015.
The LSTM networks achieved statistically and economically significant returns of 0.46% per day, outperforming Random Forests (0.43%), standard Deep Neural Networks (0.32%), and Logistic Regression models (0.26%). This demonstrated the LSTM's ability to capture complex temporal dependencies in financial data more effectively than traditional ML models.
Introduced in 2017, Transformers represented a major advancement in sequence modeling. Unlike RNNs, which process data sequentially and often struggle with retaining long-term dependencies, Transformers use a self-attention mechanism to analyze all elements of a sequence simultaneously. This parallel processing accelerates training and inference while capturing the more complex relationships between distant elements.
The paper "Temporal Fusion Transformers for Interpretable Multi-Horizon Time Series Forecasting (2019)" showcased the effectiveness of Transformers in time series analysis, where the Temporal Fusion Transformer (TFT) was able to capture long-range dependencies that are challenging for RNN-based architectures to learn, enhancing forecasting accuracy in multi-horizon time series tasks.
Although Transformers led some to declare "RNNs are dead," recent research has reignited interest in recurrent sequence models. The paper "Were RNNs All We Needed? (2024)” introduced minLSTM and minGRU, minimal versions of LSTM and GRU networks that enable parallel training without (supposedly) compromising performance.
These models demonstrate significant efficiency gains; on sequences of length 512, minGRU and minLSTM trained 175x and 235x faster per training step than their traditional counterparts. Performance-wise, they matched and sometimes exceeded models like Mamba and Transformers on tasks such as selective copying and character-level language modeling.
In finance, the ability to accelerate training via parallel processing is highly advantageous due to the massive volumes of time-series data. This efficiency enables processing longer historical sequences without prohibitive computational costs. However, merely deploying advanced ML (or in this case, deep learning) models does not guarantee alpha generation, as many market participants are employing similar techniques.
The real competitive advantage in financial AI lies in integrating these models with novel alternative data sources and unique feature engineering that captures subtle market dynamics — though this is easier said than done.
Financial forecasting is less about identifying a single superior model and more about constructing adaptive systems that respond to evolving market conditions. This could involve ensemble methods that adjust model weights based on current market regimes or meta-learning approaches that rapidly adapt to new patterns.
One example of this is has already been showcased in a new paper titled “MLP, XGBoost, KAN, TDNN, and LSTM-GRU Hybrid RNN with Attention for SPX and NDX European Call Option Pricing (2024),” where the authors incorporated attention within the best-performing LSTM/GRU RNN architecture to enhance the model’s ability to capture temporal dependencies.
👑Related Coverage👑
Blueprint Deep Dive
Long Take: Stripe and Coinbase embrace Generative AI + our AI investment index (link here)
We discuss how fintech firms like Stripe and Coinbase are positioning themselves to leverage Generative AI and blockchain technology. Stripe is partnering with leading AI companies, like OpenAI, for payment processing while also using AI to enhance fraud detection and improve user experiences.
Coinbase is building blockchain-based financial infrastructure to support AI agents, focusing on digital wallets and decentralized finance as traditional systems may not meet AI’s evolving needs. We explore the potential transition from centralized AI models to decentralized frameworks where blockchain intersects with AI, and provide a quantitative update on our Western, Eastern and Decentralized AI investment index.
🎙️ Podcast Conversation: What really happened at Synapse according to Founder Sankaet Pathak (link here)
Lex interviews Sankaet Pathak, the former CEO and founder of Synapse, and CEO and current founder of Foundation. We discuss what happened at Synapse, the issues around its bank partner Evolve and others, and the overall lessons from the BaaS space.
Sankaet is a technologist and entrepreneur with an academic background from the University of Memphis, where he earned degrees in Computer Engineering, Mathematical Sciences, and Physics. Early in his career, he worked as a research assistant, co-authoring over 25 papers on topics including the coronal loop controversy in physics. He also held roles as a tutor and student assistant, where he provided instruction in circuits, chemistry, astrophysics, and various programming languages such as Java, C++, and C.
Curated Updates
Here are the rest of the updates hitting our radar.
Machine Models
⭐ Selective Attention Improves Transformer - Google Research
⭐ Addition is All You Need for Energy-efficient Language Models - Hongyin Luo, Wei Sun
Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models - Fei Wang, Xingchen Wan, Ruoxi Sun, Jiefeng Chen, Sercan Ö. Arık
Scalable and Accurate Graph Reasoning with LLM-based Multi-Agents - Yuwei Hu, Runlin Lei, Xinyi Huang, Zhewei Wei, Yongchao Liu
Generalization Bounds and Model Complexity for Kolmogorov-Arnold Networks - Xianyang Zhang, Huijuan Zhou
On the Convergence of (Stochastic) Gradient Descent for Kolmogorov--Arnold Networks - Yihang Gao, Vincent Y. F. Tan
AI Applications in Finance
⭐ Apt or “Aipt”? The Surprising Dominance of Large Factor Models - Antoine Didisheim, Shikun Ke, Bryan T. Kelly, Semyon Malamud
⭐ Machine Learning for Interest Rates: Using Auto-Encoders for the Risk-Neutral Modeling of Yield Curves - Andrei Lyashenko, Fabio Mercurio, Alexander Sokol
Mending the Crystal Ball: Enhanced Inflation Forecasts with Machine Learning - IMF
MLP, XGBoost, KAN, TDNN, and LSTM-GRU Hybrid RNN with Attention for SPX and NDX European Call Option Pricing - Boris Ter-Avanesov, Homayoon Beigi
Computing Systemic Risk Measures with Graph Neural Networks - Lukas Gonon, Thilo Meyer-Brandis, Niklas Weber
Dissecting Machine Learning Anomalies - Yonghwan Jo, Yonghwi Kim
A Dynamic Approach to Stock Price Prediction: Comparing RNN and Mixture of Experts Models Across Different Volatility Profiles - Diego Vallarino
Infrastructure & Middleware
⭐ BlackRock and Microsoft plan $30B fund to invest in AI infrastructure - Financial Times
CMA clears Amazon’s £3B investment into AI start-up Anthropic - Computing
Microsoft invests €4.3B to boost AI infrastructure and cloud capacity in Italy - Microsoft
Microsoft to invest $2.7B in Brazil’s cloud and AI infrastructure - Verdict
Microsoft to invest $1.3B to expand Mexico's cloud and AI infrastructure - Capacity
Nebius pours over $1B in European AI infrastructure - National Technology News
🚀 Postscript
Sponsor the Fintech Blueprint and reach over 200,000 professionals.
👉 Reach out here.Check out our new AI products newsletter, Future Blueprint. (Don’t tell anyone)
Read our Disclaimer here — this newsletter does not provide investment advice
For access to all our premium content and archives, consider supporting us with a subscription. In addition to receiving our free newsletters, you will get access to all Long Takes with a deep, comprehensive analysis of Fintech, Web3, and AI topics, and our archive of in-depth write-ups covering the hottest fintech and DeFi companies.